China Bought About a Million GeForce Cards for AI, While Huawei Takes Market Share From Nvidia

Chinese AI model developers have bought about a million American GeForce gaming graphics cards and Nvidia workstation products. Yet Nvidia's share of China's overall AI accelerator market keeps falling. In the past half-year, accelerators that are not GPUs took the lead. Huawei is the main beneficiary of this shift.

Consumer hardware as an AI farm

US export rules restrict Nvidia's server accelerators in China. Gaming graphics cards and workstations fall under those restrictions far less, so demand landed on them. A million units is industrial-scale purchasing, not hobbyists with racks at home.

The logic is straightforward. Flagship GeForce cards carry 24–32 GB of video memory: the RTX 4090 has 24 GB, the RTX 5090 has 32 GB. That is enough for inference, fine-tuning and experiments with mid-sized models. It is not enough for training large models from scratch. Consumer cards lack NVLink in the form found on server accelerators, scaling across hundreds of nodes runs into interconnect limits, and the GeForce driver license formally restricts data-center use. Buyers pick these cards because almost no other legal way remains to get CUDA compute at the needed volume.

Where Nvidia is losing ground

The paradox is that purchase volumes are growing while Nvidia's share of China's AI accelerator segment is falling. A gaming card bought for an AI team does not replace a full-fledged accelerator in either cost or performance. Dominance in the non-GPU category was already established over the past half-year. That category includes neural processing units and ASIC accelerators. Huawei, with its Ascend line, plays the key role, and Cambricon, Baidu Kunlun and other local developers work alongside it.

Three factors sustain this transition:

• government and corporate procurement in China increasingly requires local chips;
• Huawei is investing not only in silicon but also in its CANN software stack and the MindSpore framework;
• large clouds and operators get predictable supply without having to worry about decisions in Washington.

The main barrier for Chinese chips is not FLOPS but the ecosystem. A decade of CUDA work cannot be ported in a quarter. But every team that has once ported training and inference to Ascend will not go back to depending on export licenses.

Key figures

• ~1 million gaming GeForce and workstation cards bought by Chinese AI developers.
• Nvidia's share of the Chinese AI accelerator market continues to fall.
• Non-GPU accelerators took the dominant position over the past half-year.
• The main winner is Huawei.

What teams and investors should do

For AI teams, the practical takeaway is simple: build portability into your architecture. If inference can run through abstractions such as ONNX or custom runtimes, migrating to Ascend costs less. Training large models remains a harder case, and CUDA will stay the standard there for a long time.

Investors should look not at Nvidia's revenue in China as such, but at the structure of demand. Growing sales of consumer cards do not offset the loss of data-center contracts, and China's segment is moving toward its own silicon. The consumer graphics card market outside China should watch prices: bulk purchasing on this scale can squeeze the availability of top-end models.

For everyone else, this is a signal that the AI hardware market is splitting into two stacks, American and Chinese. They will keep diverging, and developers will have to choose which one to optimize their code for.