Nine domestically designed AI chips from seven Chinese vendors cleared government security certification on 27 May 2026 — the first time AI accelerators have had their own category under China's Anke procurement framework, a move that formally pulls AI computing into Beijing's decade-long foreign-technology substitution campaign.
What the Anke Certification Means in Practice
The Anke framework is administered jointly by the China Information Technology Security Evaluation Centre (CNITSEC) and the National Secrecy Science and Technology Evaluation Centre. A chip that clears Level I certification is eligible for purchase by party and government agencies, central state-owned enterprises, and the broader universe of state-linked institutions covered by the Xinchuang initiative. Certification lasts three years. In effect, it is a procurement catalogue with compliance teeth: buyers in scope do not need a formal ban on Nvidia to stop buying it — they simply buy from the approved list.
Xinchuang previously concentrated on displacing Intel and AMD processors and Oracle databases from sensitive government systems. Adding a standalone "AI training and inference chips" category signals that AI infrastructure has crossed from general IT into the domain of national security-grade substitution.
The Nine Certified Chips
The full list spans the main tiers of China's domestic AI chip ecosystem. Huawei HiSilicon contributed the Ascend 310 and Ascend 910. Alibaba's T-Head semiconductor unit submitted the Zhenwu M530 and M890. The remaining five slots went one apiece to Biren Technology, Hygon Information Technology, Iluvatar CoreX, MetaX, and Moore Threads — though the specific model designations for those five have not been confirmed in publicly accessible sources. Tom's Hardware confirmed the vendor list and the four named models.
Two notable names are absent. Tom's Hardware reported that Cambricon Technologies and Baidu-backed Kunlunxin did not appear on the list, though vendors submit products for evaluation voluntarily — exclusion does not confirm a failed test. Cambricon had appeared on an earlier December 2024 Xinchuang procurement list for AI processors, a separate approval mechanism; it was not included in this May 2026 Anke certification round.
The Xinchuang Ratchet
The Xinchuang programme has operated as a technology-substitution ratchet since the mid-2010s. Each cycle adds a new hardware or software category to the approved-domestic list, narrows the window for foreign alternatives, then moves on to the next layer of the stack. CPUs and operating systems came first, then storage and databases. The 27 May certification formally pulls AI accelerators into that sequence.
Chinese semiconductor firms shipped approximately 1.65 million AI GPUs in 2025 — around 41% of total AI server shipments in China, according to IDC figures reported by Reuters. Huawei accounted for roughly 812,000 of those units. The company is projecting US$12 billion in AI processor revenue for 2026 — a vendor-stated figure based on orders already booked from Alibaba, ByteDance, and Tencent, as reported by Tom's Hardware. During that same period Nvidia's China market share had already fallen below 60%; it has since dropped to zero, according to a separate Tom's Hardware report. Anke certification does not create that demand, but it concentrates state procurement toward domestic alternatives.
What Changes for Buyers and Vendors
For government IT teams, the certified list becomes the default starting point when specifying AI infrastructure. Deviating from it requires justification. That friction, accumulated across thousands of procurement cycles, is how policy converts into market share without a formal import prohibition.
For domestic vendors, the dynamic cuts both ways. Inclusion on the list is a credibility signal in commercial tenders as well — enterprises that work closely with government or state banks often shadow public-sector procurement norms. The three-year certification clock also creates a renewal cycle, giving CNITSEC a periodic mechanism to rotate out older designs and require vendors to submit updated silicon.
For foreign suppliers, the arithmetic is straightforward. The public-sector AI hardware market in China — already partly closed to Nvidia by US export controls — has now acquired an institutional ceiling as well as a regulatory floor.
The Engineering Gap That Remains
Certification answers the question of regulatory eligibility, not performance parity. The gap in software ecosystem depth — particularly CUDA compatibility and training framework integration — remains a practical constraint for developers moving workloads onto domestic hardware. The certification removes a procurement barrier; it does not close a toolchain gap that Chinese AI firms have been working around for the past two years.
View author profile → · Editorial policy
