Like other Chinese entities, TikTok owner ByteDance cannot buy the highest-performance Nvidia GPUs and install them into its data centers in China. However, the company has found that it can still use Nvidia GPUs that are physically located in cloud data centers in other countries. Next year, the company aims to expand its usage of such GPUs and spend as much as $7 billion on access to Nvidia GPUs, reports The Information, citing its own sources. ByteDance has denied the report.
The report says ByteDance plans to invest over $20 billion in AI infrastructure, including $7 billion in accessing advanced Nvidia GPUs in the cloud, data centers, and even submarine cables. The U.S. prohibits ByteDance from purchasing Nvidia GPUs and using American cloud services. However, it cannot block ByteDance’s access to cloud services elsewhere, for example, in the Middle East or Asian countries. As a result, ByteDance can access American processors while technically adhering to U.S. sanctions against China’s AI and HPC sectors.
ByteDance has denied the report. However, if the report is correct and ByteDance invests $7 billion in cloud access to Nvidia GPUs, it will be one of the world’s largest consumers of AI hardware.
On-demand access to Nvidia’s H100 GPUs is readily available. In the U.S., the price of access to H100 GPUs starts at $1.33 per hour for longer-term commitments. Pricing in other countries should be more or less comparable, and at around $1.3 per hour, ByteDance will be able to rent a cluster of 614,682 H100 GPUs working 24/7/365 in 2025 for $7 billion.
We are not sure that there are nearly 615,000 H100 GPUs available for rent in the Middle East and Asia, and we are also not sure that ByteDance needs that many processors for its training and inference workloads, as its AI projects are fairly limited. For example, it has the Doubao AI chatbot with 51 million active users, which is believed to be its largest project. Therefore, either the company will spend less on renting its AI infrastructure, or it plans to significantly expand its AI projects and therefore needs more AI capabilities, or it intends to continue procuring Nvidia’s cut-down H20 HGX and B20 GPUs to run in its own datacenters in China in addition to renting processors from cloud providers. For instance, ByteDance has reportedly spent over $2 billion on more than 200,000 Nvidia H20 GPUs in 2024, and it is unlikely that the company will cease buying its own hardware and refocus entirely on relying on cloud providers from other countries.
Notably, ByteDance is also reportedly working with Broadcom to develop its own AI processors to reduce its reliance on Nvidia. It is rumored that the company will be working on two processors: one for training and another for inference. The chips are projected to be made by TSMC on its N4/N5 process technologies and will enter mass production in 2026. Although ByteDance will unlikely be able to make its GPUs significantly faster than Nvidia’s HGX H20 due to U.S. export control restrictions (which would prevent TSMC from shipping high-performance GPUs to Chinese entities), in-house processors will be considerably more cost-effective for the company.
