DeepSeek's CEO, Liang Wenfeng, has disclosed a number of valuable nuggets in an investor-focused conference call, spilling the beans on China's compute deficit, the superiority of NVIDIA's GB300 GPUs, Huawei's meticulous catch-up, and DeepSeek's ongoing work on data annotation.
DeepSeek's Wenfeng Says Huawei's Atlas 950 SuperPoD can fully substitute NVIDIA's GB300 NVL rack, but for every GB300 GPU DeepSeek needs 4x Ascend GPUs
DeepSeek's Wenfeng disclosed a slew of valuable information in a recent investor conference call, noting that the AI lab's available compute only allows for models that activate around 10 billion parameters at any given time vs. the 800 billion or so that today's largest models activate at a time, going on to note:
"If I wanted to train a model as big as AI['s largest], I'd need about 50,000 GB300s or 200,000 Huawei 950s. That's just training - not counting research. So our biggest gap with the US is in resources."
Elsewhere, Wenfeng disclosed that DeepSeek currently has compute resources that are equivalent to just 20,000 NVIDIA H100 GPUs, most of which arrived in the "last month or two," but that the AI lab was now expanding its compute footprint "very aggressively."
Critically, Wenfeng asserts that Huawei's Atlas 950 SuperPoD (also known as a supernode) can "fully substitute" NVIDIA's GB200 and GB300 racks in performance and price:
"I'm fairly optimistic about domestic compute. On this point, I think NVIDIA is digging its own grave. Huawei's supernode - the Huawei 950 supernode - can fully substitute for NVIDIA's GB200 and GB300 in performance and price. The price is definitely higher, but limited. Fifty percent or a hundred percent higher - a hundred percent higher doesn't matter; even two hundred percent higher doesn't matter."
He then goes on to note that "all tasks GB300 can do, the Huawei supernode can do, latency and all the same. The only cost: 4 Huawei GPUs equal 1 NVIDIA GPU, and it's two years behind."
Finally, according to Wenfeng, DeepSeek is now doubling down on high-quality annotation, where the bottleneck is not as much CapEx as time. After all, both OpenAI and Anthropic jumped on annotation a lot earlier with a lot more resources.
About the author: Writing is my one incontrovertible passion. Over the past six years, he has authored over 2,200 distinct articles on financial and tech-related topics, spanning nearly 1 million words. And he has been a member of Wcctech mobile team since 2025. As an alumnus of the University of Toronto, Rotman Commerce Program, I bring nuance, in-depth knowledge, and a unique perspective to every topic that I cover. When I'm not writing, I'm traveling the world, exploring hidden confectionaries and restaurants as an aspiring food connoisseur.
Follow Wccftech on Google to get more of our news coverage in your feeds.
