ByteDance Accounts for One-Fifth of China's AI Data Center Capacity
nashnova research
ByteDance now holds roughly one-fifth of China's data-center capacity, making it the country's largest compute-lease client, per SemiAnalysis; a chip-supply bottleneck is pushing it offshore, a dynamic that will shape China's AI infrastructure trajectory.
Why does ByteDance command a full fifth?
Unlike Alibaba, Tencent, and Baidu — which build much of their own capacity — ByteDance sources the vast majority of its data-center space through leasing, not ownership.
This means → it is the single most influential tenant in China's wholesale colocation market. In plain terms = it doesn't build the buildings, but it rents more floor space than anyone else.
By 2026 delivery capacity, Alibaba, Tencent, and Baidu rank behind ByteDance in that order, though all three have a higher self-build ratio.
Where does China rank globally?
SemiAnalysis compiled data from over 60 operators and more than 1,000 data centers across China, estimating total build-out at roughly 24 GW.
That exceeds the rest of Asia at 15 GW and EMEA at 14 GW combined, trailing only North America's 58 GW.
This reflects China's position as the world's second-largest data-center market — yet still less than half the scale of North America.
Capex is doubling — so why is vacancy still high?
ByteDance, Alibaba, Tencent, and Baidu are projected to spend a combined $100 billion this year, double last year's $50 billion.
Yet China's overall data-center vacancy rate remains at 50–60 %. In plain terms = half the rack space looks empty, even as spending accelerates.
This means → the contradiction is structural: vast numbers of small, pre-2015 facilities built by telcos are physically unsuited to AI workloads, dragging down headline utilization — while AI-ready capacity is actually undersupplied.
What broke the growth stall?
From 2022 to 2024, China's data-center market went through a prolonged growth plateau.
Expansion resumed only after generative AI took off in 2024, propelled by homegrown large models such as DeepSeek.
This reflects a demand curve that was not linear — it was activated by a specific application inflection point.
How tight is the chip bottleneck?
Nvidia products face dual export restrictions from both the US and China; domestic chip alternatives reportedly still fall short of demand.
Per Yahoo Finance, Chinese officials are weighing whether to let ByteDance and Alibaba purchase limited quantities of RTX PRO 5500 chips.
This means → even a small, model-specific carve-out requires top-level approval, signaling that the supply constraint has reached a case-by-case gatekeeping stage.
How is ByteDance routing around the shortage?
A Singapore subsidiary of ByteDance is reportedly among the largest clients of compute-cloud provider Nscale, accessing restricted Nvidia chips through Nscale's facilities in Norway.
In plain terms = chips unavailable domestically are used indirectly through an offshore entity leasing offshore rack space.
Whether ByteDance can sustain its AI infrastructure build-out before domestic chips mature is a key variable in the evolution of China's AI compute landscape.
市场有风险，内容仅供研究参考，不构成投资建议。