The answer to concentration of power in AI is not openness at all costs, but a serious research commons with the resources to compete.
Over the past few months, I’ve become increasingly uneasy about the concentration of power in the AI ecosystem.
Not because the risks of advanced AI are imaginary. They are not. More capable systems could enable cyberattacks, biological weapons development, and other civilizational dangers. If you take those risks seriously, as I do, the instinct to centralize access can seem reasonable.
But that instinct has become one of the least examined assumptions in the AI debate. In the name of safety, we are drifting toward a world in which a handful of private companies decide who can work at the frontier, which research paths are permissible, and which parts of the technology remain visible to the rest of society.
That is a mistake. We need to treat concentration of power in AI as a risk, not a solution.
The dominant conversation in AI safety has focused on misuse: how to stop bad actors from gaining access to powerful systems. That is critical. But it is not the only question. Just as important is how the power of this technology gets distributed.
Artificial intelligence is becoming a foundational productive resource. The right analogies are railroads, telecommunications, oil, electricity, the internet. These technologies did more than expand economic possibility. They reordered society around new infrastructure, shifting who could participate, innovate, benefit and accumulate power.
AI will do the same. Who gets access? Who gets to build? Who gets to participate? Who gets to decide? That is what is actually being debated, whether we acknowledge it or not.
Earlier this month, Anthropic came under criticism for a policy that could silently degrade some of their AI product’s responses when users appeared to be using the model to train competing AI systems. Many in the research community criticized the approach for altering model behavior without informing users. Anthropic reversed course, but the episode revealed something deeper.
When an institution controls access to the frontier of innovation, it can begin to see itself not merely as a builder of a foundational technology, but as governor of it. The problem isn’t that they made a bad decision. The problem is that they assumed the decision was theirs to make.
To be clear, many of the people leading today’s frontier labs are thoughtful, serious people trying to navigate genuinely difficult problems. But good intentions are not a governance system.
Democracy is built on a profound skepticism of concentrated power. Open science shares this principle. Both are built on the idea that progress and legitimacy emerge from broad, distributed participation rather than concentrated, gated authority.
What worries me most is that we’re drifting toward concentration not by design, but because of the incentives of a few organizations and individuals.
The resources required to work at the frontier of AI are unlike anything modern industry has seen before. The all-in cost to train a new version of a frontier model reaches into the billions. Access to frontier-scale compute resources is increasingly scarce. The universities that produced many of the foundational breakthroughs behind modern AI are essentially unable to access state-of-the-art systems built on that very research.
Meanwhile, the frontier itself is becoming concentrated behind the walls of a few private companies. The most capable models, the largest compute clusters, and many of the field’s most talented researchers are increasingly gathered inside a small number of closed labs.
We’ve seen a version of this story before.
The internet began as a research network built on open protocols. But that openness was not inevitable. The early analogues to today’s Anthropic, OpenAI, and Google were not consumer platforms. There were telecom and computing incumbents with every incentive to make the infrastructure of the internet a closed system controlled by a few dominant actors.
It didn’t happen. Not because everyone agreed openness was morally superior, but because specific people made specific decisions at critical moments to keep foundational infrastructure permissionless and interoperable. Vint Cerf helped develop the protocols that allowed networks to communicate with one another; Tim Berners-Lee later designed the web as a system anyone could build on. From there, researchers built on shared foundations. Entrepreneurs created products without fear of arbitrary exclusion or negotiating for permission. Entire industries emerged on top of infrastructure that no single company owned.
The result was not simply a better internet. It was an explosion of innovation no single company could have planned. Cloud computing, the modern software industry, much of the data economy, and even the training data behind today’s AI systems all stand on top of the decision to keep the underlying infrastructure of the internet broadly accessible.
This is the lesson for AI. Today, we are making different decisions, many rational, most well-intentioned, and all compounding.
If our best scientists and engineers can only reach the frontier by joining a handful of secretive labs, we do not have an open research ecosystem. We do not have a truly competitive market. We have a system in which participation increasingly depends on the permission of a few individuals at a small number of private companies.
We are moving rapidly toward a permissioned frontier.
And that matters for innovation. It matters for economic agility. It matters for national competitiveness. And yes, it matters for safety.
If advanced AI could create species-scale risks, then more of our best researchers should be working to understand them. Safety is not secrecy. Safety is error correction. Modern medicine, aviation, and engineering progressed because talented people could openly challenge assumptions, discover mistakes, and improve the state of the art. Broad participation is how we avoid catastrophic failure. It is also how we make progress.
We do not have to accept a future in which the AI frontier is accessible only inside a handful of private companies. But warning about concentration is not enough; we must build a credible alternative.
That means creating a heavyweight contender in the open: a research commons that brings together the world’s best minds across institutions and gives them access to the resources to work at the frontier. Much of the innovation that powers today’s AI emerged from universities and open scientific collaboration. But the frontier has outgrown any single university. The challenge now is to build a new commons at the intersection of academia, industry, and the public interest.
This research commons must be ambitious enough to matter. It will require frontier-scale compute, access to state-of-the-art models, top talent, operational support, public investment, philanthropic capital, deep collaboration, and companies willing to contribute to an ecosystem larger than themselves.
Our top researchers should not have to choose between independence and relevance; they must be able to reach the frontier without joining a handful of private companies. Academic contributions should not exist solely to improve the products of private firms. And the path from scientific discovery to entrepreneurship should remain open.
The goal is not to recreate the frontier labs. Even a well-resourced open commons will not match the budgets of the largest private AI labs. That relative scarcity can spur breakthrough architectures, alignment techniques, evaluations, and other research paths that frontier labs may overlook.
Yesterday, I spent a long day at Open Frontier, a meeting with many of the top minds working on open frontier AI in academic and industry labs, scientists and engineers who will not be spectators while the frontier of human knowledge is pulled behind closed doors. At the gathering, we resolved to manifest the open research commons outlined above and keep the frontier open (much of the day’s content was livestreamed here).
If we get this right, the benefits of intelligence will not be confined to the few entities that increasingly control it. They will diffuse outward, into new discoveries, new companies, new industries, and new possibilities we cannot yet imagine.
The frontier will advance. What remains undecided is who gets to advance with it.
I also posted this essay on Twitter
]]>
