Z.ai stated on 17 September that GLM-5.3-Flash — 320 billion parameters, 18 billion of them active, a one-million-token context window — is deployed across more than 100,000 Chinese accelerators. The lab claims to have tripled its throughput in under two weeks and to have reached a cost per token comparable to that of current American chips. If these figures are confirmed, it is the bet on which export controls rested that must be reassessed.
We have been following this story since the revelation that this model was running on domestic silicon, then the announcement of a cluster of that size. Here is confirmation from the operator.
The figure that matters is not 100,000
The number of accelerators is impressive and secondary. What decides the matter is the claim about cost per token.
We wrote a week ago: a hundred thousand domestic accelerators are not worth a hundred thousand cutting-edge accelerators. A larger cluster with less capable components can match a smaller cluster with better components, at the price of higher power consumption.
If the cost per token really is comparable, that means the gap in per-unit performance is being offset — by volume, by software optimisation, or by cheaper electricity. The mechanism does not matter: the economic effect is the same.
Tripling throughput in two weeks
This is the most instructive detail, and it concerns software rather than hardware.
Tripling throughput on an unchanged fleet in two weeks is not achieved by adding machines. It is achieved by optimising the way the model is served: better memory management, batching of requests, tuning of compute precision.
And that is precisely where a good part of the dominant ecosystem's advantage lay: twenty years of software tooling. Progress of that magnitude in two weeks suggests that this gap is closing faster than expected.
These figures come from the company operating the infrastructure, and they are not audited. "Comparable" is an elastic word: comparable to which chip, under what conditions, including or excluding electricity and depreciation? And above all, serving is not training. Running an existing model on domestic hardware is one thing; training a leading-edge one from scratch is another, and far more demanding in terms of interconnects between components. The demonstration covers the most visible part of the chain, not the hardest.
What this implies for export policy
The reasoning behind the controls was as follows: by limiting access to the most capable accelerators, you prevent the build-up of facilities on the scale required for leading-edge models.
That bet assumed no viable alternative would emerge in time. If a 320-billion-parameter model is served at large scale on domestic hardware, at a comparable cost, the assumption no longer holds in that segment.
That does not mean the controls were useless: they probably delayed things, and delay has value. It means they also accelerated self-reliance, which is the classic side effect of technology restrictions.
We wrote about China's five-year plan that it would take several years to know which of the two logics prevails. This week provides one data point.
What to take away
This case illustrates a point we developed yesterday in our article on the race metaphor: a technical lead does not hold like a position in a ranking.
What does hold is physical assets — chips, electricity, factories. And those are precisely what China is building, with a plan that is costed and dated.
For a European observer, the useful information is not knowing who is ahead. It is that two complete ecosystems now coexist, with frontier models available free of charge on both sides. In that configuration, the question of dependency arises differently — and rather favourably for those who know how to build on what others publish.