A single chip is no longer the unit of competition in Chinese AI hardware. The fight has moved to the cluster, and Huawei is running three races at once — one already won, one just starting, one still on the drawing board.
The Ascend (昇腾) family illustrates a strategy that is easy to misread if you flatten it into one headline number. Huawei does not publish unit shipments, so the honest frame is by generation and status: what is shipping, what just launched, and what is merely promised.
Key takeaways
- The Ascend 910C is in volume shipment to domestic customers, widely reported to be built on SMIC's N+2 (~7nm) process.
- The Ascend 950 supernode (Atlas 950 SuperPoD) moved from roadmap to launch in 2026, debuting publicly at WAIC in July and going on sale at Huawei Connect in September.
- The 960 and 970 remain on a publicly stated three-year roadmap and are not shipping.
Layer one: the 910C is the workhorse
International reporting confirms Huawei ramped 910C shipments to Chinese customers as a local substitute for datacenters and AI research labs. It is the chip most likely to be inside domestic training and inference boxes today.
Multiple reports place the 910C on SMIC's N+2 node, read as roughly 7 nanometres, made viable through multi-patterning rather than the most advanced tools. That keeps it a generation or two behind Nvidia's newest parts on a per-chip basis — but available, and improving through software.
Layer two: the 950 is the scale play
At the July 2026 World AI Conference (WAIC), Huawei showed the Atlas 950 SuperPoD (昇腾950超节点) real machine for the first time: a system supporting 1,024 NPUs linked at high bandwidth, rated at 1 EFLOPS in FP8 and 2 EFLOPS in FP4, with 256TB of unified memory address space. In September, at Huawei Connect 2026 in Shanghai, the company said the 950 supernode had officially gone on sale.
Crucially, Huawei also disclosed that its earlier 384-card supernode (昇腾384超节点) had been commercially deployed in over 750 sets, across internet, telecom, finance, healthcare, transport and manufacturing. That is the clearest evidence that Huawei's bet — "compensate for physics with mathematics," in the words often attributed to its leadership — is landing as delivered systems, not slides.
Layer three: 960 and 970 are promises
Huawei's rotating chairman Xu Zhijun (徐直军) outlined a three-year chip plan that includes the 950PR, 950DT, and the 960 and 970. The 960 and 970 sit in 2027 and beyond on that stated timeline and are not shipping. Reporting them as available would be the classic error of confusing roadmap with product.
Why the system, not the die, is the point
The throughline is architectural. Rather than chase a single flagship die, Huawei is packing many capable-enough chips into supernodes and leaning on its CANN software stack — fully open-sourced at the end of 2025 — to ease migration away from Nvidia's CUDA. The 384 and 950 supernodes are the embodiments of that idea.
This is also why unit counts mislead. A buyer does not purchase "an Ascend"; it purchases clusters measured in racks and nodes. Huawei's competitive signal is the number of supernodes deployed and the models trained on them, not a die tally it has never published.
The export-control backdrop
None of this happens in a vacuum. US export controls on advanced GPUs to China are the single biggest reason a domestic alternative has room to exist at all. Huawei's strategy — available-enough chips assembled into large domestic supernodes — is a direct answer to a market that cannot freely buy the newest Nvidia parts. That is why the unit-count debate misses the point: the value is in a working, deliverable system inside China's borders, not in matching a flagship die's spec sheet. The 384 and 950 supernodes are the proof that the workaround scales, and that scale, not a single benchmark, is what customers and planners actually buy. For planners in Beijing, a working domestic supernode is the strategic asset; the headline benchmark gap with Nvidia is a secondary concern next to availability.
Honest limitations
This piece draws on Huawei's official Ascend community announcements from Huawei Connect 2026 and WAIC 2026 coverage from reputable financial outlets, plus international reporting on 910C shipments and process node. Huawei does not disclose chip unit volumes, and no such figure is stated or implied here. The N+2 process attribution for the 910C comes from industry reporting, not a Huawei specification. Roadmap items (960, 970) are stated company plans, not shipped products; treat them as intent, not fact. Performance ratings for the 950 supernode are Huawei's own published figures.
What readers can do now
- When you see "Huawei shipped X million Ascend chips," treat it as an estimate from a tracker, not a company number.
- Judge Huawei's progress by supernodes deployed and models trained, not by die-level spec comparisons with Nvidia.
- Separate the three layers in any write-up: 910C (shipping), 950 (launched, scaling), 960/970 (roadmap).