Chinese AI companies purchased roughly 1 million Nvidia GPUs within a single year, according to a report relayed by Tom's Hardware — a volume that underscores how export-control-compliant silicon, not domestic accelerators, is currently carrying the country's compute buildout.

The figure comes from a report rather than a company filing, and it deserves that caveat. Nvidia discloses China revenue by segment, not unit shipments, so any unit count rests on outside estimation. Even so, the order of magnitude is the story: seven figures of GPUs arriving in China despite three successive rounds of US export restrictions.

Which chips are actually shipping?

The report describes the purchases as export control-compliant parts — the trimmed Nvidia products redesigned to sit under US performance thresholds. These are the successors in spirit to the A800 and H800, the chips Nvidia built for China after Washington's first restrictions, and then rebuilt again after the October 2023 rules closed the loopholes.

The compliant parts are slower than their US-market siblings. That penalty has not stopped Chinese buyers, because the alternative is worse. The report frames the purchases as a bridge across a gap in compute demand — Chinese AI labs need training capacity now, and domestic accelerators have not yet arrived at sufficient scale to provide it.

Why a million units matters

Volume compensates for per-chip performance. A cluster of compliant GPUs, strung together by the thousand, can still train large models; it simply costs more in racks, power and interconnect. Chinese cloud operators and AI startups appear willing to pay that overhead.

The number also clarifies the competitive picture. Huawei's Ascend line and other domestic accelerators are advancing, but the report's framing — compliant Nvidia chips flooding the market to bridge the compute gap — suggests buyers treat domestic hardware as a future option rather than a present solution.

For Nvidia, China remains a market it can serve only in degraded form. Each revision of US rules has forced a redesign; each redesign has still found buyers. The one-million-unit year indicates demand elasticity that Washington's thresholds have bent but not broken.

What comes next

The report gives no capacity figures for domestic Chinese accelerator production, and it does not claim the gap has closed. It claims the opposite: the compliant chips are a bridge, which implies an endpoint somewhere ahead. If Huawei and its peers reach scale, that bridge shortens. If they do not, Chinese demand keeps flowing to whatever Nvidia is permitted to sell — and the next US rule revision will decide what that is.

By Nathan Brooks