The ‘Head of Strategic Futures’ at OpenAI generated a lot of controversy with a post on X (link in comments) where he describes the negative impact he sees open-weight models having on AI accelerationism and on US-based labs. He believes the US government should create ‘large amounts of regulatory risk’ that would ‘scare off hyperscalers from serving open-weight models’.
I believe regulation to protect the interests of US frontier labs is inevitable. The reason is not national security or risk to human civilization… or whatever narrative Anthropic and OpenAI like to push nowadays to fuel the hype.
I believe it’s all about money. The hundreds of billions raised by these companies helped them manufacture a circular ecosystem around them to prop up their valuations and their investors' share prices. The entire US economy relies on AI delivery and supply chains. Investors need their exit liquidity. The current US administration has gone above and beyond to prove that stocks only go up. They can’t let companies use frontier-level open-weight models that cost 3, 5, even 10x less in inference than their US-based alternatives, or that can be hosted and served on their own hardware.
The best kind of innovation doesn’t come when you solve a problem unconstrained. Getting around limitations requires creativity and pushes you to achieve more with less. There are several reasons why I strongly believe that Chinese AI labs are going to win in the short to medium term:
- They were forced to optimize. Chinese labs don’t have access to the best NVIDIA hardware. With less compute available, they had to innovate on model architecture and token efficiency, resulting in models that punch above their weight class.
- US labs went big. They took a more brute-force approach, committing huge amounts of capex to infrastructure. Compute only became a problem for them as adoption grew.
- Unprecedented capital requirements. OpenAI has raised, to date, over $190bn; Anthropic $130bn. Their entire thesis requires them to hold incredibly high market shares. Investors need their money back. Open models are a threat.
Now don’t get me wrong, the big open models (like Kimi k3) need a lot of hardware to run at any truly usable speed. We’re talking in the region of $1m at current hardware prices. So you and I won’t have this level of model running on our personal devices anytime soon.
But what can regulation do? I don’t think open models can be banned; that would be difficult to enforce and control. What can be controlled is hardware. I think there is a non-zero chance for this to be handled through hardware scarcity:
- Tight(er) controls around who is allowed to buy high-end commercial hardware
- An embargo on hardware exports to and imports from China, as they will inevitably become competitive in the next 3-5 years maximum
- Heavy restrictions on consumer-grade hardware for AI workloads