Nvidia's H200s are flowing to China again, at 13% of the legal limit

Nvidia's H200 accelerators are reaching Chinese data centres again, and the number that matters is not the approval. It is how far below the ceiling the shipments landed.

ByteDance and Tencent have each taken delivery of roughly 10,000 H200 processors over recent weeks, according to a Financial Times report carried on August 19, 2026. Other Chinese firms are expected to receive comparable tranches. Under the licensing framework Washington issued in January, each approved customer may buy up to 75,000 units.

Ten thousand against a permitted seventy-five thousand is thirteen percent. Everything interesting about this story lives in the other eighty-seven.

The stalemate this breaks

To read the number properly you need the eighteen months behind it.

In April 2025 the United States restricted even the H20, the part Nvidia had specifically designed to comply with the previous rules, and the company took a multibillion-dollar writedown on inventory it could no longer sell. In December, the policy reversed again and H200 sales to China were approved under a new structure. By January roughly ten Chinese firms had been cleared, Alibaba, Tencent, ByteDance and JD.com among them, each with that 75,000-unit allowance.

Then nothing happened. By late February, Nvidia still had not sold the chips Washington had approved, and was publicly worried that Chinese rivals would fill the gap while the paperwork sat idle. Licences existed on one side and no hardware moved on the other. In March, we reported that Nvidia had halted China-bound H200 output entirely and redirected the TSMC capacity to its Vera Rubin line, which is what a company does when it concludes the demand is not going to convert on any useful timescale.

So the correct frame for August is not "China gets chips." It is that a pipeline which had been fully approved and completely inert for roughly seven months has started moving, at a trickle.

What the licence actually costs

The conditions attached are heavier than the word "approval" suggests, and they are worth listing because each one is a plausible brake on volume.

Approved buyers must demonstrate that adequate security procedures are in place and that the hardware will not be diverted to military use, subject to third-party review. Nvidia must separately certify that it holds sufficient inventory in the United States before it exports, which quietly converts an export licence into a domestic allocation decision. Reporting on the January arrangement also describes chips routing through the United States for mandatory testing and certification, with the US government taking a share of the sale proceeds, terms Nvidia has not detailed publicly.

Read that list again and notice what it does to unit economics. Every H200 sent to Shenzhen is one not sold to a domestic customer at a moment when advanced capacity is the binding constraint across the entire industry. It carries compliance overhead, a routing detour, and a revenue haircut. A 75,000-unit permission is a theoretical maximum, not a forecast.

Four readings of the gap

The honest position is that public reporting does not distinguish between these, and they are not mutually exclusive.

Nvidia cannot supply more. The US inventory certification is the most underrated clause in the framework. If domestic demand is absorbing everything the company can get out of TSMC, the China allocation is whatever is left, and that is a number set in Santa Clara rather than in Washington or Beijing.

The buyers are not asking. Beijing has been approving these imports while openly flagging the risk of overreliance on foreign silicon. Chinese firms have spent two years being told to build on domestic parts, and the gigawatt-scale deployments running on Chinese accelerators demonstrate that the instruction was not rhetorical. A 10,000-unit order is defensible internally. A 75,000-unit order invites a conversation about strategic dependence that no Chinese procurement team wants to have twice.

The economics are worse than the headline. Compliance costs, US routing, testing delays and a revenue share all land somewhere. If the delivered cost per accelerator is materially above what the same money buys domestically, the rational order size is small and strategic rather than large and infrastructural.

Everyone is being careful. This is the first real movement after seven months of stasis and one prior policy reversal that cost Nvidia billions. Testing a pipeline with 10,000 units before committing to seven times that is what a competent buyer does when the rules have changed three times in eighteen months.

The names that are missing

There is a detail in the reporting worth holding onto. Roughly ten Chinese firms were cleared in January. Two are described as having received chips.

Alibaba and JD.com were on the January list and do not appear in the August deliveries. That absence has at least two innocent explanations, sequencing and paperwork among them, and one that is not innocent at all, which is that some approved buyers have looked at the delivered cost and the political weather and quietly declined to place an order.

Nobody outside those procurement departments knows which it is yet. But if the next round of reporting still names only ByteDance and Tencent, the pattern stops being a queue and starts being a choice. Approvals are public and orders are not, which means the visible signal in this trade will always lag the real one by a quarter or so.

Two governments throttling the same pipe

Step back and the shape of this is unusual. Both sides are restricting the same trade, for opposite reasons, and converging on the same outcome.

Washington limits the flow to slow Chinese AI capability, and expresses that limit as volume caps, security conditions and revenue capture. Beijing limits the flow to reduce dependence on American hardware, and expresses it through industrial policy and procurement pressure. One government wants China to have fewer H200s because the chips are useful. The other wants Chinese firms to want fewer H200s because the chips are American.

A vendor facing that combination does not have a market. It has a negotiated trickle, and the trickle is the product now. That is a materially different world from January, when the first approvals landed and the working assumption was that volume would follow permission.

What is actually at stake

Before controls, China accounted for something like a fifth to a quarter of Nvidia's data centre revenue. Jensen Huang has put the Chinese AI accelerator opportunity at roughly $50 billion for 2025 alone had the company been free to sell into it, with expectations of rapid growth from there.

Against that, tranches of 10,000 units are close to a rounding error. The revenue question is almost secondary, though. The strategic question is whether Chinese AI infrastructure gets built on American silicon at all, because a generation of models trained on domestic accelerators produces a software ecosystem, a tooling stack and a set of engineering habits that do not revert when policy does.

This is the asymmetry that should worry Santa Clara more than any quarterly number. Market share lost to regulation can come back the moment the regulation lifts. Market share lost to a working substitute usually does not, because the switching cost is no longer political, it is technical, and it compounds. Every quarter a Chinese lab spends making its stack run well on domestic hardware is a quarter of accumulated engineering that an H200 shipment cannot undo.

That is the risk Nvidia named publicly in February, and nothing in the August shipments resolves it.

What would settle it

The next two or three tranches answer this cleanly, which is unusual for a story of this kind.

If deliveries scale toward the ceiling, the first round was caution and compliance testing, and the licence remains the operative constraint. Policy is still the thing setting the volume, and policy can change again.

If they keep landing near 10,000 while approvals stay open, then the constraint is no longer regulatory in any meaningful sense. It is some combination of Nvidia's American inventory obligations and Chinese buyers who have decided, or been guided to decide, that a large order is not worth the exposure. A ceiling nobody reaches has stopped being a ceiling.

The second outcome is the more consequential one, and it is not really about export controls at all. It would mean the two largest economies in artificial intelligence have each concluded, separately and for their own reasons, that they would rather not depend on the other, and have started acting on it without needing to be told.

Watch the tranche sizes, not the approvals.

Nvidia is not alone in this, which is part of why it reads as structural rather than corporate. AMD ran into comparable limits on its MI308 in the same April 2025 tightening, and the pattern since has been identical on both sides: a compliant part designed carefully to fit the rules, followed by rules that move. Any vendor building a China roadmap today is designing against a target that has shifted three times in eighteen months, which is its own argument for keeping orders small.