Research firm SemiAnalysis estimates that China has a fleet of over 24GW of datacenter capacity, based on tracking more than 1,000 facilities across over 60 players, in a report introducing its China Datacenter Model published on 25 September 2026. The figure excludes about 20GW of dated pipeline and another about 30GW of announced projects, it says.
In the same report, SemiAnalysis puts US capacity at 56GW as of the end of 2026, with about 15GW in Asia-Pacific excluding China, about 14GW in Europe, the Middle East and Africa, and about 2GW in Latin America. It says published estimates of China’s capacity differ by 15x.
Update (24 Oct): the FT maps where China is building, and Chinese models lead OpenRouter’s usage tables.The Financial Times reported on 8 October that “China is building data centres at rapid speed across its energy-rich hinterland, as Beijing seeks to turn abundant electricity and cheap land into an advantage in the global race to develop AI.” Citing SemiAnalysis, the FT wrote that China “already has 24 gigawatts of operational data centre computing capacity, more than the rest of Asia combined but less than half the 56GW in the US,” and that “a further 50GW is under construction or has been announced.” That is in line with SemiAnalysis’s own split above, about 20GW of dated pipeline plus about 30GW announced. An FT map, sourced to the National Development and Reform Commission, the Institute for Internet Industry, Tsinghua University and FT research, shows 10 planned national data centre clusters placed by demand and energy resources.
Trevor Noren, founder of SageRoad Research, quoted the FT on X and added OpenRouter figures: between 5 and 11 October, he wrote, models processed 172 trillion tokens, with Chinese models accounting for 86.74 trillion against 18.97 trillion for US models, the 24th consecutive week in which Chinese models outpaced US ones. By our arithmetic that is about 50% and about 11% of that week’s total. He also cites Wall Street Journal reporting that Chinese data centre power rates range from 3 to 5.6 cents per kWh, against 7 to 9.1 cents in the US.
OpenRouter’s own rankings page, read by FourWeekMBA on 24 October, lists the five most-used models by tokens this week as all coming from Chinese developers (two DeepSeek versions, MiniMax, Z.ai and Tencent), with Anthropic’s Claude Opus 5.5 sixth at 8.04 trillion tokens. OpenRouter says DeepSeek led with 24.4% of text requests in the week beginning 12 October. Its rankings count only traffic routed through OpenRouter.
Business Pill · LEAD TIME
A one-minute explainer of lead time: the wait between ordering something and having it ready to use. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, SemiAnalysis’s numbers describe two Chinese markets at once. Old retail racks near big cities sit partly empty, while new wholesale capacity for AI is being built inland and filled, which is how high vacancy and an AI capacity shortage can coexist.
SemiAnalysis says the combined capex of Alibaba, Tencent and Baidu reached $20B in the second quarter of 2026, more than doubling year on year, and that for the first time on record all three posted negative free cash flow.
Counting ByteDance as well, it says combined capex went from roughly $35B in 2024 to over $50B in 2025 and is on a trajectory to $100B for 2026.
According to its model, ByteDance alone occupies roughly a fifth of delivered datacenter capacity in China and rents nearly all of it.
Datacenter capacity by region, GW, as estimated by SemiAnalysis in its report of 25 September 2026: US 56 (as of end-2026), China over 24, Asia-Pacific excluding China about 15, Europe/Middle East/Africa about 14, Latin America about 2.Combined capex of ByteDance, Alibaba, Tencent and Baidu as estimated by SemiAnalysis on 25 September 2026: roughly $35B in 2024, over $50B in 2025, on a trajectory to $100B for 2026.
What Was Posted on X
FT: “China is building data centres at rapid speed across its energy-rich hinterland, as Beijing seeks to turn abundant electricity and cheap land into an advantage in the global race to develop AI…While China is struggling to secure enough advanced chips to satisfy soaring AI demand, it has been able to mobilise the land, electricity and construction capacity needed to build the data centres that house them…China already has 24 gigawatts of operational data centre computing capacity, more than the rest of Asia combined but less than half the 56GW in the US, according to SemiAnalysis, the chip-focused research firm. A further 50GW is under construction or has been announced.”
As I argue in my upcoming report on “Global Electricity Undersupply” (learn about our reports here: https://sageroadresearch.com/collections/reports), China’s electricity advantages remain an underestimated edge in the AI arms race, especially as models rapidly grow in size. Chinese models continue to widen the gap with US models in terms of token usage. Just between October 5th and 11th, AI models worldwide processed 172 trillion tokens, with Chinese models accounting for 86.74 trillion of those tokens versus 18.97 trillion for US AI models, according to OpenRouter data. It was the 24th consecutive week in which Chinese models outpaced US models in token usage. This has continued to put downward pressure on token prices, with the LLM Token Expenditure Index falling by more than half since this summer, begging even more questions about the unit economics of US hyperscalers versus their Chinese competitors. The US bet is that bigger and therefore, smarter models are their competitive advantage over China. Itโs staggering how rapidly frontier models are growing. As The New York Times reported recently:
“In 2026, GPT 6 Astra took nearly an octillion calculations to trainโover 50 million times as many calculations as GPT 1 did. Even calling this figure astronomical understates its size: An octillion is a thousand times the number of stars in the known universeโฆResearchers predict that by 2030, training the largest A.I. models may use electricity at a rate similar to New York City today.”
The bigger models get, the more electricity they demand and therefore, the bigger Chinaโs advantage becomes given their electricity supply capacity and pricing. According to Wall Street Journal reporting from late last year, Chinese data center power rates range between 3 and 5.6ยข per kWh while US data center power rates range between 7 and 9.1ยข per kWh.
Trevor Noren of SageRoad Research quotes the FT’s report that China is building data centres across its energy-rich hinterland, with 24GW operating and a further 50GW under construction or announced, per SemiAnalysis. He adds OpenRouter figures he says show Chinese models processing 86.74 trillion of 172 trillion tokens in the week of 5 to 11 October, against 18.97 trillion for US models, and argues that cheaper electricity is an underestimated Chinese edge; the argument is his.
Who Builds It
The report says GDS and VNET, the two Chinese datacenter landlords listed in the US, signed about 1.3GW of wholesale orders in the first half of 2026 with less than 10MW of retail orders, but captured only about a third of ByteDance and Alibaba orders in 2024 to 2026 so far.
State-owned carriers still own a third of national capacity, SemiAnalysis says, and China routinely delivers 100MW facilities in under 12 months.
It expects overseas leasing by Chinese hyperscalers to double from 2026 to 2029 and approach about 4GW of leased capacity.
Empty and Short at Once
SemiAnalysis says China’s market has high vacancy alongside an AI capacity shortage. Its explanation is that older retail racks around Tier-1 cities run at 30% to 50% utilization and are often unsuitable for AI densities, while new wholesale AI capacity is being built elsewhere.
It cites one colocation operator’s disclosures showing wholesale buildings back above 70% utilization while legacy retail racks sit near 60%, and says market rates for power-exclusive capacity were cut in half from about $80 per kW per month.
The Structural Read
The demand is concentrated. SemiAnalysis puts ByteDance alone at roughly a fifth of delivered capacity in China, and says it rents nearly all of it.
The spending broke a pattern. It says Alibaba, Tencent and Baidu all posted negative free cash flow in the second quarter for the first time on record.
Speed is part of the comparison. The report says China routinely delivers 100MW facilities in under 12 months.
SemiAnalysis, 25 September 2026
“Published estimates of Chinaโs datacenter capacity differ by 15x, and reports keep citing high vacancy rates.”
Three Implications
A SECOND-LARGEST MARKET SemiAnalysis estimates China’s fleet at over 24GW, against 56GW for the US by end-2026.
ONE DOMINANT TENANT ByteDance occupies roughly a fifth of delivered capacity, per SemiAnalysis.
SPENDING PAST CASH FLOW Alibaba, Tencent and Baidu capex reached $20B in the second quarter, SemiAnalysis says.
The framework’s starting point: “The structural bottleneck isnโt computing, itโs electrons.”
As we read it, SemiAnalysis’s description of China, largely free of power constraints and delivering 100MW facilities in under 12 months, is the other side of that bottleneck: where power and permits come fast, the constraint SemiAnalysis points to instead is chip supply under export restrictions.
What Is Not Established
We read the free section of SemiAnalysis’s report; the tenant-by-tenant analysis is behind its paywall and we did not read it. All capacity and spending figures are SemiAnalysis’s estimates from its own model, and the report notes that published estimates differ widely. We did not contact SemiAnalysis.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This piece is part of this week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
SemiAnalysis estimates China’s datacenter fleet at over 24GW, against 56GW for the US by the end of 2026, with combined Alibaba, Tencent and Baidu capex of $20B in the second quarter and ByteDance occupying about a fifth of delivered capacity, according to its new China Datacenter Model.
A note on sourcing. We read the free section of SemiAnalysis’s report of 25 September 2026; all figures are its estimates. We did not read its paywalled section. We did not contact SemiAnalysis. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
