DeepSeek's Next-Gen AI Model Delayed by Nvidia GPU Export Restrictions

DeepSeek, an emerging AI disruptor, has hit a roadblock in the development of its next-generation R2 model due to a shortage of Nvidia's H20 processors in China. According to a report by The Information, the company has not yet commented on the availability of the R2 model, leaving users and investors in the dark.

DeepSeek's initial R1 model, which gained rapid adoption among startups, large corporations, and government-affiliated groups, was trained using a cluster of 50,000 Hopper GPUs. This cluster included 30,000 H20s, 10,000 H800s, and 10,000 H100s, all sourced from High-Flyer Capital Management, an investor in DeepSeek. The R1 model's widespread use relied heavily on Nvidia's H20 processors, which are now in short supply due to U.S. export restrictions.

The U.S. government imposed restrictions on the sale of Nvidia's H20 processors for AI training and inference in mid-April. Despite the H20 being a less powerful version of the H100 GPU, it remains a popular choice for Chinese AI companies due to its compatibility with Nvidia's CUDA software stack. Nvidia's sales of H20 processors to Chinese entities have been in the billions of dollars per quarter, highlighting the significance of this market.

DeepSeek's AI software is optimized for Nvidia's hardware, making the company particularly vulnerable to U.S. policy decisions. The recent export curbs have exposed a critical dependency on American hardware, even as DeepSeek claims to have developed its models using significantly fewer resources than U.S. counterparts like OpenAI. The company has not addressed claims by OpenAI that it used proprietary models during the development of R1.

The shortage of H20 processors is already causing problems for the current R1 model, limiting its usage and complicating preparations for the R2 launch. If the R2 model surpasses the capabilities of existing open alternatives, its adoption could surge, potentially overwhelming Chinese cloud platforms. However, the ongoing work to improve the R2 model's performance and meet CEO Liang Wenfeng's standards continues.

As the tech industry watches, the impact of these export restrictions on DeepSeek and other Chinese AI companies remains to be seen. The dependency on American hardware and the effectiveness of U.S. sanctions in curbing China's AI advancements are key factors to monitor in the coming months.