Mark RadarMARK RADAR
中 EN
Event File AI NVIDIA AI Computing Power DeepSeek

DeepSeek Teams With Huawei to Challenge Nvidia’s CUDA Moat

2 reports · First detected 2026-07-24 · Last active 2026-07-24

Nvidia’s CUDA has long anchored its dominance in artificial-intelligence computing by tying chips to a mature suite of software, libraries and developer tools. That ecosystem remains difficult to replicate even when Chinese companies can obtain domestic accelerators. DeepSeek founder Liang Wenfeng argues that the main gap between Chinese and US AI is access to computing power rather than algorithms, making a portable software layer critical if Huawei hardware is to handle more training and inference workloads.

At a four-hour investor meeting on July 22, 2026, Liang said DeepSeek was working closely with Huawei to improve the TileLang high-level language and compiler, predicting China’s domestic-chip ecosystem could be broadly validated within a year. DeepSeek currently has about 20,000 H-series-equivalent units of compute, while Huawei can supply roughly 16,000 of its 950 chips. Liang estimated four Huawei 950s match one Nvidia GB300 and said the hardware remains about two years behind. No partnership value was disclosed.

All Coverage

2 original reports
ABMEDIA.IO 2026-07-24
解讀梁文鋒投資會議：當中國 AI 不再依賴輝達，CUDA 護城河還能保住嗎？
TECHNEWS.TW 2026-07-24
DeepSeek 創辦人梁文鋒放話華為一年就可擺脫輝達，美中差距只在算力

The Backstory

The history behind this event
DeepSeek Develops In-House AI Inference Chip to Reduce Reliance on Nvidia2026-07-08 · 3 reports · similarity 0.82

DeepSeek has primarily relied on Nvidia and Huawei chips to train and deploy its models. As the United States continues to restrict exports of advanced AI chips to China, inference costs and supply security have become increasingly important. An in-house chip optimized to run responses from existing models could reduce the company’s dependence on outside suppliers and potentially disrupt Huawei’s position in China’s AI computing market.

Reuters reported on July 7, 2026, citing three people familiar with the matter, that DeepSeek was hiring more semiconductor engineers and approaching chip design, foundry and memory companies. The project remains at an early stage, with no timetable for mass production. Reuters separately reported in June that the company planned to raise $7 billion in its first external funding round at a valuation of $52 billion to $59 billion.

Mark Radar｜MARK RADAR
Past Issues・by MarkReadFintech
All times are in Taipei time (GMT+8)