China AI Slowdown: Why Beijing and Trump Both Said No

China rejected the AI slowdown plan almost at once. Its foreign ministry called calls for restraint "fearmongering," state media labeled Anthropic CEO Dario Amodei's proposal a "Cold War playbook," and President Trump said "whoever wins AI wins." That's the hard part of any China AI slowdown. A pause only works if the leading rivals join, and neither government has said it will.

Key Takeaways

• Anthropic CEO Dario Amodei proposed slowing frontier AI on September 12, 2026, and called China the "toughest dilemma" for the plan.
• Two days later, China's foreign ministry warned against "fearmongering," and state-run Global Times called the plan a "Cold War playbook" aimed at containing China.
• President Trump also rejected a slowdown, saying "whoever wins AI wins," ahead of hosting Xi Jinping on September 24.
• Experts see no working way yet to verify a slowdown across both countries, though narrow cooperation on shared risks is possible.
• For builders, the likely result is policy volatility, not a pause, so model-agnostic systems and clear model provenance matter more than ever.

What Did Amodei's Plan Say About China?

China is the third and hardest step. We covered the full plan in our explainer on the AI slowdown plan. Here's the part that set off the diplomatic reaction.

The plan has three stages, and each one needs a wider circle to agree:

• Step 1: independent evaluators get employee-level access inside AI labs.
• Step 2: leading labs in democracies agree on shared safety standards and limits.
• Step 3: labs eventually try to coordinate pacing with authoritarian governments.

Amodei himself called that last step the toughest dilemma. It's the only step that depends on a rival government.

His reasoning is blunt. In the essay, he wrote that pacing within democracies "will be limited by the lead that US companies have over authoritarian regimes, chiefly the Chinese Communist Party." Slow down by more than that lead, he argued, and projects that aren't paced will pull ahead. He has also urged tighter chip export controls and action against what Anthropic calls illicit distillation by Chinese labs, a claim we unpacked in our look at Anthropic's distillation report.

Support at home came fast. Sam Altman and Elon Musk backed the direction. Andrej Karpathy wrote that he hopes the industry can "come together" to make it happen, and Senator Brian Schatz called it "a reasonable start." None of those backers speaks for Beijing, which is where the plan hit a wall.

Why Did Beijing Reject a China AI Slowdown?

Because it reads the plan as containment wearing a safety badge. The response came from both the government and state media within two days.

At a press briefing, foreign ministry spokesperson Guo Jiakun said that "fearmongering, confrontation and malicious competition will only disrupt the process of global AI governance and serve no one's interests." China wants all sides to promote "an open, inclusive and benevolent approach to AI," he added.

State-run Global Times went further. It called the proposal a "Cold War playbook" for the AI sector, "packed with containment provisions targeting China." Its argument: the talk of global safety hides an effort to keep China out of cutting-edge AI and out of the rules that will govern it.

Analysts say Beijing wasn't likely to accept outside limits in any case. Ja Ian Chong of the National University of Singapore told NBC News that China "wants to write the rules" rather than follow someone else's.

The honest read: China's objection has a factual hook. The same plan that asks for restraint also asks for tighter chip controls on China, so Beijing can fairly say the safety and the competition are bundled together. That doesn't prove the safety concerns are fake. It does explain why a proposal written in Silicon Valley was never going to be accepted in Beijing as written.

Why Did Trump Also Reject It?

For the mirror-image reason: he doesn't want to give up a lead. Washington's objection sounds a lot like Beijing's, just from the other side.

"We're leading China in AI," Trump said, "and frankly I want to keep it that way because whoever wins AI wins." He also criticized what he called "very negative forces" raising risks he considers unlikely. That puts the White House at odds with the CEOs who backed the plan, including some it works closely with.

Timing makes this more than talk. Trump has said Xi Jinping will visit Washington on September 24 and that the two leaders will discuss AI. Reports say managing frontier AI risks and protecting intellectual property are near the top of the agenda.

The honest read: both governments are making the same argument about each other. Each says slowing down would hand the other side an advantage. That's the classic trap of an arms race, and it's why labs can't solve the China problem on their own. Only the two governments can.

Is the US-China Gap Big Enough to Pace?

Partly, depending on what you measure. Amodei's logic depends on a US lead, so it's worth checking how large that lead really is.

Factor	United States	China	Why it matters for pacing
Private AI investment (2025)	About $285.9 billion	About $12.4 billion	A big funding lead buys some room to slow down
AI chip supply	Controls the most advanced chips and export rules	Relies heavily on imports; domestic output is growing	Chips are the most checkable part of any deal
Frontier models	Leads at the top end	Strong fast followers, often open-weight	A small model gap leaves little room to pause
Energy and materials	Power supply is a constraint	Cheap power and rare-earth advantages	Data center growth may not track chip access
Official stance on a slowdown	Rejected by the president	Rejected by the foreign ministry	Neither side has offered to go first

The funding gap is huge. According to figures cited by Al Jazeera, US companies invested about $285.9 billion in AI in 2025, compared with about $12.4 billion in China. But money isn't the same as capability.

Chinese labs have kept close despite limits on chips. As of mid-2025, economic historian Chris Miller wrote that Huawei was expected to make only about 200,000 AI chips that year, while China legally imported about 1 million from Nvidia. Yet he still described Chinese labs as, at worst, "fast followers" on model capability. Models like those we compared in Qwen vs DeepSeek vs Kimi show why.

The honest read: the US lead is real in money and chips, but thinner in models. That matters for Amodei's own test. If the lead that allows pacing is measured in months of model capability, it doesn't leave much room to slow down.

What Would a China AI Slowdown Actually Require?

Verification, first of all, and nothing close to that exists yet. A slowdown nobody can check is really just a promise.

Experts at the Atlantic Council laid out what's missing and what might work. Graham Brookie noted that only 9 of 114 AI safety institutes formally coordinate with each other. Kenton Thibaut called a major US-China breakthrough on AI safety "highly unlikely," though he sees room for narrow cooperation. Here's what they and others point to:

• Confidence-building measures. Safa Shahwan Edwards suggested steps modeled on cyber agreements, like named contacts in each country for AI incidents.
• Narrow shared goals. Thibaut pointed to preventing misuse by non-state actors, where both governments' interests overlap.
• Joint monitoring. Chinese scholars have discussed shared systems for risk warnings and emergency response.
• Chip and compute tracking. This is the most checkable lever, but no system tracks it across both countries today.
• Real evaluator power. Edwards also asked whether evaluators who "rely on frontier labs for access" can ever be strong enough.

There's one bright spot. Emerson Brooking called Anthropic's evaluator pledge "one of the most tangible transparency commitments in tech history." That step doesn't need China at all, which is why it's the only one already under way.

What Happens to the China AI Slowdown on September 24?

Probably not a deal, but maybe some useful signals. The Trump-Xi meeting in Washington is the first high-level test since the plan came out.

A broad pacing agreement is very unlikely, given what both sides said this week. Both leaders have framed AI as a contest to win, not a risk to share. Smaller outcomes would still be worth watching:

• A joint statement on AI risk. Even vague shared language would be more than both governments offered this week.
• An incident hotline. A named channel for serious AI incidents would be a concrete first step.
• Export control changes. Any loosening or tightening of chip rules will shape both countries' AI plans for years.
• Intellectual property terms. Given the dispute over distillation, any deal on model IP would matter to labs on both sides.
• Silence from the labs. Whether Chinese AI labs comment at all will show how much room they have to take part.

The honest read: judge the meeting by concrete mechanisms, not warm words. A hotline or a shared incident process would be real progress. A pledge to "cooperate on AI safety" with nothing attached would not.

Can Labs Pace the US-China AI Race Without Governments?

Only in small ways. Labs control their own release schedules, but not the forces that push them to race.

A single lab can slow down, and Anthropic says it will add outside evaluators on its own. But three things sit outside any lab's control in the US-China AI race:

• Chips. Export rules are set in Washington, and Beijing decides how hard to push its own chip production.
• Law. US labs would need antitrust relief to agree on a shared pace, and no such waiver exists yet.
• Rivals abroad. No US lab can make a Chinese lab slow down, and no Chinese lab has offered to.

That's why the China AI slowdown question keeps landing back on governments. Labs can make pacing possible by making their work checkable. Only states can make it binding.

What Does the China AI Slowdown Debate Mean for Builders?

That policy risk is now part of your model choices. You probably won't see a global pause, but you may see sudden changes in which models you can use, where, and under what rules. In the US-China AI race, those changes can arrive with little warning.

• Expect volatility, not a pause. Export rules, model access limits, and trade deals can shift fast. Plan for change, not a freeze.
• Keep your stack model-agnostic. If swapping a model is a config change, a policy shift is an inconvenience instead of a rebuild.
• Know where every model comes from. Track each model's origin, license, and hosting location, especially open-weight models from any country.
• Build governance that travels. Rules differ by region, so design controls that satisfy several at once, as we cover in AI governance and EU AI Act compliance.
• Keep evaluation running. If you're forced to switch models, you'll want an eval set ready to test the replacement, as outlined in AI agent evaluation.

The honest read: the slowdown debate may never change how fast frontier labs work. But the politics around it are already shaping the tools you depend on. Teams that can switch models and prove where their models come from will handle that far better than teams locked into one provider.

How Van Data Team Helps Teams Plan for Policy Shifts

We help teams build AI systems that keep working when the rules change. That means model-agnostic routing, a clear record of which models run where and under which licenses, and evaluation that can test a replacement model in days, not months.

Whatever comes out of September 24, the direction is clear: AI policy is becoming geopolitical, and it will keep moving. The China AI slowdown debate won't be settled in one meeting, but it shows how AI policy will be made from here. If you want your systems ready for that, our work on AI governance and AI agent evaluation is a good place to start, along with our earlier analysis of the AI slowdown debate.