ByteDance Rents a Fifth of China's Data Center Capacity

The usual way to talk about China's AI compute is through the chips it cannot buy. SemiAnalysis just published a count of the buildings those chips go into, and the number is bigger than most of the conversation assumes. On September 25, the research firm launched its China Datacenter Model, a building-by-building tracker covering more than 1,000 facilities across more than 60 operators. Its headline: China is on track for a fleet of over 24GW by the end of 2026.

That puts China second only to the United States, which SemiAnalysis pegs at 56GW. It is also larger than all of EMEA, at 14GW, and larger than the rest of Asia, at 15GW. And the 24GW figure excludes roughly 20GW of dated pipeline plus another 30GW of announced projects.

One tenant, a fifth of the market

The most striking finding is about a single customer. According to the report, written by Everlyn, Dylan Patel and Patrick Schaabi, ByteDance occupies roughly a fifth of delivered data center capacity in China, and it rents nearly all of it rather than building its own. That makes the TikTok and Doubao parent the single most important customer for every wholesale colocation provider in the country, and one of the world's largest GPU users.

The renting model matters. Alibaba, Tencent and Baidu own large amounts of their own capacity. ByteDance has instead let landlords like GDS, VNET, Chindata and the state carriers take on the construction and financing, and then absorbed the output. SemiAnalysis reports that GDS and VNET signed 1.3GW of wholesale orders in the first half of 2026, yet the two listed players captured barely a third of ByteDance and Alibaba orders between 2024 and this year. The rest went to a long tail of operators, including names like AtHub, Kehua Data and Aofei that rarely appear in Western coverage.

The effect on the landlords' economics is already visible. Wholesale capacity is back above 70% utilization, per the report, while legacy retail racks sit near 60%.

Built for retail, flipped to AI

SemiAnalysis frames the fleet's history in four eras. Before 2015, state carriers China Mobile, China Telecom and China Unicom held 60% to 70% of the market through carrier hosting. From 2015 to 2021, a cloud land grab followed, with Alibaba Cloud revenue compounding at roughly 110% a year. Then came digestion: Alibaba Cloud growth fell from 50% in 2021 to 23% in 2022, and eastern Tier-1 city pricing of around $80 per kilowatt per month was cut in half during the downturn.

Since 2024, the firm describes the market as an AI supercycle. The fleet was built for retail enterprise and cloud customers, and it is now being repurposed and extended for AI training and inference. That explains the utilization gap between modern wholesale halls and older retail racks.

A 100MW site in about a year

Speed is the other headline. A standard 100MW facility in China has compressed from around 18 months to around 12 months, according to the report. Permitting typically takes three to six months, compared with 12 to 13 months in the United States. At the extreme end, Alibaba's CUBE 5.0 design is pitched as a "100-day datacenter": 30 days in the factory, 50 days of installation and 20 days of commissioning.

The money behind it is moving just as fast. Alibaba, Tencent and Baidu spent a combined $20 billion on capex in the second quarter of 2026, double the year before, and all three posted negative free cash flow. SemiAnalysis calls it the largest capex step-up in the sector's history. Combined capex for the three was about $35 billion in 2024 and more than $50 billion in 2025, and the report puts 2026 on a trajectory toward $100 billion.

The western pull

Geography is being set by policy. The Eastern Data, Western Compute program, launched by the National Development and Reform Commission in February 2022, created eight national computing hubs implemented as ten clusters. SemiAnalysis sorts them into four western hubs aimed at training, in Hohhot and Ulanqab, Zhongwei, Qingyang and Gui'an, and four eastern hubs serving inference near population centers.

The lever is energy. Projects above 10MW need approval under an energy-consumption quota system, and Inner Mongolia, home to China's largest renewable fleet, offers power at roughly half the price of Tier-1 cities. Training runs do not need to sit next to users; they need cheap, plentiful electricity. The model also flags Shanxi, which sits outside the official plan but hosts major ByteDance capacity.

The binding constraint is still silicon

None of this means China's compute problem is solved. SemiAnalysis says its companion Accelerator and HBM Model shows growth remains constrained by chip supply. China can pour concrete and connect power faster than anyone. What fills the halls is the harder question, and it is where export controls, domestic accelerator output and memory supply still set the ceiling.

That is the useful correction in this report. Shell capacity and delivered compute are different things, and a 24GW fleet is a statement about the first. It tells you China has built the room. Whether the room is full of competitive accelerators is a separate ledger.

Why it matters

Three things stand out for anyone tracking the AI race. First, the physical buildout in China is not the bottleneck; at 100MW in about a year and permits in months, it is arguably faster than the US. Second, the market's demand is unusually concentrated. With one company renting a fifth of the country's capacity, ByteDance's model roadmap and chip access effectively set the order book for an entire colocation industry, and any shift in its strategy ripples through GDS, VNET and the long tail at once. Third, the real contest has narrowed to chips and memory. If domestic accelerator supply catches up with the shells already built or in the 50GW pipeline, the gap between China's compute and its compute ambitions could close much faster than the export-control debate assumes.