Right in the middle of Trump’s official trip to China, Anthropic decided to publish a policy paper that goes beyond just throwing caution to the wind. It tries to gauge US-China AI competition and how it will unravel by 2028, ultimately arguing for stronger American control over cutting-edge tech exports to China.
Why is Anthropic arguing for the US to limit the export of cutting-edge AI hardware to China? “Democracies, not authoritarian regimes, must lead in AI development and deployment,” according to Anthropic. They are of the opinion that only democratic countries “can shape the rules and norms that govern these systems.”
The key thing Anthropic believes is worth fighting for is AI ethics and governance. Through the policy document, Anthropic’s fear and warning to the world is that whichever political system gets to transformative AI first will be in a unique position to shape the rules and norms that govern AI for the whole world.
Anthropic’s beef in the policy paper isn’t with China as a nation state as much as it’s directed towards the Chinese Communist Party, where it claims how the CCP uses AI for large-scale censorship and surveillance in places like Xinjiang, and in adversarial cyber operations. Not the right role model for the world for deciding AI governance rules, believes Anthropic.
That Anthropic dropped their two cents on this race between US and China when Trump and Xi Jinping were meeting in China is no coincidence. In the grand race underway between the US and China for AI supremacy, advanced semiconductors play a crucial role. Cutting edge semiconductor chips are the most important ingredient, argues Anthropic in their policy paper. And currently democracies, especially the US, dominate in producing the latest and greatest AI chips. China is still second best.
Anthropic analysed NVIDIA and Huawei’s roadmaps for manufacturing silicon, and concluded that Huawei will only be able to produce roughly 4% of NVIDIA’s total AI compute capability in 2026 – and just 2% in 2027. That means China, on its own, can’t bridge the AI compute gap with the likes of NVIDIA. And if chips export restrictions are tightened, like they were a few years ago by the Biden administration, the US will have roughly 11 times more AI compute than China.
China’s trying to close the gap on US AI supremacy by two key workarounds, according to Anthropic. Export controlled or banned NVIDIA silicon is either smuggled inside the country or accessed via offshore datacentres. On the software front, Anthropic warns, China carries out distillation attacks on top US frontier models to stay competitive.
In fact, according to one China-based cybersecurity analyst, the preview of Claude Mythos sent some serious shockwaves. The China-based analyst called Mythos like facing “a fully automatic Gatling gun” while China is “still sharpening swords.” Anthropic is taking that comment as a sign that US-based, democratic AI development is ahead of China at this point in time. But how long will this lead last?
The US is staying ahead in the AI race versus China for now, but to continue that advantage for a couple of more years at least, America needs to act fast, as per Anthropic. The US continues to tighten AI chips export controls, while also cracking down on smuggled and offshore compute hubs. Distillation attacks on frontier AI models based in the US need to be disrupted as well.
This will allow US-based AI models to stay 12-24 months ahead of whatever China develops, in terms of intelligence and compute capabilities, thereby making American AI as the global choice. It will allow democracies to set AI norms, according to Anthropic.
Also read: Claude Mythos found decade old Firefox bugs that years of fuzzing missed
Failure to act will prevent this possibility, of course, allowing China to export its AI infrastructure and models to the world, especially the Global South. Huawei and Alibaba will grow and take a greater chunk of the global AI market share, if this comes to pass.
It’s important to remember that this, of course, is just Anthropic’s view on the race for AI supremacy between the US and China. There are many experts who argue against export control, suggesting DeepSeek as evidence that it actually helps China innovate around the hurdles faster. Jeffrey Ding, a prominent US-based expert on China’s AI capabilities, points out how AI supremacy among nations shouldn’t just be based on innovation but also on diffusion of technology to people at scale. Either way, let’s see how Anthropic’s predictions for 2028 come to fruit.