DeepSeek's Huawei move is real progress, but it isn't the clean break from Nvidia that China's AI industry still wants.
DeepSeek now has something it didn't have last year: a large Chinese model run through full-parameter post-training on Huawei chips without falling over. A Shenzhen Hetao Academy-led team, working with Harbin Institute of Technology's Shenzhen campus, the Shenzhen Research Institute of Big Data and Huawei teams, said in June that it completed post-training work on DeepSeek's 1.6-trillion-parameter V4-Pro model using a thousand-card Ascend 910C cluster, according to Shenzhen News and local Chinese media reports. The run lasted more than 1,500 steps with no skipped or NaN iterations, and model FLOP utilization topped 30%.
That is progress.
It also needs a clear label. This wasn't full pre-training from scratch, and it wasn't an independent benchmark of Huawei against Nvidia. It was continued training and supervised fine-tuning, the kind of work that comes after the heaviest phase of model building. If you care about whether China can replace Nvidia in frontier AI, that distinction matters more than the headline.
DeepSeek already knows the difference the hard way. The Financial Times reported in August 2025 that Chinese authorities pushed DeepSeek to train its R2 model on Huawei Ascend hardware, but the effort ran into unstable chips, slower interconnects and limits in Huawei's CANN software stack. Huawei engineers were sent in to help, according to that reporting, yet DeepSeek still switched back to Nvidia for training while keeping Ascend work focused on inference.
DeepSeek Chose Huawei's Chips Over Building Its Own In-House Silicon DeepSeek passed on building its own AI chip and doubled down on Huawei's Ascend silicon instead, timing V4's official mid-July launch with a $7 billion funding round that broke founder Liang Wenfeng's no-financing rule. The move shows China's AI self-sufficiency push is running on partnerships, not proprietary chips. - deepseek huawei chip support strategy - why deepseek rejected nvidia custom silicon
That was the low point.
The Huawei bet is getting larger #
By early 2026, DeepSeek was giving domestic chipmakers a better shot. Reuters reported in February that DeepSeek withheld its upcoming V4 model from Nvidia and AMD for performance optimization and gave early access to Chinese suppliers including Huawei instead. In April, Reuters reported that DeepSeek launched a V4 preview adapted for Huawei chip technology, and Huawei said its chips had been used in some of the model's training process.
The spending is now following the same direction, but not always for the same job. Bloomberg reported on September 4 that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at a large data center in Ulanqab, Inner Mongolia. That's roughly 1 gigawatt in scale, according to Bloomberg. The chips are meant to run finished models for users. Here's the thing: Huawei markets the 950DT as a training chip, but DeepSeek is using it for inference.
That tells you where the gap still sits. Running a trained model at scale is expensive and important. Training the next frontier model is the harder test, and DeepSeek hasn't cut Nvidia out of that stage yet.
The newest signal is more serious. The Information reported on September 21 that DeepSeek CEO Liang Wenfeng told investors that using more domestic chips for training is now a major priority, and that he expects Huawei to begin delivering training chips to DeepSeek as early as the fourth quarter of 2026, with shipments continuing into the first quarter of 2027. The same report said DeepSeek is finalizing a second funding round that aims to raise 50 billion yuan, about $7.5 billion, at a 500 billion yuan valuation.
That is not a small order book. It is a bet.
Nvidia can read the direction #
Nvidia isn't guessing about the China problem. In its fiscal first quarter of 2026, the company recorded a $4.5 billion charge tied to H20 excess inventory and purchase obligations after the U.S. government required licenses for H20 exports to China. In its latest quarterly filing, Nvidia said Washington had granted licenses for small amounts of H200 shipments to specific China-based customers, but Chinese government restrictions meant it had shipped only a fraction of the allowed volume. Those shipments were less than 1% of data center revenue in the quarter.
Huawei Speeds Up Its Ascend 960 AI Chip Launch to Take On Nvidia in China Huawei moved its next-generation Ascend 960DT AI chip launch up to early 2027 and unveiled Peerium, a new interconnect architecture meant to cluster up to a million processors into single systems. The push comes as Nvidia CEO Jensen Huang admits his company has "largely conceded" China's AI chip market to Huawei, even as Huawei itself says it... - Huawei Ascend 960 AI chip launch date 2027 - Peerium interconnect architecture challenges Nvidia in China market
Look at Nvidia's own wording. The company said it was effectively foreclosed from competing in China's data center compute market under the current rules and politics unless it can ship a system approved by both Washington and Beijing. That is the real business risk behind DeepSeek's Huawei experiments. Every successful Ascend run gives Chinese labs one less reason to wait for Nvidia's China-compliant chips.
Huawei is also moving the hardware clock forward. At Huawei Connect in Shanghai on September 17, the company said its Ascend 960DT training chip would be ready in the first quarter of 2027, three quarters earlier than the previous schedule, according to TechCrunch and the South China Morning Post. David Wang Tao, Huawei's rotating and acting chairman, also said the Ascend 960PR inference chip would follow in the third quarter of 2027.
DeepSeek still needs Nvidia where it matters most. Full frontier pre-training is the expensive, fragile, unforgiving stage, and Huawei hasn't proved it can carry that load at DeepSeek's scale. But the old argument that Chinese chips can only serve models, not train them, is weaker now than it was a year ago. You can see the shape of the next fight: Nvidia has the better stack, Huawei has the political runway, and DeepSeek has every reason to make the domestic option work.
Also read: IonQ Stock Jumps 9% After Nvidia and Oak Ridge Crack a Quantum Bottleneck, Nest Commerce Wins TikTok Grant To Build AI Creative Platform Hummingbird, Free Chinese AI Models Now Match US Rivals At Finding Software Flaws
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.