Why AI Safety Won't Make America Lose The Race With China

…

If we worry too much about AI safety, will this make us “lose the race with China” [^1]?

(here “AI safety” means long-term concerns about alignment and hostile superintelligence, as opposed to “AI ethics” concerns like bias or intellectual property.)

Everything has tradeoffs, regulation vs. progress is a common dichotomy, and the more important you think AI will be, the more important it is that the free world get it first. If you believe in superintelligence, the technological singularity, etc, then you think AI is maximally important, and this issue ought to be high on your mind.

But when you look at this concretely, it becomes clear that this is too small to matter - so small that even the sign is uncertain.

The State Of The Race

We can divide the AI race into three levels: compute, models, and applications [^2]. Companies use compute - chips deployed in data centers - to train models like GPT and Claude. Then they use those models in various applications. For now, those applications are things like Internet search and image generation. In the future, they might become geopolitically relevant fields like manufacturing and weapons systems.

Compute: America is far ahead. We have better chips (thanks, NVIDIA) and can produce many more of them (thanks, TSMC). Our recent capex boom, where companies like Google and Microsoft spend hundreds of billions of dollars on data centers, has no Chinese equivalent. By the simplest measure - total FLOPs on each sides - we have 10x as much compute as China, and our advantage is growing every day. A 10x compute advantage corresponds to about a 1-2 year time advantage, or an 0.5 - 1 generation advantage (eg GPT-4 to GPT-5).

Models: The quality of foundation models - giant multi-purpose AIs like GPT or Claude - primarily depends on the amount of compute used to train them, so America’s compute advantage carries over to this level. In theory, clever training methods and advanced algorithms can make one model more or less compute-efficient than another, but this doesn’t seem to be affecting the current state of the race much - most advances by one country are quickly diffused to (or stolen by) the other. Despite some early concerns, neither DeepSeek nor Kimi K2 Chinese models provide strong evidence of a Chinese advantage in computational efficiency (1, 2).

Applications: This is where China is most likely to dominate [^3]. They already outdo America at most forms of advanced manufacturing and infrastructure deployment (eg solar, high-speed rail). And as a command economy, they have more ability to steamroll over concerns like job loss, intellectual property, et cetera.

China knows all of this and is building their AI strategy around it. The plan, which some observers have dubbed “fast follow”, goes like this:

1. Work hard to catch up with US chip production. They are very far behind here, but also have a long history of catching up to the West on things when they put their mind to it, so they feel up to the challenge. They estimate this will take ten years.
2. For the next ten years, accept that they may lag somewhat behind America in compute, and therefore on models. But if they can smuggle in chips and steal US technological advances, they can keep this to a manageable 1-2 year gap, rather than a disastrous 4-5 year gap.
3. Leverage their applications advantage as hard as possible. They imagine that sure, maybe America will have AI that’s 1-2 years more advanced than theirs. But if our smarter AI is still just sitting in a data center answering user queries - and their dumber AI is already integrated with tens of thousands of humanoid robots, automated drones, missile targeting systems, etc - then they still win.

This is a very practical strategy from a very practical country. The Chinese don’t really believe in recursive self-improvement or superintelligence [^4]. If they did, they wouldn’t be so blasé about the possibility of America having AIs 1-2 years more advanced than theirs - if our models pass the superintelligence threshold while theirs are still approaching it, then their advantage in humanoids and drones no longer seems so impressive.

What is the optimal counter-strategy for America? We’re still debating specifics, but a skeletal, obvious-things-only version might be to preserve our compute advantage as long as possible, protect our technological secrets from Chinese espionage, and put up as much of a fight as possible on the application layer.

The State Of AI Safety Policy

It’s worth being specific about what we mean by “AI safety regulation”.