Open-Source Models Quietly Changed the Balance of Power in AI

For a brief moment it looked as though artificial intelligence would belong to a handful of well-capitalised labs, and then the weights started leaking out into the open. The narrative of the early boom was one of scarcity: only those with the largest budgets, the rarest chips, and the deepest pools of talent could train a frontier model, and that scarcity would harden into durable advantage. That story has not collapsed, but it has been quietly rewritten by the steady release of capable models whose parameters anyone can download, inspect, and run.

The moat that was assumed, not proven

The assumption underneath the scarcity story was that capability would remain expensive and concealed. Frontier performance, the thinking went, would always sit behind a paywall and an interface, accessible only as a metered service. What open releases demonstrated is that the distance between the absolute frontier and a freely available model good enough for most purposes is smaller than the leading labs hoped. For a great many tasks, good enough arrives faster than perfect, and it arrives free.

Once a competent model can be downloaded, the economics shift. A company need not rent intelligence by the token from a single supplier; it can host its own, tune it on private data, and stop worrying that a vendor will change terms or prices. The moat does not disappear, but it narrows from a chasm to a ditch.

Distribution beats secrecy

Open models spread the way useful software always has, through the unglamorous accumulation of tooling, tutorials, and shared fixes. A model that thousands of developers can poke at improves along dimensions its creators never anticipated, because the people closest to a problem are usually the ones who solve it. This is the old lesson of open-source software, returning in a new domain: a community that can read and modify a thing tends to outpace a black box, at least in breadth if not always at the bleeding edge.

What the big labs still hold

It would be a mistake to declare the leading labs defeated. They retain real advantages: the capital to train the very largest systems, the infrastructure to serve them reliably at scale, and the brand trust that risk-averse enterprises pay for. The absolute frontier still tends to appear first behind closed doors. But the frontier is a moving line, and what is closed today is often matched by an open release not long after. The premium a lab can charge for being slightly ahead shrinks as the gap behind it fills in.

A more crowded field

The deeper consequence is about who gets to build. When capable models are a commodity rather than a privilege, the interesting competition moves up the stack to the products, the data, and the judgement layered on top. Smaller firms, universities, and developers in places far from the original labs can now participate in ways that a purely closed ecosystem would have forbidden. The centre of gravity has not moved entirely, but it has spread out, and a spread-out field is harder for any single actor to dominate.

None of this guarantees a happy ending. Open weights raise genuine questions about misuse and accountability, and the resources to train the very largest models remain concentrated. But the quiet story of the past stretch is that the most important lever in this technology, the model itself, became something more people can hold. Power in a market is rarely redistributed by decree. More often it slips away from whoever assumed it was theirs, and that, more than any single release, is what has happened here.