China’s 2026 Domestic Compute Power Leaders: Top 30 Companies and Market Trends

In 2026 China’s domestic compute capacity moves from pilot‑stage substitution to a large‑scale commercial cycle, driven by the national “Six Basic Networks” project, the second phase of the East‑West computing initiative, and mandatory domesticisation for government and enterprise.

2026国产算力发展趋势

Three parallel technology routes divide the market:

CUDA‑compatible DCU route: led by HaiGuang Information, low migration cost, preferred for finance, telecom and existing enterprise workloads.

Full‑stack self‑developed NPU closed‑loop route: Huawei Ascend, Cambricon, Kunlun chips; strongest autonomy, suited for trillion‑parameter large‑model clusters, confidential government and industrial security.

General‑purpose GPGPU innovation route: companies such as BiRui, MuXi, SuiYuan target high‑end overseas GPUs, focusing on internet‑scale training and AIGC rendering.

Layered technology selection: No single “one‑size‑fits‑all” solution. Existing internet customers tend to choose HaiGuang DCU for low migration cost; large‑scale foundational models and confidential government projects favor Huawei Ascend; new internet‑scale domestic clusters adopt Cambricon; high‑end AIGC rendering gradually incorporates BiRui and MuXi GPGPUs.

Full‑stack hardware‑software autonomy becomes a hard barrier: The 15‑year‑plan compute‑network project and central‑enterprise procurement require domestically produced chips, servers and scheduling software, shrinking overseas compute purchases. Companies offering chip + system + data‑center operations (e.g., Sugon, Huawei ecosystem) enjoy stronger long‑term benefits.

Peripheral components outpace chips and servers: As domestic accelerator cards’ power consumption rises, demand for liquid cooling, 1.6 Tb/s optical modules, high‑level interposers and memory‑interface chips expands, giving component suppliers higher growth rates.

Compute‑as‑a‑service accelerates: Enterprises shift from pure hardware procurement to renting standardized domestic kilo‑card or ten‑kilo‑card clusters; integrated scheduling, model adaptation and managed operations become new growth points, raising the value of compute operators.