China Rejects Amodei's AI Moratorium: Subsidies, Foundries and the Chip Chokepoint

Washington, 1922: whoever proposes the limit already holds the lead

On 6 February 1922 the naval powers signed a treaty in Washington setting capital-ship tonnage at a 5:5:3 ratio between the United States, the British Empire and Japan. The mechanism was simple: freeze the existing hierarchy.

Tokyo signed from a position of disadvantage. In December 1934 it gave formal notice of withdrawal, and by the end of 1936 the treaty was a dead letter. Japan had meanwhile built the industrial capacity it lacked in 1922.

The pattern returns in 1946. The Baruch plan proposed international control of atomic energy while Washington held a monopoly on the weapon. Moscow refused and tested its first device in 1949.

Three precedents are enough to call it a pattern, and the third arrives in 1972: the freeze on strategic launchers became signable in Moscow on 26 May, with SALT I, once parity was close. The limit passes when the gap closes. Never before.

September 2026: the proposal and the refusal

In early September 2026 Dario Amodei, chief of Anthropic, asked the industry to slow the development of the most powerful models, writing that the measure requires Chinese cooperation. The same text calls for an end to sales of chips and semiconductor equipment to Beijing, a crackdown on smuggling, and limits on remote access to data centres outside China.

The Chinese response came within days. The day after the post, the Global Times ran an editorial on the «Cold War playbook» opened by America's tech right.

On 11 September 2026 foreign ministry spokesperson Guo Jiakun dismissed the appeal as «alarmism, confrontation and vicious competition», according to Rest of World's account of 23 September 2026^[1]. On 14 September China's cybersecurity standards body released the third edition of its governance framework, which warns about the risk of «recursive self-improvement» in systems beyond human control.

Beijing rejects other people's terms and meanwhile writes its own.

The mechanism: a moratorium photographs the hierarchy

A moratorium freezes relative positions at the moment of signature. The leader buys time; the chaser loses the only resource left to it, namely speed.

From this follows the historical rule: the power in front proposes the limit, the power behind refuses it, and the limit becomes negotiable once the margin thins. Alvin Wang Graylin, senior fellow at the Asia Society Policy Institute, puts it plainly: Beijing refuses to be held back any more than it already is. Which leaves open the possibility of regulating the pace on different terms.

The interesting data point lies elsewhere. A slowdown proposal that also asks for tighter export controls contains an implicit admission: the current chokepoint is judged insufficient by the very people who built it.

This is the structure to read. The public debate talks about model safety; the document talks about travel time.

Verification is missing, and verification is everything

The INF treaty, signed in Washington on 8 December 1987, worked for a technical reason: on-site inspections, counted, repeated, with permanent staff stationed at the facilities. Trust counted for little; physical access counted.

Frontier compute offers similar handles, in theory: fabs are few, clusters consume measurable power, accelerators carry serial numbers. The Council on Foreign Relations has mapped this possible architecture in an analysis devoted to the prospect of an AI arms control deal with China^[2].

Between theory and treaty lies the hard part. A verification regime requires inspectors inside the other side's data centres, access to training logs, declarations on the number of chips installed. Every item touches the heart of national security on both sides.

That is why the first agreements will be declaratory: shared language, communication channels, perhaps a principle on human command of nuclear weapons. The substance comes later. And it comes slowly.

The chokepoint holds on hardware and loses on capital

The US-China race on artificial intelligence is about access to semiconductors more than about model quality. Whoever controls the fabs and the lithography controls the outcome.

Chinese industrial policy for AI works on exactly that point: subsidies, state funds, guaranteed domestic demand for home foundries. The capital market does the rest. The IPO of the foundry CanSemi was oversubscribed 2,360 times, with investors chasing the AI boom, as the South China Morning Post reports^[3].

A multiple like that says one thing only: domestic capital is funding the replacement of the chokepoint at a cost of money approaching zero.

Export controls bite on hardware and leave finance uncovered. Hong Kong remains an open door to Chinese laboratories. The bottleneck is the fab; the leak is the money, and money moves faster than strategic plans assume.

Where this desk stands

The Chinese refusal should be read as an indicator on the American lead, before it is read as AI policy news. Whoever asks for a moratorium is signalling that their own margin is compressing.

This is a regime change, and it is worth saying so plainly. The 2022-2026 period saw technological control exercised administratively, through lists, licences and embargoes. The phase now opening shifts the ground to negotiation between peers, a different discipline, with different instruments and different timescales.

What would change this reading: evidence that the controls are becoming more effective. For instance a verifiable and widening delay for Chinese foundries on advanced nodes, or a collapse in the multiples on semiconductor listings in Shanghai and Hong Kong.

Evidence of that kind would shift the thesis towards the opposite hypothesis: the chokepoint holds, and the slowdown proposal becomes the luxury of a comfortable leader. The market today prices that first scenario as a certainty. That is where it is wrong.

Three implications for capital

Family offices and sovereign funds, 36-month horizon: Chinese domestic substitution on mature nodes becomes an allocation thesis with high political risk and low correlation to the American AI complex. The distinction to hold firm remains the one between advanced nodes and mature nodes.

CEOs and boards, 18-month horizon: industrial plans assume a world of two separate, stable technology stacks. A bilateral negotiation opens a third path, namely shared standards on model auditing and safety, with double compliance costs for anyone operating in both markets.

Chief risk officers, 24-month horizon: VAR models treat export controls as a slow, near-fixed variable. The absent scenario is a negotiated easing, which would reprice downwards the scarcity premium enjoyed today by those who live off fragmentation.

Any CFO taking investors a story of durable technological advantage would do well to add a line on the opposite risk: a world where the frontier closes in and the premium evaporates.

The forecast

By 31 December 2027 the United States and China will still lack a binding bilateral agreement, with verification, on the training of frontier models. They will instead reach declaratory understandings: dialogue on shared risks, common language, perhaps a principle on human control of nuclear weapons.

Confidence: 78%. Horizon: 31 December 2027, that is 458 days. Verification: the publication of a text signed by both parties.

Kill signal: a bilateral text signed by that date setting a compute threshold for training and providing for data centre inspections refutes the thesis.

What to watch

Three indicators anticipate the outcome, and they should be read in this order.

• The official language after the Trump-Xi meeting of 24 September 2026 in Washington: the appearance of the formula «shared risks» in statements from both sides.
• The subscription levels of the next Chinese foundry IPOs: falling multiples signal a chokepoint that holds.
• The fourth edition of China's AI governance framework: any reference to compute thresholds or external verification foreshadows negotiating ground.

The divergence between a world asking for rules and a world accelerating always resolves. The question is how, and who pays the bill for the transition.