Factory CEO: 80-90% of Tasks Running on Frontier Models Could Be Handled by Open-Source

Source: 20VC with Harry Stebbings | Published: 2026-06-13T13:59:54Z

Factory CEO Matan Grinberg predicts enterprise token spending will rival developer salaries within three years, yet estimates 80-90% of tasks currently run on frontier models could be done with open-source alternatives.

Enterprises Are in an AI Spending "Hangover"

Grinberg breaks enterprise AI adoption into three phases. Phase one: the board grills the CEO — "What's your AI strategy?" — and the CEO panics. Phase two: AI at all costs, token consumption baked into KPIs, mandatory usage for everyone. Phase three — where we are now — the hangover hits.

He shared a real example: a CIO discovered the company was burning hundreds of thousands of dollars a month on employees using Opus 4.8 to ask things like "What's the weather today?" and "How many calories did I eat for lunch?" Uber recently announced a $1,500-per-person AI spending cap — something Factory's clients had already been doing quietly for a while.

Near-term usage of frontier models may contract, but that's healthy. Enterprises are realizing that not every task requires the most powerful model.

80–90% of What Frontier Models Do, Open-Source Can Handle

Grinberg dropped a striking number: roughly 80–90% of tasks currently processed by frontier models could be handled by open-source alternatives. What genuinely requires frontier models is the planning layer.

This sounds bearish for frontier models, but his view is more nuanced: that 10–20% of tokens may be the most important — they're "decision tokens." Like executive leadership in a company, C-suite hours are a tiny fraction of total work hours, but those critical decisions shape the company's fate. So frontier models may actually get more expensive at these key steps, and enterprises will gladly pay.

He also flagged a psychological phenomenon: engineers have an ego about this — "The work I do can only be handled by a frontier model; open-source isn't good enough." He had the same instinct when he first started switching and found that open-source was perfectly capable in the vast majority of cases.

Token Spending Will Reach the Same Order of Magnitude as Salaries

Asked what share of developer salaries enterprises would spend on tokens in three years, Grinberg's answer: the same order of magnitude — meaning token spending could approach salary levels.

But he stressed this won't be a uniform number. Some people's token consumption will be zero; others will be tens or even hundreds of times their salary. If an organization sets a uniform token budget ratio for all engineers, its resource allocation is far too crude. The best salesperson might need zero tokens — their value comes from face-to-face client meetings. Meanwhile, another engineer might be orchestrating dozens of AI agents in parallel, burning through tokens at an extraordinary rate.

Everyone Is Trying to Commoditize Everyone Else

Model companies, application companies, and infrastructure companies are locked in a game of mutual commoditization. Grinberg said he's "saying the quiet part out loud": everyone is trying to make every layer they don't own irrelevant.

Factory is model-agnostic and wants OpenAI, Anthropic, and Google to face pricing pressure, delivering the best models at the lowest prices. Model companies want the application layer to become so trivial that anyone can build it, making the model itself the real product. Infrastructure companies have their own narrative.

"Value attribution is a time-dependent phenomenon. There's no steady state where one party captures all the value."

He sees Factory's biggest risk as one model company pulling dramatically ahead of all competitors. But he doesn't think that's likely — at least four players will remain roughly at parity, and that's good for the entire ecosystem.

Kirkland Spending $500M to Build Its Own AI Tools Proves How Hard This Is

Law firm Kirkland & Ellis announced it would spend $500 million over three to four years building its own AI tools. Grinberg's reaction was blunt: building AI technology is not a law firm's core competency. He sees this as actually bullish for specialist companies like Harvey — nothing makes you realize "this is really hard, let's leave it to the pros" faster than trying to do it yourself.

He used an everyday analogy: he knows how to walk out, order lunch, and carry the bag back, but Factory's core competency isn't the CEO fetching lunch for the team. Being able to do something doesn't mean you should. Companies need to be ruthlessly selective about what they truly own and are accountable for end-to-end — everything else should be outsourced.

"Load-Bearing Individuals" Is More Accurate Than "10x Engineers"

Karpathy recently said the future won't have 10x engineers — instead, a handful of 100x engineers and everyone else. Grinberg agrees directionally but dislikes the framing — 10x what, exactly? Lines of code? Anyone can churn out a billion lines with tools now, and it might all be garbage.

He prefers the term "load-bearing individuals": remove this person from the organization and things collapse. These high-leverage people have now been handed a tool that amplifies their leverage even further. Those who know how to wield leverage will generate outsized impact; those who don't will become comparatively less valuable.

The Generalist Era Is Back

Grinberg was obsessed with math and physics from a young age and always envied the era of da Vinci, Euler, and Newton — when fields were still shallow enough for one person to push the frontier across multiple disciplines at once. He grew up in the 2000s, when string theory had gotten so deep you might spend 50 years catching up on the literature before making an original contribution.

AI has fundamentally changed this. These tools can bring you to the frontier of a field at unprecedented speed — you won't have the same depth as a specialist in every detail, but you'll know enough to start contributing. If you're good at thinking within constraints, tackling systemic problems, tolerating uncertainty, and pushing forward through ambiguity, you can be a generalist. One person can simultaneously drive developer marketing, design token-caching strategies for software agents, and excel as a solutions engineer.

Silicon Valley's Most Controversial Take: Sales and Engineering Are the Same Product

Grinberg says Factory's most controversial position is this: the product isn't just software. From the moment a customer first hears the name Factory to their tenth renewal a decade later — the entire journey is the product. Marketing, sales, and solutions engineering matter just as much as writing code.

"Name a legendary company with a terrible sales or marketing team. You can't."

At Factory, engineers and salespeople sit together — no "engineering corner" or "sales corner." When sales closes a customer, engineers say "we closed it." When engineering ships a feature, sales says "we shipped it." In the Bay Area — especially in AI and developer tools — this is surprisingly controversial. He believes companies that don't value sales are like astronauts in zero gravity: their muscles are atrophying, and by the time gravity returns, it'll be too late.

A Security Crisis Is Coming

AI-generated code is growing exponentially, but security investment hasn't kept pace. Grinberg was blunt: "Things are about to get wild." Security incidents caused by AI-generated code have likely already happened over the past few years — people just don't want to admit AI was involved. And the most adversarial attacks haven't even materialized yet — bad actors can do extremely aggressive things with these tools.

On U.S. startups using Chinese open-source models, he sees no issue, but added that it's "kind of embarrassing" that the U.S. doesn't have its own frontier open-source model. He's not particularly worried about "trigger-word backdoor" risks — if a nation-state wanted to plant a backdoor in a model, the rational move is to do it as late as possible, because getting caught in an early model means nobody will ever trust their models again.

From String Theory to Sequoia: One Email and 72 Hours

At 12, Grinberg was so bad at math his teacher made him retake geometry. Indignant, he placed his first-ever Amazon order — a full set of textbooks from algebra through differential equations — taught himself everything over one summer, and tested out of every course when school started. He asked his father what the hardest math was. His father said string theory. So he decided to become a string theorist. Over the next 12 years, he went to Princeton, became renowned professor Juan Maldacena's first undergraduate collaborator, then headed to Berkeley for his PhD.

At Berkeley, everything fell apart. Teaching 18-year-olds who didn't care about physics made him realize this was what the rest of his life would look like. He nearly went to Wall Street to do quantitative finance, but a mentor convinced him to stay and explore. He enrolled in a seminar on "program synthesis" — now called code generation — and was captivated: code that exists for the purpose of creating itself. That kind of fundamentality was irresistible to a physicist.

He stumbled onto a YouTube video of a Stanford VC club podcast. The guest was a theoretical physicist whose papers he'd cited at Princeton, now an investor at Sequoia. He fired off an email. A 30-minute meeting turned into a three-hour walk. The next day he met his future co-founder Eno at a hackathon, and the two built a demo in 72 hours. He called the investor to show it. The response: "It's okay, I guess." He pushed back — "Are you kidding? This is going to change the world." The investor asked if he'd do it full-time. He said yes. "Then drop out. Send me the screenshot." The next day he presented to Sequoia's entire partnership — having never pitched a single investor before and not even knowing who Sequoia was. He walked out with a $1 million check.

The Traditional Enterprise Most Embracing AI? Accounting Firms

Asked which traditional company is most aggressively adopting AI, Grinberg gave a surprising answer: EY. The accounting giant is one of Factory's largest clients and the most "agent-native" organization he's encountered — more aggressive than some startups.

EY's engineering leadership lived through the last cloud computing transition and doesn't want to be caught flat-footed again. Their attitude: this will make some people uncomfortable, it won't be easy, but we're going to make this entire organization agent-native, come hell or high water.

Dario Saying "We're Coming for Your Jobs" Is Self-Serving

Grinberg is frustrated by Anthropic and OpenAI leaders repeatedly claiming AI will replace all jobs. He sees it as self-serving fundraising rhetoric: if you want to raise unprecedented capital, the best pitch is "all of capitalism will disappear and we'll be the last ones standing." But when it's time for an IPO and you need regular people to buy shares, the tune suddenly shifts to "humans matter, there will still be jobs."

He pointed out a telling contrast: the people who never say this are the ones who don't need to fundraise — Zuckerberg and Hassabis have consistently held entirely different positions on workforce displacement. Incentive structures drive narratives, not honest assessments of reality.

In the near term, workforce displacement is genuinely concerning — mass layoffs are happening, and thousands of people are losing their jobs. But Grinberg's long-term view is that the world has an enormous number of problems solvable with software, and we're currently only addressing a tiny fraction of them. More engineers freed up to tackle broader challenges across the economy — pharmaceutical research, healthcare, climate — is a net positive. The economic system just needs time to realign incentives, and the transition will be painful.