How did Chinese models get this close this cheaply?

Export controls were supposed to freeze the gap. They helped close it instead.

Chinese models are now near the frontier at roughly one sixth to one quarter the cost of comparable US systems. That is not a rounding error. The main engine is algorithmic efficiency: mixture of experts, compressed attention, the whole bag of tricks you invent when you cannot just buy more of the best chips. The export restrictions did some of that work for them. Starve a lab of easy compute and it learns to waste less.

Domestic chips plus older or smuggled high-end silicon fill the rest of the rack. Power is not the bottleneck people in the West keep assuming. Pricing is a mix of real lower serving costs and a state-backed push for share. Not pure predation, not pure profit. Adoption and lock-in first, margin later.

Western companies are already using them. Assistants, coding tools, payments firms, commerce platforms, industrial names. One widely cited estimate puts Chinese base models under about eighty percent of US startup derivative work. The hesitation you hear in boardrooms is about where the data lives and who might be listening. It is not, in the record we have, a rush toward Chinese models because people distrust Western alignment work.

And the summit will not magic this away

There is a leaders' meeting in Washington on 24 September with AI safety, IP, and chip controls on the agenda. An earlier meeting in Beijing this spring barely touched chips. The people who watch this for a living are not expecting a real bargain. A chips-for-safety trade gets rejected when one side still thinks it can close the gap. The likely text is toothless risk-management language, not a binding slowdown.

If the cheap models are already inside Western products, what exactly is the coordination plan coordinating?

That is the awkward bit. You can write communiqués about shared standards while the stack underneath is already mixed. Efficiency forced by sanctions, subsidized prices, and a market that will take a good-enough model at a quarter of the cost do more to set the pace than any clause in a summit readout. China is the weak link in every slowdown story because it has a reason not to slow down. Pretending the next meeting will change that incentive is how you get a nice press release and the same race on Monday.