The launch of DeepSeek’s upcoming model, R2, could face significant setbacks in China as US export restrictions choke the supply of NVIDIA’s H20 chips, critical for running the company’s AI models, reported The Information.
R2, the successor to DeepSeek’s widely used R1, has yet to receive a release date. The report added that CEO Liang Wenfeng is unsatisfied with its performance, and engineers continue to work on improvements before it is cleared for launch.
Cloud providers that host and distribute DeepSeek’s models warn that existing inventories of NVIDIA chips will likely fall short of meeting the demand R2 could generate, particularly if it performs better than current open-source alternatives.
These concerns have intensified following the April ban on NVIDIA’s H20 chip, which was specifically built for the Chinese market after earlier export restrictions barred the sale of more powerful Hopper series GPUs.
During the recent earnings call, NVIDIA CFO Colette Kress said the company’s outlook reflects a loss of approximately $8 billion in H20 revenue for the second quarter.
R1 and R2 are tightly optimised to run on NVIDIA’s architecture, making substitution with Chinese-developed chips difficult and inefficient.
According to the report, employees at Chinese cloud companies said DeepSeek’s models “are so completely optimised for NVIDIA’s hardware and software” that deploying them on domestic alternatives is not viable at scale.
Despite the ban, some Chinese companies have found a workaround to obtain NVIDIA hardware.
According to The Wall Street Journal, engineers from Chinese AI companies are heading to Kuala Lumpur, Malaysia, with hard drives packed with instructions and data to train AI models. They then utilise the NVIDIA chips available at Malaysian data centres to train the model and return it to China.
Meanwhile, to cope with chip shortages, some Chinese firms have resorted to using gaming GPUs like NVIDIA’s RTX 5090 and 4090, which are also under export restrictions but easier to obtain through grey markets.
DeepSeek, backed by hedge fund firm High-Flyer Capital Management, made headlines for training R1 with less compute than US competitors like OpenAI.
In response to the surge in R1 usage, major Chinese tech firms, including ByteDance, Alibaba, and Tencent, placed $16 billion worth of orders for 1.2 million H20 chips in early 2025, according to SemiAnalysis estimates. That compares with the 1 million chips shipped to China by NVIDIA last year.
Despite these efforts, the scalability of R2 in China could be limited. Companies outside China, not constrained by US chip curbs, may find it easier to deploy the model at full scale.
1 Comment
0rwzuu