Short answer: SemiAnalysis counts more than 24 GW of operating data center capacity in China, with roughly 50 GW more in the pipeline, against about 56 GW in the United States projected for the end of 2026. The Financial Times relayed the figures on October 8, 2026 under the framing that Beijing is rolling out AI infrastructure at breakneck speed in Inner Mongolia. The headline is real; the comparison is trickier than it looks, because gigawatts are not compute.
SemiAnalysis published the underlying analysis on September 25 as The Chinese AI Infrastructure Boom. Only the preview is public; much of the tenant-level detail is paywalled, so this post sticks to what the public preview and the FT summary state.
TL;DR: the numbers and the questions behind them
| Question | Answer |
|---|---|
| China operating capacity? | More than 24 GW across 1,000+ facilities and 60+ operators |
| China pipeline? | About 20 GW dated plus about 30 GW announced, roughly 50 GW |
| US capacity? | About 56 GW projected for year-end 2026 |
| Rest of world? | About 15 GW APAC excluding China, about 14 GW EMEA, about 2 GW Latin America |
| Is power the constraint? | SemiAnalysis says the Chinese buildout is chip-gated, not power-gated |
| Is it all utilized? | No: SemiAnalysis cites roughly 50 percent utilization overall and about 60 percent for legacy retail racks |
Where is the capacity, and who is paying for it?
SemiAnalysis describes Inner Mongolia as China's "Johor": the default place to put hyperscale AI capacity, in the way Johor in Malaysia absorbed Singapore's overflow. The reasons are practical. Power prices there are roughly half of tier-1 city levels, and the region is served by the Mengxi grid, the only provincial grid in China outside State Grid and Southern Grid. ByteDance runs a twin-site program at Horinger and Ulanqab. SemiAnalysis says Chengdu and Chongqing lost AI demand to Inner Mongolia on power price and grid flexibility, and Shanxi drew close to 2 GW of commitments from ByteDance, Baidu and JD.
On spending, SemiAnalysis reports that Alibaba, Tencent and Baidu together reached $20 billion of capital expenditure in the second quarter of 2026, more than double the prior year, and that all three posted negative free cash flow for the first time on record. ByteDance, which is private, holds about a fifth of delivered capacity, nearly all of it rented. Its separate newsletter coverage ranks ByteDance at about 5.3 GW, Alibaba 3.9 GW, Tencent 2.4 GW and Baidu 1.1 GW, per secondary summaries of the analysis.
The state still matters. State carriers own about a third of national capacity, and the 15th Five-Year Plan (2026 to 2030) layers a further roughly 40 percent onto the grid companies' earlier blueprint, to over $746 billion (5 trillion yuan), according to SemiAnalysis.
Why gigawatts overstate the picture
It is tempting to read 24 versus 56 as a ratio of AI capability. Do not. Three things break the comparison.
1. Chips gate the buildout. SemiAnalysis characterizes China's expansion as chip-gated. Export restrictions limit supply of the best Nvidia parts, and new capacity leans on H20-class and Huawei Ascend servers. A gigawatt filled with those delivers less training and inference throughput than a gigawatt of current-generation Nvidia racks; see our reading of Rubin NVL72's claimed 67x throughput for how fast the top end is moving. SemiAnalysis founder Dylan Patel has argued in interviews that nominal gigawatts overstate China's effective compute and that China accounts for under 10 percent of incremental AI data center capacity versus about 70 percent for the US; that figure comes from secondary summaries of an August interview, not a SemiAnalysis publication, so weigh it accordingly.
2. A lot of the shell is not AI-ready. SemiAnalysis reports that legacy retail racks often cannot support the power density of H20 or Ascend servers and that retrofitting is impractical. Developers compete hard on price, with power-exclusive rates cut roughly in half from about $80 per kW per month.
3. Capacity is not demand. Vacancy is high in places. Regulators have also enforced energy quotas: in 2023 Guangdong found two Tencent campuses running at roughly 13 to 14 times their approved scale and ordered rectification.
How fast can China build?
This is where the story is most striking. SemiAnalysis says standard delivery for a 100 MW facility has dropped from roughly 18 months to about 12, and that permitting typically takes three to six months versus 12 to 13 months in the US. Alibaba's "100-day datacenter" (CUBE 5.0) is the headline example of prefabrication, and shell-and-fit-out costs are described as a fraction of comparable US wholesale costs.
The contrast with the US is about friction rather than money. American projects are increasingly hitting local politics and grid queues; see our posts on the New York data center moratorium debate and California's data center bills. Speed of construction does not fix a chip shortage, though, so China's advantage in building shells meets its disadvantage in filling them.
What does the US side of the ledger look like?
The 56 GW figure is a projection for the end of 2026 from SemiAnalysis's global model. Inside it, a few buyers dominate. OpenAI has projected $750 billion of compute spend through 2030, Anthropic has disclosed roughly $517 billion in compute commitments, and Meta is building its own silicon and capacity, as in our post on the Meta MTIA chip and 14 GW plan. Nvidia's role as financier of its own customers is covered in Nvidia as the central bank of AI.
The pattern: US capacity is concentrated in a handful of buyers who use the best chips; Chinese capacity is spread across many operators, much of it leased, using weaker chips, with the state steering location and power.
How should you read a gigawatt number in any AI headline?
Because this comparison will be quoted for months, it helps to have a checklist. When you see a capacity figure for a country or company, ask five things.
First, is it operating, under construction, or merely announced? SemiAnalysis separates all three for China, and the gap between 24 GW operating and roughly 50 GW in the pipeline is larger than the operating base itself. Second, is it facility power or IT power? Cooling and conversion losses mean a facility figure overstates the power available to chips. Third, who owns it and who rents it? ByteDance renting nearly everything it uses is a very different risk profile from a hyperscaler owning its campuses. Fourth, which chips fill it? The same 100 MW hall can be worth many times more compute with current Nvidia racks than with older or export-limited parts. Fifth, what is the source model? SemiAnalysis tracks facilities bottom-up, while the broader press often repeats a single number from a single analyst. Further context on China's compute position is available in reporting from SCMP on how Chinese AI giants stretch each dollar of compute and in ChinaTalk's explainer on China's chip surplus.
Applied to the headline, the honest summary is this: China is building a lot of shell very quickly, concentrated in cheap-power regions, mostly for a few large tenants, and it is still limited by what accelerators it can buy or make. The United States is building fewer, larger, better-equipped sites, and it is limited by grid access, permitting and politics. Neither side's gigawatts translate one-to-one into model capability, which is why watching chip policy, utilization and actual model releases will tell you more than any single capacity chart.
What this means for builders and readers
- Do not infer model quality from GW. Chinese labs have repeatedly shipped competitive open-weight models on less compute; capacity charts do not predict that. Our post on Chinese AI tokens as an economy shows demand growing at home.
- Watch chip policy, not power. The binding variable in the SemiAnalysis framing is access to accelerators. Approvals and restrictions move the outcome more than new substations; see the approval for ByteDance and Alibaba to buy Nvidia RTX Pro 5500.
- Expect overseas leasing. SemiAnalysis expects Chinese hyperscaler overseas leasing to roughly double from 2026 to 2029, approaching about 4 GW, a way around domestic chip limits.
- Treat the 50 GW as a ceiling. It mixes dated projects with merely announced ones, and announcements in this sector have a history of slipping.
- Cross-check headline figures. SemiAnalysis notes published estimates of China's capacity differ by up to 15x, which tells you how soft the category is.
Open questions
- How much of the 24 GW is usable for frontier-scale training versus inference or general cloud?
- Will domestic chip output (Ascend and others) scale fast enough to fill the 50 GW pipeline?
- How do the tenant-level numbers behind the paywall change the picture of ByteDance's dependence on leased space?
- Does policy relief on exports change the mix?
Figures are accurate as of October 8, 2026 and come from the SemiAnalysis preview and press summaries. Full SemiAnalysis detail is paywalled.