According to insiders, ByteDance Ltd. is planning to invest $7 billion in acquiring access to Nvidia's most powerful graphics processing units (GPUs) to advance the development of cutting-edge artificial intelligence models. Instead of directly purchasing the chips, the company aims to lease these computing resources through data centers located outside mainland China.

The Information reports that ByteDance founder Zhang Yiming is in discussions with data center operators in Southeast Asia and the Middle East to secure access to Nvidia's next-generation Blackwell GPUs. It is said that Zhang has stated that ByteDance's investment in AI will exceed the combined total of all its competitors, including Alibaba Cloud, Tencent Holdings, Baidu, and Huawei.

ByteDance has already been accessing Nvidia's previous-generation Hopper AI GPUs through data centers outside mainland China, as these products cannot be directly exported to China. However, the U.S. government is considering further restrictions on such operations, potentially implementing new rules to limit Chinese companies' ability to rent chips from U.S. cloud providers. This could also involve requiring companies like Google and Microsoft to act as "gatekeepers" to oversee access to advanced AI chips.

Meanwhile, some Chinese companies are opting to optimize their existing resources to address these challenges. For example, DeepSeek recently released an open-source large language model called DeepSeek-R1, which they claim performs on par with OpenAI's GPT-4. Despite using only the downgraded H800 chips from Nvidia, the total cost for training this model was approximately $5.6 million.