Eno Reyes, co-founder and chief technology officer of Factory, talks through where the money in AI actually sits: pricing outcomes rather than tokens, the margin case against the frontier labs, the harness as the place learning accumulates, and a hiring plan in which every future employee arrives through an acquisition. Harry Stebbings works through the rounds he passed on, defends his reputation as the 996 guy, and asks him to rank Microsoft, Nvidia and Meta.
Guest: Eno Reyes, co-founder and chief technology officer of Factory, which builds autonomous software development systems
Host: Harry Stebbings
Published: 29 August 2026 on The Twenty Minute VC (20VC) feed
Watch on YouTube | Apple Podcasts | 1 hr 29 min
✅ Time saved: 51 min
Key Takeaways
The smartest model can be the cheapest system
"for many of the most intelligent, demanding tasks, I see a world where the smartest model is actually the cheapest"
A cheap model that grinds through 50 million tokens loses to an expensive one that gets the outcome first time
Trillion-dollar model valuations assume token prices can double
"Baked into 2, 3, 4 trillion valuations is an assumption that you can basically 2x the price of those tokens and people will buy them"
A $2 trillion Anthropic is a bet on the dev tools market
"you're making a $2 trillion bet on one of the most competitive application markets in one of the most finicky segments of the market which is dev tools"
80 to 90% of neolabs die in the next 18 months, and most of those deaths are good outcomes
Legal survives on proprietary workflow and data; Excel and Jira middleware does not
99% of workflows run on open models in three years, and the last 1% carries 30 to 40% of the value
The gap between frontier and open gets wider, not narrower
Calling open-source models Chinese models is a psyop by the frontier labs
He says Chinese models have shown no backdoor or security risk that American models have not
The harness, not the model, is where continual learning accumulates
"Who is the sovereign of your intelligence? Is it you or is it some other company?"
Microsoft is the best-positioned hyperscaler because it is model-independent
"No matter what model runs on top of that, Microsoft is going to win"
The data center debt is survivable for Microsoft and Google and existential for OpenAI and Anthropic
"they need to become the single greatest free cash flowing businesses in the history of technology in order for them to just live"
Talent is priced as a graph, not as nodes
"when you put all of these nodes together, the graph they make, that can be worth tens of billions of dollars"
Performative work culture is usually covering for something
"this like 996 sort of attitude is almost always correlated with making up for some other detractor or trait"
Factory allocates credits to projects, never to people
Close to seven figures of credits went into one internal evaluation in a single day, run by one person
A Bus Hit His Father at Six, and Pushed the Family Toward Computers
Mr. Stebbings said he hates background stories but asked this one anyway, after hearing it downstairs before the recording. Mr. Reyes said both his parents went to art school and were always in love with how technology intersected with creativity and media.
His father was born in San Francisco in the late 1960s and was hit by a bus at six, which he said changed the trajectory of his life in a pretty crazy way
Unable to do what he called that like traditional 1960s kids stuff, his father went to computers instead, in the era when you barely had screens
"he spent most of his life basically embracing technology as a way to extend his own reach beyond like what I'd argue his physical body could do"
Mr. Stebbings said it is hard to move from a father being hit by a bus to margins: "Like only a venture capitalist could do that in such a swift transition."
Mr. Reyes took the handoff: "Well, some margins in AI might make you feel like you've been hit by a bus."
The Cheapest Model Isn't the Cheapest System
Mr. Reyes said buyers price the input when they should price the output. The question is not what the tokens inside a code review cost, but what the code review costs. A very sophisticated model that gets the right outcome immediately on a thousand tokens beats a cheap model that spends 50 million tokens running, even when the per-token price is far higher.
"for many of the most intelligent, demanding tasks, I see a world where the smartest model is actually the cheapest"
Asked whether that means millions of company-specific models, he said the speciation of models increases very rapidly, and that the companies helping make models possible — he named Fireworks — win in that world
He splits the market three ways. Commodity task executors, which he expects open models to dominate; a small number of very high-volume specialized tasks a business only does itself; and, in between, a commodity model that a business post-trains until it is good enough
The in-between model stays private: "they probably will be the only consumer of it. So they won't even give it to the rest of the world."
"it won't be millions but it'll definitely be quite a lot"
Mr. Stebbings said company structures and teams today are simply not equipped to do post-training. Mr. Reyes agreed the recipe lives in the heads of a specialized few, "But that's also how software development was like 20 years ago."
His forecast: open a platform, click a couple of buttons, describe the task, point it at the workflows that happen in the business, "and outcomes a model"
On the labs that claim recursive self-improvement will be their domain alone: he said many businesses will have access to that technology through software services other companies sell
Verifiability Is the Single Most Important Property of Current AI Systems
Mr. Stebbings pressed on ambiguity: outcome pricing works for a code review, but marketing attribution and legal drafting have no single right answer, and he noted his girlfriend is a lawyer and that different lawyers want notes written differently.
"verifiability is ultimately the single most important property of success with current AI systems"
In healthcare and legal, he said, eval creators bring in experts to build new forms of verification — showing two examples side by side and asking which is better by the expert's judgment
"I actually would argue the frontier right now of AI is AI systems that can build verification where there is none and thus they can progress into tasks that today humans consider to be too difficult for AI to resolve"
His analogy is management, not machine learning. A novice manager judging new hires goes by gut and can be right; a large firm has to write down what good looks like and build a framework to analyze it
The catch is that the framework changes behavior: "if you say good looks like A, B, and C, you're going to get a lot of A, B, and C"
Build the wrong incentives and the system follows the pattern regardless of whether it is actually good
Mr. Stebbings applied it to his own firm: set a partner a goal of three deals a year and you get three deals a year. "I don't want three deals per year. I want a great deal."
Investors Are Underestimating AI Outcomes by an Order of Magnitude
Mr. Stebbings disclosed a position and a regret. He is an investor in Mercor at what he remembered as two or three billion, and passed on the round at 20 billion because a move to a hundred billion was only a 5x with no dilution and did not excite him. He now thinks he was badly wrong and sees a path to 200, 300 billion on the data requirements ahead.
"people are thinking about outcomes in AI and they're looking at 20, 30, 50 and they're saying that's ludicrous. That's crazy. That is underestimating by an order of magnitude how massive a transformation this is going to be."
What people get wrong is the shape of the winners, not the size. He said the businesses that become massive will not look like those of 20, 30, 40 years ago, which had a technology moat or a capability nobody could replicate
"it's basically collections of people that understand what the future looks like a little bit more clear-eyed than the other people"
On Mercor specifically: it sells data, but every person there understands how AI is going to look much more clearly than the average person, and that makes them worth more than even investors say
The TAM of Frontier Models Is Overweighted
Asked whether specialized models trained on proprietary data shrink the addressable market for frontier models, Mr. Reyes said it might.
"I think that the TAM of frontier models is frankly overweighted right now." The world assumes one to three companies get total domination over the intelligence era, which he called a silly proposition
The operative question is margin defense: how a model lab holds its margins when buyers have so many options
"Baked into 2, 3, 4 trillion valuations is an assumption that you can basically 2x the price of those tokens and people will buy them"
Mr. Stebbings put Anthropic's numbers against him — a first quarter of profitability, margins ripping, cash being thrown off. Mr. Reyes said much of that is driven by the applications on top, and that the margin profile of the models is definitely worse than the applications
Two strategies are open to a model provider: dominate the platform era by getting very good at selling inference, or move up and become an application-layer company with really good models
Asked which is right, he said the application layer is the harder battle, because being model-locked is a disadvantage when you sell outcomes
It is bad incentive alignment: a model-locked provider has to deliver its promised outcome using only its own models, while a rival can pick the best model for the task
"Anthropic can basically only deliver their model's outcomes" — and the best model is highly subjective and depends on the consumer
A $2 Trillion Anthropic Is a $2 Trillion Price on Claude Code
Mr. Stebbings framed the trade plainly: with Anthropic potentially going out at 2 trillion, an investor is essentially placing a $2 trillion price on Claude Code, which is not that difficult to switch off of. He asked what he was missing.
"that is fundamentally the risk for an investor is that you're making a $2 trillion bet on one of the most competitive application markets in one of the most finicky segments of the market which is dev tools"
He sees two routes to realizing that value, and they point in opposite directions. Regulatory capture, in which the labs scare politicians into thinking they must own the means of intelligence and become the only providers of frontier capability; or applications that hit what he called the true Pareto frontier of cost and quality, which requires opening up to more models
"You either capture it and keep the model or open up to everybody."
He said OpenAI is visibly grappling with the choice, having let more models into its harness: "They're not making it official, but they're clearly supporting an open model ecosystem in a more direct way."
The Worst Marketing Job Done by Contemporary Capitalists
Asked whether Dario Amodei's marketing message has been mistaken, Mr. Reyes widened the charge to the whole industry.
"the marketing of AI in general was probably one of the worst marketing jobs done by contemporary capitalists, in that it basically did the opposite of what you want: scare every single person, tell them it's very unreliable, and basically threaten their well-being and livelihood with the technology while you roll it out at scale"
He said the threats Mr. Amodei raises are well-intentioned and that unregulated, dangerous AI is a real risk, but the message can be delivered without embellishing the economic ends
His three-part version: the technology will be incredibly transformative, humans will have a huge role in that transformation, and there are risks if it is not done properly
"the moment you start talking about the singularity and AGI and create this godlike mythology out of AI, you're going to scare a lot of people"
Mr. Stebbings mocked the last-company-standing version of the pitch, pointing out that his guest had flown in on United
He gave Sam Altman credit for a public reversal. Relaying an interview, he said Mr. Altman had underestimated the momentum of the economy and of existing businesses, so a future he predicted has not come true
"kudos to him. It's hard to go back and say I was wrong."
The Prior He Had to Correct: Building a Business Is Reactive
Asked what he had got wrong himself, Mr. Reyes named his own planning instincts.
"building a business is much more reactive than planning"
Almost every one of Factory's best decisions came as a reaction to new information and a split-second call, rather than from a master plan forecast six months out
"the world is just this like constantly reactive feedback loop where people are just talking to each other and making decisions and no one actually has an answer to what the future is going to look like"
He said the realization is empowering rather than destabilizing: "you're basically a couple of decisions away from even greater outcomes and an even bigger business than you had prior"
Factory Won't Subsidize Users, and It Costs Them Mind Share
Mr. Stebbings asked whether AI margin profiles of 30 to 35%, against 70 to 80% for SaaS, are a buildout-phase artifact or the new normal on much larger revenue.
Mr. Reyes said not all AI businesses have bad margins, and that Factory's are good because the company sells the outcome rather than the input
The cost of that choice is distribution. Factory does not subsidize consumers in the dev tool space: "that hurts us in a lot of ways, we don't have the mind share from self-service users"
He separates self-service from product-led growth: there is plenty of PLG inside companies Factory has deployed to, but no consumer-facing public plan that is rational unless the buyer is optimizing for quality
On why he won't join the land grab: "there are two players with effectively infinite money who are trying to flood the market, and their intent is, if we flood the market we keep you"
His counter is that you do not keep customers once the subsidies are pulled back
His bet on where cost-effectiveness ends up in one to three years is open models, cheap enough to run locally on your own computer, which is why Factory optimizes for local and open models — on-premise today and for all consumers later
"eventually the self-service will come to us but not because we subsidize but because we have the best product in market"
What He Tells Investors to Look For in a Margin Profile
Asked how an investor should weigh margins in a decision, Mr. Reyes said Factory's approach is a strategy rather than the only way to win.
Lock-in through a workflow or a system of record is real, and temporarily reducing margin to buy a customer base is a classic strategy that predates AI
The disqualifier is a business with no path to a better margin. He said a lot of investments rest on the promise that a company will raise prices without adding value to match
"if you are not also raising the outcomes and the value that you get out of the product while you raise that price, people will churn and move to another thing"
On enterprise stickiness, he said multi-year contracts are treated by both sides as partnerships, and that part of what is sold is knowledge of how best to use the technology — which he distinguished carefully from consulting or professional services
"if you have a product that requires 100 FTEs to get it deployed, you just have a bad product"
Customers assume the vendor will still have that forward view in a year, and a product that behaves like a platform rather than a tool gets stickier because things are built on top of it
Stripe Didn't Pay $8 Billion for Routing Technology
Mr. Stebbings laid out the routing problem: OpenRouter sold for $8 billion, Ramp has a routing provider, his portfolio company merge.dev has one, he is invested in another called Requesty, and routing looks completely commoditized.
Mr. Reyes agreed the technology is not differentiated, and said his reading of the letter to shareholders was that Stripe made "a bet on where the direction of capital allocation is going"
His frame is that money, intelligence and energy are converging into one allocation problem. Tokens are intelligence, tokens are bought with money, and the infrastructure layer runs on energy
"It's literally like translating energy into intelligence. And you're just trading dollars along the way."
Stripe already controls the flow of money; OpenRouter gives it visibility into where intelligence is being allocated and which models people use. Energy, infrastructure and data centers could be the third
"you wouldn't pay $8 billion for the same company that had no users with better technology" — and the conclusion he draws is that "the technology is just no longer the moat"
Asked directly whether it was a good buy, he said the price is quite steep, and that it would have to be foundational to whatever Stripe's next bet on capital allocation becomes
"It feels like 8 and 10 billion is the new 1 billion."
Gateway Routing Saves a Little; Agents Need Routing Inside the Task
Mr. Stebbings asked what Factory's own routing product shows that the market does not see.
The common pattern is gateway routing, where the routing happens outside the place the task is being completed and every internal tool routes through one model gateway
He said these products deliver nice cost savings of roughly 10 to 20%
"you really need something fundamentally different when you have agentic workflows" — the agent has to understand the task dynamically and allocate intelligence in a stateful way, knowing what just happened and what is coming
"that can't happen outside of where the task is being completed. You have to be in there in the task."
Context window expansion followed the same pattern. He said it was solved inside the agent through compaction rather than at the model layer or the endpoint
"People really want the problems to be solved like sort of somewhere else like in the model or in the gateway but more and more we see it's the harness that solves these problems."
"the harness is effectively the new sort of application. It is just where all the logic happens. It's where the state is maintained."
He said buyers are now asking whether to build their own harness, and that Factory spends time teaching them what belongs inside one and what can be done outside it
Who Is the Sovereign of Your Intelligence?
Asked whether a genuinely continuous-learning model would help or hurt Factory, Mr. Reyes said the closed-loop version of continuous learning does not exist yet.
A model that held all the learning behind an API could accumulate an advantage that made it harder for others to build on. "But in reality, what has happened is quite the opposite." Model providers have acknowledged that continual learning happens at the harness layer
"that is ultimately the question of the next 5 years of AI. Who is the sovereign of your intelligence? Is it you or is it some other company?"
His worked example is a law firm that outsources every case to ten different companies
"come 5 years later those other companies can just turn around and screw you over, because they know exactly how to do your entire business"
"at least two of the largest companies that provide models today have explicitly said we are going to go after every single one of these industries and businesses that we provide like intelligence for"
He said the largest companies are wary of labs promising intelligence and luring them into a trap, and pointed to Palantir being loud about owning your intelligence and to a piece Satya Nadella wrote on the same idea
The risk either way: "if it's not your intelligence, then there is just a real risk that either A, they come after your business, or B, if they disagree with what your business is doing, they have a little bit more leverage and control than I think the typical business owner would like"
On-premise is about the option, not the technology. Factory Private is one of the company's most popular offerings, and he said many businesses who look at it still take the SaaS model because they now know exactly how a switch would work
Cursor Inside SpaceX Will Struggle to Stay Model-Independent
Mr. Stebbings asked whether SpaceX buying Cursor helps or hurts Factory, given the compute scale it brings and the model bias it creates.
Mr. Reyes called it an amazing outcome for the team and said he does not discount them as a player
The structural problem he sees is independence. "it's going to be a very hard story to become model-independent or rather stay model-independent when you're attached to a model lab. So they're going to want to push Grok."
He also flagged trust and enterprise concerns attached to the new brand
"most of the enterprises are going to have a second look at the idea of sort of ceding their software development life cycle to a provider who is, one, likely to be model-locked" — and, second, has a pattern of struggling in larger and more secure environments
New Models Keep Dropping, and the Routers Are the News Feed
Mr. Stebbings said he uses arena.ai as a discovery mechanism and finds himself using models he has never heard of and liking the output. He asked whether the release cadence holds.
Mr. Reyes said it will likely sustain for a long time, and pointed out a side effect of the routers: "A lot of people treat model routers as effectively a information or news stream about which model is next. It's kind of a free advertisement every single time a model drops."
Two forces keep it going. Building models gets fundamentally easier, and sovereign intelligence means more of them
Models will differentiate on opinion, not only on benchmarks. He compared it to buying from a founder whose views you agree with
"Models are going to be like that as well where they emit opinions and they have takes that are different from the ones that are most popular and people will gravitate towards those."
"I see this actually just getting much faster and even broader before it shrinks."
80 to 90% of Neolabs Die in the Next 18 Months
Mr. Stebbings said previous guests had put 70 to 80% of today's neolabs dying over three to five years, and asked whether that was right.
"I think it could be 80 to 90% of neolabs die in the next 18 months." He said die is a funny word for it, because many will be incredible outcomes — the businesses simply may not make sense as independent ones
His three tests for a neolab: is the business attached to a durable workflow; will that workflow change as new frontier models get better; and if the workflow were introduced to a new business, would that business figure out something better
Legal passes all three, he said: new models will not necessarily get better without access to the data, the workflow is proprietary, and there will still be a legal system in five, 10 or 20 years
What fails is the middle of knowledge work. He named intermediate tasks in Excel and Jira, and general computer use
"The workflows are very common and I think that we may not use a lot of tools like that in 5 to 10 years."
Why Coding Works and Podcast Clipping Doesn't
Mr. Stebbings laid out the unevenness he sees: coding and customer service are undeniable, legal is good but behind, and marketing copy and visuals are far off. His example was his own show — an AI clipper that misses both faces because it crops to the middle, and has no understanding of how to align an audio edit with a video edit.
Mr. Reyes said the lag is not a model problem. The handful of businesses that deal with media have not devoted themselves to taking the knowledge inside people's heads and bringing it into AI
"the moment that we start to see businesses capitalize on that delta, I think the progression will happen extremely quickly"
Clipping looks unskilled and is not: "a lot of the people couldn't even describe how they know when to do the right clip. That intuition, writing it down, is hard."
Mr. Stebbings agreed and gave the mechanism: start ten seconds early so the hook lands ten seconds in, and the chance of virality goes down significantly
Calling Open-Source Models Chinese Models Is a Psyop
Asked whether the security worries about the Chinese open-source ecosystem are justified, Mr. Reyes rejected the framing.
"I think calling open-source models Chinese models is a psyop by the frontier labs to basically trick people into thinking that they're scary and otherize them"
Open models are no different from a lab's frontier model. "They just happen to have been created by people, a couple thousand miles away."
The same three questions apply to every model, wherever it comes from: what is being censored by its creators, will it solve the problems you care about, and if it goes away in six or 12 months can you switch to something else
"the Chinese models specifically have demonstrated no examples where they have some sort of security risk or backdoor compared to American models" — they are biased toward their creators' preferences in the same way American models are
The carve-out is national security. He said anyone working on national security in the United States definitely should not use Chinese models, and that he highly recommends against them for work on American defense
His censorship example runs the other way. A company writing a 10-K whose strategy leans on recursive self-improvement in models will be blocked by one provider — "The answer is Anthropic, right?"
The general rule: "You have to be aware of who created it and you have to be careful because if the person who created it doesn't want you doing the things that you're going to do with that model, it is going to be harder."
99% of Workflows Go Open in Three Years, and the Last 1% Carries 30 to 40% of the Value
Mr. Stebbings cited a post from Guillermo Rauch of Vercel showing usage of open models rising much faster than tokens spent on closed frontier models, and asked for a three-year number.
"In three years, 99% of workflows are going to be done on open models, but 1% of those tasks is probably going to be 30, 40% of the economic value of the future of intelligence."
The split widens rather than narrows: "that difference between Frontier and Open is going to become actually larger than it is today"
He read the labs' own marketing as agreeing with him. The use cases Sam Altman and Dario Amodei describe — bio research at the frontier, advanced LLM and AI development, security and defense — fit a very specific profile
"These are use cases that really only fit a very specific profile of effectively the frontier of science and technology."
That is where he thinks OpenAI and Anthropic are genuinely differentiated, because they already have the muscle for frontier problems
Inside the Global 2000, the frontier is rarely needed. Ask any random person what they are doing today and "It's not something that needs the true frontier of intelligence like 99% of the time and so cost will dominate."
He added that a frontier lab releasing its own open model and selling inference on it could be a good business
Microsoft Is the Best-Positioned Hyperscaler
Asked whether Microsoft played a good hand by betting through OpenAI without building an extensive model layer of its own, Mr. Reyes said yes.
"I think Microsoft might be one of the best positioned hyperscalers honestly with respect to AI because of this independence."
He credited Satya Nadella with a masterful game and named Kevin Scott as sourcing much of the OpenAI deal, while saying the company now recognizes that one provider is simply not sufficient for enterprise intelligence
Azure's positioning is the payoff: a place to run inference on Anthropic, OpenAI and, most importantly to him, open models, with frontier intelligence available for those who want it
"they captured the upside with a bet and now they're capitalizing on the market as a whole"
On Mark Zuckerberg's spending on Spark, he said it is right for humanity because more American-made open models increase adoption at home and abroad, but that Meta will have to power its consumer business with those models for the investment to pay
Forced to buy only one of Meta or Microsoft, he took Microsoft: "The biggest thing that Microsoft has going for it is that infra. They own so many of these data centers."
"No matter what model runs on top of that, Microsoft is going to win."
The Debt Is Survivable for Microsoft and Existential for OpenAI and Anthropic
Mr. Stebbings raised the data center debt cycle, saying levels like this have never been seen before and that it is showing up in bond pricing for Meta.
Mr. Reyes said it should start to concern investors, because taking on massive debt without a huge amount of free cash flow is dangerous
Microsoft and Google survive a shock. Their existing businesses are durable, with huge barriers to entry, and would absorb a collapse in AI value or a smaller hit to projected cash flow
"if you're OpenAI or you're Anthropic the hundreds of billions in free cash flow that you need in order to pay back the debt that you're taking on in order to accommodate these data center buildouts in order to get the next big training run — it's totally existential for them"
"they need to become the single greatest free cash flowing businesses in the history of technology in order for them to just live"
On the labs moving into chips — Mr. Stebbings named Jalapeño and Anthropic's reported work on its own silicon — he said it is the right move: "verticalization is clearly the strongest way to unscrew yourself from taking on a massive amount of debt and burden"
The complication is the midterm relationship with current vendors, since "their explicit goal is to replace their dependency on you"
Mr. Stebbings compared it to venture, where nobody has loyalty any more: Jensen Huang is working on Nemotron and buying Poolside, and Sam Altman knows it. Mr. Reyes: "Listen, this our job is to survive."
Not 2008, but the Yahoo Era
Mr. Stebbings said older, wiser friends tell him this is peak froth, while Cursor just sold for $60 billion after four years and that money goes back to hospitals and foundations as cash.
Mr. Reyes said he is less concerned about a 2008-style financial crisis, a massive bubble or an asset crash, which "seems disconnected from the true reality of where this technology is and is going"
He said outcomes in science and in human prosperity have already started, early as it is, and that the trajectory is hard to deny
The open question is which companies turn out to be the backbone. "We are probably in like the Yahoo era where we don't actually have or at least widely recognize the Googles of the world"
He said today's frontier labs look more like Netscape — first, and finding that being first is hard to navigate
Factory's own preference comes from Apple. He said Steve Jobs had a great strategy of not being first but being the best at almost everything, and that Factory leans heavily toward best
SaaS Companies Are Movie Studios Now
Mr. Stebbings said he keeps changing his mind on outcome sizes, citing Lovable, a portfolio company now valued at 13.5 billion on 600 to 700 million, and calling the trajectory of company growth unparalleled.
Mr. Reyes said it is hard to index on AI for every industry while people are still buying speculatively, but pointed to consumer packaged goods, where brands started two or three years ago are bought for billions
"maybe it is just true that the world gets faster. It grows bigger, better than ever before."
He said people call this what the singularity will feel like — things move faster, grow bigger, and get normalized
Mr. Stebbings turned to Airtable, sold for around $2.5 billion against an $11 billion price before, and asked whether a generation of SaaS companies will exit ahead of cannibalization
Relaying a line from a fellow founder he said he resonates with: "contemporary SaaS businesses are more like movie studios now where you have to hit a blockbuster and you have to keep hitting blockbusters in order to keep the attention of the world"
Airtable made one film people loved. Rest on it, he said, and "Bending Spoons will come and eat you"
He expects a huge wave of M&A of businesses that are fundamentally good but will not capture fundamental pieces of the economy the way Stripe does
Mr. Stebbings offered the gaming parallel: a banger of a game sustains a hardcore user base for five to seven years, but you need another big hit
Every Future Hire Comes Through an Acquisition
Mr. Stebbings read back what Mr. Reyes had told him before the recording — that Factory expects 100% of its future hires to arrive through acquiring companies and bringing founders and teams in — and said founders often make bad employees.
Mr. Reyes said the profile of an organization has changed so fast that acquiring one is no longer what it was 10 years ago. Someone who quit a job to spend five to eight months building an open source project "shows so much more conviction than you can track in an interview"
These are often companies of one, already aware there is no technology moat, and willing to integrate their work in days or scrap it to build something bigger
Mr. Stebbings pushed back that mission-aligned and willing to take responsibility is not groundbreaking. Mr. Reyes agreed the words are cheap and gave his literal test: the people Factory looks at are already building harnesses for software development with tens of thousands of daily users, and may say theirs is better than Factory's
The alternative version he named — someone who worked on payments arguing payments is just like AI for software development — is not disqualifying, but it is not the target
"mission align to me isn't like a property that you can suggest or say. It's actually extremely evident in the work of the founder."
On Chamath's Claim That Silicon Valley Got Too Money-Centric
Mr. Stebbings raised the heat Chamath Palihapitiya took over the weekend for saying Silicon Valley has become too money-centric.
Mr. Reyes said the claim reflects how Mr. Palihapitiya made his money, and answered with Factory's own start
The company coined the software factory concept — a term he noted Mr. Palihapitiya likes to use as well — and spent years telling people the technology would exist and being waved away
"We're eating so much glass."
"I'm going to get super in the weeds. I'm going to have people tell me I'm wrong every single day for 3 years straight until they eventually agree with you."
He said San Francisco is full of people who care about technology rather than money, and that the people building financial instruments on top are "actually healthy and important part of the ecosystem, but it's not the only thing that happens in San Francisco"
Mr. Stebbings offered a paradox of his own: the influx of money means a small group knows they can always return to a big company and be paid well, so they choose to work on something genuinely interesting instead
Mr. Reyes gave a second reason to talk about it optimistically, and it is about training data. The stories told today become part of what the world-model systems being built now ingest
"that technology is built on the stories that we tell. So, I'm always trying to share a little bit more of the optimistic side of how I perceive the world to be because I think that that actually helps make that world occur with a higher probability."
Pedigree Is Barely a Signal
Mr. Stebbings noted that Cognition places a lot of emphasis on the chess champion and the math prodigy, and asked whether traditional certification is overweighted.
Mr. Reyes said hitting the goals other people set is "the sort of like least agentic path that you could take" — the right school, the right competitions, following rules well enough to be recognized for following rules
He was careful that plenty of smart people take that path, because it nearly guarantees a good outcome in life
"in our mind, the most important trait in an individual to hire for is how capable you are of operating outside the bounds of what today the system calls the rules" — which is very hard to measure for
"I went to an Ivy League school. I learned firsthand that that is barely a signal for competence. There are plenty of idiots who went to Ivy League schools."
The signal he trusts is someone who built something they care about and wants to tell the world about it, including one-person shows who built it in their spare time
The Graph Is Worth More Than Any Node
Mr. Stebbings asked how to put a price on small teams, citing Poolside being bought and employees moving to Nvidia at a rumored $12 billion.
Mr. Reyes said valuing talent is one of the biggest challenges of capitalism, and framed it as the Jeff Bezos question: is he worth a hundred or $200 billion, or is it the company that built it
"it's not that any one node is worth a hundred million dollars, but when you put all of these nodes together, the graph they make, that can be worth tens of billions of dollars"
Talent strategy is graph strategy. The best people make the graph stronger than it was, and companies whose graph was assembled before AI have to update it very rapidly
"that can be the difference between a $2 trillion company being a $4 trillion company. So almost anything's worth it."
Performative Work Culture Is Covering for Something
Asked what hiring mistake he sees other founders make, Mr. Reyes named the candidate who advertises the grind.
Founders read it as a signal of productivity. "this sort of performative work culture, this like 996 sort of attitude is almost always correlated with making up for some other detractor or trait that basically means that this person might not be a great hire"
The same holds at company level. He said everyone works weekends and 15-hour stretches at points in a company's life, but "trying to make that your culture points out that your business doesn't make a ton of sense without it"
Mr. Stebbings defended his own reputation as the 996 guy in the UK and Europe, saying people take it too literally
"I definitely do not mean 9:00 a.m. to 9:00 p.m. 6 days a week." What he means is that when a big client has a problem on a Sunday morning, you jump on the call rather than resuming on Monday
Mr. Reyes agreed that is the reality of building a startup, and drew the line at incentives: "Anytime you create an incentive to check to show people that you're working rather than to actually do work, you're basically incentivizing the wrong thing."
Mr. Stebbings added that senior engineering talent with families is instantly put off by performative young hustle culture, and that the leverage from infrastructure and architectural engineering is very real
Mr. Reyes tied it back to cost: "if you spend a 100 times more tokens trying to get an outcome because you just don't know as much. It doesn't matter that you worked harder."
Factory Allocates Credits to Projects, Not to People
Mr. Stebbings cited two data points from other guests: Brendan at Mercor told him the company spends more on tokens than on engineering headcount, and Jason Lemkin of SaaStr said they would give $100,000 of tokens to their best engineers.
Mr. Reyes said allocating tokens or credits to individuals is a very weird way to think about it, and is another case of measuring inputs. Factory allocates spend to projects and outcomes instead
The scale of a single research bet: on an internal evaluation the team has been hill climbing, Factory "effectively allocated almost seven figures of like credits in one day on this benchmark and that was currently being done by like one person"
The point was not the person's budget but whether the research panned out
Engineers scope a project and share the bid of what they think it will cost, and a Factory product called Agent Effectiveness tracks credits spent per project against the outcomes achieved
"in this new world, the like relationship of one-to-one mapping like agents to humans or like saying like an agent has a name is sort of a weird way to think about it when really you have an agent system and you allocate capital towards projects"
He expects that number to approach eight and nine figures for some businesses
The Good Idea He Keeps Saying No To
Asked what good idea was hardest to decline, Mr. Reyes named self-service.
"It is actually quite painful as a builder of products." The adjustments that would open it up would work against either the business or the enterprise experience
He does not see the decision as permanent, but said it nags: he cannot press a button and have 10 million people using the product, and against comparable solutions with millions of users "basically the biggest difference is economics and that's it"
Mr. Stebbings put the counter-argument that owning those consumers' outcomes and getting the data back could improve the core product by more than the cost to serve free users
Mr. Reyes said Factory already runs self-service that way. The product can be downloaded off the internet, and a steady stream of tens of thousands of people use it every day
The feedback loop saturates early: "a lot of that can be achieved with less than I'd say 250,000 people"
What is genuinely lost is community. "there is something really special about seeing the community build like sort of media and content and storytelling around your product"
Quickfire: Marry Microsoft, Shag Nvidia, Kill Meta
Mr. Stebbings ran a UK parlor game as a markets exercise — buy for the short term, buy for the long term, sell hard — on Meta, Microsoft and Nvidia.
"I would have to say marry Microsoft, shag Nvidia, and kill Meta."
Microsoft is a software company that became an everything company: "There is not a single business that doesn't have Microsoft something." He disclosed that he was a Microsoft employee for a year and a half
"Nvidia is the current kingmaker of technology. They get to decide who is currently even sitting at the table." The durability question is whether people eventually ask if one company should control the whole supply chain — to which, today, "what other choice do we have?"
Meta he called technologically accurate, on VR then and open models now, but with one cash cow in advertising, which "just makes it the weakest of the three"
On Nvidia at 10 trillion in three years: "if we let SpaceX be worth 2 or 3 trillion, then Nvidia probably is worth 10. And so I think that the answer is likely yes."
He is a buyer of Salesforce. The durable businesses are the ones that own a workflow and a system of record everybody treats as consensus, and he named Salesforce and Atlassian
His tell for a good business is a product people say they hate and buy anyway, because they are not buying the software, they are buying what is underneath it
Mr. Stebbings, an investor in Linear, noted that Linear and Atlassian are both crushing it, that the market is bigger than anyone comprehends, and that people think too much in zero-sum terms
Mr. Reyes agreed but added a warning: both sell the same workflow, agile, and a system of record that represents it. "I think that agile might be one of the things that gets hit with this new way of developing."
"it will be very challenging for both Linear and Atlassian to transform into that new way of building"
Claude Code Comes Up in Every Enterprise Conversation
Asked to rank Codex, Cursor, Cognition and Claude Code by threat level, Mr. Reyes reframed it as how often each comes up with enterprise buyers, saying he does not feel an impending threat from any of them.
Claude Code is first. He said it is brought up in every single conversation, and Factory has to explain why it sees itself as largely complementary to Anthropic's platform and suite
Codex is second and rising, because buyers are telling him they were on Claude Code and are switching
"That's actually the greatest news for us because it shows how basically unsticky this is and it gives uncertainty."
His read is that it is not Codex for coding but Codex for work, because "their work platform is better than Anthropic's"
Cognition is third, as the only other model-independent vendor in the enterprise, with a cloud offering built on imitating a software engineer as a human
Cursor is fourth, present in a lot of businesses but perceived as an IDE rather than a primary enterprise software development strategy — which he still called a great business on usage
The difference he sells against is the whole premise. He said all four go to enterprises promising an eventual human-level AI replacement for labor
"you're not going to replace human with AI as much as you're going to build a new system for developing software. And humans are going to build that new system alongside AI."
That new system will look very unfamiliar, and is "a entirely new development methodology" rather than a one-to-one labor mapping
On Chamath Palihapitiya's software venture, which Mr. Stebbings called the 1809, he said it depends on how real the software is. He has not seen examples of it working in an enterprise environment, and rated the founder as well connected with as good a chance as anyone if the focus is on software that delivers outcomes
Enterprise Selling Is Discovery, Not Persuasion
"the biggest thing that I've learned about selling to enterprises is to stop treating it like persuasion where you're trying to convince them that you're right and instead treat it as a discovery opportunity to learn about what's currently the biggest problem they care about"
The distinction is market maturity. In an established, finite, zero-sum market such as databases, where everyone knows the product is a commodity and there are a million options, he said persuasion is the right strategy — all else being equal, you buy from your friends
In a new market it is about understanding how big the opportunity is and learning alongside the customer
Buyers put a huge amount of value on people they perceive to be problem-solving with them, and that lands on a real unsolved problem "almost 10 times out of 10"
Mr. Stebbings' summary was that age-old enterprise sales does not change that much, and Mr. Reyes agreed
In Five Years, Generating Software on the Spot Will Be Ordinary
For the closing question — what looks ludicrous today and will be commonplace in five years — Mr. Reyes went after the software profession itself.
"the biggest thing that we're going to be surprised by is the fact that we let a sort of like priestly class of maybe 2 million people decide the fate of all software for all of humanity"
"in 3 to 5 years, it'll be actually like unthinkable that you couldn't just generate the thing that solved your problem with software on the fly in the moment for nearly any problem that you have in front of you that can be solved by information manipulation"
He said Lovable and Bolt are early versions of it, letting people build personal applications
His image of the end state is a boat operator on a vacation route running a fully custom interface better than the enterprise software the traveler uses at home
"this sort of total dispersement and distribution of amazing software to the entire world is going to make everything just feel way more futuristic and that I think is going to happen very, very quickly on the order of like 3 to 5 years from now"
Mr. Reyes' bottom line is that the model layer is heading for commodity economics and near-total open-model usage, so the durable value moves to the harness where the work is done and the learning is kept — which is why he treats a $2 trillion price on Anthropic as a $2 trillion price on Claude Code, in the most competitive application market there is.
Products, Companies & Tools Mentioned
Factory, Factory Private and Agent Effectiveness (The guest's company: autonomous software development, sold model-independent, with an on-premise offering he says is one of its most popular and a product that tracks credits spent per project against outcomes)
Anthropic and Claude Code (The $2 trillion question of the episode — he says the valuation is a bet on dev tools, that Anthropic can only deliver its own model's outcomes, and that Claude Code comes up in every enterprise conversation)
OpenAI and Codex (Letting more models into its harness without saying so; Codex is showing up in deals as buyers switch off Claude Code, which he reads as evidence of how unsticky the category is)
OpenRouter and Stripe (Bought for $8 billion; he says the routing technology is not the asset, the visibility into where intelligence is being allocated is)
Mercor (The host's own regret — invested at two or three billion, passed at 20 billion; the guest says its people understand AI far better than average, which is what makes it valuable)
Ramp, merge.dev and Requesty (Named by the host as evidence that routing has been commoditized)
Cursor, SpaceX and Grok (Sold for $60 billion after four years; he says staying model-independent while attached to a model lab will be very hard)
Microsoft, Azure and Nvidia (Marry Microsoft, shag Nvidia: he says Microsoft's data center infrastructure wins whatever model runs on top, and that Nvidia is the current kingmaker of technology)
Meta and Spark (Kill Meta: technologically right on VR and now on open models, but with one cash cow in advertising)
Poolside, Nemotron and Jalapeño (Chip and model verticalization by the labs and by Nvidia, which he says is the right way to escape a debt burden but complicates every vendor relationship)
Cognition (The only other model-independent enterprise vendor he sees, building a cloud offering that imitates a software engineer as a human)
Palantir (Loud about the idea of owning your own intelligence, which he says is spot-on)
Salesforce, Atlassian, Linear and Jira (He is a buyer of Salesforce on the strength of owning a consensus system of record, but warns agile itself may not survive the new way software gets built)
Airtable and Bending Spoons (Sold at around $2.5 billion against an $11 billion price, and his example of a SaaS company that made one hit film)
Lovable and Bolt (Early versions of generating personal software on the fly; the host cited Lovable's valuation as evidence outcome sizes have changed)
Fireworks (Named as a winner if the speciation of models accelerates)
arena.ai (The host's own discovery mechanism for new models)
SaaStr (Jason Lemkin's $100,000-of-tokens policy for the best engineers, cited by the host)
Excel and Jira workflows (His example of the intermediate knowledge work least likely to support a durable neolab)
Netscape, Yahoo, Google and Apple (His historical frame: today's frontier labs are Netscape, nobody has recognized the Google yet, and Factory follows the Jobs strategy of being best rather than first)
Books & Resources Mentioned
Satya Nadella's piece on owning your intelligence (Cited alongside Palantir as making the case that businesses should not outsource their intelligence)
Stripe's letter to shareholders (What he read the OpenRouter acquisition through — a bet on the direction of capital allocation rather than on routing technology)
Guillermo Rauch's post on open-model usage (The Vercel chief executive's numbers on open-model share rising faster than closed, which prompted the 99% question)
Sam Altman's recent interview (Where, in the guest's telling, Mr. Altman said he underestimated the momentum of the economy and revised a prediction)
Apple Podcasts (The episode on Apple)
Episode page (The show's own page for this episode)
Get the latest market chatter as it happens: