DeepSeek CEO Liang Wenfeng has offered a rare look into the company’s AI strategy, China’s ongoing compute shortage, and the growing competition between NVIDIA and Huawei in the race to power next-generation artificial intelligence models.
During a recent investor-focused conference call, Liang discussed several key issues shaping DeepSeek’s future, including the company’s limited access to high-end AI hardware, its aggressive expansion plans, the potential of Huawei’s Atlas 950 SuperPoD, and the importance of high-quality data annotation in training more capable AI systems.
One of the biggest takeaways from the call was DeepSeek’s current compute gap compared with leading AI labs in the United States. According to Liang, the company’s available computing power only allows it to work with models that activate around 10 billion parameters at a time. By comparison, the largest frontier AI models today can activate hundreds of billions of parameters simultaneously.
Liang explained that training a model on the scale of the world’s most advanced AI systems would require an enormous amount of hardware. In his estimate, DeepSeek would need roughly 50,000 NVIDIA GB300 GPUs or about 200,000 Huawei 950 chips just for training. That figure does not include the additional compute needed for research, experimentation, and deployment.
This, he said, highlights the biggest gap between Chinese AI companies and their US counterparts: resources.
DeepSeek is not standing still, however. Liang revealed that the company currently has compute capacity equivalent to around 20,000 NVIDIA H100 GPUs. Most of that capacity reportedly arrived within the last month or two, and DeepSeek is now expanding its AI infrastructure very aggressively.
The comments arrive at a crucial moment for the global AI hardware market. NVIDIA remains the dominant force in AI accelerators, with its GB200 and GB300 platforms viewed as some of the most powerful options for training and running large-scale models. However, Huawei is rapidly positioning itself as China’s most important domestic alternative.
Liang expressed strong confidence in Huawei’s Atlas 950 SuperPoD, also known as a supernode. He said Huawei’s 950 supernode can fully substitute NVIDIA’s GB200 and GB300 racks in terms of both performance and pricing, at least for the types of workloads DeepSeek needs to run.
He acknowledged that Huawei’s solution is more expensive, but he suggested the price difference is not a major obstacle. In his view, even if Huawei hardware costs 50 percent, 100 percent, or even 200 percent more, the availability of domestic compute could still make it worthwhile for Chinese AI companies.
Liang’s assessment is especially notable because he also admitted that Huawei’s chips are not equal to NVIDIA’s on a one-to-one basis. According to him, DeepSeek would need four Huawei Ascend GPUs to match the capability of one NVIDIA GB300 GPU. He also estimated that Huawei’s hardware is still about two years behind NVIDIA’s most advanced systems.
Even with that disadvantage, Liang argued that Huawei’s supernode can handle the same tasks as NVIDIA’s GB300, including latency-sensitive workloads. That makes Huawei’s platform strategically valuable for companies that need a domestic supply chain and cannot rely entirely on US-made AI chips.
His remarks reflect a broader shift in China’s AI industry. Rather than waiting for unrestricted access to NVIDIA’s latest GPUs, Chinese firms are increasingly adapting their model architectures, training methods, and infrastructure plans around domestic alternatives. The goal is not just to build larger models, but to build AI systems that can survive under hardware constraints.
Liang also emphasized that DeepSeek’s strategy is rooted in restraint and efficiency. Instead of trying to match the largest AI labs purely through massive spending, the company is focused on making the best possible use of the resources it has. This philosophy has become a defining part of DeepSeek’s identity, especially as the global AI race becomes more expensive and compute-intensive.
Another major focus for DeepSeek is data annotation. Liang said the company is now doubling down on high-quality annotation work, which is essential for improving model performance, reasoning ability, and reliability. Unlike hardware expansion, where capital expenditure is the main barrier, annotation is constrained heavily by time, process quality, and human expertise.
Companies such as OpenAI and Anthropic began investing in large-scale annotation earlier and with more resources. Liang appears to see this as an area where DeepSeek must work harder to close the gap. Better annotated data can help improve AI model alignment, instruction-following, coding ability, and overall usefulness, even when compute resources are limited.
The broader message from Liang’s call is clear: DeepSeek sees compute as its biggest challenge, but not an unbeatable one. NVIDIA’s GB300 remains a powerful benchmark for AI infrastructure, yet Huawei’s Atlas 950 SuperPoD is emerging as a serious domestic substitute for Chinese AI labs willing to trade efficiency for availability and supply-chain security.
DeepSeek’s path forward appears to combine three priorities: rapidly expanding compute, relying more on domestic hardware where possible, and investing deeply in high-quality data annotation. If successful, that strategy could help the company remain competitive in the global AI race despite operating with far fewer resources than the biggest US labs.
