Who Wins the Model War: OpenAI vs Anthropic vs Open-Source | Matan Grinberg

Why does it matter? Because the labs predicting your job's doom are about to ask you to fund their IPO.

Matan Grinberg runs Factory, a company that routes AI tasks across models in real time — think auction clearinghouse for intelligence. Sitting between frontier labs and the enterprises spending fortunes on them, he has a clear view of who's telling the truth. His read on the industry isn't flattering.

• The "AI will take your jobs" rhetoric from frontier labs was engineered to raise capital, not to forecast reality • Value in the AI stack — models, apps, infrastructure — migrates over time; no single layer permanently holds the premium • 80–90% of enterprise software tasks can run on open-source models today; frontier intelligence is only required at the planning step • Four or more competitive frontier labs is the win for humanity; one dominant model is a monopoly over the entire economy

The "AI will take your jobs" warning was a fundraising tactic — not a forecast

"If you're trying to raise unprecedented amounts of money — hundreds of billions of dollars — the best way to convince people to do that is to say all of capitalism is gone. The only company that's left will be me, so you better give us your dollars." Then, as IPO season arrives and those same displaced workers become potential retail investors: "Whoa, whoa, whoa — humans are pretty important. There are going to be jobs again. We like you guys."

Grinberg isn't attacking the technology. He's attacking the incentive structure. The companies that never made apocalyptic job-loss claims are conspicuously the ones that never needed to raise that kind of capital. "For all the philosophizing about AI and intelligence and all this stuff — incentive is driving the outcome."

Discount apocalyptic labor-displacement claims from frontier labs in direct proportion to how much capital they need to raise.

No layer of the AI stack permanently wins — pricing power is a revolving door

The AI stack right now looks like the Microsoft org chart meme: models, apps, and infrastructure all pointing guns at each other. Everyone is simultaneously trying to commoditize everyone else. "Everyone is trying to say, 'Oh no, this one is irrelevant. All the value is going to be here.'" But "value accrual is a time-dependent phenomenon. It's not like there is one person whose steady state gets all of the value. Maybe for this next year, this person has the pricing power. This next period of time, these people get it."

Factory's model-agnostic routing — deciding between providers task by task like a live auction — is itself a bet on this thesis. If one model runs away ahead of the others, Grinberg's own bear case materializes: a monopoly on the most powerful productivity layer in history, with no competitive pressure to improve or lower prices. Track where pricing power sits today. Plan for its migration.

Enterprises are in the AI hangover — and they're looking at a bill they can't explain

One CIO told Grinberg his company had been spending hundreds of thousands of dollars per month on employees asking Claude Opus questions like "Hey, how's it going?" and "What are my macros from the food I ate today?"

That's phase three. Phase one: board yells at CEO about AI strategy. Phase two: token maximalism — AI adoption tracked in performance reviews, everyone forced to use it, no one asking why. "Phase two was kind of like the debauchery, the long night, taking shots, having a great time. Phase three is the hangover where you go and look at the bill and it's like, 'Oh my god.'"

Uber's $1,500-per-person token cap is the public version of a reckoning happening privately across dozens of Factory's enterprise customers. A short-term contraction in frontier model usage is coming. Grinberg calls it healthy — smarter to absorb it now than crash later.

80–90% of software tasks need open source, not frontier — only the planning tokens genuinely matter

"Probably like 80 to 90%." That's Grinberg's estimate of enterprise software tasks that open-source models could handle today without meaningful quality loss. "It's typically the planning that really needs the frontier models." Once a plan exists, "the open models are typically really good."

The 10–20% that genuinely requires frontier intelligence isn't random — it's the decision layer. "Maybe 10 to 20% of the tokens, but those are really, really important because it's kind of decision-making tokens, perhaps." The analogy maps cleanly to how human orgs already work: leadership makes a tiny fraction of total hours worked, but those hours determine the company's fate. Implementation hours — most of them — don't require the same intelligence.

Cost structures should reflect that split. Enterprises burning frontier budget on execution work are paying premium rates for something a cheaper model handles just as well.

The software moat is dead — "we know how to build this and you don't" is no longer a defense

Kirkland & Ellis committed $500 million to build their own AI tools in-house. Grinberg thinks they'll eventually end up at Harvey — not because they can't build it, but because trying will teach them it wasn't worth it.

"We're so used to a world where moat in software was 'I know how to do this and you don't.'" That's gone. "The world going forward, there is going to be nothing that no one can build. Every single piece of software, anyone will in theory be able to build."

The competitive question flips entirely: not can we build this but should we. "Just because you can build a lot of these things does not mean you should. And in fact, often times you want to be really ruthless about what are the few things that you and your team own." Businesses that keep building from a sense of capability rather than strategy will waste their most irreplaceable resource: the judgment of the people in the room.

Silicon Valley's contempt for sales and marketing will haunt AI companies when the gold rush ends

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

"There's a very common Silicon Valley fallacy — research is the pinnacle, then engineers, then sales and marketing and all that dirty stuff. 'Oh, if only we could build a better product and it would sell itself.'" Grinberg calls it completely delusional. Harry pushes back — you can name companies with mediocre products and great sales teams that made it. Grinberg's point stands: legendary companies, the ones that last decades, had both.

Right now frontier labs operate in zero gravity — demand so overwhelming that sales atrophy barely registers. "It's kind of like they're astronauts in space where there's no gravity. Your muscles will atrophy. Gravity will come back." The companies that built contempt for go-to-market into their culture won't be able to compete when customer acquisition becomes an actual fight.

AI has revived the polymath era — Da Vinci-style range is achievable again for the first time in centuries

Grinberg grew up jealous of Da Vinci. Of Euler and Newton. People who could reach the frontier of multiple disciplines within a single lifetime because those fields were shallow enough to allow it. By the time he was doing his PhD in string theory, the opposite was true: "You could spend literally 50 years catching up on all of the literature and academia before you contribute anything new."

AI has collapsed that gap. "With AI, we're now completely the opposite. These tools can get you up to speed to the frontier... way faster than ever before." Not to the depth of a career specialist — but to the frontier, fast enough to contribute. "The age of the polymath is back."

Factory hires explicitly for this: range, systems thinking, agency — not depth in a single lane. The most valuable people in the next decade will push multiple frontiers at once.

Four competitive frontier labs is the win for humanity — a single dominant model is the scenario worth fearing

A year ago, Grinberg thought one or two labs might run away with the frontier. He's changed his mind. "What's pretty clear is it's probably going to be at least four that are going to probably be approximately as good. And that is a win. Like that is the win for humanity."

The bear case is the inverse: "If one model provider gets significantly better than all of the others... that's a monopoly for the entire economy to be worried about." One entity controlling the most powerful productivity layer in human history, with zero competitive pressure to improve or lower prices.

The current weekly churn — new models dropping so fast that engineers at large enterprises can't track which one is number one — is actually the good scenario. Four approximately equal frontier labs, each best at something different, is the outcome worth actively rooting for.

The companies that strip incentive distortions from their AI strategy will outlast every hangover

The quiet thread in everything Grinberg says is that almost no one in AI is making purely technical decisions right now. Labs raise capital by predicting doom. Enterprises adopt AI to satisfy boards. Engineers reach for frontier models because admitting open-source could do the job feels like a slight. When the incentive structures finally get rebuilt around outcomes rather than optics — and the routing layer enforces that discipline automatically — the companies that did it first won't just save money. They'll have built the only thing that compounds in an era when anyone can build anything: a culture that asks the right question before spending the first token.

---

Topics: AI model competition, enterprise AI adoption, open source vs frontier models, model routing, software development, AI labor market, company building, sales and engineering culture, token economics, AI infrastructure