China Certifies Nine Homegrown AI Processors, Expanding the Xinchuang Procurement List

Image credit: Getty / Wong Yu Liang

Announcement

On Tuesday, China’s two top technology‑security bodies – the China Information Technology Security Evaluation Centre and the National Secrecy Science and Technology Evaluation Centre – issued Anke V3.0 certifications for nine AI processors from seven vendors. The certifications, valid for three years, create a new “AI training and inference chips” category under the Anke security framework and become the de‑facto procurement list for the Xinchuang initiative, which aims to replace foreign hardware in sensitive state systems.

The approved chips are:

• Huawei – Ascend 310 (inference) and Ascend 910 (training)
• Alibaba (T‑Head) – Zhenwu M530 and M890
• Biren Technology – BAI‑1
• Hygon Information Technology – Hygon‑AI‑1
• Iluvatar – CoreX‑1
• MetaX – MX‑200
• Moore Threads – MT‑600

Notably, Cambricon and Kunlunxin – two of China’s most visible AI‑chip designers – are absent, though their omission does not necessarily imply a failed test; companies can opt out of the certification process.

Technical Specs and Process Nodes

Vendor	Chip	Target Workload	Process Node*	Peak FP16 Throughput	Power (Typical)
Huawei	Ascend 310	Edge inference	7 nm (TSMC)	12 TOPS	15 W
Huawei	Ascend 910	Data‑center training	7 nm (TSMC)	256 TFLOPS (FP16)	350 W
Alibaba	Zhenwu M530	Cloud inference	7 nm (SMIC N+2)	18 TOPS	20 W
Alibaba	Zhenwu M890	Mixed training/inference	7 nm (SMIC N+2)	45 TFLOPS (FP16)	200 W
Biren	BAI‑1	Vision‑centric inference	7 nm (SMIC)	22 TOPS	18 W
Hygon	Hygon‑AI‑1	HPC‑oriented training	7 nm (SMIC)	300 TFLOPS (FP16)	380 W
Iluvatar	CoreX‑1	General‑purpose AI	7 nm (SMIC)	30 TOPS	22 W
MetaX	MX‑200	Large‑scale language models	7 nm (SMIC)	120 TFLOPS (FP16)	310 W
Moore Threads	MT‑600	Gaming‑adjacent AI	7 nm (SMIC)	15 TOPS	25 W

*All nodes are reported as SMIC’s most stable “N+2” process, which industry analysts equate to a 7 nm class geometry. The node is roughly two generations behind the leading‑edge 5 nm/4 nm processes used by Nvidia’s latest H100 and AMD’s MI300X.

Performance Context

• The Ascend 910’s 256 TFLOPS FP16 figure still trails Nvidia’s H100 (≈ 1 PFLOPS FP16) but exceeds the older A100 (≈ 312 TFLOPS) in power‑efficiency, delivering about 0.73 TFLOPS/W versus H100’s 0.55 TFLOPS/W.
• Alibaba’s M890, at 45 TFLOPS, targets the mid‑range cloud market where cost per inference is more critical than raw throughput.
• Hygon’s AI‑1 chip, with 300 TFLOPS, is the highest‑performing domestic offering on the list, positioning it as a direct competitor for Nvidia’s data‑center GPUs in price‑sensitive Chinese workloads.

Market Implications

Accelerating the Xinchuang Transition

The Xinchuang program, launched in 2015, originally focused on replacing Intel/AMD CPUs and Oracle databases. Adding AI accelerators widens the scope to the fastest‑growing segment of enterprise compute. By mandating that state agencies, central SOEs, and strategic partners purchase from the certified list, the government guarantees a baseline demand of hundreds of millions of dollars annually.

Impact on Nvidia’s Domestic Share

• In 2025, Chinese firms shipped 1.65 million AI GPUs out of a total of 4 million units, capturing roughly 41 % of the domestic AI‑server market. Nvidia still leads with the remaining 59 % but faces a shrinking runway as more workloads migrate to certified domestic silicon.
• Huawei alone reported 812,000 AI chip shipments in 2025 and projects $12 billion in AI‑processor revenue for 2026. If the Xinchuang procurement volume reaches the projected $3‑4 billion per year, Nvidia could lose $1‑2 billion in annual sales within China.

Supply‑Chain Constraints

All seven certified vendors rely on SMIC for wafer fabrication. SMIC’s utilization topped 93 % in 2025, and its capex of $8.1 billion in 2024 was earmarked to maintain that level through 2026. However, the N+2 node’s throughput is limited to roughly 1.2 million die per month across all customers. Assuming an average die size of 200 mm², the fab can produce at most ~2.4 billion mm² of silicon per month, which translates to ~6 million AI‑chip die if every wafer were dedicated to AI products. In practice, SMIC must split capacity among memory, logic, and legacy nodes, leaving a tight allocation for the newly certified chips.

Forecasts and Risks

• Morgan Stanley projects China’s AI‑chip market to reach $67 billion by 2030, with domestic supply covering ~76 % of demand. The current certification list accounts for roughly 30 % of that projected volume, suggesting rapid onboarding of additional vendors or expanded product lines will be needed.
• Geopolitical pressure could tighten export controls on advanced lithography equipment, limiting SMIC’s ability to move beyond the N+2 node. A failure to advance to a true 5 nm class process would cap performance gains and keep domestic chips a step behind the latest Nvidia/AMD offerings.
• The absence of Cambricon, which announced a target of 500,000 AI‑chip shipments in 2026, raises questions about its strategic direction. If Cambricon opts out of Anke certification, it may pursue export markets or focus on niche workloads where its architecture offers a unique advantage.

Conclusion

The addition of nine AI processors to China’s Anke security certification marks the first large‑scale state endorsement of homegrown AI silicon. While the chips still lag behind the most advanced foreign GPUs in raw performance, their power efficiency, integration with domestic software stacks, and guaranteed procurement pipeline give them a competitive edge in the Chinese market. The real test will be whether SMIC can expand capacity fast enough to meet the growing demand and whether the certified vendors can continue to improve process nodes without access to leading‑edge EUV equipment. For now, the move signals a clear policy shift: China is building a self‑sufficient AI‑compute ecosystem, and the next few years will determine how quickly that ecosystem can rival the performance and scale of Nvidia’s global dominance.

---

Sources: South China Morning Post, Morgan Stanley research, SMIC 2025 annual report, company press releases.