What this is
Chinese AI chip vendors completed Day-0 adaptation of DeepSeek V4 in 2026, marking the transition from "lagging" to "simultaneous deployment" relative to NVIDIA. This page is a 2026-05-15 snapshot of the Chinese AI chip landscape.
Chinese AI Chip Vendors at a Glance
| Vendor | Flagship chip (2026) | 2026 output target | Status |
|---|---|---|---|
| Huawei (HiSilicon) | Ascend 910C; Ascend 950PR (Q1) | 600K Ascend 910C + 1.6M total dies | Government-approved supplier |
| Cambricon | Siyuan 590 / 690 | 500K accelerators (300K Siyuan) | Government-approved; targets 3x output |
| Hygon Information | DCU series | n/a published | Day-0 DeepSeek V4 adaptation |
| Biren | BR100 / BR104 series | n/a published | HKEX IPO 2026 (filed) |
| Moore Threads | MTT S5000 | n/a published | Day-0 DeepSeek V4 inference (FlagOS) |
| MetaX | C-series (C500, C600) | n/a published | Growing |
| SMIC (foundry) | 7nm process (production) | n/a | Foundry for Ascend + Cambricon |
Government Approval and Procurement
| Status | Vendor |
|---|---|
| On China government-approved AI hardware supplier list (2026) | Huawei Ascend, Cambricon |
| Excluded | NVIDIA |
| Day-0 DeepSeek V4 adaptation (Apr 2026) | Huawei Ascend, Cambricon, Hygon, Moore Threads |
| HBM supply constraint | All vendors (HBM is the binding constraint) |
Six Things the Landscape Tells You
- The "Chinese chips can't train frontier models" narrative is now wrong. Huawei's Pangu Ultra MoE (718B params) trained on 6,000+ Ascend NPUs proves the stack works at frontier scale.
- Government procurement now excludes NVIDIA in favour of Huawei and Cambricon. Structural advantage for state, SOE, and regulated-sector deployments.
- Cambricon targets 3x output expansion in 2026. 500K accelerators including 300K Siyuan 590 / 690 — meaningful supply growth.
- Ascend 910C output doubles to ~600K units. Combined with Ascend 950PR (Q1 2026 launch), total Huawei die output could hit 1.6M.
- Day-0 model adaptation is the new normal. Huawei, Cambricon, Hygon, and Moore Threads all adapted DeepSeek V4 on launch day — previously only NVIDIA achieved this.
- HBM supply is the binding constraint. Chinese vendors need HBM (mostly from SK hynix, Samsung, Micron) and limited domestic HBM production caps the absolute output ceiling.
What This Means for AI Visibility
The Chinese AI chip stack matters for AI visibility because it determines where Chinese open-source models will be deployed. As government-approved suppliers absorb state and SOE deployments, Chinese LLMs running on Ascend and Cambricon (rather than NVIDIA) become the default for a large segment of Chinese enterprise AI. Brand-visibility monitoring should treat the Chinese-stack deployment surface as a distinct channel.
Methodology
Data from Tom's Hardware on Cambricon 2026 output, RCR Wireless on Huawei Ascend output doubling, CNBC on China AI chip IPO boom, TrendForce on Day-0 DeepSeek V4 adaptation, and IEEE Spectrum on Chinese AI chips replacing NVIDIA.
How Presenc AI Helps
Brand teams selling to Chinese state, SOE, and regulated-sector AI buyers use Presenc AI to monitor brand visibility on Chinese-stack deployments (Huawei Pangu on Ascend, GLM on Cambricon, DeepSeek on Moore Threads). The Chinese-stack surface diverges from NVIDIA-stack defaults and requires separate measurement.