When the Best Researchers Cluster at Three Labs, Open Source Becomes a Strategic Necessity, Not a Philosophy

On May 19, Andrej Karpathy joined Anthropic. He co-founded OpenAI, ran AI at Tesla, and is probably the most recognizable AI researcher alive who isn't running a lab. His own explanation was simple: "the next few years at the frontier of LLMs will be especially formative." He wanted to be at the frontier, and the frontier, in his judgment, lives at a specific address.

I keep coming back to that move because it isn't news in the usual sense. It's a data point in a series, and the series is the story. The same handful of labs keep absorbing the field's best-known researchers, one announcement at a time. If you're building an AI company outside those labs, you should be paying close attention to what that series implies, because it changes the strategic case for open source from a preference into a requirement.

The market has already voted

Look at what frontier labs pay. Research scientist compensation at OpenAI, Anthropic, Google DeepMind, and Meta Superintelligence Labs is the highest in AI, and it isn't close. Multimillion-dollar packages are the baseline. Signing bonuses have reportedly exceeded $100 million in extreme cases. Total offers have reached up to $250 million over several years. Anthropic in particular pays 15 to 30 percent more in total compensation per level than its peers for research roles, which tells you exactly how it keeps landing people like Karpathy.

This mirrors something I've written about before on the capital side. In the first quarter of this year, OpenAI, Anthropic, and xAI took 67 percent of all AI venture funding. Two-thirds of the money, three companies. Now the same consolidation is happening with talent, and the two feed each other. The labs with the most capital can pay the most, the labs with the best researchers ship the best models, and the best models attract the next funding round. It's a flywheel, and it's spinning faster.

What concentration should mean, and why it doesn't

Here's the part that should worry every founder who isn't inside one of those buildings. If the best researchers in the world cluster at three or four labs, the capability gap between those labs and everyone else should widen. That's the default physics of talent concentration. The people who invent the next training breakthrough, the next architecture trick, the next data recipe, will invent it inside a closed lab, behind an API, priced per token.

And yet the gap between the best closed models and the best open ones sits at roughly three months, not three years. Why?

Open source. Published weights, published papers, published techniques. Frontier knowledge leaks, deliberately and constantly, into the whole field. Every open release compresses years of closed-lab research into an artifact anyone can download, study, fine-tune, and build on. The open ecosystem is the only countervailing force to talent concentration that actually works at scale, because it doesn't require you to hire anyone. It ships the output of the hiring you couldn't do.

We've run this experiment before

In the 1970s, elite systems talent clustered at Bell Labs and IBM the way AI talent clusters at frontier labs today. If you wanted to work on serious operating systems research, that's where you went, and the rest of the industry had no realistic way to compete for those people.

What broke the lock wasn't a hiring miracle. It was Unix escaping Bell Labs through source licenses to universities. Students learned it, ported it, extended it, and carried it into every company they joined afterward. A world-class operating system, built by talent almost nobody else could have employed, became the substrate the entire industry built on. That single leak of knowledge did more for the field than decades of everyone else's in-house OS work combined.

Open weights are playing the same role today. When DeepSeek or Qwen or Mistral or Moonshot publishes a model, they're doing what those Unix source licenses did. They're letting the output of concentrated elite talent escape into the commons, where a founder with a good product idea and no research team can pick it up and run.

The counterintuitive conclusion

This leads to a thesis I think most people have backwards. Talent consolidation at closed labs makes the open-source ecosystem more valuable, not less. Every researcher who signs with OpenAI or Anthropic increases the strategic value of the weights that are already public, and of the open labs that keep publishing. As the closed frontier gets harder to staff against, the open alternative becomes the only alternative, which is another way of saying it becomes essential infrastructure.

Think about what your options actually are as a founder outside the big labs. You cannot recruit near-frontier researchers, because you cannot pay them. That door is closed and the rent went up. What you can do is take open weights that are three months behind the frontier, fine-tune them on your domain, run them on your own hardware or your own cloud terms, and build a product on capability that would have been the world's best model half a year ago. That's not a consolation prize. For most real products, that's the whole game.

The honest part

I want to be straight about the limits of this argument, because the strongest version of a thesis includes its weaknesses. Open-source models lag the frontier, and they always will while the money is this lopsided. The labs paying nine-figure signing bonuses will keep producing the best model in the world at any given moment. Nobody publishing weights for free can match that spend, and pretending otherwise is cope.

But the bet has never been that open beats closed on raw capability today. The bet is that near-frontier and free beats frontier and metered for most real workloads. The majority of what businesses actually do with AI, summarization, extraction, classification, code assistance, agents running defined workflows, does not require the single best model on earth. It requires a model that's good enough, that you control, that doesn't reprice your unit economics every quarter, and that you can fine-tune into something the frontier labs will never build for your niche.

Strategy for the rest of us

So here's where I land. Karpathy joining Anthropic is good for Anthropic and probably good for the field, since better frontier research eventually leaks everywhere. But for founders, the lesson isn't about him. It's about the structure his move reveals. Talent is consolidating, capital is consolidating, and both trends are accelerating. The default outcome of that structure is a widening moat around three or four companies, and the only force reliably filling in that moat is the open ecosystem.

Treat open source accordingly. Not as an ideology, not as a community feel-good story, but as the one mechanism that gives you access to near-frontier capability without employing near-frontier researchers. You can't outbid a $100 million signing bonus. You can download the weights.