BEIJING, 20 AUG 2026 — Nvidia H200 accelerators are reaching mainland China again. ByteDance and Tencent have each received around 10,000 of the processors in recent weeks, and Nvidia partners including Lenovo have told Chinese customers they can resume ordering.
The restriction that had been blocking these sales was not American. It was Chinese.
What has moved
Limited batches have been approved into the mainland, in what is described as a partial easing rather than a general opening. Purchases still require separate sign-off from China's National Development and Reform Commission.
Nvidia is reported to hold roughly 500,000 H200 units in inventory, largely built for Chinese customers, with sales held up by Beijing's own restrictions. The H200 is at least two generations behind the parts Nvidia sells elsewhere, which Chinese buyers cannot purchase under United States export controls.
Two governments, one queue
The usual framing of this trade has Washington restricting and Beijing wanting. For the past several months the binding constraint has run the other way.
Washington permitted the H200 specifically because it is old enough not to matter strategically. Beijing then discouraged its purchase, on the reasoning that buying deliberately-degraded American silicon subsidises a supplier while doing nothing for domestic capability. We covered that calculation in July when Beijing was weighing whether to let its AI champions buy the chip it had spent a year keeping out.
The half-million units sitting in inventory are what that policy looks like on a balance sheet. Nvidia built for a market that was legally open and politically closed, and has been holding the stock ever since.
Approvals are now granted case by case, which is an easing, not a reversal. By retaining sign-off on individual purchases, the NDRC keeps the volume, buyer, and timing under state control — a more flexible instrument than an outright ban.
The "two generations behind" label is misleading. For serving large models, where memory capacity and bandwidth are the constraints, the H200 is still highly capable. What it lacks is the interconnect density and raw throughput that make very large training runs economic. For a company answering millions of daily requests, the gap between this silicon and the current generation is much narrower than the number implies.
Ten thousand chips is a specific quantity, not a symbolic one
Twenty thousand accelerators is significant capacity, but it is modest compared to what either firm would deploy if unconstrained or against the fleets their American competitors run.
The quantity suggests a controlled release, not a policy reversal. It is enough to relieve pressure on two companies with immediate serving needs, but not enough to change anyone's training roadmap. Both recipients run enormous consumer platforms where inference demand is continuous and non-negotiable, which is precisely the workload where older silicon remains useful and where doing without has a daily cost.
That is the shape of the compromise. Older chips go to inference for consumer products. Frontier training stays a domestic problem.
The commercial reading for Nvidia is less comfortable. Inventory built for one customer and held for a year is capital tied up in depreciating parts. Clearing it at any approved price is better than a write-down. And Nvidia's negotiating position is unusually weak: the buyer knows the stock exists, who it was for, and that no one else wants it.
Worth noting who is not in this story. Nothing published names a third Chinese buyer, and the two that are named are consumer platform companies rather than the cloud providers or research institutes that would need silicon for training. If the approvals were about building national capability you would expect a different list. This list looks like it was chosen to keep two very large consumer services running smoothly.
What it means read against everything else this month
We wrote yesterday that Samsung raised advanced foundry prices by up to 15 per cent, with Chinese buyers accepting the steepest increases. The two events tell the same story.
Chinese demand for compute is strong enough to absorb a 15 per cent price rise at the foundry and to make two-generation-old accelerators worth queueing for. The constraint on Chinese firms is not appetite or capital, but access — and that access is limited on political, not technical, fronts.
For anyone modelling the AI hardware market, that matters more than the unit count. A 500,000-unit inventory that can be released or withheld by administrative decision is not a market in any ordinary sense. It is an allocation, and its clearing price is set in two capitals.
The regional consequence is about the second-hand market
Southeast Asia sits close to this trade and is affected by it indirectly rather than directly.
Regional cloud operators and research institutions buy the same class of accelerator, and they have spent two years competing for supply against buyers with far deeper pockets. Half a million H200s moving out of a warehouse and into Chinese data centres removes a large block of inventory that was, in principle, available to everyone else.
But there is a countervailing effect. Chinese firms taking delivery of older silicon frees nothing in the current generation, so the practical impact on a Malaysian or Indonesian buyer shopping for current parts is close to zero. Where it will show up is in the used and refurbished market in two or three years, when this cohort is retired — and for a regional university or a startup, that market is often the only realistic route to serious capacity.
What we could not establish
Whether more approvals are coming. Two named recipients and a partner note about resumed ordering is not a policy statement, and the reporting does not say whether this is a pilot, a quiet normalisation, or a one-off for two firms with particular leverage.
Several points remain unestablished. We do not know what conditions the NDRC attaches to approval, whether other Chinese firms have been refused, or what the chips will be used for. The price paid, the accuracy of the inventory figure, and any interaction with domestic accelerator programmes are also unknown.
What to watch
Watch whether a third and fourth buyer are named. Two companies is an exception; five is a policy, and the difference tells you whether Beijing has changed its mind or simply made a calculation about two specific firms.
Then watch the inventory. If a meaningful share of half a million units clears over the next two quarters, that is a decision to let the domestic industry buy foreign silicon for inference while it builds its own. If it does not, this was a valve being opened briefly.
Finally, watch what Chinese domestic accelerator makers say. Every H200 delivered is a sale they did not make, and the argument for restricting these purchases in the first place was made on their behalf.
