ByteDance occupies around a fifth of data center capacity in China, according to a recent report. This makes the company the largest customer for rented compute in the country at a time when its major AI developers are racing to accelerate infrastructure buildouts and pushing against capacity limits.
A report by SemiAnalysis on China's data center buildout finds that ByteDance is the largest spender on AI infrastructure in the country. The vast majority of data center capacity used by the privately held company is leased rather than self-built, making it the wholesale colocation player with the most impact in the country.
SemiAnalysis based its conclusions on its new SemiAnalysis China Datacenter Model. It combines over 1000 data centers across more than 60 players in China. The data indicates that China's total buildout reaches roughly 24GW, more than the 15GW projected in the rest of Asia and 14GW in the region combining Europe, the Middle East, and Africa. Its figure comes only second to the 58GW built in North America.
After ByteDance, Alibaba, Tencent, and Baidu in that order are the next biggest hyperscalers by 2026 delivered capacity, although self-built sites make up a larger share of that infrastructure than with ByteDance. The report noted that the four companies' combined capex is set to reach US$100 billion this year, a doubling from the US$50 billion spent last year.
Yet despite this market experiencing its largest capex growth ever, vacancy rates among China's overall data centers remain relatively low at 50% to 60%, according to the analysis. This is due to the large supply of legacy data centers built before 2015, which were dominated by telecom providers. These tended to be smaller and are not physically suitable for today's AI demands.
Then came a period of rapid development by cloud providers Tencent and Alibaba after 2015, which led to escalating capex and a price war, followed by a period of market digestion from 2022 that saw growth slow to a halt. This downturn has since been reversed since generative AI came to the scene in 2024 and the rise of homegrown industry giants like DeepSeek.
The result, according to SemiAnalysis, is that older data centers not suitable for AI compute have dragged down the overall data center capacity rate, even as the AI sector faces a capacity shortage and is rapidly scaling the infrastructure buildout.
Breaking bottlenecks
Part of the compute shortage relates to the chips themselves. With Nvidia sales restricted by both the US and China, the country has been pushing for domestic substitution of high-end chips.
Yet despite advances by Chinese chipmakers, this has reportedly not been enough to satisfy domestic demand for now. This may indicate why Chinese officials are considering allowing ByteDance and Alibaba to purchase some RTX PRO 5500 chips, according to Yahoo Finance.
Foreign compute may also be helping to alleviate current domestic limits. Like other Chinese modelmakers, ByteDance has been found to also rent overseas clouds to train its models. Recent reporting indicates that a Singaporean subsidiary of the company may be among Nscale's largest customers, allowing it to access restricted Nvidia chips using the neocloud's Norway facility.
Article edited by Jack Wu
