Chinese chipmakers secure nearly half of local market: report

Chinese GPU and AI chipmakers took about 41% of China’s AI accelerator server market in 2025, while Nvidia’s share fell to 55% from earlier dominance, IDC data reviewed by Reuters showed.

Shipments of AI accelerator cards from Nvidia, AMD, and Chinese vendors totalled about 4 million units, with Nvidia shipping 2.2 million and AMD about 160,000 for 4%.

Huawei shipped about 812,000 cards to lead domestic vendors, followed by Alibaba chip unit T-Head at 265,000, and Baidu unit Kunlunxin and Cambricon at about 116,000 each.

The shift comes as US export controls limit Nvidia’s most advanced products, and as Beijing encourages government and corporate buyers to use local chips in new AI infrastructure projects.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

### China’s growing market share masks a persistent performance gap

Huawei shipped 812,000 cards, yet these parts do not match Nvidia’s newest GPUs. Ascend 910B lines up more closely with Nvidia’s A100 from 2020 1.
Ascend 910C still trails Nvidia’s H100 on peak performance. Nvidia has already advanced to the H200, with volume shipments starting in 2024, plus its Blackwell generation 2.
Manufacturing choices widen the divide. Huawei depends on Semiconductor Manufacturing International Corp. (SMIC), China’s largest contract chipmaker, using a 7-nanometer-class (7nm-class) process. Nvidia relies on Taiwan Semiconductor Manufacturing Co. (TSMC), the world’s leading contract chipmaker, for newer nodes, for example H100 on TSMC’s 4nm node 2, 3.
Analysts put Huawei’s total AI compute capacity at about 2% of Nvidia’s through the second half of the decade 2.

### A “silicon wall” is rising in AI

Government policy drives the shift through subsidies for buying domestic chips plus offtake agreements that push state firms to adopt them 4, 5.
One result is a parallel tech stack, meaning the combined chips and software developers rely on. Chinese firms face pressure to use proprietary software such as Huawei’s CANN (Compute Architecture for Neural Networks), framed as a domestic alternative to Nvidia’s CUDA platform (a widely used software layer for programming Nvidia GPUs) 1.
The split breaks the global AI market into a “two-track” infrastructure. AI developers must juggle incompatible systems 6.
That strain already appears in the market. Reports say DeepSeek has delayed work on its next model while trying to run more training or inference on Huawei chips 1.