Huawei
2026-07-28
DeepSeek Claims China Will Phase Out NVIDIA in One Year with Huawei Ascend 950 and Self-Developed Tile Language
Summary
Key Takeaways
DeepSeek founder Liang Wenfeng stated at an investor meeting that chip supply will no longer be a bottleneck for China's AI development within one year. DeepSeek has secured approximately 16,000 Huawei Ascend 950 chips for testing, claiming the Ascend 950 SuperPod matches NVIDIA GB200/GB300 systems in performance and price. Liang attributes the US-China AI gap to resources, not talent: US frontier models have ~800 billion activated parameters vs China's tens of billions, a 10x gap. DeepSeek estimates it trails US leaders by 12-18 months, aiming to close to 3-6 months, while producing comparable models with only 1/20 of US compute.
On technology roadmap, DeepSeek is developing its own inference chips to compete with Huawei and NVIDIA, and the Tile Language programming language to reduce dependency on CUDA. Liang expects the perception of domestic chips as difficult to use to change within a year. Market data shows NVIDIA's China AI accelerator share dropped from 59% in 2025 to near 0%, while local vendors hold 41% with 1.65 million cards shipped. Huawei's AI chip revenue is projected at $12 billion in 2026, up 60% YoY, driven by government mandates and over $1 billion in subsidies.
Ecosystem players like Alibaba, Huawei, and Moore Threads are advancing domestic alternatives: Alibaba's T-Head Zhenwu chip is deployed in thousands and open-sourced; Huawei showcased Atlas 950 SuperPoD at WAIC 2026 (connecting 8,192 Ascend chips, 6.7x compute of NVIDIA NVL144, with no US parts). Liang's one-year timeline contrasts with NVIDIA's investments in SSI, Vera Rubin production. If the DeepSeek+Huawei combo achieves performance parity and ecosystem viability within 12 months, it could fundamentally reshape the global AI compute supply chain. DeepSeek's dual bet on self-developed inference chips and Tile Language marks a critical inflection point from import substitution to an independent technology system.
Why It Matters
Despite claims of performance parity, Huawei's Ascend 950 is based on 7nm process (SMIC N+2) vs NVIDIA's 4nm, leading to lower energy efficiency and higher thermal output. In large-scale training, Huawei's HCCS interconnect and RoCEv2 network may suffer from tail latency and PFC/ECN congestion control bottlenecks compared to NVIDIA's NVLink and InfiniBand, impacting training throughput. DeepSeek's Tile Language lacks a mature developer ecosystem and toolchain, making it unlikely to displace CUDA in the short term. Government mandates and subsidies lock enterprises into the Huawei ecosystem, creating high switching costs. DeepSeek's self-developed inference chips could fragment the ecosystem further. The one-year timeline is overly optimistic given the long gestation required for software ecosystem maturity.
PRO Decision
【Vendors】Competitors like NVIDIA, AMD, Intel should exploit the lack of independent benchmarks for Huawei Ascend 950 by accelerating their own product roadmaps (e.g., NVIDIA B200/B300) and reinforcing CUDA ecosystem lock-in with migration tools. NVIDIA should lobby for export rule adjustments or offer compliant variants for China. AMD and Intel should promote ROCm and OneAPI as open alternatives to attract enterprises wary of Huawei lock-in.
【Enterprises】CIOs and architects must conduct independent benchmarking of Ascend 950 for large model training, focusing on actual throughput and energy efficiency. Assess the maturity of CANN software stack and migration costs from CUDA. Adopt multi-vendor strategies to avoid lock-in. Be cautious of government mandates; evaluate long-term operational and talent costs.
【Investors】Look beyond the PR: Huawei's Ascend revenue growth is driven by subsidies and mandates, not free market competition, raising sustainability concerns. DeepSeek's self-developed chips and language add uncertainty. NVIDIA's China share loss is geopolitical, not technological; a relaxation of controls could reverse the trend. Focus on truly competitive domestic chip makers with global potential.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)