China's leading artificial intelligence firm DeepSeek Ltd. is reportedly facing challenges in developing its next-generation R2 inference model due to insufficient access to NVIDIA graphics processing units (GPUs).
According to a report in The Information citing two anonymous sources familiar with the company's situation, DeepSeek has been working on the upcoming R2 model for several months, but CEO Liang Wenfeng remains dissatisfied with its current capabilities. The company's limited GPU inventory is hindering further improvements.
DeepSeek gained prominence earlier this year with its initial R1 inference model, which demonstrated capabilities comparable to cutting-edge models developed by US firms like OpenAI, Anthropic PBC, and Meta Platforms Inc., albeit at a significantly lower cost.
The R1 model was trained on a 50,000-GPU cluster containing approximately 10,000 H100, 10,000 H800, and 30,000 H20 GPUs optimized for the Chinese market's power efficiency requirements.
Chinese companies have never been able to legally purchase H100 or H800 GPUs, with some reportedly obtained through investor Gaofei Capital Management and others via shell companies using public cloud infrastructure services. While H20 GPUs were initially accessible through legal channels, recent US sanctions have made these chips increasingly difficult to obtain as exports to China are now prohibited.
One major factor exacerbating this shortage is the widespread adoption of H20 GPUs by DeepSeek's customers. The R1 model has been extensively adopted across Chinese enterprises and government agencies, with most implementations running on cloud-based H20 GPU infrastructure. Consequently, DeepSeek lacks sufficient remaining capacity to train its new model.
The H20 GPU scarcity is already creating limitations for R1 users. If the R2 model offers significant improvements over its predecessor, demand could quickly outstrip the capacity of China's cloud infrastructure providers, according to employees interviewed by The Information.
The H20 processor can match the performance of NVIDIA's H100 GPUs sold to Western companies, but its bandwidth and connectivity have been restricted to comply with previous export regulations regarding chips allowable for Chinese markets. However, the Trump administration determined these watered-down chips still represented excessive capability for geopolitical rivals, prompting immediate export bans in April.
This decision has created significant roadblocks for Chinese AI developers. While domestic alternatives like Huawei Technologies' Ascend 910B chipsets exist, these are less powerful than H20 GPUs and lack support for NVIDIA's CUDA software stack - a programming architecture used to optimize applications and AI models on NVIDIA hardware. This presents a major challenge since virtually all Chinese AI developers currently rely on CUDA-based workflows.
Both DeepSeek's R1 and R2 models are optimized for NVIDIA hardware. The inability to secure these chips could represent a major setback in maintaining parity with US competitors.
