The talks that are not about safety

The United States and China are preparing mid-September talks on AI safety. Both sides will speak the language of catastrophic risk because that language travels. Neither side is flying senior people across an ocean to workshop evals. The talks sit on top of a two-year grind over advanced chips, cloud access, and the right to set the standards that governments will eventually impose on labs. Safety is the door. Industrial policy is the room.

1. 01US and Chinese officials have scheduled AI safety discussions for mid-September, the first structured bilateral of this kind in the current cycle.Reuters, 5 September
2. 02US export controls on advanced accelerators remain the binding constraint on Chinese frontier training. Workarounds — smuggling, older-chip clusters, domestic silicon — have narrowed the gap without closing it.
3. 03Western labs have spent 2025–26 publishing safety frameworks in public and racing capability in private. Governments have noticed the gap between the two.
4. 04The EU AI Act is now in force in stages. China has its own algorithm and deep-synthesis rules. The United States still does not have a comprehensive statute, which makes a bilateral “safety” channel unusually valuable as a substitute for law.

We assess that both parties want a communiqué more than they want a regime. A communiqué lets each side tell a domestic audience that the other has been engaged. A regime would require verification, which neither intelligence service will offer and neither industry will accept.

The useful signal is not the statement. It is the annex, if there is one: any language on chip thresholds, cloud-service restrictions, or joint evals on dual-use models. If the annex is empty, the talks were a photograph. If it names numbers, procurement and training schedules at the major labs move within a quarter.

Labs will brief this as reputational. It is not. It is a constraint on where weights can be trained and where they can be served. Founders who treat “AI safety talks” as CSR will discover it in their cloud invoice.

• Export-control risk belongs on the same dashboard as model quality for any company training or serving frontier systems.
• A vague communiqué is still useful to Beijing and Washington. Do not over-read it as détente.
• EU rules plus a US–China channel means the labs are being regulated from three directions with no single map.

If your product depends on a US-hosted frontier model, write down what happens if serving into China, the Gulf, or the EU is restricted further. That memo is the briefing. Send it before the talks, not after the communiqué.

• Whether the mid-September readout names chips, compute, or only “safety.”
• Any pause or widening of US accelerator export licenses in the following two weeks.
• Lab public letters timed to the talks — they are positioning, not confession.

• Reuters — Mid-September US–China AI safety talks
• Public statutes — EU AI Act staging; PRC algorithm rules