The Aggressor Asks for Terms

On September 24 Xi Jinping is due at the White House, and artificial intelligence is on the agenda. Since 2022 the American position on China and AI has been a wall: cut off the advanced chips, choke the supply, keep the frontier out of Beijing’s hands. The reported agenda for the twenty-fourth, still provisional, is softer. Washington wants to talk about policing AI-directed cyberattacks, getting the labs on both sides to share threat information, and Chinese firms distilling American models. The wall has begun turning into a request for cooperation.

What Each Side Brings

The two agendas are not the same kind of thing. The American asks, reported separately by Nikkei and Reuters ahead of the summit, are defensive: help us stop AI-directed cyberattacks, get the labs on both sides to police themselves and share threat information, and raise the allegation that Chinese firms distilled proprietary American models. Every one of those is a request to limit a harm the United States feels exposed to. China’s clearest ask points the other way. Beijing wants the export controls loosened, and has separately sought relief on the high-bandwidth memory its AI firms were cut off from. One side arrives to contain its own vulnerabilities. The other arrives to collect an input it still lacks. Only one of those is the shopping list of a side that thinks it is winning.

From Eliminate to Manage

Meaningful concessions on the chip controls appear unlikely at this summit. The concession that matters is quieter, and it is a change of goal. The chip war was built to keep China off the frontier. It did not. China reached near-parity on the leading model benchmarks anyway, on a fraction of the measured private investment, which I traced in China Built the Grid. Having failed to knock China out, Washington has changed what it is trying to do. The new object appears to be guard rails, coexistence, the management of a race it could not simply win.

Chinese state media says the exclusion failed. A Global Times commentary this summer wrote that the American containment strategy was “supposed to make the other side fall behind due to a lack of resources” and had instead failed on both the hardware and the software fronts. The state press of the country the wall was built to contain is proclaiming its failure.

The Wall Came Down on Its Own

The wall did not just fail. Washington opened a gate in it. First the bans: the A100 and H100, then the cut-down H20 built to slip under the line, then the H20 put under an indefinite license requirement in April 2025. Then, in December 2025, Trump announced the reversal. A January rule set a 25% US tariff on the H200 for China, with every unit routed through American territory for testing first, and by May, Reuters reported, Commerce had cleared around ten Chinese firms, Alibaba, Tencent, ByteDance and JD.com among them, to buy up to seventy-five thousand units each, though none had yet shipped. The chokepoint had become a tollbooth. The administration sold it as a win, a way back into the Chinese market and tariff revenue for Washington. Whatever it was, it was not a wall. Selling the story of a win in place of the win is the move I traced through OpenAI’s AGI claim; here a government runs it on its own retreat.

Then the target walked away from it. Beijing told its firms to pause the orders and steered them to domestic silicon, Huawei’s Ascend and Cambricon’s parts, while some local governments subsidized data centers that switched to Chinese chips. Nvidia’s share of China’s advanced-chip market had already fallen from around ninety-five percent to effectively zero, its chief executive said in October 2025, with billions of dollars a year in sales unrealized. The United States offered to sell the chips back. Trump said China would rather build its own. To a real degree it had.

That is the state of the weapon when the two sides sit down. China’s open-weight models had already climbed to the top of the open-weight board, Moonshot’s Kimi K3 near the top of it and running at roughly a quarter of Claude Fable 5.1’s task cost. China still cannot make a full substitute for the best American chips at scale, and American labs still hold the very top of the model board. But the thing the controls were built to prevent, a Chinese model frontier that does not need American permission, exists.

The Asymmetry of a Ceasefire

A deal that constrained the race could freeze part of the board, and who gained would depend on what froze. The deal may not come. What matters is the frame the meeting fixes, signature or not. Washington sits down to negotiate coexistence, and the wall stops being the only public goal whether or not a chip changes hands. China took that before anyone sat down. On the measures that compound, cost, openness, downstream diffusion, the direction is China’s.

In Busan last October, on the sidelines of APEC, Beijing suspended for one year the sweeping rare-earth export controls it had rolled out three weeks earlier. In exchange Washington halved its fentanyl tariff, extended tariff exclusions into late 2026, and dropped a threatened hundred-percent levy. Beijing controls the bulk of the world’s rare-earth processing. It used that grip to pull the concessions out of Washington and gave up, as its side of the bargain, a restriction it had imposed only weeks before. Brookings judged that Beijing had played its cards adeptly and kept its leverage. At Busan, China dealt from a chokehold, trading a thing it controls for a thing it wants, and keeping the chokehold. On AI the leverage is a different kind. Rare earths are a gate China can close. Its AI trajectory is a door it can walk through, and there is nothing at this table worth turning back for.

The Case That This Is Not Weakness

The honest counter is that talking is not losing. Great powers manage rivalries at the table constantly, and the risks on this agenda are real and shared: an AI-directed strike on a power grid does not check the passport of its target. Brookings argues the dialogue should stay exactly this modest, a narrow technical exchange on cybersecurity and misuse, with export controls and model access kept off the table. On that reading the summit is adults managing a dangerous machine.

Grant it. The risks are real and the talking is sane. It does not touch the tell, because the tell is not in the fact of the meeting, it is in the direction of the ask. The side that spent four years trying to win by exclusion, and now asks the other to help manage shared risks, is the side that concluded exclusion failed. You convene guard rails when you have stopped believing you can keep the other driver off the road.

Why China Signs Nothing

China’s rational play does not depend on trusting the offer. Its trajectory already wins. The easy move is to take the meeting, take whatever chip relief falls out, offer cooperation in language soft enough to commit no one, and keep building. The pressure for a deal is coming from Washington’s own side of the room, a senator calling for a treaty, a pundit dreaming of a Nobel, and Anthropic, which pressed for government-backed testing over the summer and called for a slower frontier on September 12, an instinct I traced in Five Little Pigs, a call the President rejected two days later. They are drafting terms for a deal the party with the better hand has no reason to sign.

The Close

The summit will be described as statesmanship, as guard rails, as the responsible handling of a technology that frightens everyone. Strip the vocabulary and what remains is a leading power asking the challenger it could not stop for help. I understand why Washington wants the table. I do not understand why Beijing leaves it having conceded anything, and I do not think Beijing does either. It has read the same benchmarks. It will smile for the cameras. It will keep running.

Washington has named its guard rail. In early August the same administration finished a voluntary framework for reviewing frontier models before release, kept its benchmarks classified, and would not say what is in it. Pressed this week to slow the race, the President said the only control AI needs is “a STRONG AND SMART (High IQ!) PRESIDENT.”

I would rather the United States win this outright. The worry is narrower and worse. A power that trades its advantage for the appearance of one still has to fight the contest, only later, at a higher price, and on someone else's schedule. The precedent outlasts the table. Treat Beijing as an equal on AI, and it will expect to be treated as one on everything that comes after.