AI Chips in China: Ascend, Domestic GPUs and Supply

AI chips are specialized processors — GPUs and purpose-built accelerators — designed for the matrix-heavy compute of model training and inference, the physical substrate on which every model in this database runs.

Technical background

GPU computing became the AI default after the deep-learning boom of the 2010s; purpose-built accelerators followed. The market is defined by NVIDIA’s data-center line (H100-class and successors) on one side, and domestically designed Chinese accelerators — most prominently Huawei’s Ascend family — on the other. Export controls since 2022-2023 shaped this split and accelerated China’s domestic accelerator software stack.

How it works

AI workloads are dominated by matrix multiplications over low-precision formats (FP16/BF16/FP8/INT8). Accelerators provide thousands of cores, high-bandwidth memory (HBM), and fast interconnects (NVLink, HCCS) so clusters can act as one machine. Software (CUDA vs CANN) is as decisive as silicon: it determines which frameworks and kernels run efficiently.

Why it matters

Chips set the ceiling on model scale, training cost, and inference price — and, through export controls, on national AI strategy. For buyers, the practical question is usually simpler: what compute can I legally and affordably serve on?

Chinese adoption

China’s AI ecosystem runs on a dual track: NVIDIA parts where available, and Huawei Ascend plus other domestic accelerators where not. Our database does not currently track per-model hardware provenance (a data gap, not a claim of absence); the observable evidence is the ecosystem’s strong inference-efficiency engineering, which matters most when compute is constrained. That engineering is concrete and open: DeepSeek’s MLA attention (documented in the DeepSeek-V2 paper) compresses the KV cache, its open kernel libraries (FlashMLA, DeepGEMM, DeepEP) target serving efficiency, and MiniMax’s MSA sparse attention claims 9x prefill and 15x decode speedups at 1M context on MiniMax-M3. These are vendor-reported figures, but the pattern — attention compression and open inference kernels — is the visible trace of a compute-constrained ecosystem optimizing around hardware ceilings.

Major Chinese companies and models

No hardware-vendor associations are tracked per model in the China AI Hub database as of 2026-09-22. This section updates if vendors document deployment targets.

Practical applications

Training clusters for frontier models; inference serving for API providers; on-premises and regulated deployments that require domestic hardware; edge/vehicle inference.

Limitations

Hardware availability is geopolitically volatile; domestic accelerator software stacks lag CUDA maturity in framework breadth; and performance claims are hard to compare across vendors (different benchmarks, precisions and batch assumptions).

Deployment considerations

Evaluate the software stack before the spec sheet — porting costs dominate; benchmark your own workloads at realistic batch sizes; and keep deployment options portable (exportable formats, framework-agnostic runtimes).

What the available evidence actually shows

The China AI Hub database holds no per-model hardware-provenance field, so it cannot answer which accelerators run which Chinese models — that is a documented gap, not a claim that the information is absent in the world. What the database does show is indirect but real: Chinese labs lead on open inference-efficiency engineering (DeepSeek’s FlashMLA/DeepGEMM/DeepEP kernels and MiniMax’s MSA), which is consistent with an ecosystem optimizing against compute constraints. China AI Hub analysis indicates that efficiency engineering is the closest proxy the database currently offers for hardware pressure, and that a hardware-provenance field would turn an indirect signal into a directly usable one.

Future development

Expect domestic accelerator software maturity to keep closing the gap, FP4/FP8-native designs, and inference-optimized chips tuned for reasoning workloads’ long decode phases.

Labels used above: Official fact (from the NVIDIA and Huawei Ascend sources and the DeepSeek-V2 paper), Vendor-reported claim (MiniMax’s MSA speedup figures), and China AI Hub analysis (our synthesis, always introduced as such).