Huawei's Ascend line is China's main domestic alternative to NVIDIA's datacenter GPUs. The Ascend 910C is reported to reach about 60% of an H100's inference performance in DeepSeek's testing, and the newer Ascend 950 family, announced in September 2025, is Huawei's bet on closing the gap with in-house memory and very large SuperPoD clusters.
TL;DR
- Ascend 910C: widely reported at roughly 60% of an NVIDIA H100 on inference, per DeepSeek's testing as relayed by press. Its headline specs come from leaks, not a Huawei datasheet.
- Ascend 950PR and 950DT: Huawei's roadmap (September 2025) puts the 950PR in Q1 2026 and the 950DT in Q4 2026, with Huawei's in-house HBM. Huawei figures are vendor claims.
- Reuters reported in March 2026 that ByteDance and Alibaba planned to order the 950PR and that Huawei aimed to ship about 750,000 of them this year. We could not confirm actual shipments.
- The real gap is software and ecosystem: CUDA versus Huawei's CANN stack. Reuters says the 950PR is more compatible with CUDA than the 910C.
- Verdict: Ascend matters if you operate in China or study that market. For everyone else, NVIDIA GPUs are what you can use. Our H20 guide covers the China-market NVIDIA part.
Spec comparison
| Spec | Ascend 910C (leaked, unverified) | Ascend 950DT (Huawei roadmap) | NVIDIA H100 SXM | NVIDIA H200 |
|---|---|---|---|---|
| Memory | not confirmed | 144 GB HBM (Huawei HiZQ 2.0, per press) | 80GB HBM3 | 141 GB HBM3e |
| Memory bandwidth | about 3.2 TB/s (leak) | about 4 TB/s | 3.35 TB/s | 4.8 TB/s |
| FP16 | about 800 TFLOPS (leak) | not published in the sources we used | not covered here | not covered here |
| FP8 | not published | 1 PFLOPS | 3,958 teraFLOPS (with sparsity) | not covered here |
| FP4 | none | 2 PFLOPS (MXFP4) | none | none |
| Interconnect | not covered | 2 TB/s | NVLink 900GB/s | not covered here |
| Status | shipping since 2024 per press | Q4 2026 per roadmap | shipping | shipping |
The 910C column is the weakest. The 800 TFLOPS FP16 and 3.2 TB/s figures come from an analyst leak reported by press, not from Huawei, and we could not confirm a memory capacity. The 950DT column is from Huawei's Huawei Connect 2025 presentation as relayed by trade press, not from a Huawei datasheet. NVIDIA numbers are from NVIDIA's pages; the H100 FP8 figure is marked "with sparsity" there, so it is not directly comparable with a dense figure.
The 910C
Huawei says the 910C is comparable to the H100. The independent-ish data point is inference: reporting on DeepSeek's testing says the 910C delivers about 60% of the H100's inference performance, and that DeepSeek trained R1 on NVIDIA H100s but used the 910C for inference. We could not locate DeepSeek's original documentation, so this is second-hand. The same reporting says handwritten kernels and optimization could improve results.
The leak that circulated describes the 910C as two co-packaged 910B dies on a 7nm-class process. Huawei has not confirmed those details.
Reuters also reported that Huawei struggled to sell the 910C in volume to private companies before the 950PR, though that sentence comes from sourced reporting about customer sentiment, not a sales figure.
The Ascend 950 family
At Huawei Connect in September 2025, Huawei's Eric Xu laid out a roadmap, as reported by Telecoms.com, Mobile World Live and others:
- 950PR: Q1 2026. Press describes it as the lower-cost variant using Huawei's HiBL 1.0 HBM, aimed at prefill and recommendation. Reuters's March 2026 report says the base card uses traditional DDR memory and a premium version uses faster HBM, at about 50,000 yuan and 70,000 yuan. These accounts do not agree, so confirm against Huawei's datasheet when it is available.
- 950DT: Q4 2026. 144 GB of HiZQ 2.0 HBM at about 4 TB/s, 1 PFLOPS in FP8, 2 PFLOPS in MXFP4, and 2 TB/s of interconnect.
- Ascend 960: Q4 2027, promising to double the 950's compute, memory bandwidth and capacity. Ascend 970: Q4 2028.
- Atlas 950 SuperPoD: up to 8,192 950DT chips, claimed 8 EFLOPS FP8 and 16 EFLOPS FP4, 160 cabinets, Q4 2026. Huawei compares it with NVIDIA's NVL144 and claims 6.7x more compute, 15x more memory and 62x more interconnect bandwidth. These are vendor claims.
- Atlas 950 SuperCluster: 64 SuperPoDs, 524,288 chips, 524 EFLOPS FP8, expected by end of 2026.
Huawei's UnifiedBus 2.0 is the peer-to-peer optical interconnect that ties these together. It plays the role NVLink plays for NVIDIA; for the concept, see what NVLink is.
The strategy is visible in the numbers: individual chips trail NVIDIA's, so Huawei scales out with enormous clusters. That works only if the interconnect, power and software hold up, and none of it has independent benchmarks that we found.
Shipping status
Reuters reported on March 27, 2026, from unnamed sources, that customer testing of the 950PR had gone well, samples went out in January, mass production would start the next month, and full shipments would begin in the second half of 2026. It named ByteDance and Alibaba as planning orders and cited about 750,000 units for the year. Huawei, ByteDance and Alibaba did not comment. We did not find a confirmation after that, so we cannot say whether the plan was met.
The NVIDIA side and export controls
Reuters's report notes that many NVIDIA AI chips are banned from sale in China, and that the H200 was approved by US policy last year with conditions, with Chinese approval but unclear timing for arrival. The earlier China-market NVIDIA part, the H20, is covered in our H20 guide. The policy picture changes often; check current rules before building a plan on any of it.
For NVIDIA specs outside China, see the H100 and H200 pages and the Hopper-to-Blackwell lineup.
Software and ecosystem
CUDA, cuDNN, NCCL and the open-source inference stack run on NVIDIA by default. Ascend uses Huawei's CANN software and its own frameworks. Reuters says customers like the 950PR in part because it is more compatible with CUDA, which makes moving models away from NVIDIA easier. Compatibility is a claim about migration effort, not a performance guarantee; porting custom kernels is still real work.
When to choose which
- Outside China, or any workload that needs standard open-source tooling today: NVIDIA. Ascend is not offered through Aquanode.
- Inside China with US export limits: Ascend is the domestic option, and its volume depends on manufacturing capacity we cannot verify.
- Researching the market: watch for an independent benchmark of the 950DT and real SuperPoD deployments in Q4 2026.
Cost
We have no sourced Ascend pricing beyond Reuters's reported figures of about 50,000 and 70,000 yuan for the 950PR variants, which are reported, not confirmed. For NVIDIA GPUs, measure tokens per second on your model, multiply by 3,600, and compare with the live hourly price below.
Rent today
Aquanode manages and optimizes GPUs for training and inference workloads. You can rent the NVIDIA GPUs below on demand.
What's next
Huawei's own roadmap points to the 950DT and Atlas 950 SuperPoD in Q4 2026 and the Ascend 960 in Q4 2027, all vendor dates. NVIDIA's counterpart is covered in the Rubin guide. For other alternative accelerators, see Cerebras vs NVIDIA and TPU vs GPU.
FAQ
How fast is the Ascend 910C compared with the H100?
About 60% on inference per DeepSeek's testing as reported by the press. Huawei itself calls it comparable. The testing was not published in a source we could open.
What is the difference between Ascend 950PR and 950DT?
Per Huawei's roadmap as reported, the PR comes first (Q1 2026) and targets prefill and recommendation, and the DT (Q4 2026) has 144 GB of HBM and targets decode and training. Accounts of the PR's memory differ.
Has the Ascend 950 shipped?
Reuters reported plans to ship from the second half of 2026. We found no confirmation of actual volume.
Can I rent Ascend chips?
Not through Aquanode. The box above shows the NVIDIA GPUs available.
Is the H20 an Ascend competitor?
The H20 was NVIDIA's China-market part; see our H20 guide.
Sources
- Huawei Connect 2025 roadmap coverage (Telecoms.com): https://www.telecoms.com/ai/huawei-details-its-plan-to-ascend-the-ai-throne
- Huawei Connect 2025 coverage (Mobile World Live): https://www.mobileworldlive.com/ranvendors/huawei-lays-out-roadmap-to-lead-global-ai-computing/
- Huawei Connect 2025 coverage (Artificial Intelligence News): https://www.artificialintelligence-news.com/news/huawei-announces-new-ascend-chips-to-power-worlds-most-powerful-clusters/
- Reuters exclusive on Ascend 950PR demand, March 27, 2026 (Yahoo Finance copy): https://finance.yahoo.com/sectors/technology/articles/exclusive-huaweis-ai-chip-favour-065327884.html
- DeepSeek test report on Ascend 910C inference (Gigazine): https://gigazine.net/gsc_news/en/20250207-huawei-ascend-910c-inference-performance-nvidia
- Ascend 910C leaked specs (Huawei Central): https://www.huaweicentral.com/huawei-ascend-910c-alleged-specs-suggest-it-a-tough-rival-to-nvidia-h100/amp/
- NVIDIA H100: https://www.nvidia.com/en-us/data-center/h100/
- NVIDIA H200: https://www.nvidia.com/en-us/data-center/h200/
