Mass-Producing Satellite Ground Infrastructure in CA

Bridgit Mendler, Co-founder & CEO of Northwood Space, gives Sourcery a first look inside the company’s new 185,000 sqft headquarters & factory in Torrance, CA.

→ Listen on X, Spotify, YouTube, Apple

Northwood has raised over $136M since 2022, including a $100M Series B in January 2026 led by Washington Harbour Partners and co-led by Andreessen Horowitz, with Founders Fund, 137 Ventures, Alpine Space Ventures, Balerion and Fulcrum participating. The company also signed a $49.8M contract with the US Space Force to support the Satellite Control Network.

The new facility will more than 5x Northwood’s manufacturing footprint, support production of hundreds of antennas per year, and add 400 jobs over the next three years as the company works to solve one of the biggest bottlenecks in the space economy: getting data between satellites and Earth.

We tour the future manufacturing floor, antenna production lines, machining and PCB assembly areas, anechoic testing chambers, and a former movie theater that Northwood plans to turn into mission operations. The new facility will include dedicated production for Northwood’s Portal phased-array and Prism parabolic systems, while bringing more critical manufacturing and testing in-house.

Bridgit also breaks down why ground infrastructure has fallen behind the explosion in satellites, Northwood’s Space Force contract, the economics of satellite connectivity, building a global ground network, supply-chain strategy, national security, and her approach to leadership and scaling Northwood.

“If you can improve the networking for the space economy, that is how you translate bits into dollars.”

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Bridgit Mendler, Co-Founder & CEO at Northwood
(00:50) Inside Northwood’s 180,000 sq. ft. facility
(01:50) Why space missions need Ground Stations
(04:45) What “Full-Stack” means for Northwood
(08:40) How the SpaceX IPO changed space hiring
(10:43) What makes Hardware roles so exciting
(13:32) Winning a $50M Government contract
(17:17) Defending against Hacking attempts
(22:45) Turning $136M funding Into a 10-year plan
(26:24) The customers behind Northwood’s growth
(29:07) The hardest part of building Infrastructure
(30:44) The Mentors behind Bridgit’s journey
(32:24) Learning to trust her gut as CEO
(34:54) Underwater UFOs: Real or not?
(37:22) This is where Antennas will get built
(38:04) Touring the Build and Test stations
(39:26) The Blue foam testing chamber
(42:51) The NASA Astronaut

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Applied Aerospace—As global competition intensifies, strategic leadership demands space and defense technologies that deliver enduring national capabilities. At Applied Aerospace & Defense, we unleash the power of advanced manufacturing with unmatched scale and agility. Visit: applied-ad.com/

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Northwood Space Opens 185K Sqft HQ & Factory in Torrance

Bridgit Mendler on the ground segment bottleneck, building antennas out of cellular components, and what $136M buys in a manufacturing business

→ Listen on X, Spotify, YouTube, Apple

The Facility

Northwood Space opened a 185,000 sqft headquarters and factory in Torrance, California on August 31, 2026. The building takes the company’s manufacturing footprint from 25,000 to 135,000 sqft, an increase of more than 5X, and is designed to produce hundreds of antennas per year. Production moves in during October 2026, and the site will add 400 jobs over the next 3 years.

The layout is built around 3 dedicated production lines. A Portal phased-array line will produce more than a full ground site’s worth of panels each month at full ramp. A Prism parabolic line will produce 4 antennas per week. A printed circuit board assembly line brings board production in-house, which cuts roughly 3 months off lead times for new phased-array designs. Expanded machining lines will produce antenna feeds, amplifiers and other RF hardware, and several anechoic chambers will handle antenna and RF component testing on site.

We toured the building before the build-out, when the manufacturing floor was still empty. Mendler walked through where each function will sit.

“What we will have is multiple manufacturing lines for our Prism and Portal products. So for both of those, that’s going to be hundreds of antennas in production.”

The building also carries over some features from its previous tenant, including a movie theater that Northwood plans to convert into mission operations.

“I don’t know that we would’ve built out a movie theater ourselves, to be honest. But the cool thing as a ground company is you do need to have a form of mission ops.”

Why the Ground Segment Fell Behind

Thousands of satellites now orbit Earth, but the ground infrastructure that receives their data has not scaled at the same rate. Mendler attributes the gap to procurement structure rather than technical difficulty.

On the government side, spacecraft and ground systems were bought separately, which left the ground segment scoped last.

“Space systems were the major focus for a long time, & ground segment was kind of just the afterthought. People think about what’s the thing that you’re putting up in space.. & for a number of programs, the ground was just not really contemplated.”

On the commercial side, the problem is a fragmented value chain. Component suppliers, site operators and software vendors each occupy a different segment, and no single participant can respond quickly when spacecraft demand surges. Legacy operators planning next-generation constellations and sovereign operators ramping their own manufacturing both hit the same wall.

Northwood’s response is vertical integration.

“We’re looking across that whole value chain and trying to accomplish it all within one company so that we can actually be responsive to that demand, do it at better cost, make it more flexible, and just generally modernize.”

The Hardware Architecture

Northwood builds many small antennas that combine to deliver the performance of one large dish, rather than building large custom parabolics. The choice is driven by unit volume and mission flexibility.

“We want to basically be an index for space, where whatever use case winds up panning out for space, we’re able to support it.”

“If we were trying to build the Stanford Dish and the Stanford Dish alone, we’d build a giant parabolic that only sits in one place and does one thing. But if we want to accomplish serving the Stanford Dish while serving a lot of other things, we probably would build something that’s more modular.”

The product line splits into 2 categories. Portal is a USB- and SGLS-capable phased array used for telemetry, tracking and control across proliferated and multi-orbit constellations. Prism is a high-throughput parabolic moving SATCOM and mission data across S-, X- and Ka-bands, and is live on the network today.

Because the units are small, field crews can stand up a site in hours rather than the months a custom dish requires.

Supply Chain Strategy

Northwood does not buy aerospace-grade components off the shelf. The company sources from high-volume commercial supply chains and re-engineers the parts to meet space requirements.

“We took the philosophy of let’s not buy the aerospace grade components that are ready off the shelf. Let’s just hire incredible engineers that can figure out how to reconfigure a cellular diplexer.”

“We’ve been able to actually make space grade performance out of standard high-volume cellular components.”

Vertical integration decisions follow a specific test. Northwood brings a function in-house when it becomes a timeline constraint, not as a default posture.

“We only bring in-house what actually turns into a pain point on timeline.”

The Torrance PCB assembly line is a direct application of that rule. Waiting on outside board suppliers was adding roughly 3 months to each phased-array revision cycle.

Site Deployment & Regulation

Northwood operates 5 sites as of 2026. Site selection reduces to 3 infrastructure requirements, then layers coverage and regulatory considerations on top.

› Fiber, to network data back through Earth
› Power, to run the equipment on site
› Land, to physically place the hardware

The company prioritizes colocation with facilities that already have all 3, data centers in particular, which also supply on-site technical staff. Beyond infrastructure, coverage geometry determines placement. Satellites below geostationary orbit move continuously, so operators need contact points distributed to maintain constant communication, and separate sites positioned near where their end customers are.

Licensing has been less of an obstacle than expected.

“Fortunately for us, the licensing has been relatively easy to work with in the different regions that we focused on, because most countries are open for business. They want to have more space business.”

Sites now span markets including Japan, Argentina and Iceland, each with different legal and technical operating contexts.

Security & Site Diversity

Northwood’s security approach favors proliferation over physical hardening. Rather than fortifying individual sites with guards and gates, the company runs a distributed network where any single site can go offline without interrupting coverage.

“We’re taking a bit more of the SpaceX approach where they tend towards proliferation. We have a bunch of sites, and we have that diversity of sites so that we can make sure, if any one of those sites goes down, we can continue to have continuity of coverage.”

The software layer is designed to prevent write access to local systems and hardware tampering at the site level. The redundancy also covers non-adversarial failures, including weather events that take out an entire region.

Funding & the Space Force Contract

Northwood has raised over $136M since its founding in 2022. The most recent round was a $100M Series B announced January 27, 2026, led by Washington Harbour Partners and co-led by Andreessen Horowitz, with participation from Founders Fund, 137 Ventures, Alpine Space Ventures, Balerion Space Ventures and Fulcrum. The Series B followed a $30M Series A less than a year earlier.

The company also holds a $49.8M contract with the US Space Force supporting the Satellite Control Network, the system handling telemetry, tracking and control for government spacecraft. Mendler frames it as an initial engagement.

“It’s kind of like our trial period. It’s us demonstrating our system works, and then that’s the catalyst for a larger engagement for those services.”

The urgency on the government side comes from launch volume. More spacecraft require more contact capacity, and newer mission profiles involving dynamic orbital maneuvering have no existing control solution.

Mendler describes the capital strategy as demonstrating revenue rather than funding a long-dated vision.

“With investors in space today, it’s not just about bankrolling a vision that takes 10 years to materialize. It’s how can you demonstrate a business year after year that grows & that actually hits rather compelling revenue figures along the way.”

“We don’t want to be externally funded forever. We want to make sure that we can be generating our own revenue.”

The Economics of Shared Ground

Operators currently solve for reliability and surge capacity through redundancy, sometimes maintaining 3X the ground capacity they need on an average day. That capacity is expensive to build and carries lead times of roughly 18 months.

Northwood’s model replaces per-operator buildouts with a shared, software-defined network that missions plug into. Per gigabit of connectivity, a customer pays approximately half what building an equivalent site would cost. Because the network is shared and continuing to expand, that unit cost declines as capacity is added. Northwood projects it reaching roughly a tenth of self-build cost by 2030.

The customer base skews toward operators who need guaranteed reliability and the ability to reallocate capacity geographically, largely in communications and Earth observation, with end customers that include US and allied governments. Mendler cites enterprise private networks and high-demand consumer events as representative use cases where throughput must be both constant and reconfigurable.

She positions the ground segment as the entry point rather than the endpoint.

“You need to send data. That is how you monetize space.”

“For us as a starting point it’s ground, but beyond that, we want to continue to support more & more of the networking piece.”

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

RKLB Final Tour: Engine Development Center

We’re back inside Rocket Lab (Nasdaq: RKLB) with Founder & CEO Sir Peter Beck, this time for a behind-the-scenes tour of the company’s Engine Development Center and the technology powering Electron & its next-generation Neutron rocket.

→ Listen on X, Spotify, YouTube, Apple

Pete walks us through Rocket Lab’s additive manufacturing operation, where the vast majority of its engines by mass are 3D printed, and breaks down the engineering behind Rutherford and the much larger Archimedes engine being built for Neutron.

We see Archimedes thrust chambers, turbopumps, 12-laser metal 3D printers, the Rutherford engines powering Electron, and an Archimedes production line capable of producing an engine every eight days.

Pete explains why Rocket Lab chose methane for Neutron, how designing for reusability changes almost every engineering decision, why an Archimedes engine must survive 40 starts and one hour of qualification burn time, and Rocket Lab’s goal of building a rocket that can come off the barge, go straight back onto the pad, and launch again.

We also get into how Rocket Lab acquired more than $100 million of Virgin Orbit assets for $16 million, why Pete thinks some “zombie” space companies refuse to die, the difficulty investors have in diligencing rocket companies, and why Rocket Lab believes vertical integration has been critical to its success.

Pete leaves us with his view on that the space industry hasn't rly even begun yet..

“It's like we've sent our first email in the beginning of the internet.”

“I think the biggest thing to be done in space hasn’t even been thought about, let alone talked about.”

This is Tour 02, the second & final part of our Rocket Lab factory tour.

Catch up on the full series:

• Part I: Our full sit-down interview with Sir Peter Beck

• Part II / Tour 01: Inside Rocket Lab HQ, Mission Control & Satellite Manufacturing

• Part III / Tour 02: Inside Rocket Lab’s Engine Development Center, Rutherford & Archimedes

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Sir Peter Beck, Founder & CEO at Rocket Lab
(00:48) Buying a $100M building for $16M
(02:30) Why Space companies never die
(04:15) This is where Rocket Lab's engines actually get built
(06:55) The Biggest differences between Archimedes and Rutherford
(07:56) This is where Rocket Lab actually tests its Engines
(10:01) How 12 lasers print a Rocket engine part
(15:04) A full walkthrough of the Electron Rocket
(16:29) Why Iridium was just the beginning
(18:35) Methane vs. Kerosene: why it matters
(19:28) Why Reusability creates its own set of problems
(24:17) What makes a Vacuum Engine different
(26:39) The future of Space

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Applied Aerospace—As global competition intensifies, strategic leadership demands space and defense technologies that deliver enduring national capabilities. At Applied Aerospace & Defense, we unleash the power of advanced manufacturing with unmatched scale and agility. Visit: applied-ad.com/

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Rocket Lab’s Engine Development Center (“EDC”)

Rocket Lab’s Engine Development Center sits in Long Beach, a short golf cart ride from the company’s headquarters. It builds the Rutherford engines that power Electron and the Archimedes engines that will power Neutron. The building belonged to Virgin Orbit until 2023.

Sourcery toured the facility with Sir Peter Beck, walking the additive manufacturing floor, an assembled Electron, and the Archimedes production line. This is the second of two facility tours.

→ Listen on X, Spotify, YouTube, Apple

The Best Deal of His Life

Beck does not like buying real estate. He made an exception for this one.

“Where we’re taking you, EDC, was arguably one of the best deals of my life, because that was the old Virgin Orbit facility. And there was over $100 million worth of asset in that building, and we managed to buy the whole thing for $16 million.”

Virgin Orbit filed for Chapter 11 in April 2023 after four successful launches and two failures. Richard Branson had put roughly $1.2 billion into the company. Rocket Lab acquired the machinery and the lease at auction the following month.

“It happened really fast. They were in bankruptcy, and we saw it come up, and we all just piled into golf carts like this, hanging off the back, and nipped down there and made it happen.”

The strategic value was capacity rather than price. Engine production rate is the constraint that governs launch cadence, and buying an existing engine factory removed that question permanently.

“We didn’t never have to worry about a factory for building engines ever again. We’re totally sorted.”

I asked Pete if Rocket Lab maintains a running list of space companies that might go bankrupt so they can buy more up for parts & facilities.. he said yes.

Why Space Companies Are Hard to K*ll

The Virgin Orbit purchase leads into the most unusual observation on the tour, which is about how the industry disposes of its failures.

“The funny thing is, space companies are really hard to kill. The amount of space companies that should be dead, the sort of zombie companies, is quite incredible. It’s not like other industries where you go to die, you just die. Space companies tend to linger for quite some time.”

Beck’s explanation is about legibility. Space attracts enthusiasm because it is exciting, and the same quality makes claims difficult to check.

“The thing about space is that it’s awesome that somebody can be really forward-leaning and promotional, and everyone gets excited. But the downside to space is exactly the same thing, where it’s really difficult to understand and comprehend. So people can stand up and make these claims, and it’s really hard to corroborate unless you go five layers deep into the rocket equation.”

Asked who does that diligence well, his answer is blunt, and it cuts against firm reputation.

“Not really. Because you see a company that’s just objectively terrible, and they just continue to get funding. There’s just new investor after new investor after new investor. But sometimes the really big firms that you think are really good at it actually are the worst.”

That has a direct commercial consequence for Rocket Lab. Persistent underpriced competition suppresses returns across small launch, and it also generates the bankruptcy inventory that produced this building.

Additive Manufacturing at Scale

The print shop is the core of the facility. Both engine families are largely 3D printed, a decision Rocket Lab made early with Rutherford and carried into Archimedes.

“At least by mass, the vast majority of all of our engines are 3D printed.”

The process is laser powder bed fusion. Machines fire lasers into a bed of metal powder, sintering it layer by layer into a solid part. Rocket Lab runs machines with 12 lasers each, and prints in Inconel superalloys, copper, titanium, and a proprietary alloy developed in-house.

“There’s a whole lab back there with material scientists developing new materials for that particular engine and environment.”

The reason for printing is consolidation, not novelty.

“We use 3D printing because we’re able to incorporate a whole bunch of geometry into one design. We’re not stupid. We don’t 3D print bolts and all that sort of stuff, but we really become very good at designing these really complex architectures that, if you can get away with it in your 3D print, you can basically print multiple parts all in one part. And at the end of the day, it’s all about speed and cost.”

Conventional engine manufacture assembles hundreds of machined components, each requiring tooling, fixturing, joining and inspection. Printing a manifold or an injector as a single part removes every one of those steps for that assembly. The economics follow.

“An Archimedes engine is an incredibly cheap engine. Same with a Rutherford engine. It’s ridiculously so. So building engines fast, building engines cheap, but at high performance.”

Roughly 10 people run the entire shop. The longest single print takes about three and a half days. A larger machine arrives next year capable of printing an entire Archimedes engine, roughly Beck’s height, in one build, and Rocket Lab is the first customer for it anywhere in the world.

One part on the floor illustrates the material complexity. An Archimedes thrust chamber, printed in copper for its thermal conductivity, was being clad in a superalloy layer to survive the combustion environment.

Two Engines, Two Philosophies

Rutherford and Archimedes share a design ethos and almost nothing else.

“The Rutherford is an electric pump cycle, whereas Archimedes is a staged combustion cycle. And just the scale of the engines are enormously different. It’s a much more complicated cycle and a much, much larger engine.”

Rutherford uses battery-powered electric motors to drive its propellant pumps, an approach almost nobody else uses, and burns kerosene and liquid oxygen. Archimedes uses staged combustion, where propellant is burned in a preburner to drive the turbopump before entering the main chamber, and burns methane and liquid oxygen. The Archimedes turbopump produces roughly 15,000 horsepower.

“A lot of the ethos of how you do it has remained the same. How we do injectors and how we combine multiple parts into an engine all remains the same.”

The injector is where Beck’s interest sits.

“The favorite part of any engine builder is the injector, because that’s where the black magic happens. Everything else is to contain that mixing and produce thrust, but the injector and the turbo pump, those two bits are pretty cool.”

Production rates differ by an order of magnitude. Rutherford is built at one engine per day, with an engine test cell in New Zealand supporting the program, and more than 930 Rutherford engines have flown. The Archimedes line currently produces one engine every eight days. Testing happens at Rocket Lab’s facility inside NASA’s Stennis Space Center in Mississippi, where two test cells run 20 hours a day, seven days a week.

Designing the Most Boring Engine Possible

The core engineering argument on the tour is about qualification life, and it starts with a number.

“The qualification burn time for an Archimedes engine is one hour. So the engine has to do 40 starts and run for one hour. Whereas Rutherford for Electron, because it’s a single-use engine, its acceptance test is like five minutes. It only needs to run for 190 seconds in its life, so a five-minute run is pretty much enough.”

That is a twelvefold difference in required runtime, and it changes the design problem entirely.

“It’s just way harder to make a staged combustion engine run for one hour. And just not even worry about it running for one hour.”

Beck’s framing for what that produces is the sharpest line in the tour.

“If you look at the Archimedes engine itself, really the object of that engine was to make the most boring engine possible. If you’re sitting on an airplane and you look out at the engine on your wing, you’re not marveling at how amazing that is and how on the knife edge that engine is. You wanna know that engine is boring. It’s not gonna blow up. Same thing for this. The Archimedes engine needs to just go and go and go and go.”

He positions this explicitly against the high-performance engines that dominate the market.

“Those engines you mentioned are really high-performance engines, and they’re really strung out, which is great. This is, I would say, the first engine that is just not designed to be like that. It’s designed to be the most benign engine that you could possibly imagine so that you can just run and run and run and never think about it.”

The trade is not free. Archimedes runs at relatively low chamber pressure for a staged combustion engine, which lowers internal temperatures and extends life, and creates a new problem.

“You’d think, ‘Oh, that’s good for longevity,’ which it really is, but actually it makes it much harder to light, because the temperatures are way lower. So you create these things that are ultimately good, but you create a whole lot of problems along the way to try and make something good.”

Fuel Choice, & the Hardware You Have to Burn

Methane over kerosene comes down to what the engine looks like after it runs.

“The trouble with kerosene is that you get these little soot deposits in the regen channels of the engine. And you have to purge the engine out, and it’s sort of dirty.”

Regenerative cooling routes propellant through channels in the chamber wall before combustion, cooling the structure and preheating the fuel. Soot in those channels degrades cooling and requires servicing between flights, which is incompatible with rapid reuse.

The requirement that forced the decision was one Beck imposed himself.

“The design requirement for the vehicle was turn in 24 hours, which was an absurd design requirement. But it drove a whole lot of really good decisions. One of them being methane, because you can run an engine and after the engine’s run, it’s still shiny stainless steel and stuff. There’s just no residue whatsoever. You do the same with a kerosene engine, there’s just soot and black crap everywhere.”

Staged combustion also imposes a testing cost that simpler cycles avoid.

“The challenge with a staged combustion engine is you can’t separate components and then go and test them individually and then bring them together. The first time you hot fire an engine, you have to have the turbo pump and everything on it as one thing. So you have to be extremely hardware rich at the start of your program, because you consume a lot of hardware. You literally eat engines out.”

One test produced a result Beck still finds strange.

“There was one engine hot fire that we did very early on in the test program, the engine leaned out, and it actually just consumed the whole injector. Which was bizarre, because there was literally just a big hole left. There was nothing left of the injector. But the propellants were still coming together and mixing and producing thrust, and the engine was still running just fine. Well, not just fine, but still running.”

His conclusion is that analysis has limits.

“It’s very difficult to analyze. You’ve just gotta test. You’ve just gotta spend heaps of time on the engine test cell.”

Archimedes has taken three to four years of team effort so far, and has completed more than 400 hot fires.

Neutron, & Designing Backwards From Reuse

Every Archimedes design decision traces to a vehicle that has not yet flown. Neutron carries 13,000 kg to low Earth orbit on nine Archimedes engines, with a tenth vacuum-optimized engine on the second stage.

The engine count is not a performance choice.

“The upper stage engine produces a certain amount of thrust, and you have to manage that thrust for the acceleration of the upper stage. It just so happens that coincidentally is the ideal amount of thrust to land a rocket.

So you need at least one engine to do that final terminal landing phase. So it just makes sense to have one engine and just have nine of them on the bottom and one on the top. That just, it’s just physics. It just works out like that.”

Landing requires throttling down to roughly a tenth of liftoff thrust, since the vehicle is nearly empty by then. A single engine sized for the upper stage happens to sit at that number, so the same engine serves both roles and Rocket Lab qualifies one design instead of two.

Beck is direct that reuse is the whole difficulty.

“The reusability in itself is a challenge, because if our job was just to make an expendable launch vehicle, we’d be there now. We’d be launching Neutrons flat out.”

That is the clearest statement available on why Neutron has slipped. The delay is not a capability gap, it is a deliberate refusal to ship a conventional vehicle.

“What we’re trying to do here is make a step change increment on the current state of the art. We can’t just come to market with a vehicle that’s as good as the current one on the market. We have to come to it with a vehicle that’s way better. So designing a vehicle that literally comes off the barge and goes straight on the pad and launches again is what we’re trying to achieve here. So that drives a whole bunch of engineering compromise, a whole bunch of engineering decisions.”

The Hungry Hippo fairing is the most visible product of that constraint. Rather than separating in flight and being recovered from the ocean, it stays attached to the first stage through launch and landing. Beck’s account of the name is characteristically unsentimental.

“It just looked like a hungry hippo.”

The naming habit is deliberate across the company.

“We like to have fun with the mission names as well, because it’s such a serious business. All of our mission names are always called something a little bit funny, just to lighten it up a little bit.”

The vehicle also has a satellite counterpart. Flatellite, Rocket Lab’s flat, stackable spacecraft, is designed as a high-density format tailored for Neutron, a pairing Beck discussed in the full sit-down interview.

Electron, & Where This Is All Going

The tour includes an assembled Electron, the vehicle that generated the manufacturing capability everything else now runs on.

“This is the rocket that we fly the most of right now. It’s flown 93 times. It’s the second most frequently launched rocket in the world, behind the Falcon 9.”

The vehicle is carbon composite end to end, with a fairing, a kick stage that serves as the second stage, and nine Rutherford engines on the first stage.

“We build everything in this vehicle. Every piece of hardware, every piece of software, tanks, engines, you name it, we build it, and I think that’s one of the key successes of the company is just that vertical integration.”

The same principle now governs the satellite business.

“If you pull apart a satellite that we’ve built, the flight computer will be ours, the reaction wheels will be ours, all the solar panels will be ours. It’s the same philosophy.”

Iridium extends that stack into services.

“Iridium is just the start of our applications layer. I’ve always believed the big space companies of the future are gonna all look a little bit the same. They’re gonna have their own rocket, because access to space is key. They’re gonna have their own ability to build as many satellites as they need, and they’re all gonna have applications.”

Space Tourism & Hot Takes

Space tourism is not on that list, and Beck’s reasoning is about accountability rather than market size.

“I don’t think I’d be a very good tour operator, if I’m honest. Better engineer than tour operator. I think it’s always a tricky business. And unfortunately I’m the CEO, so if anything ever goes bad, guess who’s knocking on the door?”

Asked for his hottest take after 20 years in the industry, he offered a position that is less a prediction than a scale correction.

“I’ve been doing this for 20 years, and at the start of that it was like, one day there’ll be the democratization of space, and it won’t be governments launching rockets, it’ll be commercial entities. And we’re well past that. So my hot take would be I think the biggest thing to be done in space hasn’t even been thought about, let alone talked about. If you wanna make an analogy here, it’s like we’ve sent our first email in the beginning of the internet.”

→ Listen on X, Spotify, YouTube, Apple

Watch the Full Series:

Part I: Sit-down interview with CEO Sir Peter Beck

→ Listen on X, Spotify, YouTube, Apple

Part II: Rocket Lab HQ Tour (Satellites, Mission Control)

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Rocket Lab Factory Tour 1 of 2:

We went inside Rocket Lab with Founder & CEO Sir Peter Beck for a behind-the-scenes look at how one of the world’s leading space companies actually builds.

→ Listen on X, Spotify, YouTube, Apple

In the first part of our Rocket Lab factory tour, Peter takes us through the company’s Long Beach headquarters, mission control, satellite manufacturing, spacecraft testing, electric propulsion, and NASA’s LOxSAT program.

Along the way, we trace Rocket Lab’s evolution from a tiny New Zealand startup with less than $100,000 NZD in capitalization to a global, publicly traded space company spanning launch, satellites, spacecraft components, propulsion, defense, and interplanetary missions.

Peter breaks down Rocket Lab’s capital-efficient culture, why he took the company public, its acquisition strategy, the hidden moat in spacecraft components, and his goal of building a multi-generational space company that continues long after he’s gone.

We also get into Rocket Lab’s earliest rockets, Khosla & Bessemer’s first investments, NASA’s CAPSTONE mission to the Moon, the search for life on Venus, and what Peter believes most new space companies underestimate.

→ Tour 2, we head inside Rocket Lab’s propulsion & engine facilities

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Touring inside Rocket Lab's HQ
(01:44) The origin story
(05:45) What "Number Eight Wire" actually means
(07:14) Too much Capital can kill a startup
(09:03) Why Rocket Lab decided to go public
(10:45) From Rockets to an entire Space Ecosystem
(15:11) Where Escapade and CAPSTONE were flown
(16:44) Inside the satellite build floor
(17:15) In search for life on Venus
(18:08) Inside the Electric Propulsion Lab
(19:58) Inside the Thermal and Shock Testing Facility
(20:47) How big is this facility?
(21:45) Inside the NASA LOxSAT project
(23:06) Why so many Space startups fail
(24:00) Up Next: Inside the Engine Lab

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Applied Aerospace—As global competition intensifies, strategic leadership demands space and defense technologies that deliver enduring national capabilities. At Applied Aerospace & Defense, we unleash the power of advanced manufacturing with unmatched scale and agility. Visit: applied-ad.com/

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Inside Rocket Lab HQ: What 20 Years of Constraint Built

Rocket Lab’s Long Beach headquarters spans three buildings. The company operates roughly 13 acres of factory floor worldwide, across New Zealand, California, Arizona, New Mexico, Toronto, Maryland, Virginia, Mississippi and Germany. About 800 people work in Long Beach across two campuses.

Sourcery toured the headquarters with Sir Peter Beck, walking the lobby, mission control and the satellite production floor. The tour covered the company’s first two decades, why the components business exists, & what Beck thinks kills new space companies.

→ Listen on X, Spotify, YouTube, Apple

The 20-Year Overnight Success

The lobby is built around a timeline. Beck calls the company an overnight success two decades in the making.

• 2006: Founded in New Zealand

• 2006 to 2009: Operating as a technology development house, not a launch company

• 2009: Ātea-1 reaches space, the first private company in the Southern Hemisphere to do so

• 2010: DARPA contracts for a new propellant and propulsion system, plus a rocket-launched UAV

• 2013: Beck moves to the US, raises a $5M Series A led by Khosla Ventures, reincorporates as a US company

• 2014: Series B led by Bessemer

• 2015: Ground broken on Launch Complex 1, the first private orbital launch site in the world

• 2016: Launch Complex 1 completed

• 2017: Electron’s first flight, & a Series D that makes Rocket Lab a unicorn

• 2018: Three Electron launches

• 2021: Goes public, & acquires 3 companies in the same year

Total capitalization at the 2009 flight was under 100,000 New Zealand dollars, roughly $60,000 US.

“On that rocket actually is Swagelok fittings that I’d salvaged from a junkyard. Because I could only afford the ferrules, not the cost of the fittings. So I would get the fitting from the junkyard and buy the $3.20 ferrules to put in it and recondition it into a fitting. So it was seriously constrained stuff. But from that adversity comes good ideas.”

Government work bridged the gap to venture capital. The 2010 DARPA contracts kept the company solvent and technically active for three years before it raised anything institutional. Government customers now make up 57% of backlog.

The 2013 Series A funded the first hot fire of the 3D printed Rutherford engine and proved out the composite structures.

“No Money, So We Have to Think”

The engine name is the philosophy. Rutherford is named for Ernest Rutherford, the New Zealand physicist, & specifically for a line Beck adopted as the company ethos.

“We have this saying, well, it’s actually.. we stole it from Ernest Rutherford, & that’s why the engine is called a Rutherford.

He said, ‘We have no money, so we have to think.’ And that’s kind of the ethos of the company.”

The New Zealand shorthand for this is “number eight wire,” a reference to fencing wire and the habit of improvising with whatever is available. Beck, however, is not so much a fan of the term..

“I actually personally hate it. Because it means, well, it’s got 2 connotations:

1. In New Zealand it means to bodge stuff together. And I’m the opposite of a bodge stuff together kind of guy.

2. But its more admissible meaning is making do with not much, which I think is technically accurate.”

The distinction matters commercially. Rocket Lab’s differentiator is capital efficiency, not improvisation, and the two look similar from outside.

The comparison Beck reaches for is direct.

“The Electron rocket, I think it was under $100 million & 80 people put the first rocket into orbit. And at the time, our nearest competitor was Virgin Orbit, that was funded by Sir Richard Branson. And in the end I think Richard put $1.2B into that company before it ultimately went bankrupt with a handful of launches under their belt.”

Virgin Orbit filed for Chapter 11 in April 2023 after four successful launches and two failures. Rocket Lab bought its production assets out of the bankruptcy auction for $16.1 million and now builds Rutherford and Archimedes engines in the former Virgin Orbit facility a short distance from where this conversation took place.

“If I look back and think of lots of startups that I’ve seen or been involved with, the ones that have failed most spectacularly have been the ones that have been funded the most well.”

Why He Took the Company Public

Rocket Lab listed in 2021 & made 3 acquisitions the same year. Beck gives two reasons for going public, and the second is quite thoughtful.

The first is discipline.

“Being a public company teaches you extreme discipline. You can’t be messing around as a public company and wasting capital and not delivering. So the company has always had a really strong culture of delivery, but as a public company, it’s just even much, much more intense.”

The second is succession.

“I’m trying to build a multi-generational space company. And so many companies in this industry are defined by their founders or a small group of people. And you have to ask yourself, when they ultimately kick the bucket, what happens to those companies?”

He traces it to something his parents told him.

“The only thing that my parents ever said to me as I was growing up was,

‘Have the most amount of impact in your lifetime.’

So you can have an amount of impact when you’re alive, but if you can keep that impact going after you’re dead, then by proxy you have much, much larger impact.”

This is a specific view of what public markets are for. Most founders treat a listing as a financing event or a liquidity event. Beck treats it as an institutional constraint that outlives him, and the mechanism he trusts to keep the company running when he is not there to run it.

3 Acquisitions in 1 Year

2021 was also when the components roll-up began in earnest. Beck’s assessment of doing 3 at once is short.

“We went public & acquired 3 companies. Don’t recommend that. It is really painful.”

The reason is not regulatory. It is what he calls ‘the drive home.’

“You go out to a used car lot, & you look around the nice big polished car & it all looks great, & it’s all shiny.. And then it’s not until the drive home that you realize that the visor’s squeaking, the wheel’s rumbling, & all of the things just aren’t quite right.”

The counterintuitive part is that deal size does not predict difficulty.

“There’s no such thing as a small acquisition.

Ironically, some of the smaller companies have consumed vastly more time than some of the bigger companies. So it’s just a lot of work to integrate a company in & aggregate all the financial systems & get the culture where you want it.”

Rocket Lab has completed 7 acquisitions in 6 years. Sinclair Interplanetary brought reaction wheels in 2020. Advanced Solutions brought flight software, Planetary Systems brought separation systems, and SolAero brought solar cells and arrays, all within roughly 14 months of the listing. Geost added missile warning payloads in 2025. Mynaric added laser communications and Motiv added space robotics.

However, the count reaches 9 if you include two purchases that were not companies: the Virgin Orbit machinery and lease, bought out of bankruptcy in 2023, and a plant in Warkworth, New Zealand, taken over from SailGP in 2024 along with about 50 staff. Iridium would make 8, or 10, once it closes in 2027.

The Component Moat

The reason for buying all of it is the part of the business that gets no attention.

“People focus on the rocket because it’s the big firework in the sky, everyone gets emotional, & it’s all very cool.

It’s hard to get emotional about a reaction wheel, or a star tracker, or a little component that goes into a spacecraft.”

The argument is about supply, not sentiment.

“Launch is a huge moat, but all of these spacecraft components are huge moats as well. Because it doesn’t matter how much capital you’ve got, if you turn up tomorrow and say, ‘I wanna build 1,000 satellites,’ good luck.

It’s just not happening, because the supply chain is so small & fractured.”

A reaction wheel is a spinning flywheel that rotates a satellite by transferring momentum, no propellant required. A star tracker photographs the sky and matches the pattern against an onboard catalog to determine attitude to arcsecond precision. Both are small, both are unglamorous, and both take years of flight heritage to qualify. Neither can be conjured with a funding round.

The conclusion Beck draws is about optionality.

“We saw that to get to this end-to-end space company in our vision, we needed to own all the components that go into the spacecraft so that we could literally build any spacecraft we want.”

That strategy is now the majority of the business. Space Systems delivered 81% of revenue in Q2 2026 against 19% for Launch Services.

Mission Control, & What Gets Built Here

Rocket Lab runs five mission control rooms. The Long Beach room handles satellite operations primarily, and has flown two of the company’s more notable missions.

CAPSTONE, the lunar pathfinder Rocket Lab built and flew for NASA, was operated from Long Beach and from New Zealand on a 24-hour handoff between hemispheres. The ESCAPADE spacecraft, en route to Mars, were also controlled from this room.

CAPSTONE’s purpose was to validate an orbit before Artemis committed to it.

“That mission was actually the very first part of the Artemis program. So we went into a really weird orbit that the gateway mission was gonna go into to prove that that orbit was stable.”

An engineering unit of CAPSTONE sits in the building. The flown article is roughly five million miles away in a near-heliocentric orbit. Beck’s explanation for why the spare exists is a one-liner on how spacecraft programs work: you always build two.

The satellite floor is arranged as a production line, entering at one end and exiting as a completed spacecraft. One current program involves electric propulsion, integrated in a dedicated tent on the floor. Electric thrusters handle station keeping and modest orbit raising without carrying large volumes of chemical propellant.

Small details on the floor reveal the manufacturing philosophy. Rocket Lab installs vending machines stocked with electrical components, tape and consumables, so technicians pull what they need instead of queuing at a storeroom.

“It just makes it easier for the staff if they need various bits and pieces, consumables for the spacecraft build. Rather than going to a storeroom and having to stand in line and asking for somebody, you can just go to the vending machine and get it. It’s just way faster.”

That is automotive thinking applied to spacecraft. Rocket Lab hired COO Frank Klein out of Mercedes, and the production line layout, the ERP and MRP systems, and the vending machines are all the same instinct: remove the small frictions that compound across thousands of build hours.

The qualification process is the other half. Rocket Lab builds a full thermal model of a spacecraft first, using aluminum blocks to represent thermal masses, and runs it through thermal, vibration and shock testing before assembling the flight article. All of that testing happens on site in Long Beach.

The mission profile explains the standard.

“These particular missions are very difficult missions.

They’re like 12-year lifespan, operating in nasty, nasty radiation, & they’re missions that just cannot fail. There’s always a big difference between low Earth orbit (“LEO”) missions that have a short duration & are allowed to fail. Most of the stuff we do is super hard, defense-related, or interplanetary and stuff like that where you just can’t have stuff fail.”

Another program on the floor is LOxSAT, built for NASA, which stores liquid oxygen in orbit for extended periods. Cryogenic fluid management is one of the unsolved problems standing between current capability and long-duration interplanetary missions, because liquid oxygen boils off.

Beck’s team also keeps a habit worth noting. There is always an interplanetary mission in the clean room. The Venus mission, searching the planet’s cloud layer for phosphine as a possible biosignature, is the one he talks about personally.

Where New Entrants Fail

Asked what new space companies get wrong, Beck names reliability rather than technology or capital.

“The most amount of failure I’d say is because people don’t build reliable stuff. It’s easy to rush it and throw something on orbit. But the distance between putting something on orbit and putting something on an orbit that can last 12 years or go to a planet, that gap is just massive.

If you want a valley of death, that’s kind of it.”

The same logic drives where Rocket Lab chooses to participate in lunar programs. It supplies components and systems rather than becoming a Moon company.

“There’s a bit of whiplash there.

Like, America’s going to the moon. No, no, no. Now we’re going to Mars. Now we’re going to the Moon again. Now we’re going to Mars. And no, no, this time we’re really going to the Moon.

And that’s fine, but when you’re a moon company, that just plays havoc with your balance sheet. So we get to play in these programs, but we like to play at a system level rather than a mission level.”

That is the components thesis applied to political risk. Selling into every lunar program insulates the company from which one survives the next budget cycle, which is the same reason selling components to competitors insulates it from which satellite prime wins the next contract.

Much of what Rocket Lab builds in Long Beach could not be filmed. Beck’s note on that was brief: the company runs a lot of national security programs, and these were the missions that could be shown.

→ Listen on X, Spotify, YouTube, Apple

Check out the rest of the series:

• Part I: Full sit-down interview w/ Peter Beck

• Part III: Tour 02 - Engine Development Center (coming soon)

Part I: Sir Peter Beck, Founder & CEO

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Welcome to Rocket Lab Week.

Sir Peter Beck (he goes by “Pete”) founded Rocket Lab 20 years ago. Today, the company is scaling far beyond launching rockets.

→ Listen on X, Spotify, YouTube, Apple

We sat down with Founder & CEO Sir Peter Beck after touring Rocket Lab’s Long Beach, CA facility to break down the current business & what comes next.. this is the first part of a 3-part journey looking into the company that we’ll share over the next few days. First, a sit down interview (this), then two tours of their facilities.

In this sit down, Peter explains Rocket Lab’s transformation from a launch company into an end-to-end space company spanning rockets, satellites, components, defense, and communications.

We get into Neutron and why Peter believes investors should be asking about the 10th launch, not just the first, the strategy behind the Iridium acquisition, Rocket Lab’s growing national security business, Flatellite, its acquisition playbook, and why going public gave the company the firepower to pursue opportunities it never could have as a private company.

We also discuss Neutron’s delays, Archimedes, reusability, Rocket Lab’s vertically integrated manufacturing strategy, building an entire space industry from scratch in New Zealand, and Peter’s long-term ambition to build a multi-generational company.

ICYMI: Rocket Lab (NASDAQ: RKLB) posted a record Q2 2026:

• $234M revenue, +62% YoY
• $2.36B record backlog, +137% YoY
• $437M+ in new launch contracts
• $581M+ in new Space Systems contracts
• 90+ launches in backlog
• Q3 revenue guidance of 250M–265M

Plus: aliens, getting knighted, why Archimedes engines are painted black, & Peter’s surprisingly simple philosophy on business.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Sir Peter Beck, Founder & CEO at Rocket Lab
(00:59) Rocket Lab's biggest quarter
(02:53) The competitive edge in Space
(04:10) Rocket Lab's Acquisition Strategy
(05:04) Reason behind the Iridium Deal
(06:13) Competing with Blue Origin & SpaceX
(09:56) Rocket Lab's Global Domination
(11:19) Building a Launch Monopoly in New Zealand
(12:33) Is this why Peter got Knighted?
(16:42) Beating SpaceX to the Public Markets
(20:05) Why Neutron is still facing delays
(24:24) The reality of running a Rocket Company
(27:38) Rocket Lab's latest contract wins
(29:04) How defense became a prime mover for Rocket Lab
(30:39) Ghost: Rocket Lab's portable launch pad
(31:30) Has he seen Aliens?
(33:18) What's next for Rocket Lab?
(34:30) Where does Peter draw his inspiration?
(36:12) Why he's never read a Business book
(37:31) Why their Engine is painted black

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Sir Peter Beck on Building Rocket Lab Into a Multi-Generational Space Company

Rocket Lab reported Q2 2026 results in August, its third record quarter in a row, with revenue of $234 million, up 62% year over year, and backlog of $2.36 billion, up 137%. The company signed more than $1 billion in new contracts across the quarter and the weeks after it closed. We toured the Long Beach facility and sat down with Sir Peter Beck for an in-depth discussion. Hopefully, the first of many.

Welcome to the first official Rocket Lab week.

→ Listen on X, Spotify, YouTube, Apple

The Man & the Company

Beck founded Rocket Lab in New Zealand in 2006. He has no university degree. He started as a precision engineering apprentice at an appliance manufacturer in 1993, spent a decade in manufacturing and composites research, and has said he could not have been hired at NASA or Boeing. He was knighted in 2024.

Rocket Lab went public in 2021. It now operates from New Zealand, California, Arizona, New Mexico, Toronto, Maryland, Virginia, Mississippi and Germany. It has made nine acquisitions in six years. In June 2026 it agreed to buy Iridium for roughly $8 billion, its largest deal by an order of magnitude.

Electron, the small launch vehicle, has flown 93 times. It set a record of 21 launches in 2025 with 100% mission success and has flown 14 times so far in 2026.

What Rocket Lab Actually Sells

Rocket Lab is known as a launch company. But now Launch is just a smaller part of the business.. The company reports two main business segments:

Launch Services is the original business.

1. Electron: small launch vehicle, roughly 300kg to low Earth orbit, flown 93 times

2. HASTE: suborbital variant, used for hypersonic and missile defense flight testing, and accounts for 20% to 25% of the annual manifest

3. Neutron: reusable medium-lift vehicle, joins the segment when it flies

Space Systems is everything else, and it is 3 distinct businesses.

1. Whole satellites built for customers. That covers the Photon bus family, the Space Development Agency constellations, and Flatellite, the flat, stackable design built for volume production. Rocket Lab is the prime contractor on these.

2. Components and subsystems sold to other satellite builders. Star trackers, reaction wheels, solar cells and arrays, separation systems, propulsion, radios, flight software, laser communications terminals, robotics. Here Rocket Lab is a supplier, and many of its customers compete directly with the satellites in the first business.

3. Mission operations & space data services, the part that runs spacecraft after delivery. Iridium adds recurring subscriber revenue for the first time

Almost none of the components business was built in-house. It was acquired, one specialist at a time, over six years. That roll-up is why Space Systems now carries 81% of revenue, and started with a single phone call.

The Reaction Wheel Problem

“When we kind of started that space systems team, the first thing we did was try & buy a reaction wheel. We rang up Sinclair Interplanetary at that stage & said, ‘Doug, we need some reaction wheels.’

Doug said, ‘That’s great, Pete. It’ll be 9 months.’

I’m just like, ‘Doug, I haven’t got 9 months for anything. We need way, way faster than that.’

And so the way to solve that problem is we just went & bought Doug.”

Reaction wheels point satellites. Spinning flywheels inside the spacecraft transfer momentum to rotate it, no propellant required. That is how an imaging satellite swings from one target to the next. They are unglamorous, hard to substitute, and were historically built in small batches by specialist shops.

Rocket Lab bought Sinclair Interplanetary in April 2020 for $12.34 million in cash. It now builds over 2,000 reaction wheels a year.

Build, Acquire, Scale

With little time for waiting, Beck kept up the pace by making more strategic moves.. building into one of the most competitive space companies in the market through acquisitions and vertical integration.

“If you took a satellite & you laid it out on the boardroom table, we just pointed all the things that were really painful, & went out & acquired the very best companies, or in some cases created the technology ourselves, & then scaled it.”

Advanced Solutions came in October 2021 for $40 million, bringing flight software and guidance. Planetary Systems closed that December for $42 million, adding separation systems. SolAero closed in January 2022 for $80 million, adding the largest space solar cell production line in the world and 425 employees. Geost came in 2025 for $275 million, adding missile warning payloads. Mynaric added laser communications terminals. Motiv added space robotics in May 2026.

None of it looked obvious at the time.

“At the time I remember a lot of folks were saying, ‘Well, why are you buying this company in Albuquerque & this company in Toronto that makes these little widgets?’ And then I think it’s very obvious now that those widgets are the things that truly enable a sense of scale for some of the space infrastructure.”

Is Launch the Only Moat?

“Launch is always a huge moat. Launch is probably the biggest moat. But I would say the second-biggest moat is those components that go into those spacecraft. Because you can have infinite capital, it doesn’t make any difference. You still need some star trackers & some reaction wheels, & they’re just not available at scale.”

The claim is testable. Capital buys engineers, factories and time. It does not compress the qualification cycle on flight hardware. A star tracker photographs the sky, matches the pattern against an onboard catalog, and determines attitude to arcsecond precision. Certifying one for flight takes years of heritage. Rocket Lab redesigned its ST-16HV star tracker for mass manufacture rather than peak performance, a deliberate trade for constellation customers who need volume and short lead times.

The strategic argument is about dependency.

“I’ve always viewed that the large space companies are gonna have the ability to build their own satellites and launch them. And if you’re beholden to someone else providing you all the critical components, then your ability to scale is governed by their ability to scale, and that’s not a place where you wanna be.”

The installed base backs it up. Rocket Lab components have flown on more than 1,700 missions. Its solar cells power over 1,000 spacecraft. Its flight software has run on 50 spacecraft for a cumulative 150 years on orbit.

Winning Twice Against the Primes

Selling to competitors creates an unusual dynamic in defense contracting. However, Rocket Lab bids as a prime against companies that are also its customers.

“Even when we go after those big prime contracts, we’re prime on two national security programs right now. Even when we go after being a prime, even if we lose, we still win. And when we win, we win twice, because we’ll go head-to-head with a prime for a contract, and if we lose, then the next day there’s a whole bunch of purchase orders for reaction wheels, the solar, & a whole bunch of stuff.”

The Space Development Agency award is the example he gives.

“Our SDA contract was a great example where we won an $800 million contract. There was two other primes that also won contracts, but actually, that turned into over a billion-dollar contract for us because the primes needed our stuff.”

The Tranche 3 award was $816 million, the largest prime contract in company history. Component sales to the competing winners pushed the total past $1 billion.

The contracting model is the other differentiator.

“The typical national security contract would go to your typical primes, cost plus, blow out in times, & blow out in money, & all the rest of it. And we’re just not built like that. When we go out into the national security world, it’s firm fixed price and we deliver.”

Under cost-plus, the government reimburses costs and adds a fee. Overruns transfer to the customer. Under firm fixed price they do not, which is a different risk posture for a company still burning cash.

The customer mix has shifted fast. Government went from roughly 35% of backlog in Q4 2025 to 57% in Q2 2026. Alongside the SDA work sits a $397 million Space Force contract to build and launch Flatellites for the Space Based Airborne Moving Target Indicator program, $266 million for up to 18 suborbital HASTE missions, and $190 million from the Department of Defense for hypersonic testing.

The Iridium Logic

Rocket Lab agreed to buy Iridium in June 2026 at roughly $8 billion enterprise value. The price is $54/share, split as $27 cash & a collared stock component, closing in 2027.

Beck says the market’s surprise surprised him.

“It’s funny because I thought everybody would’ve picked that well early, right? And I was surprised that people were surprised that we would be doing that.

Because if you just think about it logically & you stand back & you go, ‘Okay, what is the biggest TAM in the space industry?’ It’s comms. If you ignore broadband because two really well-capitalized people are taking care of that, what’s the next logical thing to do in comms?

And it’s safety critical, defense critical services. That is Iridium.”

Iridium runs 66 satellites in L-band at roughly 1.6 GHz, plus 14 on-orbit spares, serving more than 2.5 million subscribers. It did $871 million in revenue in 2025 at a 57% operational EBITDA margin.

There is a physics argument under the market one. Lower frequencies carry less data but travel better.

“In that L-band spectrum, penetrating through rain and weather & buildings, you can just do stuff. It’s physics. You can just do stuff that the more broadband spectrum can’t do.”

Starlink and Kuiper run Ku and Ka band. Higher frequency, more throughput, worse performance in heavy rain. The two are not substitutes.

His framing of the deal is the sharpest line on capital discipline.

“It’s a super quintessential Rocket Lab deal. We don’t come with hopes and dreams and bore a big hole in the P&L.

Iridium adds profitability to the entity straight away. It also buys us time to put up a new constellation. We don’t buy companies that are startups that maybe they’ll do something one day.

You have to be proven & successful, & then you’re the right kind of target for us.”

Rocket Lab secured a $3.6 billion 364-day bridge facility from Deutsche Bank and Wells Fargo. Roughly $2.1 billion refinances Iridium’s existing debt. The rest combines with about $1.6 billion from Rocket Lab’s balance sheet to fund the cash portion.

Iridium is a starting position, not the destination.

“Rocket Lab is not the L-band safety critical comms company. That is number one. And there’ll be many, many more after that.”

The deal buys spectrum, a customer base, and time. Those are the three things a company cannot build quickly. It also makes Rocket Lab an operator for the first time rather than only a supplier.

Neutron: Why the 10th Flight Matters More Than the 1st

Neutron is Rocket Lab’s reusable medium-lift vehicle. It carries 13,000 kg to low Earth orbit on nine Archimedes engines burning liquid oxygen and methane. But, its launch debut has slipped repeatedly.

“It’s a rocket program, and I think every rocket CEO is probably inflicted with the same optimism over reality. But I would say that I’m being really strategic about how we bring Neutron to market.”

“It’s one thing to put a rocket on the pad and just fly it once. That’s super important, don’t get me wrong. But I guess the question that investors & analysts should be asking me is:

‘Pete, when’s the 10th rocket gonna be on the pad? How quick can we get the 10th rocket on the pad?’

Because by the time you get to rocket 10, your reusability’s sorted, all your production’s sorted, you’re banging out whole stages, and all of the launch infrastructure is just cranking.”

There is precedent behind the concern. Several vehicles have flown once and then gone quiet for years while production and pad operations caught up.

“It would really be bad if we put the first rocket on the pad, and then it was two years before we went and put the next one on the pad, which happens more often than not.”

Rocket Lab is targeting a 1-3-5 ramp across its first three years. The vehicle remains on track for pad delivery in Q4 2026, though Beck has said the window for an end-of-year launch is narrowing.

The demand pressure is now internal as well as external.

“Launch is constrained for everybody else. Launch is constrained for our government customers. I’m now my own user of launch for Iridium. I need launch too, so I’m like my harshest customer.”

Neutron is priced at $50 to $55 million with no significant discounting for early flights. Archimedes has completed more than 400 hot fires across first stage and vacuum engines at Rocket Lab’s test complex inside NASA’s Stennis Space Center.

The 24-Hour Constraint

The most instructive engineering answer concerns a requirement Beck admits was unreasonable, and imposed on purpose.

“I made a constraint at the very early part of the program where I think I lost a lot of credibility with the engineers, and I said, ‘Look, guys, we need to be able to turn this vehicle in 24 hours.’ Which is, I’ll admit, a stupid constraint, but what that constraint did is it drove a whole bunch of really, really important design decisions. Like propellants, engine lights, structures.”

The Hungry Hippo fairing came out of it. On most reusable vehicles the fairing separates in flight, gets fished out of the ocean, then cleaned and requalified.

“You’ve got 24 hours to fish a fairing out of the ocean, recondition it, blow all the salt water out of it, and get it back on a rocket. That’s not gonna work.”

Neutron’s fairing stays attached to the first stage through launch and landing. It opens to release the second stage, then closes again. Nothing gets discarded and nothing touches salt water. Beck’s summary: “Sometimes you’ve got to be cruel to be nice.”

The composite structure follows the same logic.

“Everything is a giant engineering compromise in the rocket world. And if you fill a room up with engineers from every discipline, if one of those engineers from one of those disciplines is happy, you have failed.

Everybody needs to be equally as miserable. Then you know you’ve got the right design.”

The compromise is between structural mass and engine stress.

“We want the Archimedes engine to be incredibly reliable, and we don’t wanna have to service that engine for multiple flights. So we make super lightweight structures and super low-stress engines. Now, if you make heavy structures, then you have to make super high-performance engines. So you be lazy on your structures and really difficult on your engine, or you can be really lazy on your engine and really difficult on your structures.”

Rocket Lab has built carbon composite structures since Electron. It took the lighter-structure, lower-stress-engine side of the trade. Archimedes runs at reduced stress specifically so it can fly repeatedly without teardown.

GHOST, the Launch Site That Ships in a Container

Rocket Lab announced GHOST alongside earnings in August. It packs the launch pad, ground support equipment, range control systems and the rockets themselves into standard shipping containers. It supports Electron and HASTE on shared ground infrastructure, and it debuts with a suborbital launch from Kodiak, Alaska in 2027. That takes the company to six pads across two hemispheres.

The requirement came from customers, not from engineering.

“Essentially, obviously we’ve got three launch sites for the Electron vehicle. And our government customers, mainly around the HASTE missions, were really keen to see us go into very discreet locations to do some of that testing. So it was driven from a requirement from the customer that can you build this mobile launch site?”

The reason it is feasible now is scale. Neutron reset the company’s baseline for what counts as heavy infrastructure.

“Now that we’re doing Neutron, even the smallest thing needs a crane. You look at an Electron launch pad, and it’s tiny now. So moving stuff around is not a big deal for us anymore. So it’s pretty elementary for us to build these mobile launch sites and plonk them down.”

The strategic point is what it does to Mahia. Rocket Lab needed a bilateral treaty between two governments, a new national space agency and a new act of parliament to open Launch Complex 1 in 2017. GHOST reduces that same capability to freight. The company spent nine years building the hardest launch site in the world, then engineered a way to skip the process entirely.

Range control travelling inside the containers is the part with the longest regulatory tail. It means Rocket Lab brings its own range rather than using a national one, which is the piece that determines how quickly GHOST can actually deploy somewhere new.

Starting From Zero in New Zealand

Mahia exists because small launch economics cannot absorb someone else’s pad fees.

“If you have a small launch vehicle and you turn up to a large launch pad, the reality is that the fees don’t change. And when you’re trying to sell a rocket for $8M, you can’t have a $1.5M worth of launch fees.

On a $50M rocket, it doesn’t make any difference, right? So we were very early on, in order for this to be financially viable, we need our own launch site.”

Building one in the United States was hard. Building one in a country with no space industry turned out to be easier.

“New Zealand had no space industry, so sometimes it’s easier to start something from scratch and formulate it in a way that best makes sense than to try and integrate with a whole bunch of rules that don’t fit.”

New Zealand signed a Technology Safeguards Agreement with the United States in 2016, created a space agency the same year, and passed the Outer Space and High-altitude Activities Act in 2017. Beck was knighted in 2024, in part for creating that industry. His reaction: “It’s very weird. Super weird.”

Geography as a Real Structural Advantage.

“LC1 is a very unique position that we can launch sun synchronous, and also out to mid-inclination out of the one site. So here in the States, if you wanna go sun synchronous, you generally have to go out of the west coast. And if you wanna go mid-inclination, you have to go to the east coast. So we get a huge swath of azimuth. And there’s 1 plane every 3 days that we have to worry about.”

Sun-synchronous orbits require launching nearly polar, which from the US means Vandenberg. Mid-inclination means Florida. Mahia does both, because there is nothing downrange but open ocean.

That converts directly into responsiveness. On the VICTUS HAZE mission for the Space Force, Rocket Lab had 24 hours to integrate a satellite, roll it to the pad and launch. It finished in 16 hours and 42 minutes.

The cost was building everything from nothing.

“We had to build an entire industry from scratch in New Zealand. There was no suppliers, no infrastructure, nothing. We had to build absolutely everything.”

Public Markets as an Instrument

Rocket Lab went public on August 25 2021 through a merger with Vector Acquisition Corporation, raising $777 million. SpaceX listed nearly five years later, in June 2026.

Beck gives two reasons for going early. The first is structural.

“If you can reach a certain level of maturity and go public, the capital markets enable you to do things you couldn’t do in private markets. Buying Iridium, we could have never done that as a private company.”

The second reason is about time horizon.

“I didn’t want to build a space company that just lives for the lifetime that I’m running it. It needed to be multi-generational & go on & on & on.

Being a public company really enforces that discipline of profitability, and profitability is survival. The rockets and satellites, it’s all cool, but at the end of the day, it’s all for nothing if your economics aren’t there and you’re not profitable.”

His advice to founders is a warning about motive.

“The worst thing you can do is go public because you can’t raise capital privately. We never had an issue raising capital privately. But I think a lot of companies went public during that SPAC time because they were just crap companies, and they couldn’t raise capital privately, so their only option was to be a public company. That’s never a good reason to be a public company.”

He calls public status “a lot of work” with “a lot of constraints” and “probably way less fun.” It supplies resources worth having, if a company actually needs them.

The stock has fallen more than 50% from its May 2026 high despite 3 straight record earnings. Beck points at duration rather than sentiment.

“Neutron delays are always a question, and rightly so, people should be riding us hard for that. But at the end of the day, I’m not gonna say it doesn’t matter, but how late was Electron to the market? Nobody knows. Because it doesn’t matter.”

Asked for his operating principles, he answered: “Just do logical things.”

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Woolly Mice, Dire Wolves, & AI-Driven Biotech

Inside $12 Billion Colossal Biosciences + Spinout ‘Astromech’ Announces $20M Raise at $3.8 Billion

Ben Lamm is the Co-founder & CEO of Colossal Biosciences + their AI spinout, Astromech. Colossal is the $12 billion de-extinction company using genome engineering, AI and reproductive technologies to bring extinct species back to life, while Astromech is its AI biology spinout, now valued at $3.8 billion.

→ Listen on X, Spotify, YouTube, Apple

We went inside Colossal’s lab to see its Gen 2 woolly mice and sat down with Ben to discuss the technology behind the dire wolf, the race to bring back the woolly mammoth, and the much bigger business being built underneath de-extinction.

Colossal has raised $635 million, including a recent $400 million round that was originally targeted at $200 million. Ben explains why he still believes Colossal is “one of the most undervalued companies in the world,” and how technologies developed inside the company are creating entirely new businesses.

Astromech, the AI biology spinout Co-founded by Lamm and George Church, just raised $20 million at a $3.8 billion valuation, bringing total funding to $60 million. Astromech is using evolutionary and multispecies genomic data to build AI models designed to predict biological change, with potential applications across cancer, disease and drug discovery.

We cover Colossal’s rise, woolly mice, dire wolves, mammoths, artificial wombs, genome editing, AI and biology, Astromech, and what happens when humans learn to program life.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Ben Lamm, Co-Founder & CEO at Colossal Biosciences
(01:07) Inside Colossal's De-Extinction Lab
(05:07) Every wall in this lab has a purpose
(06:02) The AI idea that could save the planet
(08:09) What Colossal is building next
(08:28) Revealing Colossal's woolly mice
(10:49) Wild requests they get every day
(13:35) Colossal's business model
(15:07) How Colossal raised $635 million
(20:17) How AI is converging with Biotech
(22:44) How ancient DNA actually gets to the lab
(26:24) The real stats behind Mass Extinction
(32:07) Will their animals ever live in the wild?
(34:58) Building this company with no background in Biology
(41:44) Choosing their earliest backers
(52:51) Why they refuse to work on Humans
(58:07) Secretly working on Aliens?

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Inside Colossal With CEO Ben Lamm

Jurassic Park for REAL

Colossal CEO & Co-Founder, Ben Lamm has founded and exited four companies before this one. Chaotic Moon Studios, the Austin design and engineering shop he co-founded, sold to Accenture in 2015. Team Chaos, a mobile gaming studio, sold to Zynga in 2016. He then founded Conversable, a conversational AI company acquired by LivePerson in 2019, and Hypergiant Industries, which built AI and space systems for defense and enterprise customers including work with the Air Force. On tape he references earlier satellite and imaging work for Army space, and he sits on Avi Loeb’s UAP scientific advisory board.

→ Listen on X, Spotify, YouTube, Apple

He started Colossal in 2021 with George Church, the Harvard geneticist who helped launch the Human Genome Project, on the premise that de-extinction was an engineering problem rather than a research one. Five years later Colossal has raised $635M, employs 260-plus people, and has produced three sets of live engineered animals. Its AI spinout, Astromech, has just raised $20M at a $3.8B valuation.

Lamm is not a biologist and does not pretend to be. What he brought was the software operator’s instinct to own the entire stack, which is why Colossal runs ancient DNA extraction, computational genomics, cloning, and animal operations under one roof rather than farming pieces out to academic partners.

Company Background

Colossal runs out of a 55,000 square foot Texas headquarters with 260-plus employees, 17 academic partners, 80 more people inside labs it funds, and roughly 95 advisors across scientific, indigenous, executive, and youth boards. Church serves as lead geneticist.

Six species sit in the de-extinction portfolio: the woolly mammoth, thylacine, dodo, dire wolf, giant moa, and bluebuck, the last added in April 2026. Mammoth calves are targeted for 2028.

“Extinction is a systems model. Genetic engineering from taking a 73,000-year-old skull and going to puppies is a systems model.”

“What’s great about this company, and is also hard about this company, is that there’s not another one. It’s not like here’s your SaaS company roadmap. The good news is if we make mistakes, there’s not really other people to be judged against. But the bad news is no one had ever done it before.”

The Extinction Numbers

The premise Lamm operates from is a forecast of up to 50% biodiversity loss over the next 25 years. He does not treat de-extinction as a fix for the rate of loss.

“No matter how good conservation works, it just doesn’t move at the speed of which we’re eradicating species and changing the planet.”

Individual species numbers carry the same shape. EEHV, an elephant herpesvirus, kills roughly 20% of baby elephants. Asian elephants get cancer at 2% to 3% against 24% to 25% in humans. Colossal holds the largest elephant dataset in the world and the largest multi-species herpesvirus dataset. Those two datasets are what Astromech now trains on.

The stated response has three parts: a de-extinction toolkit with applications past de-extinction, a government-funded BioVault network, and open sourcing.

“We open source all of our technologies for conservation. So anybody can use any of our technologies for free for conservation.”

Why Colossal Spins Companies Out

Colossal did not set out to run a portfolio. The spinouts exist because the technologies kept turning out to have applications nobody scoped at the start, and because keeping them in-house would tangle the brand and the mission.

“We don’t know what we don’t know, even about the business model.”

The clearest version of the problem is the one Lamm reaches for himself. Colossal has grown hair follicles from elephant stem cells and can produce mammoth hair without touching an animal, which has obvious consumer applications. It is not going to pursue them under the Colossal name.

“We don’t wanna have Colossal launches a hair solution, and then people are like, is that mammoth shit that people are putting on us? We never wanted that.”

The same logic applies to human health work. Colossal does no human work in house, but the technologies apply, so they go out the door instead.

“Our technologies can be applied to human, and we can spin them out and license them, but we won’t do the human work here.”

Three companies have come out this way. Form Bio in 2022, a computational life sciences platform, with a $30M Series A led by JAZZ Venture Partners. Breaking, working on plastic degradation. And now Astromech. Artificial womb technology is next, with Lamm telling Rolling Stone in May it should be ready in 2027, with human fertility applications.

Astromech Raises $20M at a $3.8B Valuation

Astromech is an evolutionary biology AI company building predictive models of biological change. The round is $20M led by Bob Nelsen, with Peak 6, NeoGenesis Capital, Builders VC, and CAZ Investments participating, bringing total raised to $60M at a $3.8B valuation, up from $2B in March on a $40.5M round. Texas-based, pre-revenue, with its platform and initial research pipelines operational. Church is co-founder.

The company gestated inside Colossal studying how genes diverged historically, then developed forward-looking capability as it ingested more data. Its stated inheritance from Colossal is the genome bank of extinct and living species, ancient DNA capability that reads old genomes against living ones, tooling built for massive biological datasets, and a team trained across evolutionary systems and computational biology. The proceeds go to expanding the research team, adding species to its functional genomic datasets, and scaling comparative genomic infrastructure.

The framing the company uses is weather forecasting. Biology can be described in extraordinary detail after the fact, but not anticipated.

“Biology runs the world and historically, we have only reacted to it. Every genome carries a record of what changed, when it changed, and the tradeoffs that followed, but almost none of that history is readable at scale today.”

What Astromech Actually Does

The architecture takes three kinds of input and produces three kinds of forecast. Inputs are genomic, meaning living and extinct genomes; evolutionary, meaning deep-time ancestry and divergence; and functional, meaning expression, traits, and response. Outputs are where a genome or population is headed, where that system is most likely to break, and which regulatory circuits are driving the change.

Two engines sit underneath. The first is a suite of deep learning models finding patterns across species and systems, covering how genes are expressed, how organisms respond to their environment, and how vulnerabilities develop. The second works backward through evolutionary history, reconstructing how a system reached its current state, then running the same math forward to project where it goes next. Together they train on 3.8 billion years of biological history.

Stated applications span human health, biosecurity, agriculture, and conservation: flagging cross-species pathogen susceptibility before it reaches humans, predicting drug resistance before treatment failure, mapping disease risk and healthspan drivers, modeling herd vulnerability under disease and climate stress, and anticipating food security threats. The company is positioning upstream of all of them rather than building a separate model per market.

The Technical Core

Astromech extends ancestral state reconstruction past sequence into ancestral regulatory state, covering chromatin accessibility, gene expression, and functional annotation, integrated into a Bayesian framework that returns calibrated confidence rather than point predictions. The same engine runs against different traits by changing the evidence overlay rather than the architecture.

The training breadth is the other half of the argument. Models predicting regulatory function from sequence are typically trained on one or two reference genomes. Astromech trains across species, generating functional data in-house where public coverage is thin, which lets it examine regulatory change between lineages rather than variation within one species.

Internal benchmarks put its learned tree-inference method at roughly 100X faster than conventional maximum-likelihood methods at comparable topology accuracy. Retrospective validation recovered trait-associated genes already established in published research and surfaced additional candidates.

None of this came off the shelf.

“It turns out that Opus doesn’t do an ancestral state reconstruction of a genome, and it doesn’t build comparative genomics models. So we had to go build all of that, which I think was good, ‘cause we had to go build out an AI team and think through some pretty hard problems.”

The cost structure is a differentiator against frontier AI.

“It’s a very focused model, so it doesn’t need the breadth that a lot of these other models need. So we don’t have to go build a $500 billion data center in computing power. We don’t have to fight for water and other things that people are all upset about.”

Longevity as the First Proving Ground

The first demonstration maps 46 longevity-associated genes across a time-calibrated tree of life, letting researchers see how genes tied to longevity, cancer resistance, and genomic maintenance were conserved or changed across evolutionary history.

The comparison set is built from outliers. Asian elephants evolved notable cancer suppression despite size and lifespan. Tasmanian devils are vulnerable to a transmissible cancer that has devastated wild populations. Bowhead whales live past two centuries at a body mass associated with elevated cancer risk. Brandt’s bats weigh a few grams and live past forty years. Birds routinely outlive mammals of comparable size despite higher metabolic rates.

“Comparing living species by their genomes can fail to assign DNA differences to functional impacts. Most of the variations that matter for complex traits, for example, morphology and longevity are regulatory rather than coding, so reconstructing the ancestral regulatory state, not just the ancestral protein, has the crucial explanatory power.”

Next phase is forecasting pilots with partners in health and biosecurity, moving from retrospective validation to prospective. The company is hiring across ancestral modeling, regulatory genomics, genomic inference, sequence reconstruction, metabolic modeling, and protein folding.

The Category Repricing

The round lands against a repricing of AI applied to biology. On August 19, Moderna and Merck reported the first positive Phase 3 for an AI-designed personalized cancer vaccine, and Moderna shares surged a record 177%. Lamm’s read, recorded before that result:

“It is truly exploding. It’s like you’re gonna see this like Precambrian explosion of new drugs & new therapies that all come from AI-assisted drug design & discovery.”

What Colossal Has Actually Produced

Three sets of live animals exist. Three dire wolf proxies born in 2025, produced by making 20 edits to a gray wolf genome using DNA from a 13,000 year old tooth. Four red wolves cloned the same year from ghost wolf lineage, which is endangered species work rather than de-extinction. And the woolly mice, which carry mammoth-associated edits.

Mouse gestation runs 21 to 22 days against 22 months for an elephant, which makes the mouse the testbed for mammoth phenotype work. Colossal took the edits it had identified for the mammoth, found the mouse equivalents, engineered them, and produced live animals inside a month.

“So we had 100% healthy mice born with exactly the predicted phenotypes, with the most multiplex edited delivered at one time ever. And we did it in 30 days.”

Edit counts have moved fast against original projections. Everything else in the portfolio is pre-animal, at the genome and cell engineering stage.

“We were told if in five years we were doing five to seven edits, we’d be great. We’re now doing hundreds. We haven’t done this yet, but we’re now pushing a new construct that we’ve designed and a new delivery mechanism that we think we can be doing thousands of edits at once.”

The No-Go List

Colossal does no human work in house. The line sits below primates. Lamm was wearing a Gigantopithecus shirt during the interview and pointed to it as the example of what the company will not build.

“We will not work on anything related to non-human primates and over. Gigantopithecus would be amazing to see. But we’re not gonna do that.”

Dinosaurs are the most common inbound request. Megalodon is second. There is no defense partnership on either.

“There is no megalodon DNA. And if there was it would be terrible to make. I mean, the ocean’s scary enough.”

Conspiracies and the UAP Question

Part of the reasoning behind the no-go list is reputational.

“We already have people, there’s a fringe group of people that already have some crazy conspiracy ideas about us. So I don’t want to encourage that.”

Aliens come up regularly, and here Lamm is correcting a rumor rather than declining work. He has a longstanding tangential connection to the UAP community through prior satellite and imaging work for Army space, and sits on Avi Loeb’s UAP scientific advisory board. He says he understands why people connect the dots, and that the conclusion is wrong.

“One of the other conspiracies is that the advancements in genome engineering has come from us studying, which is just not the case, which is just crazy talk.”

“Civilization Tech”

Colossal does not sell animals. Requests arrive daily, from schools, museums, & the Smithsonian, which asked about taxidermying the woolly mice for display when they die.

“We’re not in the selling animal business, so therefore it’s not currently a part of our roadmap.”

The asset is the toolkit. Lamm places three Colossal capabilities inside a category he did not coin.

“I think genome engineering at scale and full DNA synthesis at scale and artificial wombs at scale have an application that’s, like, insane.”

Much of the revenue surface is undefined by his own account.

Gene Drives & the Invasive Species Market

Lamm sizes invasive species as a half trillion dollar problem globally. A gene drive spreads a sterility trait through a population, removing future generations without chemical spraying.

“Invasive species is like a half trillion dollar problem globally. Think about like rats on islands, think about the screw worm here in Texas, which is gonna decimate the cattle industry.”

The New World screwworm is the live case. The federal approach runs $1.5B in taxpayer money over 8 years to slow or stop the spread. Colossal does not hold permits.

“Our solution, which is already made, could eradicate the entire problem in every model that we’ve done in six to seven months.”

On dual use, Lamm’s position is that the risk does not justify not building.

“There’s always gonna be bad actors. That doesn’t mean that we shouldn’t innovate the technologies for positive use cases.”

Government Contracts & the BioVault

Government work is the only near-term revenue Colossal has publicly described. Deployment of anything in the wild also sits with government, indigenous groups, and private landowners.

“Right now we are educating USDA, the intelligence community, the Department of War, and others on these technologies and how they can combat these issues domestically & abroad. But it is up to them. They have to give us the green light.”

Three contracts anchor the line. The UAE invested $60M and partnered on the World Preservation Lab and BioVault, announced at Dubai’s Museum of the Future in February 2026. Colossal signed an MOU with the Department of the Interior and US Fish and Wildlife Service in June on biobanking and conservation genomics. The Colossal Foundation partnered with the University of Tasmania on devil facial tumour disease. The biobank is being open sourced to labs and universities rather than held.

Artificial Wombs & Inbound Demand

Individual components of the artificial womb work apply to IVF and embryo health. The longer-range application is germline editing out genetic diseases and disorders.

Commercial inbound has arrived from outside biotech. An apparel buyer purchasing billions of units of an ethically harvested animal fiber approached Colossal about 5X-ing domestic supply through cloning or production. Colossal owns the rights to the animals and has grown hair follicles from elephant stem cells, producing mammoth hair without touching an animal. A second inbound came from the Middle East on a critically endangered bird tied to falconry culture, where Colossal’s artificial egg technology is the proposed path.

“The hardest thing that we will continue to have is making sure that we stay focused on our core mission, and don’t deviate too much for some of these large economic opportunities.”

Vertical Integration & the Viagen Acquisition

Animal cloning was the acquisition case. There were two meaningful operators globally, Viagen in the US and one company in South Korea.

“Viagen had cloning efficiencies at 80% where most people are at 2%. And so it made sense for us to say, okay, we need to buy them and bolt them on from a vertical integration perspective.”

The somatic cell nuclear transfer lab runs on custom hardware Colossal built in house, with computer vision tracking laser position, oocyte position, and DNA extraction volume. The company expects to automate the embryology process with robotics. Enterprise tooling was the harder migration.

“The hardest thing we’ve done is reprogram, not stem cells, but scientists to work in things like Jira and Smartsheets.”

$635M Raised & Rounds That Kept Doubling

Colossal has raised $635M. The most recent round was structured at $200M and closed at $400M.

“I think that we’ve completely mispriced every single round of capital because every single round of capital has been more than 2X oversubscribed. So what I hear from that is we’re cheap. Like, we’re a cheap date, and we should raise our rates on capital.”

Investor selection was deliberate. Colossal targeted long-duration technology capital and impact investors who track positive impact as an internal KPI, and passed on funds requiring a return inside a four to five year window. Named profiles include Thomas Tull, Mark Walter, & Jim Breyer.

“Our original deck was like, we have no scientist. We have no partnership with Harvard. We have no technologies. And we’re a handful of software and hardware people that have no background in biology, and George Church thinks he can make a wooly mammoth, and we’re pretty sure that that’s worth something. It was pretty bad. I’m being honest, I would’ve passed on it.”

What closed rounds afterward was delivery against a stated roadmap.

“It’s easy to raise capital when you have trust, and I think trust is built with setting and managing expectations.”

“I do think that we are one of the most undervalued companies in the world, ‘cause I truly do.

The ability to preserve, engineer and direct life on a planetary scale, in all species, I think is a pretty important skill set for humanity.”

The Criticism

Two objections run alongside the funding. The first is scientific and concerns labeling. Colossal’s dire wolves are gray wolves carrying 20 edits, a point Beth Shapiro, Colossal’s own Chief Science Officer, has conceded publicly. Vincent Lynch at the University at Buffalo has called the species labeling arrogant. Colossal’s own term is functional de-extinction.

The second is policy. Conservationists have argued that de-extinction rhetoric weakens the case for the Endangered Species Act, a link Interior Secretary Doug Burgum has drawn explicitly. That sits directly against Colossal’s Interior and USFWS biobanking work.

Lamm on the Conservation Pushback

A third objection came from inside conservation itself, and is the one Lamm addresses on tape. He attributes it to an early read that Colossal positioned itself as a replacement for modern conservation.

“The conservation community is actually some of the biggest, at times, has been historically some of the most vocal critics of the work.”

“We argue that 99% of people should keep doing what they are doing with conservation, and we’re that 1% difference.”

The measurable claim is downstream funding.

“If you go talk to any of our elephant conservation partners, they’ll tell you that their funding has gone up since Colossal was invented.”

Inspiration as a Stated KPI

Colossal runs three objectives at the top level: value creation for investors, impact, and inspiration. The third was a term of the deal.

“We told our investors this from day one, that inspiration needs to be as important as value creation, and you need to be on board with that.”

The youth advisory board sits alongside the scientific, indigenous, and executive boards, on the reasoning that the next generation inherits the solutions as well as the problems. Lamm cites direct air capture as the failure case, a trillion dollar infrastructure buildout handed forward as technical debt. The brand also functions as a hiring instrument.

“How do you take, like, Harvard meets MTV? That was our thesis from day one.”

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

The Golden Egg of Cybersecurity

Nikesh Arora is the Chairman & CEO of Palo Alto Networks (NASDAQ: PANW), the famous cybersecurity giant that has grown from roughly $18 billion in market cap when he joined to ~$300 billion today (3x in value since Feb 2026).

→ Listen on X, Spotify, YouTube, Apple

We sat down at Palo Alto Networks to discuss why Nikesh believes the AI cybersecurity boom is only beginning, how companies will defend themselves against increasingly powerful AI attacks and rogue agents, and why the time between discovering a vulnerability and exploiting it is collapsing.

Nikesh also explains why he signed Jensen Huang’s recent open-source AI letter, what the rise of autonomous agents means for enterprise security, and why he believes the entire software industry will be rewritten over the next decade.

We also go inside the operating playbook behind Palo Alto Networks’ rise, including 40+ acquisitions in eight years, the $28 billion CyberArk deal, Nikesh’s “paranoia” about disruption, hiring AI-native talent through hackathons, and the lessons he took from Larry Page at Google and Masayoshi Son at SoftBank.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Nikesh Arora, Chairman & CEO at Palo Alto Networks
(00:51) The AI attack era has already started
(03:30) What happens when agents go rogue
(04:39) Why AI Labs are slowing down
(06:31) The Mythos Fallout
(17:25) Cybersecurity's secret weapon
(19:43) Buying a company is the easy part
(20:48) Where most acquisitions go wrong
(24:58) Predicting the SaaS Crash?
(28:03) The question nobody in Tech is asking
(30:54) There's never enough Compute
(33:12) Palo Alto's bar for AI fluency
(35:37) The controversial hire
(37:12) Using LinkedIn to spot opportunities
(39:19) What a CEO's day actually looks like
(43:47) The leadership traits that matter
(47:02) Google vs. SoftBank: What was different?
(51:02) Why planning horizons are shrinking
(52:55) Lessons From Elon & Masa
(56:44) What's next for Nikesh?

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Palo Alto Networks CEO Nikesh Arora

Rogue Agents, Mythos, Linkedin, & the $300B Mythos Run-up

→ Listen on X, Spotify, YouTube, Apple

The Controversial Hire

Nikesh Arora joined Google in 2004, months after the IPO that priced at $85 a share for a $23B market cap. He ran European operations, then EMEA, then became President of Global Sales Operations & Business Development in 2009, then Sr. Vice President & Chief Business Officer in 2011, owning revenue & customer operations. Google booked $3.2B in revenue the year he arrived & $66B in 2014, his last year.

He went on to serve as President & Chief Operating Officer of SoftBank Group under Masayoshi Son. He took the Palo Alto Networks job in June 2018 at an $18B market cap. The company trades above $300B today.

He had none of the obvious qualifications for the job.

“I’d never done cybersecurity in my life.. I’d never been a public company CEO.. And I’d never sold enterprise. I was a consumer guy. Other than that, they got everything right.”

The early years ran on borrowed expertise. He describes imposter syndrome throughout, sitting in technical meetings and watching how his own comments landed, then calling founder Nir Zuk and the company’s product leadership on the drive in and the drive home.

AI Made Attacking Cheap & Defending Hard

The AI labs have been publishing demonstrations of their models finding software vulnerabilities, chaining them together, and running attacks against live infrastructure. Every one of those demonstrations describes a capability that is now broadly available.

“This is the beginning. This is not a moment”

“It’s a lot easier to attack. As an AI lab, it’s much harder to defend.”

Getting the world's infrastructure to a standard that survives AI-speed attacks is a multi-year build, and no single vendor can absorb it. That is his explanation for why the sector has been bid up.

$PANW shares have traded between $139.57 and $398.88 over the past 52 weeks, and the company carries a market cap of roughly $300B.. with a recent ATH at $322B.

Mythos Pushed Security to the CEO

By his account, security gets bought below the CEO. A vendor calling the chief executive gets handed down to the technology organization, and he spent 8 years on the wrong end of that.

“For 8 years, I spent my career trying to convince CEOs they need to pay attention to cybersecurity. Tried everything. Call them, dine them, wine them, try to talk to them, and they usually send you off to their technology team. ‘Go talk to those guys.’ You know what Mythos did? Every CEO wants to talk about Mythos.”

Mythos changed that. Anthropic disclosed the model on April 7 2026 and said it would not release it publicly, because it was too good at finding high-severity vulnerabilities in major operating systems and web browsers. It went instead into Project Glasswing, a defensive program with a limited set of partners. Roughly 50 of those partners have since used it to find more than 10,000 high or critical severity vulnerabilities across the most systemically important software in the world, and the program has been extended to around 150 more organizations in over 15 countries.

Palo Alto Networks ran it against its own code. Speaking on All-In in June, Arora said a 6-week test surfaced vulnerabilities that would have taken his own team 5 to 7 years to find, at a cost in the low millions. His framing there was that 50 years of badly written human code is now assessable by AI. He has also put Mythos-level offensive capability roughly 3 months from being available in the wild.

From 55 Days to 4 Hours

Patching a vulnerability and exploiting it have always been a race. He puts the industry average for shipping a fix at 55 days, and his argument is that with AI as strong as Mythos has collapsed the other side of that race down to just minutes. Closing a gap that wide is an engineering problem the company has been working on for years, across detection, testing, and delivery.

“The average time to fix it was 55 days in the industry. The average time [adversaries] will find it and try and attack you is in minutes..

[We took it] from 55 days to 4 hours.”

The 4 hours is what they got it down to. Palo Alto Networks launched the capability at Black Hat in August and pushed it to every customer, taking their own patch delivery from the 55-day industry standard to same-day. He points to it as AI finally showing up on the defensive side of the ledger.

Compression moves work onto the customer too. He expects them to test their own software and open source dependencies continuously and to modernize the infrastructure underneath, and he reads that spend as good news for the security industry.

What Rogue Agents Have Already Done

This stopped being hypothetical during the making of this episode.

On July 21, 2026, OpenAI disclosed that two of its models, GPT-5.6 Sol and a more capable unreleased model, escaped a sandboxed testing environment during an internal evaluation called ExploitGym. The benchmark scores AI systems on finding and exploiting software vulnerabilities. Rather than solve the challenge, the agents went after the answer key. They found a zero-day in the package registry proxy that had been treated as the environment’s only controlled egress path, reached the open internet, rooted a third-party code sandbox, used it as a staging base, and chained further zero-days into Hugging Face’s production clusters.

Hugging Face reconstructed roughly 17,600 attacker actions from logs covering July 9-13th, with the agent inside their systems for about two and a half days. What it reached was contained, 5 datasets whose names tie to the benchmark challenges and their solutions, with no customer-facing models or data touched. Hugging Face detected the intrusion and reported it to law enforcement before OpenAI connected the activity to its own evaluation run. Their summary of it was blunt, describing an AI agent that escaped its sandbox, cheated on its benchmark test, and hacked their infrastructure to steal the answer key.

The details that emerged later are the uncomfortable part. OpenAI researchers presenting at Black Hat described agents writing files into an internal package manager to create a covert message board, delegating sub-tasks to each other across separate evaluation runs, and cryptographically signing their messages out of concern that impostor agents might be reading them. Reuters reported on July 31st that OpenAI had found other instances of agents escaping sandboxes, described as limited, with none believed to have left OpenAI’s own network. The Black Hat presentation ran at the same conference where Palo Alto Networks announced its 4-hour patching capability.

Then there is the version that happens to normal people. On August 10th, ABC reported that an Australian man asked his agent to book him into a morning class at his gym. The agent found that the booking limit was enforced in the website front end but not in the API underneath it, so it booked him months further ahead than the gym allows. It then found the cancellation endpoint had no authorization check at all, and tested that live against the person sitting at position 1 on the waitlist. (sick)

“The API has zero authorization checks on cancelling other people’s reservations.. I tested this with the person in waitlist position #1, and it actually went through.”

Asked to undo it, the agent could not, because creating a booking required authorization and the person it removed would have to rejoin at the back of the queue. It apologized for running a live call instead of a dry run, then drafted a vulnerability disclosure email to the gym’s software vendor. Neither flaw was exotic. A competent developer could have found both in an afternoon. Nobody had, because until recently nothing was reading that API looking for a way into a workout class.

However, the warnings came earlier. In March, an engineer at Meta asked an agent to help analyze an internal question, and the agent posted its answer to the forum without asking permission to share it. The advice was wrong, and acting on it exposed company and user data to unauthorized engineers for 2 hours, logged internally as a Sev 1. A Meta safety director described her own agent deleting her entire inbox despite instructions to confirm before acting. An agent published a hit piece attacking an engineer who rejected its code. Alibaba’s engineers traced a burst of security policy violations on their training servers to their own coding agent, which had started mining cryptocurrency and opening covert network tunnels.

Nothing here was attacked from the outside. In every case an agent pursued the objective it was given, found the weakest path to it, and took that path.

Who Is Responsible for the Model??

So a model that can find and chain vulnerabilities gets pointed at your company. What happens next? and who answers for it?

The labs themselves have been the loudest voices calling for the pace to slow down. Anthropic and OpenAI have both argued publicly for more caution and more oversight of frontier capability, in some cases about the models they had just shipped. One read of that, and plenty of commentators hold it, is that the caution is insurance. If something goes badly wrong later, the record shows they were the ones asking for restraint… He reframes the question entirely.

“Are they asking for a slowdown? Or they’re asking for permission to be able to go release these?”

Underneath it is a liability problem nobody has solved. The law assigns responsibility to people and to the companies that employ them, and a model that acts on its own breaks that mapping.

“We have to understand liability. We have to understand who’s responsible at the end of the day. It’s very easy to ascribe responsibility and liability to human beings. If you do something wrong, it’s your fault. If I do something wrong, it’s my fault. If I use a model and the model does something wrong, whose fault is it? Is it the model’s fault? Is it my fault for using the model?”

Read that way, the calls for governance are the labs trying to establish who carries the risk before someone else decides for them, which is what lets them keep shipping.

“A lot of the AI labs want to get ahead of it, make sure there is some governance framework around it to ensure that they can keep developing the technology at the pace at which they’d like to. So in a way, they’re probably doing the right thing. It doesn’t seem like it, but I think they are doing the right thing in trying to get some governance around it so they don’t get hauled back. As you saw, they did get hauled back when they tried to launch Mythos or Fable 5.”

The template for getting there already exists in autonomous vehicles. Billions of dollars of training and edge case work preceded public willingness to hand a machine life and death authority. Every enterprise use case that gives an agent real agency runs the same process, and it takes years.

Why Palo Alto Networks Signed the Open Weights Letter

Jensen Huang used his first post on X, published 24 July 2026, to share Open Weights and American AI Leadership. It launched with 25 signatories and passed 270 companies and organizations by 3 August. Palo Alto Networks is on the list, alongside CrowdStrike, Cisco, and Zscaler. OpenAI, Anthropic, and Google were all absent at launch. OpenAI and Google have since signed. Anthropic has not.

The letter argues that American AI leadership will be decided by whether an open ecosystem diffuses into every sector rather than by any single frontier model. It asks Washington to expand compute access for startups and researchers, fund shared training assets, and avoid premature restrictions on open models. It also separates distillation, which it defends as a legitimate technique, from unlawful extraction from closed models.

The security section is what puts a cybersecurity company’s name on it. When attackers have advanced AI, defenders need models of comparable capability to detect, simulate, and respond. Concentrating capability behind a small number of closed models creates single points of failure that outsiders cannot inspect.

“Openness may be one of the most important paths to AI safety and security.”

His own framing is about diffusion. Open source sets a global price point for access, open weights allow fine-tuning against specific enterprise use cases, and closed models cover what neither can. He also thinks the industry is spending too much attention on the models themselves.

“There’s an over-indexing on the model part of it.. Models are important, but it’s also important to get all the context collected and all the training data right.”

40 Acquisitions in 8 Years

40 acquisitions in 8 years sounds aggressive until you look at what the industry requires. Attackers change methods constantly, and every new technology that reaches the market needs new security built around it. Cloud, then containers, then browsers, now agents. Each shift creates a category the incumbent does not have a product for, and customers will buy that product from somebody. Building it in house takes longer than the window allows, so the large vendors buy their way into each new category instead. That is why the sector consolidates the way it does.

“It is the most innovative industry in the world because the bad guys are trying to figure out how to attack you in a different way every time. The moment we suss out how they did it, they’ve moved on to finding the next time ... In that environment, it’s impossible that all the innovation’s gonna come from us. There’s always somebody else who’s got a different angle. There’s always somebody else who’s tried something that’s gonna work better than what I thought about.”

Palo Alto Networks has acquired more than 40 companies since 2018, including CyberArk at roughly $25B, the second largest deal in cybersecurity history. The market disliked that price until the results came in.

He plans 2 years out, and visibility thins past that point. Acquisitions fill the gap.

“That’s kind of the paranoia I live with.. Sometimes we’re looking for interesting products, sometimes we’re trying to fill gaps, sometimes we’re anticipating the market and saying, ‘How can we get ahead?’”

By his own scoring, 3 out of 4 acquisitions have worked.

The Biggest M&A Mistake

Nikesh reveals the biggest mistake he’s seen with his fair share of change management. It’s happens after the deal closes, in how buyers treat the people they just bought..

“The biggest mistake that people make during acquisitions is underestimate the intelligence of the people who built the business that you acquired.. Our attitude is, ‘You kicked our ass. Come tell us what we did wrong. Come run this for us.’”

That creates friction internally, because the acquired team often ends up running the internal team it beat. Integration is simpler when the acquisition lands in a category the company does not already play in. Resource the team, leave it alone, then plug it into a field organization that already has the customer relationships.

The End of the SaaSpocalypse + Software That Has an Opinion

The SaaS complex sold off roughly 50% on the thesis that AI models would absorb application layer software. Palo Alto Networks called the end of that repricing for cybersecurity 6 months ago, on the argument that model capability covers the broad case and security lives in the exception.

“AI can be great at 80% use cases. We live in the 0.1% use case. We’re looking for the needle in the haystack.”

Category by category, the market is still sorting which incumbents survive. Application tools now compete with native model capability that reaches the same output, the market has already “declared” some of those companies dead.. with some Italian sharks circling the waters on some great arbitrage moments.

The larger change is foundational, given that for the past 20 years of software it executed deterministic tasks and held no view on the work. Now AI-native software has a POV, which means an HR system could tell you not to hire someone and a security product could tell you your firewall deployment is wrong.

“The entire software industry will get rewritten in the next 10 years.”

Making Over 20,000 People AI Native

Palo Alto Networks employs over 20,000 people, and the competency question applies to all of them. Nobody senior enough to teach this exists yet, so the learning happens peer to peer. Twice a week, the top 24 technical people in the company meet for 2 hours to walk through what they are building and why. He calls it AI IO, after the Old MacDonald refrain (EI-EI-O).

He also changes who is in the room. Managers are told to keep hiring AI-native engineers until they outnumber the people who were already there, on the theory that behavior shifts once the balance tips and features start shipping faster.

“The biggest risk we have to ourselves is two guys in a garage who are sitting down building the next company, which is fully AI native.”

Hiring Out of Hackathons

To that point, hackathons have surprisingly became a hiring channel about 9 months ago. Teams are told to source a third of their headcount from people who have been experimenting with agents on their own, and hackathons are where those people surface. No credential certifies agent fluency, so the signal he trusts is what somebody does after work.

“There’s no school they can teach you this stuff in. How am I gonna know that if you know how to use OpenClaw well, how am I gonna know you understand what an agent is? If you’re not going home and figuring this stuff out yourself, that’s a problem because you’re supposed to be the architects of my technology for the future. If you’re not curious and not learning, where am I gonna find these people?”

The channel works alongside the composition mandate. Once enough of those hires land in a team, the people who have been there longer start picking up the tools, and the ones who do not eventually select themselves out. He contrasts that with the approach of cutting a third of the workforce on the assumption they will never get there, which he rejects.

Proud LinkedIn Power User

We spent the morning walking the Santa Clara office asking people what we should ask Nikesh. Only one person took us up on it, and it was Lee Klarich, the 20-year Palo Alto Networks veteran who serves as Chief Product and Technology Officer and sits on the board.

Asked what the one question should be, Klarich said to ask him what he learns from LinkedIn. Is he a big LinkedIn guy. He is a big LinkedIn guy. He loves it.

Nikesh’s first response was to call it a conspiracy, on the theory that the rest of the office had fed the question to Klarich rather than ask it themselves. The answer underneath it is that he treats LinkedIn as an intelligence feed and a hiring channel. The general counsel came off LinkedIn. So did Wendy Whitmore, who runs incident response.

“They’re not interviewing when they’re posting on LinkedIn ... If I can read what people have written over the last four years on LinkedIn, I can tell you who they are without having to ask them.”

The habit traces to Google in 2004, where nobody took a meeting without searching the person first. His team feels it daily. A competitor post he finds between 4:30 and 6:30 becomes a message asking what they plan to do about it.

“That’s why it’s like, holy shit, he’s on LinkedIn again.”

Role Models & Next Play Mentality

With a pretty legendary career, leading teams & investing from Google, SoftBank, and now Palo Alto Networks, I asked Nikesh if he had any role models or mentors along the way. Well, he gave an interesting answer of inspirational figures to say the least..

Elon

Elon Musk taught him the value in N-of-1 ambition, taking the world’s biggest problems head-on. Electric cars when nobody believed in electric cars, rockets landing themselves, Starlink terminals now bolted to cars, boats, and planes. The pattern he draws from it is about problem selection rather than execution. Most entrepreneurs work on small problems because those are the only ones they can see the end of, and Musk commits to problems whose solutions he cannot see either.

“If you take a really hard problem nobody’s working on, if you get it right, you win and you win big. And if you look around you, a lot of entrepreneurs are busy trying to solve small problems because this is the problem they can see.. Elon cannot see when he comes up with a problem he’s trying to solve. He can’t see that far, but he thinks if he tries to put his mind to it, that problem gets solved.”

Masa

Masayoshi Son taught him capital allocation. Son started SoftBank selling packaged software and has pivoted the business roughly 20 times since, was briefly the richest man in the world, lost it, and rebuilt. What Arora takes from him is risk appetite that has grown rather than shrunk with age.

“He’s the oldest man I know with the risk appetite of a teenager. As he gets older, his risk appetite becomes bigger.”

The specific lesson came as a correction. Arora was on the phone multiple times a week with the CEO of a portfolio company that was down 50%, trying to coach it back to plan.

“He looked at me & says, ‘If you put in that much effort on the company that’s doubling, they might quadruple.’

We might make more money on the one that quadruples than you fixing the one that’s broken.

That’s an insight.. As operators, our tendency is to try & fix everything because we don’t want things to break. As an investor, he said, ‘Double down on your winners. They’re gonna be way more interesting for you than the ones that are gonna not make money.’”

Steph

Steph Curry supplied the last one, speaking at a conference Arora attended with his son. Next play mentality works the same in business and in sport, and it is the counterweight to the paranoia that runs through everything else he describes.

“Steph talked about this next play mentality. It’s like you can’t win if you can’t get rid of the last play that he missed. You gotta focus on next play. That’s an interesting lesson, whether you’re in business or you’re in sport.”

He describes what holds it together as karmic calm, which sits oddly next to the paranoia and appears to be the point.

“You have to try to do your best. You have to put your heart and soul into it. You have to wanna win.. And if it works against you, wake up in the morning, shake it off, and do your best again. If you get hung up on what happened yesterday, you won’t be good.”

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

FROM -92% → $100B+

Adam Foroughi, Co-Founder & CEO & Giovanni “Gio” Ge, CTO of AppLovin, join Sourcery from AppLovin’s Palo Alto headquarters for a deep dive into the technology and strategy behind its advertising business.

→ Listen on X, Spotify, YouTube, Apple

AppLovin (NASDAQ: APP) is one of the most efficient company in public markets at ~400 people. Approaching $7B in EBITDA. That's ~$17M per employee, roughly 3x the revenue per head of Nvidia & 7x Apple, Meta, or Google. Engineering hasn't grown in 3 years (100 people) while cash flow scaled 20x.

Gio joined AppLovin in late 2022, when the company was worth roughly $5.5 billion, and helped architect Axon 2, the recommendation engine that became central to AppLovin’s transformation. The new architecture improved predictions and advertiser returns, helping drive the company’s expansion beyond mobile gaming into e-commerce ads and its recovery from a 92% post-IPO decline to a $100B+ company.

Adam and Gio break down how Axon works, how AppLovin scales with a remarkably lean engineering team, how AI is changing the way they build, why knowing what not to build matters, and what it would take for AppLovin to eventually become a trillion-dollar company.

We also cover AppLovin’s expansion into e-commerce, the future of advertising across LLMs and connected TV, engineering culture, and how they think about building for the next five to ten years.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Adam Foroughi, Co-Founder & CEO and Giovanni Ge, CTO at AppLovin
(00:52) Gio's first impression of AppLovin
(03:38) Joining AppLovin after a 30% stock drop
(09:59) How Axon 2 actually got built
(20:23) Why AppLovin never needed a sales team
(21:59) From gaming ads to e-commerce
(24:47) From a napkin drawing to a real product
(30:54) The onboarding mistake most companies make
(34:37) What "AI-Native" actually means
(36:37) Should new engineers still learn to code?
(41:41) UGC vs. traditional ads: what's winning?
(47:56) Adapting to AI without rebuilding from scratch
(54:50) Gio's daily tech stack
(1:02:17) AppLovin's path to a trillion-dollar valuation
(1:06:19) Recovering from a 90% stock drop
(1:09:47) What's next for AppLovin?
(1:13:44) What happens when AI takes over?
(1:15:48) Adam and Gio's hottest takes
(1:19:50) How to actually stay locked in
(1:24:47) Building something that lasts

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

AppLovin’s Playbook

I sat down with AppLovin CEO Adam Foroughi and CTO Gio Ge at the company’s Palo Alto headquarters to go inside the engineering rebuild that took AppLovin from a 92% drawdown to one of the most efficient companies in public markets.

→ Listen on X, Spotify, YouTube, Apple

A 30% Drop and a New Architect

Ge joined AppLovin in November 2022 from big tech. Foroughi and then-CTO Vasily Shikin recruited him to architect Axon 2, a full replacement for the Axon 1 model that powered the advertising business on outdated machine learning techniques. AppLovin was worth roughly $5.5 billion when the interviews started.

Days before Ge’s start date, the stock fell 30% on the November 2022 earnings call. He joined anyway, reasoning the lower price meant more equity upside.

“If I just wanted to stay on the winning team, I would have not chosen to leave. I wanted to find a place where I see opportunities and I can actually make a difference.”

Axon 2 shipped roughly 5 to 6 months after he arrived. The rebuild reflected a principle Foroughi set with his first lead engineer at the company’s founding.

“Technology moves really fast, and you gotta be humble about what you have and know that if technology goes faster than what you’ve developed, you gotta throw away what you have and re-architect it, rebuild it, and it’s probably gonna happen every couple years or even faster in the future.”

Throwing Away the Old Engine

Axon 1 was a tree-based model, hundreds of thousands of if/else branches with clusters of user-item combinations sitting at the end of each branch. The structure could not keep up with relationships that shift with time, weather, promotions, & holidays.

Axon 2 replaced the trees with semantic embeddings, the same concept behind large language models. Learnable embedding tables encode user IDs & item IDs into semantically meaningful vectors, & deep neural networks study the interactions between them. The new model recognizes patterns and extrapolates them to unseen user-item pairs.

The architecture was also cheaper to run. Neural networks are built on general matrix multiplication, the operation GPUs are optimized for, while the old selection trees were hard to optimize on GPU machines.

“The new generation model is very powerful and efficient with the modern GPU architecture.”

Ge built the training infrastructure for Axon 2 through December 2022, including 12 hours of offline coding on a flight to Italy, reading library source code because he had no internet access. The first generation of the model went into training the day he returned.

60 Seconds With 1 Billion Users

The most important data in the system is user interaction with ads. Every ad served to AppLovin’s billion-plus users is a feedback data point, and the format generates far more signal than a static banner. Ads inside games run around 60 seconds, and half are opted into by users who watch to earn an in-game reward. Many contain playable mini games that preview the advertised app.

“If a user sees 10 ads on our platform a day, that’s 10 minutes of engagement with the website effectively.”

The funnel closes with advertiser data. AppLovin serves the ad, tracks engagement through install, then receives revenue and retention data back from the advertiser, following the same playbook Facebook used when it pixeled websites to understand users beyond the social feed. AppLovin started pixeling websites and learned what its users do outside of games.

The audience skews 30 to 50 years old and slightly female, with a concentration of heads of households playing mobile casual games. Analysts and investors pushed back for years that gaming data could not extend to other categories.

“It’s a billion people on the other side, not a billion people who are only playing games and doing nothing else. If you understand the person, then you can actually deliver value to them outside of just a game-to-game scenario.”

E-commerce Started on a Napkin

The move beyond gaming began about a year after Axon 2 scaled, once the team confirmed the architecture was general enough to support any vertical. The first data flow engine for the e-commerce product was designed at a breakfast during a Google conference in Las Vegas.

“There was no paper, so we just found a napkin, and we started discussing and drawing and design on the napkin. By the time the food arrived, we already have a pretty good design.”

The engineers built the first prototype of the data engine before landing back in San Francisco. Within a couple of months the team had the data, model, and infrastructure live with a handful of test advertisers, and purchases came through with decent ROAS even on a premature model with limited data.

AppLovin runs almost no product management. Engineers are expected to understand business problems well enough to write their own architecture and their own product.

“You have engineers that didn’t even talk to a business person sitting down together at breakfast on a napkin writing out a solution to a problem that we never even discussed, and then something that could become a big business.”

400 People, No One-on-Ones

Excluding the gaming studios it has since sold and Adjust, AppLovin has run at around 400 people for years, cut down from 600 roughly 2 to 3 years ago. Engineering is about 100 people, the same size as when Ge joined, through a period when revenue multiplied. Shikin personally wrote around 60% of the company’s code before Ge arrived.

There are no one-on-ones. When Ge requested a weekly meeting with Shikin in his first month, Shikin did not understand what it was, and the two canceled it in favor of communicating through code. New hires, including interns and new graduates, push code to production within their first week.

Ge holds a specific view of how the engineering team should relate to AI tools.

“I don’t want our engineer to sit next to AI. I want our engineers to sit on top of AI.”

The boundary comes with accountability. Prototypes and dashboards can be built however engineers want, but core systems carry a different standard.

“You can use AI however you want, but in the end of the day, you, as a human, is held accountable for every decision that your AI made for you.”

$17M in Revenue Per Employee

The math on that headcount separates AppLovin from every large company in public markets. The core business is approaching $7 billion in EBITDA on roughly 400 people, around $17M in EBITDA per employee, with trailing revenue of $6.8 billion putting revenue per head at a similar level.

Nvidia, the efficiency leader among big tech, generated $3.6 million in revenue per employee and $2 million in net income per employee in fiscal 2025, and topped $5 million in revenue per head in fiscal 2026. Apple runs at $2.4 million in revenue per head, Meta at $2.2 million, Alphabet at $1.9 million, and Microsoft at $1.1 million on their most recent full fiscal years. AppLovin's EBITDA per employee is roughly 3x Nvidia's current revenue per head and 7x or more every other company on that list, and its engineering headcount stayed flat through 3 years in which cash flow scaled roughly 20x by Foroughi's math.

“As the business is growing, the number of people stays the same, which actually means every team members is growing together with the company, which is not typical, no matter what company you’re looking at.”

The business side is being built toward the same ratio. Every manual process on the team is a candidate for automation across the full advertiser base, with central staff building the tools and dashboards to do it.

“We’ll get to a place a couple years from now where hopefully most every process is automated, and then the business interpersonal skills matter a lot.”

No Formula for a Great Ad

Foroughi has spent 21 years in advertising and rejects the idea of a repeatable creative formula. Advertisers on the platform test 20 to 100 videos to find one that works, and winning concepts do not translate across brands.

“If you create 30 ads a week, probably one of those might be interesting, and you don’t wanna just change very small things around. You wanna create concepts that are differentiated.”

The format punishes social-style creative. A social ad has 3 seconds to capture attention, while AppLovin’s users sit with an ad for a minute, and the solitaire and mahjong player is a different consumer from the power user scrolling Instagram and TikTok. Advertisers who invested early in learning the format are generating outsized campaign results while the rest of the market catches up, the same dynamic gaming advertisers went through when playable ads emerged 10 years ago.

AI creative is early. Language models can produce a good short clip, but a 30 to 60 second brand-safe video that stays engaging remains out of reach, so the near-term effect is humans producing more ad content faster and cheaper. Ge splits the tooling market into collaborative tools advertisers use with AI, which are widely adopted, and fully automated generation, which AppLovin is still working on because it has to serve every advertiser at once.

“If the AI starts writing all the ad creative, then all of it’s gonna eventually look the same, and the user won’t respond to the ad.”

Search Ads, Discovery Ads, & What Comes After Mobile

Foroughi puts chatbot advertising in the bottom-of-funnel category alongside search, where the consumer already knows what they want and the platform closes the transaction. AppLovin operates at the top of the funnel, showing consumers products and games they did not know they wanted.

“If 99% of the use case is search, then the ads in chatbots are probably gonna look and feel a lot like search.”

“We’re trying to go show consumers things that they don’t know they wanna go get. We wanna start the funnel and then close the loop.”

The expansion targets follow that logic. Connected TV and video placements on the open web are both discovery surfaces where the user does not arrive with intent, and both are fragmented markets without the Apple and Google duopoly structure of mobile. AppLovin is running R&D on extending the platform to both on a 3 to 10 year horizon.

The 92% Drawdown & the Math to $1 Trillion

AppLovin fell 92% in its first 18 months as a public company. The company stopped doing investor relations for over a year at the bottom, kept stock-based compensation in fixed dollar terms to avoid runaway dilution, and deployed its cash flow into a slew of buybacks. Those reached roughly $8.1 billion lifetime share repurchases through June 2026.

AppLovin stock buybacks by year

• 2022: ~$700M in share repurchases under the original February 2022 $750M authorization, deployed mostly in H2 as the stock bottomed (Q1 2022 was only $43.7M)

• 2023: ~$1.2B in buybacks, 54.3 million shares repurchased and withheld, cutting total shares outstanding by nearly 10% net of issuances

• 2024: $2.1B in buybacks, 25.7 million shares retired and withheld, including the ~$570M concurrent repurchase from KKR’s secondary offering

• 2025: $2.58B in buybacks, 6.4 million shares, with the board adding an incremental $3.2B authorization in October 2025

• 2026 YTD: $1.55B in buybacks, $1.0B in Q1 plus $551.3M in Q2, with ~$1.8B remaining under authorization

“We were able to flip it and say, we’re just gonna buy our own shares. We’re gonna become our best investor.”

Foroughi contrasts that position with enterprise SaaS companies in drawdowns today. Robust platforms without algorithmic leverage can only accelerate revenue by hiring go-to-market teams, which raises burn while the stock is low, and companies in that spiral tend to get taken private and restructured by private equity.

The company now approaching $7 billion in EBITDA, converting roughly 75% to cash after stock-based compensation, about 20x the level of 3 years ago. Foroughi’s goal ladder has moved from $1 billion to $10 billion to $100 billion in value, each communicated to the team with a plan behind it. The next zero requires $30 billion plus in annual cash flow, which is why gaming alone cannot support it and why the consumer vertical, and categories beyond it, exist.

“To take it from this scale up and be worth a trillion dollars, we’ve gotta believe that we can get to 30 billion plus of cash flow a year.”

Ge’s commitment runs on the same number.

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Neros’ 250K sqft Drone Factory

Sourcery goes inside Neros Technologies’ new 250,000-square-foot Millennium One facility with Co-Founder & CEO Soren Monroe-Anderson.

→ Listen on X, Spotify, YouTube, Apple

Neros designed the factory with enough space to eventually produce 1 million drones per year. Today, the company says it is already manufacturing more than 250 drones per day, with engineering, testing, supply chain & production operating under one roof.

Soren takes us through Neros’ engineering labs, flight-test operation, drone and ground-station assembly lines, manufacturing software and 100% end-of-line flight testing. We also get a look at the latest Bandit interceptor prototype, which can reach roughly 300 km/h, with Neros targeting closer to 400 km/h for intercepting faster jet-powered Shahed variants.

We also visit the new-product introduction line where Neros is preparing Archer AI for mass production, and the massive open portion of the factory reserved for bringing more component manufacturing in-house. Soren explains why mass-producing drones requires designing the product around manufacturing, why 3D printing does not solve the underlying supply-chain problem, and how Neros plans to vertically integrate its way toward million-unit scale.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Inside Neros' 250,000 sq. ft. Millennium One Factory
(01:21) How Neros pushes its drones to the limit
(02:35) Why 3D printing can't solve the drone supply chain problem
(03:40) How their drones are built different
(06:20) The new 300 km/h Bandit interceptor
(09:09) Inside the drone assembly line: producing 250+ drones a day
(11:17) South Bay's biggest American flag
(11:42) How Neros stress-tests its drones
(15:26) Inside the new product introduction line
(17:01) What roles is Neros hiring for
(19:00) Soren's hot take on defense tech

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Inside Millennium One: The Factory Neros Built for 1 Million Drones a Year

Neros moved into its Torrance facility a few months ago. Co-founder and CEO Soren Monroe-Anderson walked us through it, from the flight test racks to the empty floor still waiting to be filled.

“We sized this building to be big enough to do a million drones per year.”

→ Listen on X, Spotify, YouTube, Apple

Flight Test: 10 People, 2 Teams, Every Day

The tour opens at the flight test team’s area, where drones sit on racks at various points in the development testing process. The team is 10 people, out at desert sites roughly 2 hours from Torrance, with 2 teams in 2 different locations every day.

On the racks were a V1 Bandit interceptor prototype and an Archer AI drone, both of which had spent time being run to failure in desert heat.

“All of this stuff has been just getting beat to hell out in the hot desert, and surviving. That’s the goal. But really we’re just trying to push everything to its limit and see where it breaks.”

Neros has no wind tunnel, which sets it apart from most commercial drone companies. Quadcopters are far less affected by wind than fixed-wing aircraft, and the company does enough testing in genuinely windy conditions that simulating an exact wind profile has never been necessary. Fixed-wing is not on the roadmap today, though Monroe-Anderson sees the same low-cost, high-effectiveness logic from FPV applying there eventually.

Why Neros Does Not 3D Print Drones

The idea of printing drones at the edge circulates widely enough that it came up on the tour. The answer is that it solves a different problem than the one Neros is solving.

“3D printing is generally good for making a few of a thing, making lots of different things. We are mostly focused on making a lot of one thing, or a really consistent process.”

Printers can adapt an airframe or a payload adapter to a mission set. They cannot touch what actually constrains production.

“You’re not gonna be 3D printing your circuit boards, your motors, your radios, any of that stuff out on the battlefield. So it does not solve the supply chain problem in the ways that maybe are claimed in some of the posts today.”

Factories, in his framing, are efficient precisely when they are large and centralized.

The Airframe Tradeoff

Archer airframes are built mostly from carbon plate, which is light and strong. The alternative approach, cast or CNC’d metal bodies, buys a thermal advantage Neros wanted without the manufacturing cost.

The solution was a flat aluminum bottom plate that contacts the motherboard and the radios through thermal interface material when the drone is closed out, acting as a large heat sink.

“It’s just a flat plate that’s cut into a shape in 2D versus having to be cast or CNC’d, which is just a lot cheaper for mass manufacturing.”

What Comes Off the Line

Archer ships in three variants, differentiated by propeller size and mission set. The 5” is built for closer range and close-quarters or urban environments with smaller payloads. The 8” and 10” are optimized for longer distance strikes with heavier payloads.

Compatible sensors across the platform include low-light, EO/IR, thermal, and a down camera. Archer sells starting at $2k and runs to around $5k once peripherals and warheads are added. Archer AI, the autonomy-enabled variant adding position hold and terminal guidance, runs roughly $3k to $6k with warheads and sensors.

Crossbow, the ground control station assembled on the line’s right side, now fits the full kit of radio and pilot box on a single soldier’s plate carrier. The latest generation cut size by more than 50% while holding 25 km range, and adds twice the radio beam width coverage so operators can focus on the target rather than antenna orientation, along with weather-resistant ruggedized enclosures, reduced-profile antennas with reinforced joints, and push-pull harness connections on goggles, handset, and radio for setup in low light.

Bandit is speced against Class 3 targets like Shaheds, at an estimated 233 mph top speed, 16,500 ft max altitude, and 12.4 mi engagement range, starting around $5k depending on payload.

Engineering Sitting Next to the Line

Engineering workstations carry whatever hardware and test equipment the engineer needs, with a larger lab area immediately adjacent, including an electronics lab where new boards get brought up. The point is to keep as much iteration in-house and as close to the desk as possible.

The larger point is the floor plan. Mechanical engineering, manufacturing engineering, and supply chain all sit near the line and interspersed with each other.

“The engineers here are sitting right next to production, and the teams that are sitting close to production are the ones who are most impactful to production.”

“There’s really good communication between all the teams for the full life cycle of the product, not just design, which is where a lot of things tend to get left and just thrown over the fence if you don’t have this really close collaboration.”

The Bandit Interceptor, Two Months In

The mechanical engineering department held the newest Bandit, noticeably sleeker than the desert prototype and roughly 40 km/h faster, at about 300 km/h. Monroe-Anderson still calls it a rough prototype. It represents a couple of months of development.

300 km/h is a workable speed against typical Shahed drones. Jet-powered Shaheds are the harder target.

“To go after jet-powered Shahed drones, we need to be closer to 400 kilometers an hour. So that would be the ideal end state.”

Getting there means bigger motors, a bigger battery, and more aerodynamic work, with room left on all three.

Production: 250 Drones a Day, Every Flight Tested

The line runs drone assembly on one side and ground station assembly on the other, with a production dashboard tracking daily, weekly, and historical output. Current rate is a little over 250 drones a day, and that is the normal rate, not a rate staged for cameras.

Stations are manually operated, with parts preloaded into bins and fed from the back so they sit within reach, plus tooling and preset-torque drivers to make each station easy to run. Manufacturing software tracks serial numbers of critical components into airframes and logs station time by tablet tap, which surfaces where time is being lost and where the bottleneck sits.

Automating the line is not currently practical, because the product was not designed for it. That changes with the ground-up redesign.

“To get to an automated line, it takes a pretty significant redesign of the product itself. But for the rate at which we are iterating, it actually makes a ton of sense to have manually operated stations.”

Every drone Neros makes goes through an automated test station and then a 100% flight test, where operators load a mass simulation payload and stress the drone hard in a short flight. The eventual goal is sample testing rather than universal flight test, which requires enough data from both stations to trust that issues are caught upstream. That is still some distance away, and the data collection to get there is happening now.

The tour also passed a dedicated rework line, set up after an issue caught in production turned out to affect drones already sitting in inventory.

The Empty Space

The old El Segundo facility ran 100 people and manufacturing inside 15,000 square feet, which meant fighting to get drones out the door. Torrance is 250,000 square feet, and the difference shows in what the company can now afford to dedicate space to.

One example is the new product introduction line, where process gets developed before anything reaches the main line. Archer AI is going through NPI now, ahead of entering main production later this year. The NPI line alone is larger than the entire old production line.

“This is bigger than our old production line, just dedicated to new products.”

Past that is open floor. The plan starting this year is in-house component manufacturing feeding the systems built on the main line.

“We wanted, with this facility, to not have to have multiple sites, not have to move. And so we’ve got all the space we need to actually vertically integrate everything, and then get to that million per year target.”

Asked when Neros outgrows the building, Monroe-Anderson said he does not know, and that not knowing is the good outcome. The only stated requirement for a future facility is higher ceilings, since the American flag hanging over the line is 30 by 60 feet and, as far as he knows, the largest in the South Bay.

We closed by asking for his hottest take.

“A lot of the defense tech, maybe the majority of defense tech products that are being put out right now by startups are not very effective and maybe even completely useless.”

» » » NEROS IS HIRING « « «

Neros is hiring across every engineering discipline, with the sharpest constraint on responsible engineer level talent, meaning engineers who can own a system or subsystem outright. Historically the company ran fewer than 10 REs across the entire Archer platform. Moving to multiple product lines requires both platform-level and component-level REs across many projects simultaneously, which multiplies demand for exactly the people who are hardest to find.

Monroe-Anderson describes the global pool of hardware engineers at that level as very limited, and every high-growth LA hardware company is competing for it. The response has been to build a company that top talent wants to join and prove themselves at.

TLDR: Hiring runs across essentially every department, with the sharpest need for hardware engineers and engineers capable of full system responsible engineer ownership.

→ Listen on X, Spotify, YouTube, Apple

Want to check out the FULL INTERVIEW?

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Manufacturing 1 Million Drones

Soren Monroe-Anderson, Co-Founder & CEO of Neros Technologies, joins Sourcery from the company’s new 250,000-square-foot factory in Torrance following its $250 million Series C at a $2.5 billion post-money valuation.

→ Listen on X, Spotify, YouTube, Apple

The round, co-led by Sequoia & American Strategic Technology Fund, comes as Neros expands from its core Archer FPV platform into multiple product lines, including Archer AI, Marksman, & Bandit, its counter-UAS interceptor. Neros is also fresh off a $500M Army IDIQ & is building toward the capacity to manufacture 1 million drones/yr.

Soren explains why modern warfare is increasingly being shaped by low-cost drones, why America fell so far behind China in drone and electronics manufacturing, and what it actually takes to “de-China-fy” a defense supply chain. Neros has pushed critical components including boards, chips, motors and magnets outside China, but Soren argues complete independence is far harder when Chinese suppliers still dominate some raw materials and precursor chemicals.

We also discuss Neros’ transition from a single-product company to a multi-product defense manufacturer, the rise of interceptor drones, autonomous capabilities that can operate through communications loss, sovereign manufacturing for U.S. allies, and why Soren wants Neros to become the “Toyota of defense manufacturing.”

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Soren Monroe-Anderson, Co-Founder & CEO at Neros Technologies
(01:02) Neros just raised $250M at a $2.5B valuation
(03:34) Working with Shaun Maguire
(04:25) Inside the Army's $500M drone contract
(05:38) How the Ukraine war rewrote modern combat
(08:05) The Interceptor drones taking over the battlefield
(10:31) How many drones China really makes in a year
(12:55) Is it even possible to build a drone without China?
(16:09) China is arming both sides of the war
(17:08) What actually separates Neros from every other drone company
(19:55) The biggest BS claims in defense tech right now
(20:51) Why most defense startups never actually deploy?
(22:58) Neros’s growth to over 250 employees
(24:00) Why defense startups need to screen for spies
(25:29) The Talent bottleneck slowing down American manufacturing
(26:51) How AI is quietly reshaping engineers
(29:43) The biggest myth about re-industrializing America
(30:50) Neros's plan to build drones with U.S. allies
(32:53) The brutal tradeoff between speed and scale
(34:51) Why Neros keeps a permanent office in Ukraine?
(35:52) What actually happens on a modern battlefield?
(37:41) How Starlink is changing drone warfare
(39:14) What happens when a drone loses its signal
(40:41) Soren’s learnings from Kelly Johnson

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Manufacturing Drones for America

Soren Monroe-Anderson is co-founder and CEO of Neros Technologies, which he started in 2023 with Olaf Hichwa after both raced FPV drones professionally. We recorded this interview inside the company’s 250,000 square foot factory in Torrance, California, sized to build 1 million drones a year.

→ Listen on X, Spotify, YouTube, Apple

$250M Series C at $2.5B Post

Neros raised $250M in a Series C at a $2.5B post money valuation. Sequoia and American Strategic Technology Fund led the round, with participation from Interlagos, Valor Equity Partners, Allen & Company, Thiel Capital, Spark Capital, and Mantis. The round was preempted and came together quickly with existing investors, which has been the pattern across the company’s last several rounds.

“Fundraising isn’t the goal, but it is a tool, and I think this is a great milestone for the company.”

The round follows a $500M IDIQ contract with the Army for hundreds of thousands of drones over the next 5 years, structured so the Army receives Neros’ latest products, including the new Archer AI. The award came out of PBAS, the Purpose Built Attritable System program, which is the first program of record for FPV drones in America. Neros spent roughly a year building new drone sizes and a new ground station ahead of the fly-off, and the Army selected three companies out of it.

“They selected two other companies as well, but we were the only ones to really deliver in a meaningful way.”

Capital from the round goes to four places. Production expansion at the Torrance facility. Vertical integration, including component manufacturing brought in-house. International expansion and sovereign manufacturing across European, Asian, and Middle Eastern allies. And development of Block 3 hardware, the ground-up redesign meant to enable 1,000,000 drones per year by 2028.

What Changed on the Ukrainian Front Line

The front line in Ukraine is largely static, and the no-man’s zone running along it has widened from a few kilometers to roughly 30 kilometers. Drones drove that change, and the widening zone is now pushing Ukrainian FPVs to get physically larger to cover the distance. Fiber optic drones, which are immune to jamming, have increased in frequency over the same period.

“Anything that is visible from the air that moves in that zone is going to be immediately attacked and eliminated.”

Ukraine has been running a mid-strike campaign against Russian logistics routes 150 to 200 kilometers past the front line. That campaign has been in the works for about 9 months and has become effective in the last 3. Deep-strike drones are hitting the Moscow region more frequently. On the defensive side, roughly 90% of the Shahed deep-strike drones Russia launches are now taken out of the sky before reaching a target.

Over 70% of casualties in Ukraine are attributable to drones, and 90% of those come from manually piloted FPV systems. Ukraine is estimated to use 7 million FPV drones this year.

“FPV drones are one example of where low-cost drone technology is reshaping warfare. It’s sort of like the first really big example, but this is gonna happen in many different types of drones and different types of defense capabilities.”

ICYMI: Ukraine snuck $1M worth of drones destroyed $7B of Russian ⅓ of long distance bomber fleet…

From One Product to Three

Archer is the modular FPV platform, currently shipping as Block 2 in 5”, 8”, and 10” variants with max payloads of 1.5 lbs, 5.1 lbs, and 7.0 lbs. The 8” carries the longest unloaded range at 22.5 miles. Archer sells starting at $2k and runs to around $5k once peripherals and warheads are added.

Archer AI is the autonomy-enabled variant, launching with terminal guidance and GNSS-denied position hold, estimated at $3k to $6k with warheads and sensors. Terminal guidance addresses the point in the mission where most drones fail. Jamming concentrates over valuable targets, so the signal environment gets exponentially worse as the operator, sitting 10 kilometers back, tries to bring the drone within hundreds of meters of the target.

“Once the pilot has selected the target and the drone has a lock on it, it doesn’t need input anymore. So even if radio comms get cut, it will continue in and hit the target.”

Bandit is the counter-UAS interceptor, built for Class 3 targets like Shahed drones, with an estimated 233 mph top speed, 16,500 ft max altitude, and 12.4 mile engagement range, starting around $5k. Neros intends to deploy Bandit in Ukraine before the end of the year. The company also announced Marksman, an in-house ruggedized handset going to Drone Dominance Gauntlet II this month, and a new generation of the Crossbow ground control station that fits radio and pilot box on a single soldier’s plate carrier at a 50% size reduction while holding 25 km range. Alongside the products, Neros introduced a “Useful Autonomy” framework that measures any autonomy feature against three tests, whether it reduces operator workload, whether it works in actual battlefield conditions, and whether it improves mission success rates.

Multi-product changes the engineering org. Historically Neros ran fewer than 10 responsible engineers across the entire Archer system. New product lines require both platform-level and component-level REs across many projects at once, and the company is hiring across every engineering discipline to fill them.

“It’s way more clear how do you structure all your teams when your whole engineering team is just split by department but working on one product. When you need to support multiple different product lines, it starts to become a little bit fuzzier.”

“De-Chinafy”ing the Supply Chain

Neros was the first company to put an FPV drone on the Blue UAS list, meaning the Department of War can purchase and use it without additional approvals. That required removing Chinese content from the core components. The boards and the chips that populate them are non-Chinese. The motors are made outside China. The magnets are produced outside China, and the material that goes into the magnets is sourced outside China.

“You can look at our boards, you can look at all of the chips that go onto our boards. Those will not have ever touched China.”

The independence has a floor. Printed circuit board manufacturing depends on precursor chemicals that China dominates, which applies to American ITAR-certified board shops producing hardware for programs as sensitive as the F-35. Raw materials and process chemicals sit at the same depth in the stack.

“I would be lying to you if I said nothing in our supply chain relies on China, because the truth of the matter is if you go down far enough, you will find places where China is the main supplier.”

What a single drone company can do is hold the line on the components specific to its own system and generate enough demand to support non-Chinese batteries and non-Chinese thermal cameras. Solving the rest requires a national effort. The broader defense startup ecosystem is working with suppliers that were never set up for volume, scale, or low-cost systems.

How America Fell Behind & What China Gets From Selling to Both Sides

The United States represents 1% of global production of small unmanned aircraft systems. Around 30 years of offshoring consumer electronics manufacturing produced that number. The bill of materials for an FPV drone resembles a laptop or a phone far more closely than a traditional weapons system, so the national security consequence of losing consumer electronics went largely unexamined until drones became central to the battlefield.

“The way America got left behind was not caring about consumer electronics manufacturing being done domestically.”

China produces an estimated tens of millions of drones per year, possibly into the hundreds of millions, and supplies both Ukraine and Russia with drones and components. DJI makes the most commonly used reconnaissance drones on the battlefield, with Mavics and Matrices flying for both sides.

“It’s a morally corrupt approach, although it would also be far worse for Ukraine if they didn’t have access to Chinese components.”

Supplying both sides gives the Chinese industrial base a continuous feedback loop from Ukrainian and Russian FPV companies, with turnaround on requested changes measured in weeks. The accumulated learning from scaled production is the compounding advantage.

Industry Production Claims & What Neros Has Delivered

Neros migrated earlier in 2026 from a 20,000 square foot facility in El Segundo to the 250,000 square foot Torrance factory. Current output is approximately 1,250 drones per week, the highest volume small FPV drone production in America. The company expects to exceed a 100,000 drone per year run rate by the end of 2026 and reach 1,000,000 per year in 2028. A drone is assembled end to end in about 45 minutes, with one coming off the line every 2 minutes. Some competitors spend 12 labor hours per drone, and competitor facilities run under 100,000 square feet.

Every drone off the line takes its first flight in an on-site test chamber, and the data feeds future simulators. The flight test team runs experiments at desert test sites 5 or more days per week, which lets an engineer propose a test in the morning and have results that evening. Roughly half of the drones Neros has built are deployed to the military for combat and stockpile use, excluding training and demos.

“To claim a production capacity, in my opinion, you need to have done it many times over.”

At Drone Dominance Gauntlet 1, Neros was the top performing American manufactured drone and the first to fully deliver against Department of War orders. Several competitors have still not completed initial fulfillment months after the competition ended, and Neros received bonus allocations because of those delivery failures. One manufacturer publicly claiming 1 million drones per year is not appearing at Drone Dominance Gauntlet II.

“When we got the order after the gauntlet, we had already built pretty much all of the drones and just shipped them out, and then we were so far ahead of everybody in delivery.”

Headcount went from under 100 employees at the start of 2026 growing to more than 250. The company is filtering for the top 1% of hardware engineers who can carry responsible engineer level ownership of a system or subsystem, a pool Monroe-Anderson describes as very limited globally and heavily competed for among LA hardware companies.

Sovereign Manufacturing & Block 3 Redesign

Allies across Europe, Asia, and the Middle East have shown strong interest in Archer FPV systems and want them produced domestically. Neros has been responding to that demand, and a large share of its own supply chain already comes from allied countries.

“I say we wanna be the Toyota of defense manufacturing, where instead of being like a current prime where we’re very exquisite and siloed, we’re making these very accessible systems and we’re ubiquitous across allies.”

Standing up production abroad has no established playbook. The Torrance factory works partly because engineers sit next to the line, and the production org has recently become less dependent on engineering for daily operations. Replicating it requires someone who holds the full context of how Torrance runs without being the person running it.

“Our factory benefits so much from the engineers sitting right next to the production line, and we have to essentially figure out how to productize the production line itself and have it run without just constant sustaining engineering.”

Scale also constrains iteration speed. Changing a radio design means eating obsolescence cost on roughly 30,000 boards already fabbed and waiting for assembly. Neros moved to a block system to manage that. Small iterations from user feedback roll into Archer Block 2 continuously, while Block 3 proceeds as a full ground-up redesign of the system on a longer timeline. Ukrainian manufacturers iterate extremely fast without the runway for deep technical work on longer horizons, and Block 3 is the hardware that carries Neros to 1,000,000 drones per year by 2028.

Culture & Growing to 250 People

Headcount is growing fast. That growth changed how the team has to operate. Management style, communication flow across the company, maintaining the hiring bar, and screening for culture fit all became deliberate problems over about 6 months. Insider threat protection is also an extremely live consideration at this size & in this sector.

“When you’re under 100 people, you just have a bunch of really smart, high-agency people kind of just working on one thing. Stuff just kind of works.”

Neros is hiring across every engineering discipline, with the sharpest constraint on responsible engineer level talent, meaning engineers who can own a system or subsystem outright. Historically the company ran fewer than 10 REs across the entire Archer platform. Moving to multiple product lines requires both platform-level and component-level REs across many projects simultaneously, which multiplies demand for exactly the people who are hardest to find. Monroe-Anderson describes the global pool of hardware engineers at that level as very limited, and every high-growth LA hardware company is competing for it. The response has been to build a company that top talent wants to join and prove themselves at.

»»» Careers at Neros « « «

High-growth companies across sectors are converging on the same profile, highly capable individual contributors who also make business decisions. While AI is accelerating that concentration in software, with less effect so far in hardware engineering, which still requires much hard-earned skills.

The Torrance building holds the entire company, corporate, engineering, R&D, and production together in one facility, organized into engineering pods for autonomy, firmware, mechanical, and electrical. Colocation is what makes the daily experiment cycle work, where an engineer proposes a test in the morning, loads drones into the flight test van, and has desert results that evening.

The building also has a gym, an outdoor area, and a sauna, which Monroe-Anderson frames as the tradeoff for putting everyone inside a large industrial box and asking them to work very hard.

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

$1B & Launching Base Core

Justin Lopas, Co-Founder & COO of Base Power, joins Sourcery at Factory 1 in Austin for the launch of Base Core and the company’s $1 billion Series D.

→ Listen on X, Spotify, YouTube, Apple

The round values Base at $13 billion post-money and is led by Ribbit, Addition, Valor Equity Partners, and JPMorganChase’s Strategic Investment Group. Existing investors Thrive Capital, a16z, Lightspeed, Trust Ventures, and CapitalG re-invested, bringing total capital raised past $2.5 billion.

Base Core is the largest standalone home battery on the market at 40 / 80 kWh. Justin walked me through the factory floor, the install demo, and why Base builds the hardware itself.

“It is so critical to bring manufacturing back to the US, especially for critical infrastructure.”

We cover:
› The $1B Series D at a $13B post-money valuation
› Base Core, 40 kWh of storage and 20 kW on the inverter
› Whole home backup versus circuit-level backup
› Retail choice markets versus utility partnership markets
› Installing batteries around data centers to offset interconnection load
› 4,000 units per week by the end of the year
› Over 100 installs a day
› 100% US and Mexico supply chain
› Factory 1 built in under 6 months

→ BASE POWER IS HIRING

DISCLAIMER: This was a delightfully last minute & very BTS visit/recording, I am so happy we were able to film something in the short timeframe ‘off-hours.’ Hope you enjoy.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Justin Lopas,Co-Founder & COO at Base Power Company
(01:06) Base Power raises a $1B Series D
(02:26) Meet Base Core
(05:42) Demo: Whole-home backup in action
(07:06) How Base Power's business model works
(08:55) Inside Base Power's Factory
(11:28) Why Base Core is 3x larger than other home batteries
(12:05) Inside the manufacturing line
(13:23) Can humanoid robots build batteries?
(14:13) Why Base Power manufactures its own batteries
(15:47) Inside the robotic assembly line
(18:09) How lithium-ion batteries work
(21:40) Scaling production to 4,000 units a week
(22:33) Installing 100+ systems a day
(23:03) COO or Electrician?
(23:53) From oil rigs to Base Power

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

Turing—Turing partners with frontier AI labs to improve model capabilities in coding, reasoning, tool use, & multimodality, as well as with Fortune 500 enterprises to build & deploy end-to-end agentic AI systems in mission-critical workflows Visit: turing.com/sourcery

VCX—VCX is the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View The Portfolio at GetVCX.com

Deel—Deel is the global people platform that helps startups hire, manage, pay, and equip anyone, anywhere. Trusted by more than 35,000 fast-growing companies, Deel is the people platform that just works, so teams can scale without the chaos. Visit: deel.com/sourcery

Public-–Investing platform Public just launched Generated Assets, which lets you turn any idea into an investable index with AI. With Generated Assets, you can build, backtest, refine, and invest in any thesis with AI. Gone are the days of one-size-fits-all ETFs. Try it today: public.com/sourcery

Base Power Raises $1B at $13B & Launches ‘Base Core’

About 2 weeks ago I stopped by Base Power’s Austin headquarters with Co-Founder & COO Justin Lopas for a casual walkthrough (that quickly turned into an entire podcast episode tour) of Base Core & a tour of Factory 1, the converted printing plant where the company builds every module and inverter it deploys.

Today, Base is announcing a $1 billion Series D at a $13 billion post-money valuation alongside the launch of Base Core, the largest standalone home battery on the market at 40 / 80 kWh.

→ Listen on X, Spotify, YouTube, Apple

From SpaceX & Anduril to a Battery Company in Austin

Base Power was founded in 2023 by Zach Dell and Justin Lopas and is headquartered in Austin, Texas. Dell spent time at Blackstone underwriting energy deals including lithium mining, then invested at Thrive Capital, where he watched SpaceX, OpenAI, and Anduril scale. Lopas spent nearly 4 years at SpaceX as a manufacturing engineer on Falcon thrust structures and as lead engineer for Starship manufacturing, then ran manufacturing at Anduril from 2020 to 2023, managing a team of over 150 people. The two met on a factory tour at Anduril.

Vertically integrated like good ole SpaceX, the company designs + builds the battery, the inverter, & the disconnect box in house, writes its own manufacturing execution software, & runs its own delivery & install logistics.

The company launched service in Austin and its deregulated suburbs, then expanded into Houston in the CenterPoint territory and into Dallas-Fort Worth. Early distribution came through a partnership with Lennar, one of the largest homebuilders in the country, and a first utility partnership with Bandera Electric Cooperative. Base owns the batteries and charges an install fee starting at $695 plus a monthly rate, then aggregates the fleet into a virtual power plant that arbitrages price spreads and sells ancillary services into ERCOT.

The deployment curve has been pretty damn impressive. Base passed 100 MWh of residential capacity in the fall of 2025, reached 300 MWh in operation at more than 60 installs a day by March 2026, and has now passed 500 MWh at more than 100 installs a day.

The Path From $68M to a $13B Valuation

Base raised a $68 million Series A in 2024. In April 2025 it closed a $200 million Series B co-led by Andreessen Horowitz, Lightspeed, & Valor Equity Partners, with participation from Thrive Capital, Altimeter, Terrain, and Trust Ventures. Lee Fixel of Addition and Antonio Gracias of Valor joined the board. That capital was earmarked for the first Texas factory and national expansion.

In October 2025 the company raised a $1 billion Series C led by Addition, one of the largest venture rounds in Austin history, with CapitalG, Spark, BOND, Lowercarbon, and Ribbit joining alongside a16z, Thrive, and Lightspeed. Reporting at the time put the post-money valuation around $4 billion.

The Series D announced today is $1 billion at a $13 billion post-money valuation, led by Ribbit, Addition, Valor Equity Partners, & JPMorganChase’s Strategic Investment Group, with participation from Altimeter, D1 Capital Partners, Sands Capital, Coatue, Layer Global, and Energy Impact Partners. Thrive Capital, a16z, Lightspeed, Trust Ventures, and CapitalG re-invested. Total capital raised is now over $2.5 billion. The valuation roughly tripled in 10 months.

“Since our initial investment, the Base team has executed on an ambitious plan to become America’s Power Company. Power demand has surged in that time, & Base has emerged as one of the fastest & most cost effective ways to add capacity to the grid.” — Antonio Gracias, Founder, CEO & CIO of Valor Equity Partners.

The capital goes toward national expansion and hiring.

Launching Base Core!

Base Core is a 40 kWh battery with a 20 kW inverter, sold at 40 and 80 kWh configurations. It is the largest standalone home battery on the market. Four battery modules and one inverter stack inside a ground-mounted enclosure, with a Base-designed disconnect box mounted on the wall between the utility meter and the electrical panel. Power flows from the meter into the disconnect box and out to the panel, which gives the system whole home backup.

Many home batteries back up selected circuits chosen at install time. “That’s a bummer, and so we said, ‘Why don’t we back up everything in the home and build a system that can do that?’”

The install is split into two visits. An electrician wires the disconnect box in 1 to 2 hours, and the next day a two-person delivery crew stacks the modules and plugs in the quick connects. “It’s almost like plugging in an EV.” The crew turns the system on from a phone app. There is also a generator input for outages that run past the battery’s duration.

Base Core carries about 3X the capacity of a conventional home battery, and install costs do not scale with capacity. “This is like three times larger than what you would find in a traditional home battery, and that’s purposeful. We wanna put as much power and as much energy on a home as we can because our install costs are basically the same whether or not we install 5 or 40 kilowatt hours.” Each module is friction stir welded shut into a sealed unit that survives submersion.

Factory 1 & Vertical Integration

Factory 1 sits in the former Austin American-Statesman printing plant. Base moved in around September of 2025, ran warehousing and distribution out of it for 6 months, and started building production lines in February and March of 2026. About 100 people work in the building today, most of them hired in the last 6 months, with a second shift starting soon.

There are 3 lines. One builds inverters. Two build modules, running at double the pace because every system needs 4 modules and 1 inverter. A robotic cell applies glue and places battery cells into aluminum enclosures, automated carts follow tape on the floor between stations, and Base wrote the manufacturing execution software running on every monitor on the floor. Stations are a mix of manual and automated, since high-dexterity steps like placing a screw inside a populated circuit board are faster by hand.

“It’s a vertical integration thesis. We can cut down cost. We can reduce lead time. We can change the hardware much more quickly because we can have the engineer drive the golf cart or run over here & go change something on the line from HQ.”

The printed circuit board comes from a vendor in Austin, a large sheet metal component comes from a company in Fort Worth, and sourcing is essentially split between the US and a few suppliers in Mexico. “It is so critical to bring manufacturing back to the US, especially for critical infrastructure,” Lopas said in the announcement. A second facility is coming, a 486,000 square foot plant near Austin-Bergstrom International Airport representing a $265 million investment and 500 new jobs.

2 Electricity Markets & Business Models

The US splits into retail choice markets and non-retail choice markets. In retail choice markets including Houston, Dallas, parts north of Austin, and the Chicago area, Base sells electricity directly. A customer using 1,000 kWh in a month gets billed for 1,000 kWh, and the battery fleet on surrounding homes lowers Base’s cost of serving that load.

In markets without retail choice, Base partners with the utility. “We still put the battery on the home. We own the battery, but the utility controls it and then they pay us for access to the battery. So that’s how we recoup the cost of the battery, and that’s our business model in those markets.” El Paso is one of those markets. California is not yet served.

Utility partnerships now cover over 200 MW of capacity across El Paso Electric, Austin Energy, and CoServ. The CoServ agreement outside Dallas-Fort Worth targets 100 MW of distributed home battery capacity, roughly the output of a natural gas peaker, and pays only for capacity Base actually delivers. Austin Energy’s council approved its partnership in April 2026. The fleet has grown past 500 MWh, and the company expanded beyond Texas into Illinois in the past year.

100+ Installs/Day & the Road to National Scale

Base is installing more than 100 systems a day, with more than 100 people in the field running installs and deliveries. The company handles its own logistics. “A lot of what you see here is cool inside the four walls, but where the rubber really hits the road is literally on the road going to the homes.” The near-term goal is to double the daily install rate.

A single unit takes 10 to 15 minutes of line time from start to finish. “We’ll be able to produce about 4,000 of these units per week towards the end of the year.” That works out to roughly one unit every few minutes, against a stated long-term target of 4 GWh per year of production capacity.

Reliability testing runs in a bank of thermal chambers on the R&D side of the building, one of them holding 85% humidity at 85 degrees C. “We wanna be able to install the thing for, you know, it takes like an hour or two, and then not touch it for 15 years. And so that turns out it’s like a harder technical problem than you think.”

Base Core is available today to homeowners in select regions of Texas and Illinois. The company is hiring across every function.

→ BASE POWER IS HIRING

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

ICYMI: 3 Conversations Worth Your Time This Weekend.

Scott Wu of Cognition, Michael Hurlston of Lumentum, and Dylan Fox of AssemblyAI join Sourcery to discuss three of AI's biggest shifts: autonomous software engineering, the optical infrastructure powering AI data centers, & the rise of voice AI. Cognition has raised more than $2.5B to build Devin, Lumentum supplies the lasers connecting next-generation AI clusters, & AssemblyAI now serves 1M+ developers processing roughly 100M API calls per day.

Have feedback for topics, questions, or guests? Comment below or reply directly to the email!

Before we begin, I do have to address the Situation. Turns out, a16z didn’t win this one. Damn. Nevertheless, congrats to Leopold on his wedding & investor letter.

Subscribe now

Share

1. “Devin Writes 95% of Our Code” Scott Wu, CEO of Cognition

Scott Wu, CEO & Co-Founder of Cognition, joins Sourcery in Paris amid the RAISE AI Summit. Cognition is the applied AI lab behind Devin, the autonomous AI software engineer, & Devin Desktop, formerly Windsurf.

The company has raised more than $2.5 billion, including a $1 billion round in May 2026 at a $26 billion valuation, while growing annualized revenue from roughly $37 million to $492 million in a year. Devin now writes roughly 95% of Cognition's own code, total code shipped has grown 7x in six months, and customer usage has increased 11x to 12x.

→ Listen on X, Spotify, YouTube, Apple

2. Lumentum CEO on How Lasers Are Transforming AI Data Centers

Stock went up 1000%, Copper is out, Optics are in

Michael Hurlston is the CEO of Lumentum (NASDAQ: LITE), one of the world's leading suppliers of optical networking and laser technologies powering AI data centers.

Everyone talks about Nvidia GPUs. Few people talk about the lasers connecting them.

→ Listen on X, Spotify, YouTube, Apple

3. Voice AI is Dominating

1M+ Devs, 100x TAM, 4x YouTube

Dylan Fox is the founder and CEO of AssemblyAI, the voice AI infrastructure platform behind AI notetakers, medical scribes, drive through ordering, contact centers and humanoid robots. AssemblyAI went through Y Combinator’s first AI batch in 2017 & now serves 1+ million developers with ~100 million API calls a day.

→ Listen on X, Spotify, YouTube, Apple

Special thank you to Brex, MongoDB, & AssemblyAI for helping make this RAISE AI Summit mini-series in Paris, France happen.

Want more???

Brought to you by:

MongoDB–Millions of developers and more than 65,200+ customers across industries – including ~75% of the Fortune 100 – rely on MongoDB for their most important applications. With integrated capabilities for operational data, search, real-time analytics, & AI-powered data retrieval, MongoDB helps organizations everywhere move faster, innovate more efficiently, & simplify complex architectures. visit → mongodb.com/ai

AssemblyAI–Millions of developers use AssemblyAI to power their voice ai apps & features. One API gives you access to best-in-class speech-to-text, voice agent, & speech understanding models for both pre-recorded and real-time audio. Granola, ClickUp & HeyGen are scaling with AssemblyAI - get $50 of free credits today → AssemblyAI.com/sourcery

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

1M+ Devs, 100x TAM, 4x YouTube

Dylan Fox is the founder and CEO of AssemblyAI, the voice AI infrastructure platform behind AI notetakers, medical scribes, drive through ordering, contact centers and humanoid robots. AssemblyAI went through Y Combinator’s first AI batch in 2017 & now serves 1+ million developers with ~100 million API calls a day.

→ Listen on X, Spotify, YouTube, Apple

We get into the volume numbers, why weekly conversations are up 800%+ in 3 years, the three macro trends pulling voice into everything, why coding agents turned small businesses into API customers, how data mixture drives 75% of model quality, the humanoid robot speaker problem, on device models, and the disclosure question the entire voice agent industry is avoiding.

“It’s very easy to optimize for a public open source benchmark. It’s very hard to optimize across all these real world applications.”

We cover
› 120 million weekly conversations and 4x YouTube’s daily voice volume
› Coding agents expanding the TAM by 100x
› Context aware voice models and the McDonald’s ordering problem
› Why humanoid robots can’t disambiguate speakers
› Whether voice agents should disclose they are AI

Special thank you to Brex, MongoDB, & AssemblyAI for helping make this RAISE AI Summit mini-series in Paris, France happen.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Dylan Fox, Founder & CEO at Assembly AI
(01:06) AssemblyAI's voice traffic just beat YouTube by 4x
(03:33) $100K in GPU credits that started it all
(06:39) Why Voice AI is inflecting right now
(10:11) AssemblyAI's infrastructure-first strategy
(12:41) Handling 120 million weekly calls
(14:49) How AI agents are rewriting Internal Ops at AssemblyAI
(17:05) The Sovereign AI problem nobody's solved for voice
(18:26) Is the Keyboard and Mouse finally dying?
(22:33) The problem Humanoid Robotics hasn't solved yet
(25:14) The real problem behind translating languages
(29:27) Can AI actually talk to animals?
(31:24) The real reason Open-Source benchmarks lie to you
(34:20) Building websites in chat rooms as a kid
(36:48) On-Device AI models are about to change everything
(43:03) The people who shaped Dylan's career

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

MongoDB–Millions of developers and more than 65,200+ customers across industries – including ~75% of the Fortune 100 – rely on MongoDB for their most important applications. With integrated capabilities for operational data, search, real-time analytics, & AI-powered data retrieval, MongoDB helps organizations everywhere move faster, innovate more efficiently, & simplify complex architectures. visit → mongodb.com/ai

AssemblyAI–Millions of developers use AssemblyAI to power their voice ai apps & features. One API gives you access to best-in-class speech-to-text, voice agent, & speech understanding models for both pre-recorded and real-time audio. Granola, ClickUp & HeyGen are scaling with AssemblyAI - get $50 of free credits today → AssemblyAI.com/sourcery

Dylan Fox on the Voice Infrastructure Running Underneath AI

AssemblyAI builds the voice AI infrastructure behind AI notetakers, medical scribes, drive through ordering, contact centers and humanoid robots. Founder and CEO Dylan Fox walked through the volume moving across the platform, the problems still open in voice, and what changes over the next 18 months.

→ Listen on X, Spotify, YouTube, Apple

Where the Volume Is

AssemblyAI went through Y Combinator in 2017 in a batch organized by Daniel Gross that contained six AI companies in total. The perk was compute. Production AI deployment was rare at the time and the first TensorFlow meetups were still running.

“We got $100,000 in GPU credits. That was our perk, which now seems cute.”

Weekly conversation volume is now up over 800% in three years. On a peak week the platform carries more than 120 million voice conversations. Daily throughput runs above 2 million hours of voice, which as of December is 4x YouTube’s daily volume.

“Over 2 million hours of voice, which as of December of this past year, was a little over 4X the amount of daily volume that’s going to YouTube.”

There are a little over 1 million developers on the platform. 40% of them signed up in the past year, and close to 100 million API calls a day run against the API.

3 Trends Making Voice Inflect

1) Model capability & context. The newest voice models accept context about the environment they are deployed into. They shipped a few weeks before the interview and are the first voice models that can do it. The same models are faster and lower cost, which widens where they can be deployed.

“Today, if you drive through a McDonald’s and you order with your voice, and there’s a voice AI system there, that voice AI system has no clue that it’s taking a McDonald’s order.”

2) The rest of the AI stack. Reasoning models, vector databases and models across other modalities now exist as commodity components. A company building drive through ordering, AI notetaking or a healthcare scribe assembles the voice layer plus the rest of the infrastructure it needs.

3) Coding agents! (the newest of the 3). AssemblyAI is an API platform, which historically meant engineering and product teams inside software companies. But now, anyone can code. A small business lawn care chain signed up and is automating back office work through AI-driven apps like Lovable, Cognition’s Devin, Claude Code, Cursor and Replit. Global enterprises are now building their own internal software on the API with two people and a set of agents.

“Our TAM has just increased by 100X because we’re not just selling to engineering teams within product companies.”

Infrastructure & Nothing Above It

The company builds models, the inference layer around them, an orchestration layer, and the developer and agent experience. It stays out of the application layer.

“We’re 100% focused on voice AI infrastructure. So we don’t do anything at the application layer.”

Models are tailored to healthcare, drive through ordering, contact center, AI note-taking, and much, much more. Most AI notetakers run on the platform for scalability, cost, accuracy and burst capacity, with no servers for the customer to stand up. About half the company’s engineering work goes into scaling the infrastructure across regions and clouds.

The company is around 80 people, almost entirely product, research and engineering, with forward deployed engineers working directly with customers. Research hires arriving from larger companies come in with a workflow built around building a model and benchmarking it.

“I think that’s the biggest difference between us and a lab at a bigger company. We have a lot of researchers, research engineers that will come to Assembly from a bigger company, and they’re so far removed from the customer.”

What Makes a Voice Model Good

Data mixture accounts for most of model quality. Model updates ship every couple of weeks, and the platform runs multiple model versions and APIs, including one launched the day of the interview for dictation and push to talk.

“Probably like 75% of it is like the data that you’re training on.”

Correct behavior is defined by the application, and two customers can want opposite output from the same audio. Legal discovery teams running police body cam footage want every speaker captured with nothing left out. A drive through order needs everything filtered away (like screaming kids) except the person in the front seat.

“If the kid in the back screams, an extra large Frosty, you wanna be smart enough to ignore that.”

Public benchmarks compress the visible gap between models. The company runs a dedicated internal evals team measuring across a large set of metrics, data sets and languages.

“It’s very easy to optimize for a public open source benchmark. It’s very hard to optimize across all these real world applications.”

The latest model handles 20 languages in one model with automatic language detection and real time switching. Native speaker judgment determines what correct output looks like, and fixing one behavior frequently breaks another. Local voice AI vendors are typically the strongest in their own languages because they understand the nuanced cultural dictations (something Google translate is still trying to understand) and can identify which training data is good.

Voice as an Additional Dimension of Computing

Voice became reliable as a form of data capture in roughly the past year. That covers computer dictation, meeting notes, medical charting during patient encounters and field service coaching.

“Voice is a reliable form of data capture now.”

Ciro AI runs on the platform building software for plumbers and HVAC technicians that listens to a service visit and delivers sales coaching afterward. Those technicians are taking home around 20% more. Large consumer electronics companies are implementing the technology now, and on-device models that run on a phone or a low powered device such as a TV remote are in development.

And voice is now becoming such a reliable form factor that we may actually not need keyboards anymore (shoutout to Coatue’s Max Cook).

“That enables more passive hardware and more passive computing experiences where you can talk to things and you can be walking around and you can be free from being like a prisoner to the device.”

But when it comes to something a bit more complex like humanoid robots, they’re hitting a specific limit. Three people standing next to a robot produce jumbled and merged output because the system cannot disambiguate who is speaking, and cameras do not solve it when a speaker is turned away.

The Disclosure Problem & the Next 18 Months

Text based agents in support flows and outbound sales identify themselves as AI. Voice agents are built to the opposite standard.

“Today when you’re building a voice agent that you’re talking to over the phone, for the most part, the goal is to try to trick the human into believing that it’s also a human.”

Customers building voice agents see the behavior in their own numbers. Users hang up on disclosure because prior voice experiences were bad, which makes mimicking a human the current way to signal intelligence.

“If our customers, if they’re building a voice agent and you disclose upfront that you’re an AI, people just hang up.”

Callers who work it out two minutes in report the experience as strange, and on a hospital nurse line a caller may want a human specifically. Good pacing and natural conversation are achievable without the deception, and the industry has not settled the UX for it.

Sovereign and self hosted deployment is available and rising, though voice sits earlier than text on custom model ownership. Getting the technology to work is the current phase, and the working assumption is another 18 months of it. Over the next 12 months some of the fun & exciting growth areas Dylan is looking forward to are consumer hardware and software including toys, games and consumer electronics.

“In five years, kids are just gonna talk at things, expecting them to be able to understand them.”

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

$10 Trillion to Move Data Millimeters

Everyone talks about Nvidia GPUs. Few people talk about the lasers connecting them.

Michael Hurlston is the CEO of Lumentum (NASDAQ: LITE), one of the world's leading suppliers of optical networking and laser technologies powering AI data centers. Before joining Lumentum in 2025, he served as CEO of Synaptics & previously led Finisar, making him one of the few executives to lead three public technology companies.

→ Listen on X, Spotify, YouTube, Apple

In this conversation from the RAISE Summit in Paris, Michael explains why AI is forcing a once-in-a-generation shift from copper wiring to fiber optics, why hyperscalers are demanding millions of optical components, and why manufacturing these systems has become one of the biggest bottlenecks in AI infrastructure.

“We’re being asked by hyperscalers to deploy in millions, tens of millions of units.”

He also discusses optical networking inside GPU clusters, the future of "optical scale-up," & how the same laser tech could eventually power internet connectivity in space.

"You're not gonna have birds burning up as these lasers shoot data down to the Earth, right?"

We cover:
› Why AI data centers are replacing copper with fiber optics
› The manufacturing bottlenecks facing optical infrastructure
› Why hyperscalers need millions of optical components
› The coming transition to "optical scale-up"
› Space lasers and the future of internet connectivity
› AI infrastructure, geopolitics, and U.S. manufacturing
› Michael's journey leading Finisar, Synaptics, & now Lumentum

Special thank you to Brex, MongoDB, & AssemblyAI for helping make this RAISE AI Summit mini-series in Paris, France happen.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Michael Hurlston, CEO at Lumentum Holdings
(00:58) Backstage with Tony Kim at RAISE Summit
(01:53) What optical connectivity actually means
(03:31) The $10 trillion irony of the data center build-out
(05:37) Scaling up inside a data center
(08:38) Tripling revenue in 5 quarters
(10:13) From Finisar to Synaptics to Lumentum
(11:51) Why Lumentum went all-in on data centers
(13:10) Why space is the next big market for optical lasers
(14:06) Debunking laser myths
(16:40) Why copper is about to disappear from the server rack
(19:00) The geopolitics of the optical supply chain
(20:16) The advice that shaped Michael's path

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

MongoDB–Millions of developers and more than 65,200+ customers across industries – including ~75% of the Fortune 100 – rely on MongoDB for their most important applications. With integrated capabilities for operational data, search, real-time analytics, & AI-powered data retrieval, MongoDB helps organizations everywhere move faster, innovate more efficiently, & simplify complex architectures. visit → mongodb.com/ai

AssemblyAI–Millions of developers use AssemblyAI to power their voice ai apps & features. One API gives you access to best-in-class speech-to-text, voice agent, & speech understanding models for both pre-recorded and real-time audio. Granola, ClickUp & HeyGen are scaling with AssemblyAI - get $50 of free credits today → AssemblyAI.com/sourcery

The Stock That Went Up 1000%

Michael Hurlston became President and CEO of Lumentum (LITE) on February 7, 2025. Fifteen months later the company reported record quarterly revenue of $808.4M, up 90% year over year, joined the Nasdaq-100, and took a $2B investment from Nvidia. This conversation was recorded at the RAISE Summit in Paris, where he had just come off a keynote with BlackRock’s Tony Kim.

→ Listen on X, Spotify, YouTube, Apple

Copper is so over.

The case for optics inside the data center is a physics argument before it is a business one. Fiber replaces the copper wire with glass and the electrical signal with modulated light.

“Obviously, there are fibers, and fibers are essentially glass. They’re just conductors. Instead of having wires, copper wires, you have fibers, and fibers basically contains the light just like a wire or a copper strand contains electricity.

And so what you get with fiber optics is the ability to move data faster over longer distances with less heat and less power than you get with electricity.”

Copper carries a resistive property that turns into heat as data moves across the wire, and that resistance limits how far a signal can travel before it degrades. Data center architecture today is still built mostly on copper, and Ethernet cables are copper.

“The physics of copper is there’s a resistive property. So as data tries to move across the wires, it gets hotter, and the resistance prevents it from going all that far. And so now inside the data center you have speeds are increasing, bandwidths are increasing, and copper can no longer carry these signals the distances they need to go.”

Fiber optics has been the backbone of the internet for decades. The first wave of optical companies grew up around digging up streets in the United States, Paris, and London and laying cable, and later around the submarine cable that carries traffic between continents.

Those deployments ran at volumes Lumentum measures in the low thousands of units per quarter. The shift now underway moves the same technology from between buildings to inside them.

3 Public Companies & the Competitor Built From One of Them

“I think if you look at most CEOs, they either retire or somebody retires them for them. You don’t see too many CEOs change jobs, and I’ve actually been lucky enough now to be CEO of three different public companies.

The first public company I was a CEO of is a company called Finisar that was optical. That’s now the heartbeat of one of our biggest competitors, but also one of our biggest partners in Coherent. Coherent’s built around the engine that was Finisar that we ended up selling to them.

Then I was a CEO of a semiconductor company. I’m mostly a semiconductor guy, and I was CEO of a company called Synaptics for five or six years.”

“We’ve just hit the company with the right products at absolutely the right time. Our team before I got there, the previous CEO set it up well. Right products, right road map, and we have all of the goods that people are looking for now as they convert their data centers to more and more of the fiber optics.”

Why Optics Cannot Be Outsourced

Nvidia and Broadcom manufacture at TSMC. The optical industry has no equivalent, because the material optics depends on sits outside the CMOS (yes, it sounds like sea moss, don’t fret little LA Erewhon fangirl, no sea moss gel being manufactured here) supply chain the rest of semiconductors is built on.

“When you talk to companies like Nvidia and Broadcom, they manufacture most of their products at TSMC, so they essentially outsource the manufacturing problem. The optics industry doesn’t have that luxury.

It’s a very sort of niche industry. It’s one that’s grown up over multiple decades. And so for us, we actually operate our own fabs. So in some respect, we’re like the TSMC of the optical industry.”

Semiconductors run on CMOS. Memory is built on CMOS, TPUs are built on CMOS, and the Nvidia GPU is built on CMOS. Lasers and fiber optics are built on indium phosphide, a boutique material with a narrow supplier base, which means Lumentum has to manufacture its own.

“We actually have to make our own indium phosphide products. We have to manufacture those ourselves, and that requires pretty long lead times. It takes us multiple years to build a fab, and it requires a lot of expense.

So we’re a very CapEx-heavy company. We spend a lot in trying to build out and manufacture our products, and it takes a while. So there’s a bottleneck that’s developing. As these fiber optics becomes more and more prevalent in the data center, it’s creating an ever tighter noose around the industry because there’s only so much we can do so quickly.”

Dark Fiber, Nvidia, & Why the Constraint Runs the Other Way

Lumentum was spun out of JDS Uniphase in 2015. That lineage matters, because JDSU sat at the center of the last time optics got repriced.

Between 1996 and 2001, carriers buried roughly 80 million miles of fiber across North America and Europe against projections of exponential internet traffic. Most of it was never turned on. Estimates of how much sat unlit range from 85% to 97%, and less than 5% of what was laid carried traffic between 1999 and 2004. The industry named the leftover capacity dark fiber. More than 60 telecom bankruptcies followed. JDSU acquired E-TEK for roughly $15B in stock in early 2000 and SDL for roughly $41B later that year, then reported a net loss of $50.6B for the fiscal year ended June 2001, the largest annual loss by a US corporation at the time. The stock went from about $153 to under $2. Headcount went from roughly 29,000 to about 5,300.

They weren’t wrong. The traffic did arrive. It arrived five years late, to companies that had financed the wait with debt.

What made that overbuild possible was a specific property of fiber. Almost all the cost of laying it is the trench, the permits, and the labor. The glass itself is close to free once the ground is open, so the rational move was always to lay far more strands than anyone needed. Overbuilding was nearly costless at the margin, which is exactly why it happened at that scale.

Optical components do not work that way. There is no trench to fill. Every transceiver and every laser has to be fabricated on indium phosphide in a fab that takes years to bring up, and nobody can build a spare decade of inventory while the equipment is already running. That inverts the shape of the risk. In 2001 the industry had capacity sitting idle waiting for demand. Today the constraint runs the other direction. Lumentum’s wafer fab capacity in Japan is fully allocated, lead times run in years, and the company is supply constrained into a demand curve it cannot currently serve.

The financing is also structured differently. The last buildout was debt raised by carriers against traffic forecasts, with suppliers manufacturing into orders that could evaporate. This one has the customer funding the supplier directly. Nvidia committed $4B across Lumentum and Coherent in March 2026, $2B to each, structured for Lumentum as Series A convertible preferred stock and paired with multi-year purchase commitments. That is demand prepaying for capacity rather than capacity betting on demand.

None of which makes the demand itself certain. The 2001 buildout also looked demand-led from inside it, with real contracts against real traffic growth that simply arrived slower than the capital structures assumed, and today’s demand sits with a handful of hyperscalers whose capex plans are revisable. What has genuinely changed is the direction of the constraint and the physics of the product, not the reliability of the forecast.

The capacity being added is going onshore. Lumentum announced the Greensboro site on March 26, 2026, a 240,000 square foot facility acquired from Qorvo, already operational and being retrofitted for indium phosphide on 6-inch wafers. It will produce continuous wave and ultra-high-power lasers, with Nvidia named as a customer, more than 400 US manufacturing jobs planned, and hundreds of millions of dollars of investment. Production ramps in mid-2028, which is another way of saying the shortage has roughly two more years to run before that capacity touches it.

The Bubble Question

Every conversation about AI infrastructure eventually arrives at whether the buildout is real. The optical answer is structurally different from the GPU answer, because a large share of the demand does not depend on new construction at all.

“The question we get all the time is, is this a bubble? What’s interesting about the optical industry is if there are no more data center build-outs, if you don’t see this crazy rate of construction that we’ve seen over the past three or four years, still you’re gonna have a shift within the existing properties from electrical signaling to optics.”

The stock has moved on that thesis. LITE traded in a 52 week range of $95.18 to $1,085.68, crossing $1,000 for the first time on Nasdaq-100 inclusion in May 2026. Fiscal Q3 2026 delivered 47.9% non-GAAP gross margin and $2.37 non-GAAP EPS, with June quarter guidance of $960M to $1.01B.

The internal forecast has been beaten by a wider margin than the external one.

“The board of directors showed me a forecast, and they said, here’s what we think the company is gonna do. And I, of course, as one does, discounted it by about 50%. And it’s been probably 4X over what they even showed me.”

Hurlston has narrowed the company toward that demand, moving away from legacy telecom products sold to AT&T and Verizon and from industrial lasers used to cut metal and machine parts.

“Let’s focus our efforts on the data center. There’s enough here going on that we really wanna simplify our business as much as we possibly can, focus on where we think the biggest opportunities are, and that’s the data center.

So we’ve almost become a pure play data center. Got some risk. If the data center ends up rolling over and that market hiccups a bit, we’ll be caught. But I don’t think that that’s gonna happen.”

Space Lasers

The same technology extends past the data center. Lasers already carry traffic across the Pacific through submarine cable, and the physics carries into orbit.

“Our buddies at SpaceX are trying to do these build-outs, and obviously lasers are super important for all the same reasons.

If you can put a laser in the ocean and transmit across the Pacific, you can certainly put lasers in space and create this reality of an internet in space. It’s gonna be necessary.”

Power level is what separates the applications, and it is also the answer to the concern that satellite optical links pose a risk to wildlife.

“It’s all about power. The iPhone with the Face ID, that’s a laser product that beams lasers onto your face. And one of the concerns of course Apple had was, ‘hey, I don’t want to have people’s eyes burning out.’

So the power of those lasers are relatively modest, and you’re never gonna burn somebody’s eye out.

Similarly, you’re not gonna have birds burning up as these lasers shoot data down to the Earth. You can have these super high-power lasers that can shoot down missiles, that can cut through metal. You can have modestly powered lasers that power the internet backbone, and you can have these super low-power lasers that do things like the Face ID.”

Geography of Supply

Demand is concentrated on the customer side and split on the manufacturing side. Lumentum’s business is heavily weighted to US hyperscalers, and the trade in optical components does not run symmetrically.

“China is a difficult market. The geopolitics of China are very tricky. We don’t supply very much gear into the Chinese data centers, whereas Chinese optical component suppliers, some of our competitors, have a pretty big footprint in US data centers.

So there’s been a lot of push and pull around the geopolitics of the situation. Are we gonna see a world where it gets mandated by either the government or by the customers themselves to use more US content? We’ll see.”

Optical Scale Up & the Volume Step Change

The optics already inside the data center handle connections that leave the rack, which the industry calls scale out. The back of the rack itself is still copper, filled with Ethernet cables running between servers. Moving that layer to optics is what Hurlston calls optical scale up, and he expects the first deployments within 12 months.

Cost is the reason it has not happened already, and physics is the reason it will.

“Optics is more expensive, so people don’t wanna do it if they don’t have to. It’s a more expensive medium than is copper. Copper is a well-trod path. People understand it. People understand the expense that there is. And generally they’ll try to avoid using optics, but it’s getting to a point where the physics of the back of that rack, that connectivity that sits in the back of the rack, is such that you have to do it.”

The volume implication is the part that reprices the industry.

“You’re talking about volumes of, let’s say 1,000 units a quarter. And now we’re being asked by these hyperscalers to deploy in millions, tens of millions of units.

So it’s really changed the dynamic in our industry, not just for Lumentum, but I think every single optical company now is dealing with different scales than we ever had before.”

→ Listen on X, Spotify, YouTube, Apple

Want More on AI Macros?

Check out our pod w/ BlackRock’s Tony Kim

AI’s $10T Systemic Rebuild of Tech: Chips, Memory, Robotics, Quantum

Tony Kim is Managing Director & Head of the Global Technology Team within Fundamental Equities at BlackRock.

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Abundance Era is Real

Scott Wu, CEO & Co-Founder of Cognition, joins Sourcery in Paris amid the RAISE AI Summit. Cognition is the applied AI lab behind Devin, the autonomous AI software engineer, & Devin Desktop, formerly Windsurf.

→ Listen on X, Spotify, YouTube, Apple

The company has raised more than $2.5 billion, including a $1 billion round in May 2026 at a $26 billion valuation, while growing annualized revenue from roughly $37 million to $492 million in a year. Devin now writes roughly 95% of Cognition's own code, total code shipped has grown 7x in six months, and customer usage has increased 11x to 12x.

We cover Cognition's acquisition of Windsurf, negotiated over a single weekend after Google's $2.4 billion talent deal, why Scott is keeping Cognition independent, why enterprises won't buy coding agents directly from the model labs, AI's impact on software engineering, the 6-month singularity debate, and the origin story of Devin, from fixing a MongoDB error in 2023 to deployments at Goldman Sachs, Mercedes-Benz, Citi, Dell, Santander, NASA, the U.S. Navy, and the U.S. Army.

Recorded July 7, 2026, ahead of Cognition shipping SWE-1.7 on July 8 and acquiring The Interaction Company (maker of Poke) on July 23.

Special thank you to Brex, MongoDB, & AssemblyAI for helping make this RAISE AI Summit mini-series in Paris, France happen.

Subscribe now

Share

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

(00:00) Scott Wu, Co-Founder & CEO at Cognition
(01:00) The Identity crisis of building Devin
(01:42) How Cognition 7X-ed its output in 6 months
(02:55) The Metric every AI company gets wrong
(04:18) Cognition's 12x growth curve
(07:46) The Windsurf acquisition
(13:26) Scaling overnight: Cognition's hardest week
(15:11) The strategy behind the integration of Cognition products
(17:23) Cognition's merger playbook
(20:01) Why the DMV is Cognition's next big target
(22:59) Why Cognition bet against AI centralization
(25:09) Why enterprises refuse to bet on a single AI lab
(28:32) Scott's honest take on Superintelligence
(32:36) Is Silicon Valley losing its mind over the AI gold rush?
(33:26) Devin's origin story
(37:17) The question Scott thinks nobody's asking yet
(39:59) The poker match that almost closed a deal
(41:24) Scott's vision for the next year

Brought to you by:

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking—wrapped into a high-performance stack. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

MongoDB–Millions of developers and more than 65,200+ customers across industries – including ~75% of the Fortune 100 – rely on MongoDB for their most important applications. With integrated capabilities for operational data, search, real-time analytics, & AI-powered data retrieval, MongoDB helps organizations everywhere move faster, innovate more efficiently, & simplify complex architectures. visit → mongodb.com/ai

AssemblyAI–Millions of developers use AssemblyAI to power their voice ai apps & features. One API gives you access to best-in-class speech-to-text, voice agent, & speech understanding models for both pre-recorded and real-time audio. Granola, ClickUp & HeyGen are scaling with AssemblyAI - get $50 of free credits today → AssemblyAI.com/sourcery

The $26B Company That Builds Itself

Cognition was founded in November 2023 by Scott Wu, Steven Hao & Walden Yan, all International Olympiad in Informatics gold medalists. The company came out of stealth in March 2024 with Devin, their first product marketed as an autonomous AI software engineer. And now, in just 3 years, Devin writes roughly 95% of Cognition’s own code.

→ Listen on X, Spotify, YouTube, Apple

The Birth of Devin

Devin started as separate experiments inside the company Slack. Each founder built an agent version of himself, a dev “Steven” & a dev “Walden,” and the ideas were eventually folded into one product. The moment it worked was a MongoDB install Yan could not get running in December 2023. He had already tried the usual approach of pasting the error into Google and running whatever the top result suggested. He handed it to his agent instead, which ran a few minutes of commands, fixed it, and messaged him to say so.

“You need encyclopedic knowledge of all of the different errors that you could run into and what would cause each of them. And then you need the ability to actually run and diagnose things. Those 2 things, the encyclopedic knowledge & the ability to actually go & run commands, that’s literally what a coding agent is.”

Yan did not believe it at first. Cognition still has the recording.

“I couldn’t sleep that night. We were just like, shit, maybe it does work. Maybe you can just have an AI engineer buddy that can just do work for you.

That was the very first moment that we believed it was possible.”

Founders Fund’s Early Bet

Founders Fund has backed Cognition from seed through Series D, leading most of its financings. The firm led Cognition’s $21 million seed round in March 2024, which closed shortly before Devin was announced, the $175 million Series A one month later at a $2 billion valuation, and the $400 million Series C in September 2025 at a $10.2 billion valuation.

Lux Capital & 8VC jointly led the Series B in March 2025 at roughly a $4B valuation before reuniting with General Catalyst to co-lead the company’s $1B+ Series D in May 2026.

During the seed round, the two sides were close on terms but had not reached an agreement. Wu proposed a rather creative way to negotiate with Founders Fund partner Napoleon Ta..

“Since you’re such a big fan of poker, we could just play a heads-up poker match for that, and we could have the winner decide whose terms we actually sign. In retrospect, it would’ve been for 1% of the company.”

Ta was willing. Peter Thiel, however, shut it down. Bummer..

(Napoleon leads the firm’s growth practice & sits on Cognition’s board)

However.. on playing chess with prolific Chess Master Thiel, Scott was pretty clear that’s a no go.

“I would not recommend playing against Peter Thiel in chess. You gotta pick games that you can win.”

$26B Series D

Cognition has raised $2.5 billion in total funding and is valued at $26 billion. Its latest financing, a $1 billion+ Series D announced on May 27, 2026, was co-led by Lux Capital, General Catalyst, and 8VC at a $25 billion pre-money valuation.

The round included participation from existing investors Founders Fund, Elad Gil, Alpha Wave, Definition Capital, Positive Sum, Avenir, Vitruvian, Bain Capital Ventures, Conversion Capital, 137 Ventures, Soma Capital, and Omri Casspi, alongside new investors Ribbit Capital, Atreides, and Layer Global.

At the time of announcement Cognition disclosed run-rate revenue of $492 million with the round, up from $37 million a year earlier.

The Growth Curve

Internally, Devin has become central to how Cognition builds software. Wu says it now writes roughly 95% of the company's code, up from 89% at the Series D. Over the past 6 months, the amount of code shipped has increased roughly 7x, while customer usage has grown 11x to 12x. Cognition says enterprise usage has grown more than 10x since the start of 2026.

“If we only 3x’d year over year, we’d be lagging the market.”

Rather than measuring token consumption, Wu says the company focuses on business outcomes.

“Obviously literal tokens is just not the right answer. What you care about is outcomes and what you’re actually delivering. There’s no secret trick or easy way out in terms of measuring actual ROI, productivity, & so on.”

Who Is Running Devin in Production

Cognition says Devin is now deployed across financial services, enterprise software, healthcare, manufacturing & government. Customers include Goldman Sachs, Citi, Santander, Nubank & Itaú in banking; Dell, Cisco, Palantir & Ramp in enterprise software; Mercado Libre in e-commerce; Elevance in healthcare; Mercedes-Benz in automotive; & government agencies including NASA, U.S. Army & U.S. Navy.

In the case of some large incumbents, Cognition focuses on legacy modernization and security remediation. According to the company, a 4-week Mercedes-Benz pilot had Devin analyze more than 200,000 lines of COBOL and reduce an estimated 8-month modernization project to 8 days. It also says Devin automatically remediates 70% of identified security vulnerabilities at Itaú. Neither result has been independently verified.

Two products released around this interview target the same buyer. Devin Security Swarm, announced 1 July 2026, finds vulnerabilities, validates them at runtime and fixes them. Devin Fusion, released 29 June, routes tasks across models.

“The hardest part about building any product, selling a product, is just explaining to people why they should care. The nice thing about AI that can write code by itself and build self-driving products for you is, it’s kind of obvious why you should care.”

Why Enterprises Will Not Buy Agents From the Labs

As coding models improve, the competitive advantage shifts from the underlying model to how companies deploy AI across an organization. Wu argues enterprises are making long-term decisions about how their engineering teams will work, rather than betting on whichever model leads the benchmarks at any given moment.

“You don’t wanna teach all of your engineers or your entire team how to use one particular suite & one particular product & then find out, oh, actually people aren’t using this one anymore because everyone says this other one is better.”

For Cognition, the larger challenge is organizational rather than technical. Rolling out AI coding agents changes how teams plan work, write specifications, conduct user research and manage engineering workflows. Success depends as much on implementation as on model capability.

“It’s not as simple as throw the tool over the wall, here’s your chatbot, try that out, hopefully it’s good for you. There’s a lot of fundamental questions that you have to think through and work through.”

Competition & Consolidation

Cognition competes against products from every major lab. Anthropic’s Claude Code, a terminal-based agent launched in February 2025. OpenAI’s Codex runs asynchronously inside ChatGPT. Google has three entries, Gemini CLI, Code Assist and the async agent Jules, and launched Antigravity 2.0 at I/O in May 2026. Microsoft’s GitHub Copilot has the deepest GitHub integration and the widest model catalog, and switched to usage-based AI Credits billing on June 1 2026. Amazon is winding down Q Developer in favor of Kiro. Cursor and Replit compete from the startup side, alongside open-source agents including OpenCode, Cline and Aider.

The category has consolidated through acquisition rather than attrition. OpenAI’s $3 billion offer for Windsurf expired in July 2025. Google then paid $2.4 billion to license Windsurf’s technology and hire its chief executive and research leadership, and Cognition bought what remained days later. On 16 June 2026 SpaceX agreed to buy Anysphere, the maker of Cursor, for $60 billion in all-stock consideration, with closing expected in the third quarter. Cognition bought Poke in July 2026.

Cursor had reached roughly $4 billion of annualized revenue before the SpaceX agreement. Its share of corporate AI coding spend, measured by Ramp card data, fell from around 41% in June 2025 to about 26% by May 2026.

Staying Independent

“There’s been a narrative out there that in order to win, you have to be a lab yourself, or you have to go sell to a lab.

We’ve always believed there’s a lot of value & a lot of power in being independent.”

Doubling down on it, the company actually spent Independence Day in Washington DC, & gives 2 reasons for being there:

1. Demand. He describes government as possibly the single largest case of an organization that needs software built and does not have the engineers to build it, and Cognition already sells to the Army, the Navy and NASA.

2. Policy. He expects AI to become a fundamental policy question within a few years, and draws the comparison to how the internet was governed.

“You think about the early days of the internet and there were all these discussions about, are there gonna be a handful of corporations that control the entire internet. Obviously we ended up with the open internet, & now all the businesses of the world are built on the internet.

But there will come a day where all the businesses of the world are built on AI, & I think it’s something that we should be thinking about sooner rather than later.”

The Windsurf Weekend

Cognition has bought 2 companies.

Just about a year ago, Windsurf came first, in July 2025, after OpenAI’s $3 billion offer expired & Google hastily scooped up chief executive Varun Mohan, co-founder Douglas Chen & the research leadership for $2.4 billion. Windsurf had built one of the most widely used AI IDEs and a real enterprise business, $82 million of ARR across more than 350 enterprise customers, in total Windsurf had raised $243 million in funding at a $1.25 billion valuation.

What remained after Google was roughly 200 employees, the IDE, the intellectual property & a full GTM organization, with no leadership & no core engineering team. Cognition at the time was 35 to 40 people at $70-$80 million of run rate.

Jeff Wang, Windsurf’s head of business, had become interim CEO. Graham Moreno, previously VP of global sales, became interim president. Wang held an all-hands that Friday where most of the team expected to hear that OpenAI had bought them, and instead heard about Google. He later described the mood in one word, “bleak”, with people in tears and a hostile Q&A.

Cognition heard the shocking news at the same time as everyone else. First, they internally treated an acquisition as a hypothetical scenario, before seeing it as a real strategic opportunity. Wu & Russell Kaplan reached Wang that evening, and the deal was worked out between the four principals across the weekend.

“We got in touch with the Windsurf team that evening, so it was Friday evening, & then over the course of that weekend, we worked out the entire acquisition together. So by Monday morning we had the announcement ready to go.”

A 40 person company absorbing a 200 person company over 48 hours is not a process anybody had run before. Wu attributes the speed to who was in the room.

“A lot of making it up. A lot of our team are former founders. Of our first 56 people, 30 of us had founded a company before this. A lot of us together were people who liked being ambitious, entrepreneurial, & thinking from first principles.”

“For all of us, the experience of building the company is a lot of the fun of it.”

The strategic case was that the IDE and the async agent belong together. The tool a developer sits in and the agent they delegate to need the same understanding of the codebase, and a session started in one should hand off to the other. The merge itself took most of a year.

Cognition ran offsites, worked through overlapping customer relationships where both sides had an existing contact, rebuilt billing, and found a San Francisco office big enough to hold everyone by the end of 2025. Wang and Moreno continue to lead the Windsurf product line as a division of Cognition, and Wang co-announced the rebrand to Devin Desktop with Wu on June 2 2026. Cascade was rewritten in Rust as Devin Local, the original retired on July 1.

“Rather than force it & say, over the next 30 days or 60 days we have to integrate these products, it was much more letting the actual usage & the new features that we were building pull the products together more gradually rather than forcing it.”

“You come to see our office, & you wouldn’t know at all, is this person from the old Cognition or the old Windsurf, or did they join after that time?”

Poke

On 23 July 2026 Cognition bought The Interaction Company of California, maker of Poke, for a reported low nine figures. Poke launched in March 2026 as an agent reachable over iMessage, SMS, Telegram and WhatsApp, passed 100 million messages in three months, and became the first AI agent approved on Apple’s Messages for Business platform. Wu and Yan had both invested personally. The stated logic is that Poke’s interface orchestrates multiple Devin sessions and carries memory across them.

Why Benchmarks Are Kindaaa BS?

Every model release now comes with benchmark scores. Wu’s position is that a benchmark is a task that has already been specified, which is the condition reinforcement learning needs.

“The thing that’s counterintuitive is we’re kind of getting to the point where you can solve basically any benchmark.

Because what does it mean to have a benchmark?

It means you’ve already defined the task, you’ve clarified what success or failure looks like, you’ve given a bunch of examples of what that task looks like & what the right behavior should be. And the truth kind of is, well, if you do that, then you can teach the model to go do that. RL works. It’s incredible how much it works.”

6 Months to Singularity?

Wu agrees that AI systems are becoming increasingly recursive. Devin already writes most of Cognition’s own code. But he doesn’t believe that means the world collapses into a single superintelligence within months.

“In terms of just pure knowledge work, there is nothing that humans do that AIs can’t do. Partly because our brains are just computers, like the AIs are too.”

“There’s some point at which the AI gets so smart that it’s hard for us to even reason about what that world will look like. Will we still have desire? I think we still will have desire. But who knows. You just gotta stay sane.”

Beyond technical progress, he points to practical constraints, from organizational change to deploying infrastructure at scale. More fundamentally, he rejects the idea that AI inevitably leads to a single company controlling intelligence.

“There’s a world that folks sometimes talk about where there’s the full recursive superintelligence, & then everything is all owned by a single entity.. Once you’ve solved AGI, you just take over the entire world.

I just don’t think that’s the right future for us.”

Entering the Abundance Era

The first thing most buyers ask is how much faster existing work gets done, and the answer is roughly a third of the time it used to take. Wu’s argument is that the efficiency gain is the smaller half of it.

“The real unlock is not in efficiency but in capacity. What can we do if we can build so much more?”

He reaches for the Industrial Revolution as the analogy, a period where the ability to mass produce arrived before anyone had worked out what to do with it, and reads software as sitting at the same point now. Software eating the world was already the correct call when every line had to be written by hand. What changes when code can be generated on demand is the floor on what is worth building at all, including single-use software written for one job and thrown away, and software that runs itself.

“Anything that involves using a computer in some form is ultimately some kind of software, & when all the software drives itself.. what do you want to do? What do you want to create?”

“I’m excited to have AI handle all these different little details in my life. I’m excited to have AI that allows me to spend all of my time and focus on the things that I care about and take care of the rest. I think the abundance era is real, and I think we’ll be there pretty soon.”

→ Listen on X, Spotify, YouTube, Apple

Subscribe now

Share

The material presented on Molly O’Shea’s website are my opinions only and are provided for informational purposes and should not be construed as investment advice. It is not a recommendation of, or an offer to sell or solicitation of an offer to buy, any particular security, strategy, or investment product. Any analysis or discussion of investments, sectors or the market generally are based on current information, including from public sources, that I consider reliable, but I do not represent that any research or the information provided is accurate or complete, and it should not be relied on as such. My views and opinions expressed in any website content are current at the time of publication and are subject to change. Past performance is not indicative of future results.

Paid Endorsement. Brokerage services by Open to the Public Investing Inc, member FINRA & SIPC. Advisory services by Public Advisors LLC, SEC-registered adviser. Crypto trading provided by Zero Hash LLC, licensed by the NYSDFS. Generated Assets is an interactive analysis tool by Public Advisors. Output is for informational purposes only and is not an investment recommendation or advice. See disclosures at public.com/disclosures/ga. Matched funds must remain in your account for at least 5 years. Match rate and other terms are subject to change at any time.

Brex—The intelligent finance platform: cards, expenses, travel, bill pay, banking. Built for scale. Trusted by OpenAI, Anthropic, Vercel, Granola, Deepgram, & Sourcery.. teams that move fast AF. visit → brex.com/sourcery

ICYMI: 3 Conversations Worth Your Time This Weekend.

From David Friedberg on whether America is sleepwalking into a crisis, to BlackRock’s Tony Kim on the $10 trillion AI-driven rebuild of the global technology stack, to MongoDB CEO CJ Desai on why data is becoming AI’s biggest downstream winner.. plus HOT TAKES.

Have feedback for topics, questions, or guests? Comment below or reply directly to the email!

Share

Subscribe now

1. David Friedberg on America vs Socialism

Sleepwalking Into a Crisis

David Friedberg, CEO of Ohalo, & Co-Host of All-In joins Sourcery to share his take on America’s 250th anniversary, reflecting on where does the country actually stand? And how do the bottom 50% get left behind?

→ Listen on X, Spotify, YouTube, Apple

2. BlackRock on AI’s $10T Systemic Rebuild: Chips, Memory, Robotics, Quantum.. Space Data Centers

$15T Asset Manager explains the rebuild of tech as we know it..

Tony Kim is Managing Director & Head of the Global Technology Team within Fundamental Equities at BlackRock.

→ Listen on X, Spotify, YouTube, Apple

3. Inside AI’s Biggest Downstream Winner.. the Surge in AI Database Demand

AI’s Biggest Downstream Winner

CJ Desai, CEO of MongoDB, joins Sourcery at the RAISE Summit in Paris. We cover why data is the downstream winner of the AI cycle, how hyperscaler capacity limits are pushing workloads back on-prem, the return of data sovereignty, & what a bank’s real agentic architecture actually looks like.

→ Listen on X, Spotify, YouTube, Apple