DeepSeek’s CEO believes Huawei is only 2 years behind $NVDA

Liang Wenfeng claims Huawei’s Atlas 950 SuperPoD can run essentially the same workloads as NVIDIA’s GB200 and GB300 systems, with comparable latency
However, the underlying efficiency gap remains enormous:
Four Huawei Ascend 950 accelerators are required to replace one NVIDIA GB300.
Liang estimates that NVIDIA accelerators can remain economically useful for around five years, while Huawei’s chips may last closer to three years before their weaker energy efficiency makes them unattractive
Huawei’s strategy is to compensate for weaker individual accelerators through massive system-level scale
The Atlas 950 SuperPoD is expected to connect up to 8,192 Ascend 950DT chips through an all-optical interconnect, with availability planned for Q4 2026
This is why Liang argues that even if Huawei’s system costs 50%, 100%, or even 200% more, the difference may not matter
For Chinese AI labs, availability matters more than price
⸻
Liang says restrictions on NVIDIA hardware are forcing Chinese companies to finance, develop, and adopt an alternative ecosystem that might otherwise have struggled to gain traction
When NVIDIA GPUs are available, Chinese companies naturally prefer them. When they are unavailable, those same companies are forced to optimize Huawei hardware, rewrite software, and solve the ecosystem problems holding domestic accelerators back
DeepSeek says it trained V3 on NVIDIA hardware but reduced its dependence on CUDA by using TileLang, a higher-level programming framework
Liang believes AI-generated code and higher-level compiler tools will make it progressively easier to port workloads across different accelerators