DeepSeek CEO Liang Wenfeng plans early adaptation to Huawei AI chips
China's most talked-about AI lab is betting big on domestic hardware, with plans to deploy 160,000 Huawei chips and train models up to 8 trillion parameters
DeepSeek, the Chinese AI startup that rattled Silicon Valley earlier this year with its efficiency claims, is making its most aggressive move yet toward cutting the Nvidia cord. CEO Liang Wenfeng has laid out a strategy to adapt Huawei’s Ascend AI chips for full-scale model training, a significant escalation from using domestic hardware only for lighter tasks like inference.
The plan involves deploying at least 160,000 Huawei Ascend 950DT chips at a new data center in Inner Mongolia, with initial chip deliveries expected between Q4 2026 and Q1 2027. Liang has described the transition as a crucial bet for DeepSeek’s future, one that could reshape how China’s AI ecosystem sources its most critical resource.
The numbers tell the story
Liang has been unusually candid about the performance gap. By his own estimates, achieving frontier-scale training would require roughly 50,000 Nvidia GB300 chips. The Huawei equivalent? About 200,000 Ascend 950 chips. That’s a 4-to-1 ratio.
DeepSeek plans to train a 2-trillion-parameter model using the Huawei hardware, with eventual aims to scale up to 8 trillion parameters.
Huawei’s next-generation Ascend 960DT training chip is reportedly on track for release in early Q1 2027, which would land right in DeepSeek’s delivery window.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Why go domestic now
US export controls have progressively tightened access to Nvidia’s most advanced chips for Chinese firms. DeepSeek previously relied on Nvidia hardware for its core training runs and had been developing software alternatives to Nvidia’s CUDA ecosystem, the programming framework that makes Nvidia GPUs dominant for AI workloads.
Liang has maintained that DeepSeek will continue using some Nvidia hardware in the near term, suggesting this isn’t a clean break but a gradual migration. The Inner Mongolia deployment will initially focus on inference rather than training.
The CEO’s public stance has been notably optimistic about the trajectory. He believes Huawei’s AI chips could catch up with Nvidia’s within a few years.
What this means for the AI hardware landscape
The 4-to-1 chip ratio matters for economics as much as performance. Running four times as many chips means roughly four times the power consumption, cooling requirements, and physical space. The Inner Mongolia location, with its cold climate and relatively cheap electricity, starts to make a lot of sense through that lens.