There is a competition running through the AI world that gets less attention than the race between the big closed labs, but that I think may matter more for how the technology actually spreads. It is the steady, relentless catching-up of open models — systems whose weights are released for anyone to download, run, and build on — against the most advanced closed, proprietary systems. That gap is narrowing, and the narrowing has consequences the industry is not fully reckoning with.
For a long time, the assumption was that the best models would always be closed, available only through a provider's service, kept proprietary as the source of competitive advantage. Open models were seen as interesting but second-tier — fine for experimentation, not for serious work. That assumption is eroding. Open models have been improving so quickly that for a growing range of tasks, the freely available option is good enough, and "good enough and free to run yourself" beats "marginally better but rented and controlled by someone else" for a lot of real use cases.
I think this dynamic is profoundly destabilising for the economics of the field, and here is why. A great deal of investment assumes that frontier capability can be monetised because it is scarce and controlled. But if an open model that you can run yourself reaches eighty or ninety percent of the frontier's capability at a fraction of the cost and with none of the dependence, then the premium for that last increment of capability shrinks dramatically. Why pay a premium and accept lock-in for a marginal improvement when a free, controllable alternative does most of what you need? The open option puts a ceiling on what the closed frontier can charge.
This is the same pattern that has played out in software before. Proprietary systems that once dominated were repeatedly undercut by open alternatives that became good enough, free, and controllable. The open option does not have to be the best. It only has to be good enough to make the proprietary premium hard to justify for most users. And once that happens, the value migrates away from owning the model and toward the things open models do not provide on their own: integration, support, trust, and the surrounding services.
There is a geopolitical wrinkle here too that I find striking. Efforts to restrict access to advanced AI by controlling hardware run into the reality that capable open models can be downloaded and run by almost anyone, almost anywhere. You cannot embargo a file that has already been released to the world. The spread of capable open models quietly undermines strategies that assume capability can be bottled up and controlled, because the capability is, increasingly, just out there.
I want to be balanced. Open models bring real risks — they can be used by anyone, including for harmful purposes, with fewer of the safeguards that providers can build into a controlled service. The same openness that distributes power also distributes the ability to misuse it. That tension is genuine and unresolved, and I do not think anyone has a fully satisfying answer to it. The case for open models is strong, but it is not cost-free, and pretending otherwise would be dishonest.
Still, the trajectory seems clear. The gap between open and closed is narrowing, and a narrowing gap reshapes the entire competitive landscape. The companies betting everything on permanent frontier exclusivity may find that exclusivity is more temporary than their business models assume.
So the question I would pose to anyone building in this space is this: if a freely available model is good enough for most of what your customers need, what are you actually selling — and is it the model, or is it everything around the model that the model cannot provide?
The above reflects my personal views only and is intended for informational and discussion purposes. It does not represent the position of any employer or organisation.