Matan Grinberg on AI Coding, Token Costs and Building Factory

Factory co-founder and CEO Matan Grinberg joins Harry Stebbings to discuss AI coding, model costs, enterprise adoption and building a company. Stebbings' 20VC is an investor in Factory, which sells software-development tools; both have interests relevant to the discussion.

Grinberg expects AI tools to raise productivity and eventually economic growth, but says organizations need time to adjust resources. A company might use the same people to solve more problems or use fewer people to solve the same ones; he says the balance is difficult to predict.

Rather than defining a '10x engineer' by lines of code, he looks for people whose judgment and work are load-bearing for an organization. He expects tools to amplify people who know how to use that leverage, while warning that more generated code can also be poor code.

Grinberg frames tokens, money and headcount as a resource-allocation problem. He suggests leaders start with outcomes such as customer satisfaction, revenue or market share rather than assuming that shipping more features is valuable in itself.

Stebbings cites a reported $500 million, five-year AI investment by law firm Kirkland and discloses his relationship with Harvey elsewhere in the discussion. Grinberg questions whether building all the technology internally is the firm's core strength and thinks specialist vendors may benefit. The reported amount and project scope are not independently verified here.

Grinberg argues that even if software becomes easier to build, a company must decide whether building it is the best use of its time. He uses the example of a CEO picking up lunch: the task may be possible, but outsourcing it can free time for the organization's core work.

He describes model providers, application companies and infrastructure suppliers as trying to commoditize one another. As Factory's CEO, he wants several model providers to compete on price, performance and speed; he acknowledges that this is also in his own company's interest.

The bear case for Factory, he says, would be one model provider becoming dramatically better than all others, making customers want to standardize on it. His current expectation is that several models will be competitive, though the relative quality of models changes and the outcome is uncertain.

Grinberg predicts model improvements may eventually feel continuous rather than like a series of discrete releases. He says application vendors can help enterprises choose models by task, cost, quality and latency instead of asking every engineer to track each release.

He sees open models as a counterweight to expensive frontier models and says many tasks may not need the most capable option. This is his view as the seller of a model-agnostic application; the best choice depends on the workload, security requirements and evaluation.

Grinberg describes an enterprise adoption cycle: leaders first demand an AI strategy, employees then experiment heavily, and management later asks whether the bill produced a return. He cites a customer anecdote about expensive model use for simple or non-work questions, without identifying an audited dataset.

He thinks use of the most expensive models could contract briefly as firms manage costs, and describes Factory customers setting user limits after unexpectedly high bills. He says limits can differ by team and task rather than applying one allocation to everyone.

Stebbings asks what token spending will be relative to developer salaries in three years. Grinberg expects a wide range across individuals and tentatively suggests a median on the same order of magnitude as salary. This is a speculative forecast, not a current company budget or established industry norm.

Some people may create more value through customer meetings or analysis than by using large numbers of tokens, he says, while others may delegate many tasks to agents. He warns against applying one spending ratio to every employee.

Grinberg estimates that 80% to 90% of tasks currently using frontier models could use open models, while important planning decisions may still warrant stronger models. The percentage is his broad estimate, not a benchmark, and he notes that a small share of tokens can still be economically important.

Asked about expensive AI hiring, Grinberg says Factory's distinctive culture helps attract people motivated by more than compensation, though hiring remains competitive. He describes the whole customer journey—from marketing and sales to renewals—as part of the product, not a hierarchy with engineering above other work.

Factory seats salespeople and engineers together and encourages shared ownership of deals and features, he says. Grinberg warns that companies which neglect sales and marketing during a boom may discover the weakness when demand becomes harder to win; this is his management judgment.

He expects strong engineers increasingly to own business outcomes and work with customers and sales, not only implement features. He values initiative and demonstrated work over credentials alone, while acknowledging that competition achievements can be useful evidence in context.

Grinberg predicts more engineer roles resembling general managers who own a result end to end, including product metrics and launch communication. Stebbings observes that similar end-to-end ownership may apply across other functions.

He calls this a return of the polymath: AI can help people learn across fields faster, though he acknowledges uncertainty and that they may lack specialists' depth. His optimism about reaching a frontier more quickly is a view about the tools, not a guarantee of expertise.

Asked about agent operations, Grinberg can imagine people identifying inefficient workflows and building agents, while expecting employees in many functions to do some of that themselves. He predicts tasks such as drafting release notes and documentation will increasingly be automated.

The panel turns to code review. Grinberg says easy code generation can overwhelm reviewers with low-quality changes. He recommends current documentation, executable environments, tests, linters and deployment checks to make agent output easier to validate; these measures help oversight but do not guarantee correctness.

He argues that well-designed APIs, data structures and documentation help agents as well as humans. Standards are still needed to prevent bloated output, and human engineers must make judgments about maintainability and technical debt.

The name Factory reflects his view that engineers will increasingly design the systems that produce software, much as humans design an automated assembly line. He expects their work to shift toward scaffolding, review and outcomes rather than disappearing.

Grinberg is concerned about short-term labor displacement but is optimistic that engineers could eventually work on more unsolved problems. He says the economy must create appropriate incentives for that transition; the number and timing of future jobs remain uncertain.

He argues that better software and AI could help health research and mentions dementia. His assertion that AI will solve dementia is an aspiration, not established medical evidence or a timeline; the interview presents no clinical results.

On government intervention, he sees a role for safety and for incentives where markets underinvest, but prefers a high threshold for intervention. He also speculates that faster AI development could help address climate change despite near-term emissions; neither the benefit nor the trade-off is demonstrated in the interview.

Asked whether AI infrastructure is in a bubble, Grinberg allows for short-term corrections but is confident in long-term demand. This is the forecast of a founder whose company depends on AI adoption, not a valuation assessment of the sector.

He sees behavior change as a major adoption bottleneck. Experienced engineers may be accustomed to existing workflows but skilled at delegation; newer engineers may adapt quickly while having less management experience. He presents this as an observation, not a rule about age or ability.

On enterprise sales, Grinberg says his main lesson is to meet customers, understand their problems and determine whether the product fits, rather than begin with a pitch. He describes learning from different engineering practices across banks, consultancies and pharmaceutical companies.

He recounts pursuing physics from childhood, studying mathematics intensely and working with physicist Juan Maldacena at Princeton before starting a Berkeley doctorate. Teaching and broader reflection led him to reconsider whether a research career was what he wanted.

At Berkeley he took computer-science courses and became fascinated by program synthesis: code designed to generate code. He concluded that building a company might be the way to pursue the problem and began learning about entrepreneurship through books and videos.

A video led him to contact a former physicist who had become a Sequoia investor. Grinberg says a planned short meeting turned into a three-hour walk; soon after, he met co-founder Eno at a hackathon and they improved a demo together.

According to Grinberg, the investor asked whether he would work full-time, and he left his PhD before presenting to the Sequoia partnership in April 2023. He says the firm offered $1 million at a $5 million post-money valuation. These financing terms are his recollection.

He says the connection and trust mattered more to him than shopping for a higher valuation, and generally believes founders should consider taking less favorable terms for a partner they trust. His advice is a personal view, not a guarantee that a particular firm or deal is best for every founder.

Grinberg warns that investors can make a founder feel special when a company is hot. He values board members with conviction when success is not obvious and who remain helpful through difficult periods.

He explains that employee Francesca's investment-network relationship introduced Factory to Ivanka Trump and her firm. Asked about her value as an investor, Grinberg praises her network, time and assistance. This is the founder's account of an investor relationship, not an independently assessed contribution.

On the coding market, Grinberg argues that separating applications from models can encourage model providers to compete and give customers more choice. This is directly aligned with Factory's business model; integrated providers may have different incentives and cost structures.

He says enterprise customers are wary of vendor lock-in and may prefer one application that can choose between models. Stebbings notes that providers' integrated coding tools are taking significant market share, and Grinberg does not provide a forecast of exact shares.

Asked about consumer-oriented tools such as Lovable and Replit, Grinberg says their future is not obvious to him because he has limited firsthand exposure. He sees a possible use case for non-engineers creating sales or marketing materials, while emphasizing enterprise controls for code and data.

Grinberg expects generated code to grow faster than security efforts and worries about serious incidents. This is a risk prediction, not evidence that a particular breach was caused by AI-generated code.

He supports US startups using Chinese open models and distinguishes sending data to an external service from hosting weights locally. He discusses hypothetical malicious triggers and says organizations can mitigate exfiltration, but provides no evidence of a specific backdoor; local hosting and safeguards do not eliminate all model or supply-chain risk.

Grinberg thinks Europe faces challenges building frontier models but sees opportunities in infrastructure and energy. He contrasts decision-making systems and believes US states may experiment with different data-center policies; his claims about economic benefits and construction speed are broad predictions rather than measured comparisons.

Asked about forward-deployed engineers, he says they should accelerate adoption, not be the only reason a product works. His view reflects Factory's preferred sales model and does not establish that services-led implementations are inherently unsuccessful.

On 'grind' culture, Grinberg argues that hours worked are an intermediate metric and output matters more. He favors adequate sleep and says he gave the team Eight Sleep products during a sprint; his personal experience of sleep benefits is not medical evidence.

He compares supporting teams with supporting athletes and imagines investing in recovery and decision-making. Stebbings pushes back playfully on extreme optimization, while Grinberg suggests periods of intense work can alternate with rest; these are personal management preferences.

Asked to choose between Anthropic and OpenAI as an investment, Grinberg considers both well positioned and gives a tentative preference based on perceived volatility. This is an opinion, not investment advice or a confirmed IPO opportunity.

He criticizes public warnings about AI job displacement and speculates that fundraising incentives influenced them. The claim about other leaders' motives is his interpretation, not established fact; his own optimism does not settle the risks of displacement.

Grinberg names EY as a major Factory customer and praises its adoption of agents. That customer relationship and assessment are his statements, without independent deployment or performance data in the interview.

He says he has moved away from expecting one or two labs to dominate and now predicts at least four roughly competitive providers. He sees diversity as beneficial to users; the eventual market structure remains uncertain.