Chinese Tech Firms Still Buying Nvidia Chips in Bulk; Domestic Substitution Gap May Persist Until 2030
nashnova research
Chinese tech companies have purchased over 1 million Nvidia gaming and workstation chips this year and repurposed them for AI, while domestic chip supply of roughly 3.5 million units still falls far short of total demand exceeding 5 million — the industry expects the compute shortage to persist until at least 2030.
Four years of export controls — why is Nvidia still flooding in?
Chinese tech firms have bought over 1 million Nvidia workstation and gaming chips this year — models such as the RTX Pro 6000D, RTX Pro 5000, and RTX 5090, according to the Financial Times. None were designed for AI, but all comply with current U.S. export rules.
This means → export controls shut the front door on high-end AI chips, but left a side door open through the consumer and workstation product lines. Companies are running parts of their AI workloads on gaming-grade silicon.
Nvidia responded that U.S. firms "continue to be constrained by outdated export controls" covering gaming products released nearly five years ago, as well as by China's own import restrictions.
How many chips can domestic makers produce — and is it enough?
China's domestic AI chip output this year is estimated at roughly 3.5 million units (including about 1 million fabbed overseas), but total demand exceeds 5 million. In plain terms = domestic supply covers only about 70% of demand; the remaining gap is the market Nvidia still fills.
Breakdown by maker: Huawei is on track for 1.6 million chips; Alibaba's and Baidu's chip units will produce a combined ~1 million; Cambricon is expected to deliver 300,000.
Morgan Stanley analyst Gary Yu noted: "We believe the constraint on China's AI compute expansion has been more about the availability of high-end compute than demand itself."
Can Nvidia's gaming chips actually run AI?
These workstation chips trail Huawei's flagship Ascend 910C on most benchmarks, but they handle simpler inference-stage tasks such as "prefilling" — the step where a model reads and processes user input before generating a response.
This reflects a pragmatic strategy: use whatever compliant, lower-end chips are available to plug the compute gap now, and wait for domestic high-end chips to mature.
This means → even locked out of the high-end market, Nvidia can still extract revenue from China by letting its product line "move down-market."
Why is the software ecosystem harder to replace than the chips themselves?
Nvidia's CUDA — a software platform that lets developers tap GPU compute power — has been built up over more than a decade. Most Chinese tech companies have their codebases and toolchains built on it.
Huawei's alternative platform, CANN, is widely regarded in the industry as harder to use. In plain terms = swapping the chip is like swapping an engine; swapping the software platform is like rewriting every manual on the production line — the migration cost of the latter far exceeds the former.
DeepSeek is working closely with Huawei on joint hardware-software optimization, and Huawei is using AI tools to accelerate CUDA-to-CANN code migration. A state-owned cloud executive said: "No mainstream large Chinese model has completed pre-training on Huawei's stack yet, but given the progress made, that breakthrough is inevitable."
Demand is accelerating — will the gap keep widening?
Alibaba, ByteDance, and Tencent are all aggressively pushing advanced models and AI-agent tools this year, driving explosive growth in demand for tokens — the basic data units models process.
The three companies' combined capex this year may more than double year-on-year to roughly RMB 700 billion (about $104 billion). Morgan Stanley estimates China's total compute spending will reach $193 billion this year, roughly 21% of the U.S. level.
Alibaba CEO Eddie Wu said on the August earnings call: "The industry generally believes that the current AI compute shortage will not ease until at least 2030." This means → the demand curve is climbing steeply, and whether domestic chip capacity can catch up before 2030 will be the single most critical test of the substitution effort.
市场有风险，内容仅供研究参考，不构成投资建议。