DeepSeek's AI ambitions hit wall: chip shortage stalls model training

2025-06-27 nvidia

Beijing, Friday, 27 June 2025.
DeepSeek, a rising AI powerhouse, faces significant setbacks in training its latest models. A critical shortage of NVIDIA H20 GPUs in China is delaying progress. Export restrictions are impacting access to these high-performance GPUs. DeepSeek’s earlier R1 model relied on 50,000 GPUs, including 30,000 H20s. The current scarcity is not only hindering the development of the R2 model, but also affecting deployment of the R1 model. This could reshape the competitive landscape.

NVIDIA’s market position in china

The GPU shortage stems from U.S. export restrictions designed to limit China’s access to advanced technologies [1]. These restrictions directly impact NVIDIA’s sales and market position in China [GPT]. DeepSeek’s R1 model is widely used by Chinese companies and government agencies, primarily running on H20 GPUs in the cloud [1]. The scarcity of H20 chips affects not only DeepSeek’s R2 development, but also the deployment and functionality of the existing R1 model [1][3]. This situation creates an opportunity for domestic alternatives like Huawei’s Ascend 910B chipset, though it is less powerful and lacks CUDA support [1].

Revenue potential impact for NVIDIA

The restrictions on H20 exports could reduce NVIDIA’s revenue potential in the Chinese market [GPT]. DeepSeek had been using 30,000 H20 GPUs for its R1 model [1][3]. This indicates a substantial demand that NVIDIA can no longer fully meet due to export limitations [alert! ‘precise revenue loss not specified in sources’]. The delay in DeepSeek’s R2 model, combined with deployment issues for R1, could lead Chinese customers to explore alternative solutions, further impacting NVIDIA’s sales [3]. Holger Mueller from Constellation Research Inc. notes that failure to regularly update AI models can diminish their accuracy compared to competitors [1].

Competitive advantage at risk

NVIDIA’s competitive advantage is threatened as Chinese AI firms struggle to access its top-tier GPUs [GPT]. DeepSeek’s models are optimized for NVIDIA chips, making the H20 shortage a critical bottleneck [1]. The lack of access to advanced GPUs could slow the progress of Chinese AI companies, affecting their competitiveness in the global AI landscape [1][6]. While Huawei offers an alternative, its chipset’s weaker performance and lack of CUDA support pose challenges for companies like DeepSeek, whose models are designed for NVIDIA’s platform [1]. This creates a dilemma between utilizing less effective domestic options or facing delays [1].

Bronnen

GPU Shortage Export Restrictions