Eno Reyes on AI Model Economics, Enterprise Control and Factory's Strategy

Factory co-founder and CTO Eno Reyes joins Harry Stebbings to discuss AI model economics, enterprise software and the competition to build coding tools. Stebbings identifies Factory as an investment of his firm. Their discussion includes company-reported metrics, unverified transaction descriptions and forecasts that should be understood as the speakers' views.

Reyes traces his interest in technology partly to his parents, both of whom attended art school. He says his father turned toward computers after a childhood accident limited his ability to participate in some physical activities. Technology became a way to extend what he could do and to explore creativity.

Reyes's central pricing argument is that the cheapest model per token is not necessarily the cheapest system. A buyer should consider the cost of the desired outcome, such as a correct code review, including repeated attempts, errors and time. A more capable model can cost less overall if it completes a demanding task more efficiently.

That does not mean the most expensive model wins for every task. He expects a mix of broadly useful commodity models and specialized systems adapted for high-volume workflows. A company may post-train an open model for work where a general model is insufficient but a frontier model is more expensive than necessary.

For routine work, Reyes expects businesses to run broadly available models that are inexpensive enough to use at scale. A specialized task can sit between that commodity layer and the frontier: a company may adapt an existing model rather than train everything from scratch. The right choice depends on actual quality and cost.

Stebbings asks whether enterprises have the skills to build and maintain such models. Reyes says the expertise is currently concentrated, but expects software services to make adaptation more accessible. His example is a platform where a business specifies a task and connects its workflows; the ease and quality of that future process remain predictions.

Measuring the result is another challenge. A code review may be testable, while marketing judgments or legal analysis can be ambiguous. Reyes calls verifiability a key property of successful AI systems and expects people to build better evaluation methods for work that does not have a single obvious answer.

He describes experts comparing examples and articulating what makes one result better. Models might then help apply or develop similar evaluation frameworks. For sensitive areas such as law or healthcare, the interview does not establish that an automated evaluator is reliable enough to replace expert judgment.

His analogy is a manager evaluating new hires. An intuitive manager may make good calls informally, but an organization operating at scale often needs to write down what it values and how it judges performance. Reyes expects AI systems to benefit from similarly explicit descriptions of good and bad outcomes.

Explicit metrics also change incentives. If a company defines success using only A, B and C, people and systems may optimize those visible measures even when the real goal is broader. Stebbings compares it to rewarding a venture investor for completing a fixed number of deals rather than finding a genuinely good one.

Stebbings says this challenge is familiar in venture investing. If a manager rewards a fixed number of deals, investors may deliver the count without improving the quality of the portfolio. The example reinforces Reyes's point that evaluation criteria influence behavior as well as measurement.

Stebbings discloses that he invested in a data company and wonders whether its market could become much larger than he previously assumed. Reyes agrees that AI-related outcomes could be very large and argues that understanding a changing market can matter as much as a conventional technology moat. Both are offering investment judgments, not verified valuations.

Asked whether company-specific models reduce the opportunity for frontier labs, Reyes says their addressable market may be overestimated. His question is how a provider protects margins when customers have alternatives. He thinks some valuations assume substantial future pricing power; those assumptions and his conclusion are contested forecasts.

He distinguishes a model platform from an application built on top of it. Reyes believes applications can have different margin profiles from model serving, and says a provider tied to its own model may have difficulty selecting the best combination of cost and quality for every customer task. He sees model choice as an advantage for an independent application.

A company selling a model-locked application, Reyes argues, may only be able to offer its own best model, not necessarily the best available model for a specific task. Stebbings observes that 'best' itself varies by user, task and appetite for risk. Reyes agrees that there is no single model choice for all situations.

Stebbings asks about a hypothetical Anthropic valuation and the risk of relying heavily on a competitive developer-tool market. Reyes agrees that investors face uncertainty about application adoption, pricing and choice. His comments are a competitor's analysis and do not establish a future listing, valuation or returns.

Reyes speculates that model labs may seek to protect their position through regulation or by building strong applications, and characterizes some of their advocacy critically. He does not provide evidence establishing the private motives of any company or executive. He also acknowledges that dangerous or poorly governed AI can pose real risks.

He criticizes AI messaging that emphasizes unreliable systems and threats to livelihoods while the technology is being introduced. In his view, communicators could explain both potential benefits and risks without relying on a mythology of inevitable superintelligence. This is his judgment about messaging, not a determination that the risks are absent.

Asked what has surprised him as a founder, Reyes says business decisions have been more reactive than his earlier plans assumed. He describes important choices as responses to new information rather than the execution of a six-month master plan. That realization has made him see the future as both uncertain and open to influence.

Stebbings asks whether AI margins will eventually resemble those of conventional software. Reyes says not every AI business has the same margin profile and reports that Factory has good margins, without disclosing audited figures. He distinguishes growing adoption within an enterprise from subsidizing a broadly available consumer product.

Reyes says Factory has chosen not to heavily subsidize consumer self-service and expects local or open models eventually to improve the economics of that experience. He recognizes that the choice can reduce consumer awareness today. His expectations about future model costs and product adoption are strategic forecasts.

Reyes says Factory is trying to make its product work well with local and open models, including for customers that want private deployment. He expects that to support a less-subsidized consumer experience in the future. The strategy is a statement of intent, not evidence that future prices or adoption will follow his projections.

For investors, he says a temporary subsidy may make sense if a company can build a durable workflow and later improve margins. The risk is assuming customers will accept a higher price without a corresponding increase in value. He argues that competition gives users alternatives if price rises while outcomes do not.

Enterprise contracts can involve more than access to a model, he says: buyers also value a supplier's knowledge of how to use the technology. He distinguishes that guidance from labor-intensive consulting and argues that a product requiring a very large implementation team may not be designed well enough.

The conversation moves to model routing and a reported acquisition involving OpenRouter. Stebbings asks whether routing technology itself is becoming commoditized. Reyes says the technology may be replicable and offers his own interpretation that usage, customer relationships and information about demand can be strategically valuable.

He frames the reported transaction as a potential way for a payments company to understand how businesses allocate spending on AI. He does not claim insider knowledge of the buyer's reasoning or establish the deal's price. He says a similar technical product without users would not necessarily be worth the same amount.

Reyes compares the strategic value of a routing service with its access to information about where customers use models and spend money. He says that network and distribution can matter even if competitors can reproduce the routing technology. His interpretation does not establish the private data practices or investment rationale of a specific buyer.

Reyes distinguishes gateway routing, which directs requests outside an application, from decisions made inside an agent while it works. He says a gateway can save costs, but an agent may need to know the current task, prior steps and likely next steps before selecting a model. His quoted savings are illustrative company observations, not a universal benchmark.

A gateway outside the application can select among models, but Reyes says an agent making a multi-step decision may need to account for what already happened and what remains. He calls that a stateful allocation problem. Whether it improves outcomes depends on implementation and testing, not merely its location in the stack.

He calls the harness the place where application logic and state can accumulate. Techniques such as compacting context may let an agent manage longer workflows without solving everything at the model or gateway layer. He says Factory spends time explaining these choices to customers considering whether to build or buy.

Stebbings asks whether continuous learning helps or hurts Factory. Reyes distinguishes a model that learns entirely within a closed provider system from learning captured in the surrounding workflow or harness. He argues that companies should ask who owns the data, decisions and successful processes produced by their work.

His example is a law firm that outsources all cases and learning to other organizations: over time, those suppliers could gain knowledge the firm itself needs. He says some large companies worry about dependence on model providers that may also enter application markets. This describes perceived risk, not proof that any named provider will misuse customer information.

The ownership question is not simply where a model is hosted, according to Reyes. It also includes the records of what worked, the business context and the ability to reuse those workflows elsewhere. Customers must still examine actual contracts, data handling and technical controls when choosing a provider.

For Reyes, on-premises options can offer customers confidence that they could take more control if needed. He says Factory offers a private deployment but that some customers choose its hosted software after understanding the alternative. The interview does not independently verify adoption rates or the security properties of either option.

Stebbings raises a reported Cursor transaction involving SpaceX and asks how it might affect Factory. Reyes praises the team but speculates that an attachment to a model provider could make future model independence harder and raise enterprise questions. The transaction and his predictions about a competitor's choices are not independently established in the conversation.

He expects new model releases to continue at a rapid pace as building models becomes easier and providers pursue different use cases. He also thinks gateways serve as discovery channels: a new listing can attract attention to both a model and the platform. The pace and diversity of future releases remain uncertain.

He suggests model discovery is influenced by more than a benchmark score: businesses and users may also care about a provider's perspective, reputation or relationship. That could support multiple offerings, in his view. His prediction does not imply that every new model or provider will find a sustainable market.

Asked how many new AI labs will survive independently, Reyes says it is plausible that 80%–90% will not remain independent within 18 months. He clarifies that an acquisition could be a good outcome for some teams, so 'not independent' is not the same as failure. The percentage is his speculation, not measured attrition.

His proposed test is whether a lab is attached to a durable workflow, whether better general models could erase its advantage and whether another business could reinvent the process. He sees specialized legal work as potentially durable because of proprietary data and enduring demand, while expecting some common intermediate tasks to become less differentiated.

Stebbings observes that AI has advanced unevenly across coding, customer support, legal work and media editing. Reyes agrees that recording expert knowledge and translating it into evaluation criteria can be hard. He expects progress when businesses invest in those workflows, but the rate of improvement across sectors is an open question.

Stebbings uses podcast clipping as an example of a task where today's systems can miss framing or timing. Reyes says experienced editors have tacit knowledge that may be difficult to write down, including how to choose a strong opening. He expects progress as that expertise becomes easier to represent and test.

Asked about models developed in China, Reyes criticizes what he sees as broad-brush treatment of open models as inherently suspect. His suggestion that frontier labs deliberately promote that framing is an allegation about motives for which the interview supplies no evidence. He says all models should be evaluated for behavior, suitability and dependence.

He says users should ask about possible censorship, whether a model meets their needs and whether they can switch if it disappears. He reports that he has not seen examples of particular security backdoors in the models he discusses; that personal observation does not establish that any model is free of vulnerabilities or risks.

Reyes says he has not personally seen evidence of a backdoor in the models he is discussing, while acknowledging all providers can encode restrictions or preferences. His statement is limited to his knowledge; model weights, hosting, supply chains and deployment practices can each introduce different security considerations.

Reyes distinguishes ordinary business tasks from national-security use, where he says the origin of a model may warrant different choices. He argues that restrictions and preferences of any provider can affect a workflow. The discussion does not provide a security audit or establish equivalent risk across all models and deployments.

He predicts that in three years open models could handle 99% of workflows, while a small share of frontier tasks could still represent substantial economic value. He also expects frontier capabilities to remain differentiated in areas such as advanced research. These percentages and market outcomes are his projections, not present usage data.

He expects inexpensive open models to serve many routine workflows while frontier labs pursue a smaller set of difficult tasks. A small number of tasks can still have high economic value in his scenario. The predicted division of usage and revenue is a hypothesis and depends on future model capabilities, prices and customer choices.

Reyes views Microsoft as well positioned because customers can access multiple model providers through its infrastructure while it also has an investment relationship with OpenAI. He thinks model choice could matter to enterprise buyers. His assessment of a company's position is an opinion, not an investment recommendation.

He also regards investment in open models by Meta as potentially valuable, while saying a business must apply those models to useful operations. Asked to choose a hypothetical investment between Meta and Microsoft, he favors Microsoft. The comparison reflects his personal thesis, not verified future performance.

On data-center debt, Reyes says substantial borrowing can be dangerous for businesses without durable free cash flow. He contrasts larger companies with established businesses against frontier labs that may need strong future economics to support infrastructure commitments. The interview does not establish any company's complete financing obligations or future solvency.

Reyes believes established cloud companies may be better able to absorb an infrastructure setback because they have other sources of cash flow. He views the burden as more acute for businesses that depend on financing future growth. The argument is qualitative and does not verify a particular debt amount.

He sees chip development and greater vertical integration as possible responses to infrastructure dependence, while acknowledging that they can complicate relationships with existing suppliers. Those tradeoffs involve cost, execution and financing; building a chip does not by itself establish that debt or operating costs will fall.

Stebbings asks whether current valuations indicate a bubble. Reyes says he is less concerned about a crisis resembling 2008 than about identifying which businesses will ultimately sustain value as the technology develops. His macroeconomic view and claims about future prosperity are opinions rather than guarantees.

He compares early AI leaders with earlier technology pioneers that did not necessarily remain dominant. Being first can help, but he argues execution over time matters. The analogy is a way to think about competitive risk, not a prediction that any particular current company will repeat a historical outcome.

Stebbings says rapid company growth has changed his expectations for possible outcomes and discloses an investment in Lovable. Reyes agrees that business creation and growth may accelerate, while acknowledging that many current AI purchases are exploratory. Neither speaker establishes that today's growth rates will persist.

Discussing reported software acquisitions, Reyes expects some established SaaS companies to be sold as they face pressure to keep creating products customers want. He compares software companies to movie studios that need repeated successes. The acquisition examples raised by the host are not independently verified in the conversation.

Stebbings asks about Factory's interest in bringing in founders and small teams through acquisitions. Reyes says a person who spends months building an open-source project demonstrates conviction that an interview may not capture. He looks for teams already solving related problems and able to take responsibility independently.

By mission alignment, he means evidence in the work itself: a founder may already be building a software-development harness or measuring AI outcomes with customers. He says talented people from unrelated backgrounds can also contribute, but the closest match makes it easier to evaluate how they would work within Factory.

The pair debate whether Silicon Valley has become too focused on money. Reyes says many people he meets care deeply about building technology, while financing and investing are also legitimate parts of the ecosystem. He prefers to share an optimistic account of the culture, which is his interpretation rather than a survey of everyone in it.

Reyes says building Factory initially involved repeated conversations with people who did not believe the proposed software would exist. He associates that persistence with a technology-building culture rather than financial speculation alone. Stebbings adds that financial security can also give some founders more freedom to pursue an interesting problem.

On hiring, Reyes cautions against treating school prestige or competition credentials as the only evidence of ability. He says those paths can include highly capable people, but may overlook others who build useful projects outside formal systems. His own Ivy League experience informs his skepticism about pedigree as a sufficient signal.

Asked how to value a small team, he says value can arise from how people work together rather than from assigning a fixed price to each person. Adding someone can strengthen the network of skills and decisions within an organization. His illustrative dollar values are speculative, not a valuation method with guaranteed results.

He warns against performative work culture: signaling that someone is always busy can be mistaken for productive work. He acknowledges that startup teams may sometimes work weekends or long days, but argues that making visible exhaustion the goal can reward the wrong behavior and exclude strong candidates.

Stebbings clarifies that his own emphasis on availability means responding to an important customer problem when needed, not literally prescribing a fixed schedule. Reyes agrees that the goal should be the outcome. He says experienced people may save time and tokens by directing agents toward the right problem from the start.

The two draw a distinction between a real customer emergency and rewarding the appearance of constant work. Reyes accepts that responding outside normal hours may sometimes be needed. His objection is to incentives that value showing activity over delivering a useful result.

Asked whether engineers should receive fixed token budgets, Reyes says Factory instead allocates spending to projects and desired outcomes. He describes an exceptionally costly evaluation run led by one person as a project investment, not a standard individual allowance. The amount and results are his account, not independently audited.

He says teams can estimate a project's scope and expected spending, then compare the credits used with the outcome achieved. In his view, an agent system is better understood as resources applied to work than as a one-to-one replacement for a named human employee. He expects some company budgets to grow substantially.

Reyes says self-service is a difficult opportunity for Factory to limit because he wants more people to use the product. He distinguishes having a downloadable product and individual users from heavily subsidizing a mass consumer offering. He says the company receives feedback from tens of thousands of daily users in that segment; the number is self-reported.

He thinks a sufficiently large user group can supply useful product feedback without subsidizing millions of accounts, but recognizes the loss of grassroots storytelling and community that broader reach might create. The right balance between feedback, distribution and cost is a company-specific strategic choice.

In a quick-fire comparison, Reyes favors Microsoft for long-term durability, discusses Nvidia's current role in AI infrastructure and is more cautious about Meta's reliance on advertising. He discloses past employment at Microsoft. His company and valuation views are personal forecasts, not financial advice.

Reyes says Microsoft has broad enterprise distribution, Nvidia occupies a strong current infrastructure position and Meta must continue to translate its investments into business value. These are subjective assessments made in a hypothetical comparison. He does not provide evidence sufficient to predict the future price of any security.

He says established systems of record can be durable because customers depend on the workflow beneath the interface, while AI may change how software itself is developed. Stebbings discloses an investment in Linear as they compare it with Atlassian. Both discuss potential changes rather than demonstrating a future market outcome.

Asked about competing coding products, Reyes describes which names arise in Factory's enterprise conversations and how he positions Factory as model-independent. His ranking is his own account as a competitor, not an independent comparison of product adoption, capabilities or customer satisfaction.

He argues that AI may lead to a new software-development system built by people and agents together, rather than a simple one-for-one replacement of each engineer. He presents this as Factory's strategy and his interpretation of competitors; the future division of work and actual product capabilities remain uncertain.

His advice for enterprise sales is to treat conversations as discovery: understand the customer's hardest problem and work toward a solution together. In an emerging market, he says this can be more useful than trying to persuade a buyer that a predetermined product is right for every situation.

Reyes contrasts emerging markets, where a seller and customer can discover an unsolved problem together, with established markets where buyers may already understand their choices. He thinks customers value a partner who helps them define a useful outcome. The approach does not guarantee a sale or a successful implementation.

For a final prediction, Reyes imagines custom software being generated on demand for many more people and small businesses within three to five years. He sees broader access to useful software as a major opportunity. That timeline and degree of adoption are speculative, and the conversation closes with Stebbings thanking him.