Like other Chinese companies, TikTok’s owner ByteDance is unable to purchase and install the highest-performance Nvidia GPUs in its data centers in China. However, the company discovered that it can still use Nvidia GPUs that are physically located in cloud data centers in other countries. The company aims to expand its use of such GPUs next year, spending up to $7 billion on access to Nvidia GPUs, The Information reports, citing independent sources. ByteDance denied the report.
According to the report, ByteDance plans to invest more than $20 billion in AI infrastructure, including $7 billion in access to advanced Nvidia GPUs in the cloud, data centers, and even undersea cables. Masu. The US prohibits ByteDance from purchasing Nvidia GPUs and using US cloud services. However, you cannot block ByteDance’s access to cloud services elsewhere, such as in the Middle East or Asian countries. As a result, ByteDance will have access to U.S. processors while technically complying with U.S. sanctions on China’s AI and HPC sectors.
ByteDance denied the report. But if this report is correct and ByteDance invests $7 billion in cloud access to Nvidia GPUs, the company will become one of the world’s largest consumers of AI hardware.
On-demand access to Nvidia’s H100 GPU is available immediately. In the US, access to H100 GPUs starts at $1.33 per hour on long-term contracts. Prices in other countries should be more or less comparable, at around $1.3 per hour, meaning ByteDance could rent a cluster of 614,682 H100 GPUs running 24/7 in 2025 for $7 billion .
I don’t know if there are nearly 615,000 H100 GPUs available for rent in the Middle East and Asia. Also, since ByteDance’s AI projects are fairly limited, we don’t know if training and inference workloads will need that many processors. For example, there is the Doubao AI chatbot, which has 51 million active users and is considered to be the company’s largest project. Therefore, the company either plans to reduce the cost of renting AI infrastructure, or significantly expands its AI projects, which requires more AI capabilities, or Nvidia’s reduced H20 HGX and B20 GPUs. Either you intend to continue sourcing to run it yourself. In addition to renting processors from cloud providers, we also support data centers in China. For example, ByteDance reportedly spent more than $2 billion on more than 200,000 Nvidia H20 GPUs in 2024, indicating that the company will once again stop buying its own hardware and rely entirely on cloud providers from other countries. It is unlikely that you will focus on it.
In particular, ByteDance is reportedly working with Broadcom to develop its own AI processor to reduce its dependence on Nvidia. The company is rumored to be working on two processors, one for training and one for inference. The chip is expected to be manufactured by TSMC based on its N4/N5 process technology and is expected to go into mass production in 2026. However, due to U.S. export control regulations, ByteDance won’t be able to make its GPUs significantly faster than Nvidia’s HGX H20 (which would prevent TSMC from having to ship high-performance GPUs to Chinese companies). An in-house processor will be much more cost-effective for the company.
