ByteDance and Alibaba Place Massive GPU Orders with NVIDIA, Fueling the AI Race

According to recent reports from LatePost, major Chinese internet companies that have a cloud computing business have placed substantial orders with NVIDIA following this year’s Spring Festival. ByteDance, for example, has ordered over $1 billion in GPUs from NVIDIA this year.

One source close to NVIDIA stated that ByteDance has so far received a total of 100,000 A100 and H800 accelerator cards, including both delivered and pending orders. Notably, the H800 only began production in March of this year. It was revealed that the orders from ByteDance alone this year might approach the total number of commercial GPUs NVIDIA sold in China last year.

A ByteDance insider shared that in June 2020, after the release of GPT-3, the company had trained a generative language model with several billion parameters using the V100 computing card. However, due to the limited parameter scale, the model’s generation capabilities were average, and ByteDance did not see its commercial potential at that time.

Another source from Alibaba Cloud revealed that Alibaba had also been actively purchasing GPUs in 2018-2019. At that time, Alibaba‘s purchase volume reached at least tens of thousands of units, including V100 and T4 computing cards. Approximately one-tenth of these were sent to the Damo Academy for AI technology research and were used in the trillion-parameter large model M6 released in 2021.

SEE ALSO: ByteDance’s Douyin Releases Rules that Prohibit Public Welfare Accounts from Participating in Live Streaming Tipping

The same source mentioned that major companies like ByteDance and Alibaba mainly negotiate purchases directly with NVIDIA. The demands of these companies are so large that distributors and the secondary market struggle to meet them. NVIDIA, however, offers certain discounts based on the scale of the purchase.

The current price of an A100 on NVIDIA’s official website is $10,000, while an H100 costs $36,000. The A800 and H800 models, which were launched in the Chinese market last year, are slightly cheaper. However, due to supply and demand issues, the transaction price is bound to be higher.