DeepSeek Bets on Huawei Chips, Eyes Training Chip Supply as Early as Q4
nashnova research
DeepSeek founder Liang Wenfeng told investors behind closed doors that training models on Huawei chips is one of the company's biggest bets — and it must work. He expects the first batch of training-grade Ascend chips as early as Q4 this year, pushing China's AI compute self-reliance from aspiration toward real-world proof.
What did Liang say — and under what conditions?
Last Sunday's meeting required investors to visit DeepSeek's Beijing or Hangzhou office in person, hand over phones and devices, and take notes with pen and paper only. Liang joined remotely.
He called training on Huawei or other Chinese-made chips one of the company's top strategic priorities — and said it "must succeed."
This means → DeepSeek has elevated domestic chips from a fallback option to a core strategic track, not merely a reactive response to sanctions.
How large are these models — and why is chip demand surging?
DeepSeek is training a 2-trillion-parameter model, larger than its current flagship V4 at 1.4 trillion parameters, with plans to scale further to 8 trillion parameters.
In plain terms = more parameters make a model smarter, but compute demand grows in lockstep — that is the direct driver behind the chip scramble.
In a May investor call, Liang estimated that training a system on par with OpenAI's largest models would require roughly 50,000 Nvidia GB300 chips or 200,000 Huawei Ascend 950 chips. This reflects a roughly 4× per-chip performance gap between domestic and Nvidia hardware — a gap that must be closed with sheer volume.
When will Huawei deliver — and what is the bottleneck?
Liang expects Huawei to deliver a new batch of training-capable chips as early as Q4 this year or Q1 next year, but did not disclose the volume.
Huawei's capacity constraint sits mainly in high-bandwidth memory (HBM) — a key component that feeds data to the chip — and other upstream part shortages, not chip design itself.
This means → even if Huawei's chip architecture is ready, upstream memory and packaging are the real chokepoints setting the delivery pace.
Where is the money coming from — and are Nvidia chips still in use?
DeepSeek is still training new models on Nvidia chips, reportedly sourced through secondary-market channels.
The company is spending roughly RMB 30 billion to expand compute capacity. It closed a RMB 50 billion round in June and is now pursuing a second round of the same size, at a RMB 500 billion valuation.
In plain terms = DeepSeek is pouring cash into Nvidia stockpiles to keep current training on track while betting on Huawei chips as the long-term replacement — running on two legs, but the Huawei leg has to start sprinting.
What does this mean for China's AI compute ambitions?
Whether Huawei's Ascend chips can ship on schedule and perform as expected in large-scale training is the critical proof point for China's AI compute self-reliance agenda.
Success would mark the first time domestic chips are battle-tested in frontier model training at the trillion-parameter scale.
This means → this is not just one company's gamble — it will answer a fundamental question for China's entire AI supply chain: can domestic chips actually handle the most demanding training workloads?
市场有风险，内容仅供研究参考，不构成投资建议。