Huawei Ascend Holds Half of China’s Domestic AI Chip Market

Nine domestic AI processors cleared Chinese state procurement certification in May 2026, the clearest sign yet that Beijing’s push for silicon self-sufficiency has moved from policy into procurement reality. Export controls cut off access to advanced Nvidia silicon, forcing Chinese cloud providers and model developers to build around what domestic fabs can actually produce. The strategic emphasis has shifted from matching individual chip performance toward integrated rack-scale supernode systems capable of sustaining large-scale model training.

Huawei Ascend Leads

Huawei‘s Ascend series holds the leading position in China’s domestic AI chip market, accounting for roughly half of domestic AI chip shipments in 2025. The Ascend 910C, released in 2025, delivered approximately 60% of Nvidia’s H100 inference performance according to DeepSeek’s own testing, workable for many inference workloads, if not a like-for-like replacement. Huawei’s roadmap, disclosed by the company in September 2025, moves directly from the 910C to a new Ascend 950 series launched in Q1 2026: the 950PR, optimized for inference prefill and recommendation workloads, and the 950DT, optimized for inference decode and model training, each paired with Huawei’s own proprietary HBM (HiBL 1.0 and HiZQ2.0 respectively). The 910C, still Huawei’s current shipping product, is the chip actually filling the gap left by Nvidia’s restricted H20 today, with Huawei reportedly targeting 600,000 units of it in 2026.

Huawei’s CloudMatrix 384 supercomputer reportedly delivered higher computing power than Nvidia’s GB200 NVL72 platform on key metrics, according to reports citing SemiAnalysis. That result illustrates what Huawei can do when it controls both the chip and the system integration layer rather than optimising either in isolation.

Baidu’s Inference Focus

Baidu‘s Kunlunxin subsidiary is carving out a distinct position in inference. Its Kunlun Core 2 chips showed strong cost-effectiveness on Baidu’s own large language models, according to the company. GPU Cloud revenue, which includes Kunlun-powered inference workloads, grew 184% year-over-year in Q1 2026, per Baidu’s own Q1 2026 earnings release. At the 2026 World Artificial Intelligence Conference, Baidu showcased its next-generation Kunlun M100 chip and the Tianchi supernode series, with 32- and 64-card units already in mass production supporting domestic LLMs. Kunlunxin was absent from the May 2026 state procurement certification list, but its deep integration across Baidu’s cloud and model stack keeps it relevant for internal deployments and select external customers.

Cambricon Ramps Volume

Cambricon Technologies is targeting around 500,000 AI accelerator deliveries in 2026, including as many as 300,000 units of its Siyuan 590 and 690 processors, more than triple its 2025 output. The Siyuan 590 runs on SMIC’s N+2 (7nm-class) process, with volume production starting in Q3 2024, and carries an estimated 15-20% performance gain over the earlier Siyuan 290. Demand is real: ByteDance and other large Chinese operators are actively sourcing domestic alternatives as access to foreign hardware tightens. Cambricon’s financials for the first half of 2026 show revenue of RMB 5.996 billion, up 108% year-over-year, with net profit rising 123%, driven primarily by its cloud accelerator line.

Alibaba T-Head Scales Up

Alibaba‘s T-Head subsidiary secured state procurement certification for its Zhenwu parts in May 2026 and has been working with SMIC since December 2025 on a new 5nm chip aimed at AI inference. At the 2026 World Artificial Intelligence Conference, T-Head demonstrated the Zhenwu M890 × Panjiu AL128 supernode, pairing the in-house Zhenwu M890 chip with its ICN Switch 1.0 interconnect. The system is designed to support clusters of up to 100,000 cards. T-Head also released the SAIL open-source software stack alongside the hardware, pointing toward a vertically integrated approach: custom silicon, custom interconnect and a software layer built to match, rather than relying on third-party tooling.

SMIC’s Foundry Ceiling

SMIC is China’s designated foundry for strategic AI hardware, and the gap between what it can produce and what the industry needs is where the self-sufficiency story gets complicated. The 7nm N+2 process underpins both the Ascend 910C and Cambricon’s Siyuan 590. The deeper structural constraint is EUV lithography. ASML’s extreme ultraviolet systems, the only viable tool for economically manufacturing below 7nm at volume, remain inaccessible to SMIC under current export controls. Without EUV, the path to cost-effective leading-edge nodes stays closed regardless of how well the design teams perform. The design gap between Chinese and leading-edge global chips is narrowing; the fabrication gap is a separate and harder problem, as the effective closure of Nvidia’s China market makes plain.

China’s domestic AI chip industry has made real progress. Huawei, Baidu, Cambricon and Alibaba are all shipping competitive silicon, and the loss of Nvidia supply has given domestic vendors a procurement runway they would not otherwise have, but SMIC’s yield constraints and the EUV embargo mean cost-effective production at leading-edge nodes remains out of reach. The design work is credible; the fab economics are not yet.