Open-weight AI models are much cheaper, and here's the temptation to ban them
A free software developer highlights how much more affordable open-weight models are compared to proprietary alternatives, warning against the risk of using fear of China to ban them.
Free software developer and technology project manager James O'Claire has published an article emphasising how much cheaper open-weight artificial intelligence models are compared to proprietary and commercial alternatives. The cost difference is not marginal and affects who can utilise this technology and who cannot.
An open-weight model is one that is distributed in such a way that anyone can download it, run it on their own infrastructure, audit, and adapt it. This means you do not pay for each query to a provider, you do not rely on a company to maintain the service, and there is no need to send sensitive data externally. For administrations, media, universities, and small businesses, this autonomy often makes the difference between being able to use the technology or being excluded due to cost.
In the same text, the author articulates a very sensible concern: the potential use of fear narratives, particularly against China, to ban open-weight AI models, which often originate from the Asian country, in favour of OpenAI, Google, and Claude. That is, a national security argument could effectively serve to protect the commercial position of a few closed providers.
This is at the heart of the matter. Open models act as a counterbalance that prevents artificial intelligence from being concentrated in the hands of three or four companies that set its prices and conditions. Banning or restricting them in the name of a geopolitical risk would have the effect, intentional or not, of eliminating this counterbalance. It is important to distinguish between the real dangers of a technology and arguments that, under the guise of security, primarily defend a market.
Ultimately, the question is not whether China produces good models, but whether we want an artificial intelligence ecosystem where it is still possible to choose an alternative that you can inspect and host yourself. Losing this would be far more costly in the long run than any risk currently cited to justify its ban.