Companies / Huawei Ascend
AI chips
Huawei Ascend
State of play as of
Huawei’s Ascend line is the chip China reached for when Nvidia was cut off: a domestic AI accelerator, its own memory, its own software stack, now good enough for inference at scale.
Ascend is the AI-accelerator line from HiSilicon, Huawei’s chip-design arm. It exists because US export controls took Nvidia’s top parts off the table, and it matters because it has stopped being a fallback. In 2026 a frontier model, DeepSeek V4, shipped with Ascend as a primary deployment path, hyperscalers placed bulk orders, and the regulator judged the chips competitive enough to bar domestic firms from buying Nvidia. We track what Huawei ships, who runs it, and where the numbers are self-reported.
Fast facts
- What it is
- The AI-accelerator line from HiSilicon, Huawei’s in-house chip-design division. The answer to a compute cut-off, not a from-scratch bet.
- Flagship chip
- Ascend 950PR, unveiled 20 Mar 2026, mass production from April. Huawei says 2.87x an Nvidia H20 on single-card compute, and it is the only chip in China supporting FP4.TrendForce, via CAD
- In-house memory
- The 950PR uses Huawei’s own HiBL 1.0 HBM: 112GB at 1.4 TB/s. The inference-focused 950DT uses HiZQ at 144GB and 4 TB/s. No Samsung, SK Hynix or Micron content.TrendForce; SemiAnalysis via InfoQ
- Software stack
- CANN, Huawei’s answer to Nvidia’s CUDA, open-sourced August 2025. Operator coverage rose from below 30% in early 2024 to 80 to 90% by mid-2026.InfoQ, Leiphone, via CAD
- Shipments
- IDC figures show Huawei shipped 812,000 Ascend accelerator cards in China in 2025, leading the domestic market.IDC, via CAD
- Market position
- Huawei holds about half of the roughly $50 billion China AI-chip market; the field behind it (MetaX, Alibaba, Baidu, Cambricon) is filling in fast.CAD reporting
- Roadmap
- A published cadence to 2028: the 950 series in 2026, the 960 in 2027, the 970 in 2028, one generation a year.Huawei roadmap via InfoQ, CAD
The state of play
DeepSeek still trains on other firms’ hardware, Nvidia H800s at the start and Huawei Ascend since, and will for a long time. Huawei holds about half of the roughly $50 billion China AI-chip market, while Nvidia’s share drifts toward zero as domestic firms fill in behind it. The direction is not ambiguous, even as the top of the line stays early.
A SemiAnalysis teardown, walked through in full by InfoQ, found part of DeepSeek V4’s architecture was co-designed for the Ascend 950DT inference accelerator, which carries Huawei’s in-house HiZQ memory at 144GB and 4 TB/s aimed at the decode stage. When V4 launched, only two software stacks ran it optimally on day zero: Nvidia’s CUDA and Huawei’s CANN. Reuters reported ByteDance booked half of the Ascend 950 production line, with Alibaba and Tencent each taking hundreds of thousands of units and China Mobile specifying 776 Ascend nodes.
A joint team from Huawei’s GTS division and two Shenzhen universities completed what it calls the first third-party, full-parameter post-training of DeepSeek V4-Pro on a domestic Ascend SuperNode cluster: more than 1,500 training steps with zero NaN failures, at 34.9% model-flops utilization. In the same period Ascend’s CANN operator coverage had risen from below 30% in early 2024 to 80 to 90%. It was a public proof that a frontier-scale model could be post-trained end to end on Chinese silicon.
IDC figures show Huawei shipped 812,000 Ascend accelerator cards in China in 2025, leading the domestic market by a wide margin. The Ascend 950PR runs DeepSeek V4 at roughly one-quarter the inference cost of Nvidia’s China-specific version, and Huawei’s published roadmap runs to 2028: the 950, 960 and 970 series, one generation a year.
China’s internet regulator formally barred major domestic tech firms from buying Nvidia AI chips, after a technical assessment concluded China’s own processors match or exceed the still-permitted Nvidia parts. Regulators summoned domestic chipmakers including Huawei and Cambricon. The move made Ascend not just an alternative to Nvidia but, for the largest buyers, the mandated one.
Reuters reported Huawei plans to ship around 750,000 units of the Ascend 950PR in 2026, with mass production from April and full-scale shipments in the second half, as Alibaba, ByteDance and Tencent competed for allocation. DeepSeek V4 had launched days earlier with day-zero Ascend support, the first frontier model built to show Chinese chips could replace Nvidia in production at scale.
Huawei unveiled the Ascend 950PR at its China Partner Conference: 2.87x an Nvidia H20 on single-card compute, the only chip in China supporting FP4, and Huawei’s self-developed HiBL 1.0 memory at 112GB and 1.4 TB/s, with no Samsung, SK Hynix or Micron content. Huawei’s stated aim was not to catch Nvidia’s frontier but to be good enough for inference at scale.
Our reporting on Huawei Ascend
- The Tell
DeepSeek is secretly building its own chip, and it targets inference, not training. You do not build an inference chip to out-Nvidia Nvidia; you build one to own the compute bill that recurs.
- The Co-Design
A teardown found part of DeepSeek V4’s architecture was co-designed for Huawei’s Ascend 950DT inference chip. The model and the silicon are now being built for each other.
- Closed in Code
A team completed what it says is the first third-party full-parameter post-training of a 1.6-trillion-parameter model on about 1,000 Huawei Ascend cards, with no instability. The penalty is efficiency, and it is shrinking.
- The Boarding
The full Nvidia-to-China chip sequence, from the A800 and H800 to the H20 halt and a 4.5 billion dollar write-off, one of the largest write-downs in semiconductor history.
- The Exit
China’s Cyberspace Administration barred major domestic firms from buying Nvidia AI chips, citing an assessment that domestic parts now match them. The control now cuts both ways.
- The Floor
Reuters: Huawei plans to ship around 750,000 Ascend 950PR units in 2026, mass production from April, full shipments in the second half. Alibaba, ByteDance and Tencent are competing for allocation.
These link to the full issues on the newsletter. New pieces go out in the daily first.
Common questions
What is Huawei Ascend?
Ascend is the AI-accelerator line from HiSilicon, Huawei’s in-house chip-design division. It is China’s leading domestic alternative to Nvidia’s AI GPUs, built out after US export controls cut off Nvidia’s top parts. The current flagship is the Ascend 950PR, unveiled in March 2026, with an inference-focused 950DT and a roadmap running through the 960 and 970 to 2028.
How does the Ascend 950PR compare to Nvidia?
Huawei says the Ascend 950PR delivers 2.87x the single-card compute of Nvidia’s H20, the China-legal part, per specs reported by TrendForce, and it is the only chip in China supporting FP4. It runs DeepSeek V4 at roughly one-quarter the inference cost of Nvidia’s China-specific version, per IDC-era reporting carried by CAD. It is not built to match Nvidia’s training frontier; the stated goal is inference at scale.
Does Huawei Ascend use its own memory?
Yes. The Ascend 950PR uses Huawei’s self-developed HiBL 1.0 high-bandwidth memory at 112GB and 1.4 TB/s, and the inference-focused 950DT uses HiZQ at 144GB and 4 TB/s. Both carry no Samsung, SK Hynix or Micron content, per TrendForce and a SemiAnalysis teardown reported via InfoQ.
What software does Ascend run?
Ascend runs on CANN, Huawei’s answer to Nvidia’s CUDA, which Huawei open-sourced in August 2025. When DeepSeek V4 launched, CUDA and CANN were the only two stacks that ran it optimally on day zero. CANN’s operator coverage rose from below 30% in early 2024 to 80 to 90% by mid-2026, per InfoQ and Leiphone reporting carried by CAD.
Who uses Huawei Ascend chips?
DeepSeek deploys V4 inference on Ascend and uses it as a parallel training stack; Reuters reported ByteDance booked half of the Ascend 950 line, with Alibaba and Tencent each taking hundreds of thousands of units and China Mobile specifying 776 Ascend nodes. Zhipu trained its GLM-5.1 model on 100,000 Ascend 910B chips. In May 2026 China’s regulator barred major domestic firms from buying Nvidia, pushing the largest buyers toward Ascend.
Every dated fact above traces to a named source, and where it comes from our reporting, to the dated issue that carried it. Where a claim is self-reported or disputed, we say so.