Huawei Ascend

China's de-facto domestic NVIDIA alternative — Ascend 910B/910C ships in volume to Chinese hyperscalers and labs, capped by HBM scarcity and US export controls.

1. Core Product / Service

The Ascend family is Huawei's AI accelerator line, designed by HiSilicon and originally fabricated by tsmc before US export controls cut that channel.

• Ascend 910B — N+2 SMIC node (rough TSMC 7nm equivalent), HBM2e; the workhorse of Chinese AI training in 2024–2025. Roughly 60% of H100 performance in DeepSeek's internal benchmarks [5].
• Ascend 910C — dual-chiplet 910B successor; began shipping to Chinese customers from May 2025 [4]. Targeted at Llama-class training and inference.
• Ascend 910D — added to US export-control prohibition list in 2025.
• Ascend 920 — announced 2025/2026 as the next-gen part [4].
• CloudMatrix 384 — rack-scale 384-Ascend cluster Huawei pitches as a system-level alternative to NVL72.

Software stack: CANN (low-level), MindSpore (framework, less adopted than PyTorch), PyTorch via Ascend NPU plugin. Ecosystem maturity is years behind CUDA but has improved meaningfully through 2025–2026 as Chinese labs (DeepSeek, Moonshot, Alibaba) co-optimize.

2. Target Users & Pain Points

Effectively a single buyer pool: Chinese cloud providers and AI labs forced off NVIDIA by export controls.

• Alibaba Cloud, Tencent Cloud, Baidu Cloud, China Mobile/Telecom/Unicom — bulk Ascend deployments for sovereign AI capacity
• deepseek, kimi (Moonshot), Zhipu, Baichuan, Qwen team — frontier-lab training at scale on Ascend clusters
• State-sponsored "AI compute centers" across Chinese provinces

Pain solved: a domestic alternative to embargoed nvidia silicon. Pain unsolved: software portability, HBM supply ceiling, single-vendor risk inside China.

3. Competitive Landscape

NVIDIA's H20 reportedly outsold Ascend in China through 2024; the gap is narrowing as Huawei ramps and successive US controls tighten.

4. Unique Observations

• TSMC violation is the load-bearing supply story. Investigations indicated tsmc manufactured ~2 million 910B logic dies that ended up with Huawei via shell-company orders — the entire 2024–2025 Ascend ramp ran on those die banks [1]. Going forward, Huawei must depend on SMIC at N+2 (yield-poor 7nm-class), which materially caps volume.
• HBM is the binding constraint, not logic. SMIC could in principle produce ~1M 910Cs + ~500K 910Bs in 2026, but HBM availability gates how many become full ASICs [1]. With YMTC/CXMT domestic HBM still immature and Korean/Micron HBM cut off by US controls, Huawei realistically ships <1M Ascends in 2026 [1].
• Performance gap to NVIDIA is real but workable for Chinese labs. deepseek researchers found Ascend 910C performs ~60% as well as H100 [5]; sufficient for second-tier training and most inference. The gap to B200/Blackwell is wider — closer to ~25–35%.
• Captive market is strategically captive. Every 910C sold inside China is a sale Huawei doesn't have to fight nvidia for; every Chinese lab forced off CUDA contributes to a parallel software ecosystem (CANN + MindSpore + Ascend-PyTorch) that becomes more sticky over time. From the perspective of tsmc and NVIDIA, the loss isn't 2026 revenue — it's a permanent forking of the global AI stack.
• US enforcement now extraterritorial. May 2025 Commerce Department guidance prohibited anyone worldwide (US or non-US) from using, selling, transferring, or servicing Ascend 910B/C/D — a meaningful escalation that prevents non-Chinese customers from shopping Ascend even where they could obtain it [6].
• Ascend 920 at TSMC 7nm-equivalent or N+3 SMIC is the strategic question for 2027 — if Huawei can credibly close to ~80% of NVIDIA per-chip performance with adequate HBM supply, the China market becomes structurally NVIDIA-light.

5. Financials / Funding

• Parent: Huawei Technologies (private, employee-shareholder structure)
• Ascend revenue: estimated $5–10B run rate within Huawei's enterprise/cloud business in 2025; not separately disclosed
• HiSilicon (chip subsidiary) — non-public; Ascend, Kirin, and Kunpeng all in scope
• Capital: reinvested from Huawei's overall ~$100B+ revenue base; no external funding rounds
• Government support: extensive — China's "Big Fund III" (~$47B) backstops domestic semis broadly, with SMIC and HBM suppliers as the immediate beneficiaries that flow through to Huawei

6. People & Relationships

• Founder/CEO of Huawei: Ren Zhengfei
• Rotating chairs: Eric Xu, Ken Hu (rotation handles Ascend cloud/enterprise messaging)
• HiSilicon (chip arm) President: He Tingbo
• Foundry: SMIC (post-TSMC); some pre-control inventory still in service
• HBM: domestic CXMT, YMTC; smuggled SK hynix/Micron parts (subject of US enforcement actions)
• Major customers: Alibaba Cloud, Tencent, Baidu, China Mobile/Telecom/Unicom, deepseek, kimi, Zhipu, Baichuan, Qwen team
• Direct alternatives: nvidia H20 (China-compliant SKU), Cambricon, Biren, Moore Threads