Chinese buyers chase three million workstation GPUs at once

Three internet giants have booked over $35B of a card Washington still allows them to buy, and supply covers barely a tenth of demand.

ChipNews Staff
2 Min Read

Three of China’s largest internet groups have booked more than three million of Nvidia’s RTX PRO 5500 workstation cards, a combined order worth over $35B at roughly $11,400 each. Reuters reported the orders from ByteDance, Alibaba and Tencent on Oct 20.

The headline value is not booked revenue. Nvidia has told customers and partners it cannot guarantee delivery because Chinese demand is outrunning planned output. Two people said buyers may receive only 5 to 10 percent of the quantities they asked for.

A workaround card, not a frontier part

The RTX PRO 5500 is a workstation accelerator, not one of the data center processors used to train the largest models. Nvidia’s Blackwell-based B300 sits behind US export restrictions, and other advanced parts need licences. The 5500 remains sellable, which is exactly why it has become the load-bearing card for Chinese AI services.

The three companies intend to run inference on the hardware, generating answers and completing tasks for users rather than pretraining models. Inference demand scales with the number of people using a product, so a successful consumer rollout converts directly into accelerator orders.

Total Chinese demand for the part has passed six million units, the people said, while Nvidia has indicated global production may top out near three million during 2027. That gap explains the queue.

Beijing’s domestic push meets CUDA gravity

Chinese firms have steadily increased purchases from homegrown suppliers as Washington tightened rules and Beijing encouraged local substitution. Huawei has become the most visible domestic AI processor supplier, and DeepSeek has been working with it on chip design.

Even so, the order book shows the pull of Nvidia’s software stack. Porting inference workloads to a new architecture costs engineering time, and CUDA remains the default. None of the five companies has commented publicly on the report.

Share This Article