Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Like other Chinese entities, ByteDance, TikTok’s owner, cannot purchase high-performance Nvidia GPUs and install them in its data centers in China. However, the company found that it could still use Nvidia GPUs physically located in cloud data centers in other countries. Next year, the company aims to expand its use of these GPUs and spend up to $7 billion on access to Nvidia GPUs, according to reports. information, Quoted from its own sources. ByteDance has denied this report.
The report says ByteDance plans to invest more than $20 billion in AI infrastructure, including $7 billion to access advanced Nvidia GPUs in the cloud, data centers and even undersea cables. The US bans ByteDance from purchasing Nvidia GPUs and using US 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 US processors while technically complying with US sanctions against China’s AI and HPC sectors.
ByteDance has denied this report. However, if the report is true and ByteDance invests $7 billion in cloud access to Nvidia GPUs, it will be one of the largest consumers of AI hardware in the world.
On-demand access to Nvidia’s H100 GPUs is readily available. In the US, access to H100 GPUs starts at $1.33 per hour for long-term commitments. Prices in other countries should be fairly comparable, and at around $1.3 per hour, ByteDance will be able to lease a pool of 614,682 H100 GPUs running 24/7 in 2025 for $7 billion.
We’re not sure that there are approximately 615,000 H100 GPUs available for rent in the Middle East and Asia, nor are we sure that ByteDance needs that many processors for its training and inference workloads, as its AI projects are fairly limited. For example, its Doubao AI chatbot has 51 million active users, which is believed to be its largest project. So, either the company is spending less on leasing AI infrastructure, or it plans to expand its AI projects significantly and therefore needs more AI capabilities, or it intends to continue purchasing Nvidia’s H20 HGX and B20 GPUs to run them itself. . Data centers in China as well as leasing processors from cloud service providers. For example, ByteDance reportedly spent more than $2 billion on more than 200,000 Nvidia H20 GPUs in 2024, and the company is unlikely to stop buying its own hardware and completely refocus on relying on cloud providers from other countries.
Notably, ByteDance does as well It is said to be working with Broadcom to develop its own AI processors To reduce its dependence on Nvidia. The company is rumored to be working on two processors: one for training and one for inference. The chips are expected to be manufactured by TSMC using its N4/N5 processing technologies and will enter mass production in 2026. Although ByteDance is unlikely to be able to make its GPUs much faster than Nvidia’s HGX H20 due to control limitations US exports (which would prevent TSMC from shipping high-performance GPUs to Chinese entities), in-house processors would be more cost effective for the company.
