Nvidia ships the A800 to China to stay inside export rules

Background

In October 2022 the United States restricted exports of Nvidia's A100 and H100 data-center GPUs to China ^[1]. Those are the company's chips for training and running large AI models, and China had been a significant market for them. Nvidia said the earlier round of US export restrictions would cost it about US$400 million in potential China sales in the third quarter of 2022 ^[1]. That figure set the scale of the problem: either find a product that could legally be sold there, or give up the sales.

What Nvidia changed

The answer was the A800. It entered production in the third quarter of 2022 and went on sale in China in late 2022 ^[1]. Technically, the change was narrow. Nvidia's only modification was to cut NVLink bandwidth from 600 GB/s to 400 GB/s, while leaving the chip's compute untouched ^[1][2].

NVLink is the high-speed connection that lets several GPUs in one server exchange data directly, so that a group of them can work on a single large model as if they were one bigger machine. Lowering its bandwidth means the chips pass data to each other more slowly, which matters when a workload is spread across many GPUs — as large-model training usually is. But the arithmetic each chip can do on its own is the same. In effect, the restriction was met by slowing the connections between chips rather than by making the chips themselves weaker.

On the market

The card did not stay a paper announcement. By November 2022, Taobao stores in China were listing A800 cards for sale ^[1]. Global Times reported that one seller based in Shenzhen claimed to hold 1,000 A800 GPUs in stock and ready for delivery ^[1]. Listings on a consumer marketplace, and stock counts quoted by individual sellers, are a rough signal rather than a measure of total supply, but they show the product reaching buyers quickly.

Why it mattered

Analysts quoted by Global Times said that reconfiguring products to avoid trade restrictions had clear commercial logic given the size of the Chinese market ^[1]. The A800 is a concrete example of that logic in practice: rather than exit a market, a chip designer adjusts the specification that the rules actually touch, and keeps selling a part that performs the same calculations but communicates more slowly.

For buyers, the trade-off is specific. Work that fits on a single GPU, or on a few GPUs that do not need to talk to each other constantly, is barely affected. Work that depends on shuttling data between many chips — the pattern typical of training large models — is affected, and the gap widens as clusters get bigger. It also showed how export controls on computing hardware would work in the following years: rules written around measurable specifications, and vendors redesigning products to sit just inside them.