On a conference stage in Shanghai this September, a Huawei executive made a claim that would have sounded like science fiction three years ago: the company's homegrown AI accelerators are no longer prototypes or pilots. They are shipping, by the thousand, into real data centers. The room was full of engineers who had reworked their software to run on chips they could not buy from abroad.
The number that mattered was "1,000."
The bet: "super-node + cluster"
At Huawei Connect 2026 (华为全联接大会2026), rotating chairman David Wang (汪涛) said Ascend (昇腾) super-nodes — dense racks of tightly interconnected accelerators — had been deployed in more than 1,000 sets, and that the newer Ascend 950 super-node (昇腾950超节点) had entered large-scale commercial use. His framing was blunt: Huawei's AI strategy centers on compute power (算力), it intends to "monetize the hardware," and the formula is "super-node + cluster" (超节点+集群) to build what he called a solid Chinese compute base and "a new choice for the world."
That phrasing is the real story. Huawei is not selling individual chips to compete with Nvidia on a spec sheet. It is selling systems — cabinets of Ascend silicon stitched together with proprietary interconnects — because that is the only way domestic hardware can close the gap left by export controls.
What shipped, and what comes next
The roadmap Wang laid out runs years ahead of where most outside observers placed it:
- Ascend 960 is slated for the first quarter of 2027, well ahead of its original schedule, with double the performance of the 950.
- Ascend 970 is targeted for 2028, and 980 for 2029, both built on a new "Tao" (道) architecture.
- Huawei reaffirmed an annual cadence for the Ascend lineup — a direct echo of Nvidia's own yearly refresh discipline.
One Huawei-provided figure worth flagging: China's daily AI inference token volume has already reached around 500 trillion, a number the company expects to climb to the quadrillion scale by the end of the decade. That is Huawei's own estimate, not an independent measurement, but it signals where the company believes demand is heading — and why it is racing to fill the gap.
Why this is also a software story
Hardware was never the hard part. The hard part is the software stack that lets a model actually train and serve on non-Nvidia silicon. Huawei's answer is CANN (Compute Architecture for Neural Networks, 异构计算架构), its counterpart to CUDA. Over the past year Huawei has progressively opened CANN to outside developers and aligned it with mainstream frameworks including PyTorch and MindSpore, so that teams can port workloads without rewriting everything from scratch.
According to Li An, chief scientist at BrainNet (脑机网), the real test for domestic accelerators is not single-chip throughput but cluster stability at scale — a system that drops nodes mid-training is worthless no matter how fast each chip runs.
The honesty gap remains. Most public benchmarks for Ascend come from Huawei or its partners, and independent, reproducible comparisons against Nvidia's Blackwell-class hardware are scarce. Enterprise buyers should treat vendor performance claims as directional until third-party audits appear.
The Nvidia-shaped hole
The backdrop makes the push legible. Under U.S. export controls, Nvidia's most advanced data-center GPUs cannot be sold on the Chinese mainland. Nvidia CEO Jensen Huang told investors the company had "largely conceded" China's AI accelerator market — its China data-center revenue, once a meaningful slice of the total, has been written down toward zero in the company's own guidance.
That vacuum is what Huawei is filling. The strategic risk for customers is vendor lock-in: committing to Ascend means committing to CANN, to Huawei's interconnect topology, and to a toolchain that, while improving fast, still lags CUDA in operator coverage for multimodal and audio pipelines. The brute-force workaround — buy more silicon and scale horizontally — works for inference, but frontier model pre-training remains where the gap bites hardest.
Honest limitations
This article relies on Huawei's own announcements at Connect 2026 (reported by City News Service and Chinese financial press including 财联社/科创板日报) and on previously published third-party analyses of the Ascend roadmap. I have not independently benchmarked any Ascend hardware, and Huawei does not publish per-chip utilization or cluster reliability data for outside verification. The "500 trillion daily tokens" figure is a Huawei estimate, not an audited statistic. I have excluded specific market-share percentages attributed to single research notes because they could not be double-sourced. The piece does not cover Huawei's smartphone or networking businesses, and it does not assess the geopolitical durability of the export-control regime that created this opening.
What readers can do now
- If you run inference workloads aimed at the China market, pilot a single Ascend 950 super-node against your serving stack before committing — budget real time for CANN porting overhead.
- Track independent benchmarks (academic labs, cloud providers' published numbers) rather than vendor slides when judging whether Ascend meets your training needs.
- Watch Huawei's 2027 cadence: the Ascend 960's on-time, double-performance delivery is the cleanest signal of whether "Chinese compute" is maturing or merely scaling.
