Washington and Beijing Negotiate AI Four Days Before Xi Lands

U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng meet in New York today, September 20, on artificial intelligence, trade and critical minerals. The meeting exists to prepare agreements ahead of a Trump–Xi summit, with Xi's U.S. visit planned for September 24.

The AI talks themselves are an outcome of the Trump–Xi meeting in May. Four months of working-level preparation lands in a single day in Manhattan, four days before the principals meet.

Expectations are modest, and the people setting them are the ones in the room.

What is actually plausible

The two concrete items analysts point to are narrow and specific: an emergency AI crisis deconfliction hotline, and a commitment to keep AI out of nuclear command and control.

Neither constrains capability. Neither touches export controls, compute access, model weights or the commercial competition. Both are drawn directly from the Cold War arms-control playbook — the hotline is literally the 1963 Moscow–Washington direct line rebuilt for a different failure mode.

That is not a criticism. It is what is available. The two governments do not agree on what AI risk is, who should govern it, or whether the other side is acting in good faith. What they can agree on is that neither wants an automated system to start something neither intended. Deconfliction mechanisms are what you build when you cannot build trust: they assume the relationship stays adversarial and try to keep the adversary reachable at 3 a.m.

Analysts expect limited measures demonstrating willingness to avoid escalation, with broad agreement unlikely. One commentator's read — that the likely outcome is a handful of toothless risk-management clauses — is probably right on substance and wrong on significance. Toothless clauses that establish a channel are how every subsequent agreement gets negotiated.

The loophole sitting on the table

The talks are complicated by a live enforcement problem that surfaced days before the meeting.

An investigative report on the American subsidiary of a blacklisted Chinese server maker described the free export of Nvidia's most advanced AI chips through a gap in Entity List rules — adding fresh uncertainty to a summit that was already fragile.

The timing is awkward for both sides in different ways. For Washington, it undercuts the claim that export controls are working, at the exact moment U.S. negotiators need leverage. For Beijing, it makes any concession look like a payment for access it was already obtaining.

The gap is structural, not incidental. In May 2026, BIS issued guidance requiring export licenses before selling advanced computing products to any company whose ultimate parent is based in an arms-embargoed country or Macau — even when that company is located elsewhere. That guidance exists because the previous rules were drawn around geography and the evasion was drawn around corporate structure. Every tightening generates the next workaround: subsidiaries in permitted jurisdictions, and remote access to Nvidia compute sitting in Southeast Asian data centers, which moves the workload rather than the chip and is nearly impossible to police at the border.

Beijing already answered

The thing most likely to shape these talks is a decision China already made.

China's internet regulator barred major domestic tech firms from purchasing Nvidia chips in September 2025. In November 2025, Beijing mandated that state-funded data center projects use only domestic AI chips — a directive that covers even the export-licensed H20 and H200.

That materially reduces what Washington has to trade. Export controls are leverage only if the other side wants the goods. Beijing has spent two years building the industrial and policy apparatus to not want them, treating supply as a vulnerability rather than a benefit. The domestic-silicon mandate is the clearest possible statement that China intends to lose access on its own schedule rather than have it revoked on someone else's.

So the chip-access bargaining space is narrower than it looks from the outside. Which is another reason the deliverables are hotlines and nuclear carve-outs.

What the summit is really for

The reason this matters beyond the communiqué is that AI governance currently has no shared international structure. The EU has the AI Act, the U.S. has executive action and export controls, China has its own regulatory stack, and China chartered a rival institution for global AI governance earlier this year. Those are four separate systems with no arbitration layer.

A Trump–Xi summit on September 24 that produces even a deconfliction hotline creates a bilateral channel where none existed — and the U.S.–China relationship is the one that determines whether any multilateral framework is possible at all. Everything else is downstream of whether these two governments can talk.

The honest assessment is that today's meeting is a preparatory session for a summit that will likely produce symbolism. But symbolism at this level is infrastructure. The hotline installed in 1963 was also symbolic, right up until the first time someone had to pick it up.