DeepSeek Partners with Huawei to Design AI Chips, Taking Aim at Nvidia
nashnova research
DeepSeek is jointly designing future AI chips with Huawei, aiming to end China's top-tier model training dependence on Nvidia entirely. This means China's AI supply chain is shifting from buying chips to building them — and the first results could arrive as early as next year.
How deep does this partnership actually go?
Huawei engineers have been stationed inside DeepSeek's offices long-term, directly shaping chip architecture decisions — including how models compute and move data across processors.
This means → this is not a normal "vendor takes client feedback" arrangement. DeepSeek is at the design table before the chip is finalized.
In plain terms = Huawei used to build chips and DeepSeek used them. Now they are drawing the blueprint together, so the chip is tailor-made for DeepSeek's models from day one.
When do the chips arrive — and who gets them first?
DeepSeek plans to begin pre-training on Huawei's latest flagship chip, the Ascend 950DT, as early as late this year. Pre-training is the stage where a model learns foundational capabilities from massive datasets. Testing is still ongoing; the timeline could shift.
When the next-generation Ascend 960DT ships, Huawei plans to allocate most of the supply to DeepSeek first, rather than spreading it across other Chinese AI developers.
Huawei's disclosed roadmap: 960DT in Q1 2027, 970 in 2028, 980 in 2029. This reflects a deliberate bet — concentrating scarce capacity on a single marquee customer.
How wide is the gap with Nvidia?
Research firm Epoch AI estimates that Nvidia's B300 chip has roughly seven times the raw compute of Huawei's Ascend 950.
Epoch AI expects Huawei to trail Nvidia by three to four years in single-chip performance, a gap projected to persist through 2030.
In plain terms = what one Huawei chip can do, one Nvidia chip can do seven times over. That gap will not close soon — but DeepSeek's bet is that "good enough and sanction-proof" beats "best but cuttable."
Why would DeepSeek accept a slower chip?
Bernstein senior analyst Qingyuan Lin frames the choice: DeepSeek is opting to "move slower, but move further" — trading slower training speed for supply-chain stability free from U.S. export controls.
Context: Beijing requires domestic AI to run on homegrown chips and has capped how many Nvidia processors major Chinese tech firms can buy. Nvidia CEO Jensen Huang told CNBC in May that Nvidia has "basically ceded the China market" to local rivals.
This means → for DeepSeek, the performance gap is a known cost. The real risk is that one day even Nvidia's older chips become unobtainable.
Where are Huawei's own bottlenecks?
U.S. export controls cut Huawei off from TSMC in 2020, forcing it to rely on domestic foundries.
Since 2024, Washington has further restricted China's access to advanced high-bandwidth memory chips (HBM) critical for AI, slowing the capacity ramp of Huawei's foundry partners.
In plain terms = Huawei's problem is not chip design — it is that the factories and key components needed to manufacture the chips are constrained. A great blueprint means little if production cannot keep up.
Has DeepSeek given up on Nvidia entirely?
No. DeepSeek is still seeking more Nvidia Blackwell chips and older processors, and has already purchased Blackwell servers through intermediaries in countries where sales are permitted, using them to train its V4-series models.
This means → DeepSeek is running a dual-track strategy: Nvidia for near-term frontier competitiveness, Huawei for a long-term channel that sanctions cannot cut.
This reflects a pragmatic judgment — Huawei's chips cannot yet support a full training run for a top-tier model. Whether this partnership can produce a "good-enough domestic alternative" is the critical test of China's AI self-sufficiency push.
市场有风险，内容仅供研究参考，不构成投资建议。