The short answer. China's data center market is projected to reach 362.1 billion yuan, about $52 billion, in 2026, up 13.9% from 318 billion yuan in 2025. AI data centers account for 176.3 billion yuan, about $25.3 billion, or 48.7% of the total, and are growing at 30%. National intelligent compute capacity hit 2,185 EFLOPS in June 2026, up 177% year over year. Source: China Industrial Research Institute via AskCI (January 2026), MIIT via Economic Daily (July 2026).
This page collects the numbers that keep getting quoted without attribution, and puts the date on each one. It is updated as new figures are published. Where a number is a forecast, it says forecast. Where a number is a single-firm estimate rather than an official statistic, it says so.
China data center market size
The headline number comes from the China Industrial Research Institute, published through AskCI in January 2026.
| Metric | 2025 | 2026 (forecast) | Change |
|---|---|---|---|
| Total data center market | 318 billion yuan | 362.1 billion yuan (~$52B) | +13.9% |
| AI data centers (AIDC) | 135.6 billion yuan | 176.3 billion yuan (~$25.3B) | +30.0% |
| AIDC share of total | 42.6% | 48.7% | +6.1 pts |
| National rack count | — | approaching 1 million units | — |
Source: China Industrial Research Institute via AskCI, January 2026. 2026 figures are full-year forecasts, not audited results. Currency conversion at the January 2026 rate.
The composition number matters more than the total. When 48.7% of a market is being built for one workload type, the market's cost structure changes with it. AI facilities carry higher power density per rack, higher cooling capital cost, and a demand profile tied to inference volume rather than to construction schedules.
The AI data center segment (AIDC)
China's industry uses a specific term for AI facilities: AIDC, for AI data center. It is not a marketing label, it appears in provincial planning documents and subsidy rules, which is why the segment can be measured separately at all.
AIDC grows at 30% against 13.9% for the overall market. The gap is the transition from building general-purpose colocation to building power-dense inference capacity.
The driver is inference. China's inference compute demand now runs four to five times its training demand, according to industry estimates. Training is episodic: a model is trained once, then deployed. Inference runs continuously for as long as the product is live. When inference demand exceeds training demand by that multiple, the market stops being a construction market and becomes an operations market with recurring electricity and cooling costs.
Compute capacity and the supply gap
| Metric | Value | Period | Source |
|---|---|---|---|
| Intelligent compute capacity | 2,185 EFLOPS (FP16), +177% YoY | June 2026 | MIIT via Economic Daily |
| National compute utilization rate | 71.4% | June 2026 | MIIT via Economic Daily |
| AI compute demand growth | +417% YoY | Q1 2026 | CAICT |
| AI compute supply growth | +128% YoY | Q1 2026 | CAICT |
| Total compute available to Chinese entities | 4.6M H100-equivalents | 2026 | Institute for Progress, Oct 2026 |
| Total compute available to Chinese entities | 8.8M H100-equivalents, +91% | 2027 forecast | Institute for Progress, Oct 2026 |
| Share from US-designed hardware | ~65% | 2027 forecast | Institute for Progress, Oct 2026 |
| Share delivered by remote cloud access | ~one third | 2027 forecast | Institute for Progress, Oct 2026 |
Demand grew three times faster than supply in the first quarter of 2026. That gap is why the three largest Chinese cloud providers all accelerated capital spending in the same period rather than waiting for the cycle to soften.
Data center electricity consumption
The Energy Institute's 2026 Statistical Review of World Energy counted global data center electricity consumption for the first time in the report's 75-year history.
| Region | 2025 consumption | Global share |
|---|---|---|
| World | 787.8 TWh (up from 658.2 TWh in 2024, +20%) | 100% |
| United States | 312.6 TWh | 39.7% |
| China | 205.7 TWh | 26.1% |
| Europe | 144.6 TWh | 18.4% |
Source: Energy Institute, 2026 Statistical Review of World Energy, August 2026. Global data center electricity demand has grown about 92% since 2020.
Gartner projects global data center electricity consumption at 565 TWh in 2026, up 26% from 447 TWh in 2025, and above 1,200 TWh by 2030. That is roughly Japan's entire annual electricity consumption.
Why data center power is cheaper in China
Data centers in Inner Mongolia pay about 0.33 yuan per kilowatt-hour. The equivalent US industrial rate in Ohio runs near $0.60. The gap is not a subsidy line item, it comes from how the tariff is set.
Chinese grid operators use regulated tariffs rather than market-clearing prices, and carry roughly 4 trillion yuan in debt doing it. When a data center in Ulanqab draws power from a wind farm 41 kilometers away over a dedicated 220 kV line, the transmission cost is a regulated number chosen to make the project viable, not a price discovered in an auction.
US industrial electricity varies by more than a factor of four within the same country: New Mexico around 5.43 cents per kWh, Texas near 8.26 cents, California averaging 21.53 cents. For a 100-megawatt facility, annual power cost runs about $736,000 per megawatt in New Mexico and $2.873 million per megawatt in California. Location decisions follow those electrons.
PUE and the efficiency floor
New data centers in China's national hub nodes are required to meet a PUE below 1.2. That is a regulatory floor, not a target, and it shapes which designs get built.
The economics explain why. A facility at PUE 1.1 spends about 10 cents on overhead for every dollar of server electricity. At PUE 1.5 it spends 50 cents. At PUE 2.0 it spends a full dollar. For comparable compute, the difference between PUE 1.1 and 2.0 is roughly $7 million versus $79 million a year in overhead energy.
The metric itself is close to exhausted as an improvement lever. It has been the industry's primary efficiency benchmark since 2007, and the easy gains from air-side economizers and hot-cold aisle containment are already banked. What remains is cooling architecture, which is why liquid cooling moved from pilot to default in AI facilities.
Cooling capital cost: China versus North America
| System | China (yuan/kW) | China (USD/kW) | North America (USD/kW) |
|---|---|---|---|
| Cold plate liquid cooling | 4,000 to 5,000 | ~$550 to $690 | >$1,380 |
| Air cooling | 2,000 to 3,000 | ~$275 to $415 | ~$550 to $690 |
Source: research compiled by Hansen Fluid, a Chinese thermal management company, 2026. The capex gap is one-time. The PUE savings from a better cooling design compound every year for the life of the facility, which is why the capex number alone understates the difference.
Grid capacity behind the build-out
China built 46 ultrahigh-voltage transmission projects in two decades: 22 AC lines and 24 DC lines, more than 62,000 kilometers of route length, expected to pass 70,000 kilometers by the end of 2026. West-to-east transmission capacity reached 340 million kilowatts.
The US grid was built in the 1960s and 1970s. The average large power transformer in service is 38 to 40 years old, and more than 70% of transmission lines are past 25 years. Transformer lead times surpassed 160 weeks in the first quarter of 2026, more than three years, against US data center capacity projected to grow from roughly 24 gigawatts in 2025 to 110 gigawatts by 2030.
That is why US operators started building generation behind the meter. Nearly 60 behind-the-meter gas projects have been reported since the start of 2025, representing roughly 90 gigawatts of planned capacity, with data center tracking firm Cleanview identifying around 60 data centers and 97 gigawatts of on-site generation. Grid interconnection queues in the US run four to seven years; data centers need power in 18 to 24 months.
Curtailment and direct green supply
China wasted 360 billion kilowatt-hours of wind and solar generation in the first half of 2026, more than its net new electricity demand. The cause is geographic: western China holds more than 78% of the country's wind and solar resources but carries only about 30% of its electricity load. New energy installations grew 22% in 2025 to 430 gigawatts while grid investment grew only about 5%.
The response has been to move demand to the generation instead of moving generation to the demand. On May 2, 2026, a 500-megawatt solar plant in the Tengger Desert began sending power directly to a data center cluster one kilometer away over four dedicated 110 kV lines, bypassing the public grid. The project cost 8.7 billion yuan and serves the Zhongwei cloud base in Ningxia. It was China's first large-scale computing-power coordination green direct-supply project, and at least half a dozen similar projects are operating or under construction across western China.
For contrast, PJM's capacity auction cleared at $28.92 per MW-day for the 2024/2025 delivery year, $269.92 for 2025/2026, and $329.17, the price cap, for 2026/2027. The grid's independent market monitor attributes about 63% of the 2025/2026 increase to data centers.
Who is building
China's market is more distributed than the US one. The three state telecom carriers, China Mobile, China Telecom and China Unicom, remain significant operators. They are joined by provincial state-owned enterprises, specialized AIDC developers, and the internal infrastructure arms of Alibaba and ByteDance.
| Player | Commitment | Period |
|---|---|---|
| Alibaba | More than 380 billion yuan for cloud and AI infrastructure | Three-year plan |
| Tencent | 84.7 billion yuan capex | First half 2026, above full-year 2025 |
| Top nine global cloud providers | Above $886.7 billion combined capex, +~90% | 2026, TrendForce forecast |
Source: Alibaba and Tencent financial disclosures (2026); TrendForce, August 2026. The nine providers in the TrendForce figure are Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba and Baidu.
Policy is explicit about the direction. China's 15th Five-Year Plan, released in March 2026, lists the national integrated computing network as one of six major infrastructure networks and one of 109 major projects, and calls for cloud-edge-device coordination and reducing the cost of compute for society as a whole. Of the 1.3 trillion yuan in ultra-long special treasury bonds issued in 2026, 800 billion yuan is directed at major national strategy and security capacity projects, supporting 1,417 projects across the year.
Where the capacity sits
The build-out is not centered on the coasts. Power availability and land cost pushed new capacity inland, and the projects that get cited as firsts are increasingly in the west and offshore.
- Ulanqab, Inner Mongolia. Power at roughly 0.33 yuan per kWh, with wind farms connected by dedicated lines. One of the largest clusters in the country.
- Zhongwei, Ningxia. The 500 MW Tengger Desert solar plant feeding the cloud base directly over four 110 kV lines, online May 2026, PUE below the 1.2 hub-node requirement.
- Gui'an New Area, Guizhou. A 40-foot container data center with four immersion-cooled racks and 80 kilowatts of IT load, assembled in a factory and connected to a substation in three months.
- Lingang, Shanghai. A 24-megawatt commercial underwater data center lowered into the East China Sea in May 2026: 1,950 tons, 192 racks, roughly 2,000 high-performance servers, cooled by seawater, powered by an offshore wind farm 500 meters away. Cost 1.6 billion yuan, about $226 million. Microsoft abandoned the same concept in 2024.
- Orbit. Twelve satellites launched in May 2025 carrying 5 POPS of in-orbit computing power, in regular experimental service since August 2026.
China versus US, side by side
| Dimension | China | United States |
|---|---|---|
| Data center market size, 2026 | ~$52 billion | Larger in absolute dollars; no single official figure |
| AI share of domestic market | 48.7% | Majority AI for new capacity, not separately measured |
| Market structure | Three state carriers plus provincial SOEs plus private AIDC | Concentrated among hyperscalers |
| Data center electricity, 2025 | 205.7 TWh (26.1% of world) | 312.6 TWh (39.7% of world) |
| Industrial power price | ~0.33 yuan/kWh in Inner Mongolia | $0.054 to $0.22/kWh depending on state |
| Ultrahigh-voltage lines | 46 lines, 62,000+ km | No UHV network at comparable scale |
| Large transformer lead time | Domestic supply | Above 160 weeks, Q1 2026 |
| Interconnection queue | Regulated allocation | Four to seven years |
| Capacity, 2025 to 2030 | Not published on a comparable basis | ~24 GW to 110 GW |
Frequently asked questions
How big is China's data center market in 2026?
362.1 billion yuan, roughly $52 billion, up 13.9% from 318 billion yuan in 2025. Source: China Industrial Research Institute via AskCI, January 2026.
How fast is China's AI data center segment growing?
30% year over year, reaching 176.3 billion yuan, about $25.3 billion, in 2026. That is more than double the overall market's growth rate.
What share of China's data center market is AI-specific?
48.7% by the end of 2026, up from 42.6% in 2025.
How much computing power does China have?
2,185 EFLOPS in FP16 as of June 2026, up 177% year over year, with a national utilization rate of 71.4%. That is intelligent compute capacity as measured by MIIT, not total compute.
How much electricity do China's data centers consume?
205.7 TWh in 2025, 26.1% of the global total, against 312.6 TWh for the United States. Source: Energy Institute, 2026 Statistical Review of World Energy.
Why is data center electricity cheaper in China?
Regulated tariffs rather than market-clearing prices, plus grid operators absorbing roughly 4 trillion yuan in debt. Inner Mongolia data centers pay about 0.33 yuan per kWh.
What PUE do new Chinese data centers have to meet?
Below 1.2 in the national hub nodes. The Zhongwei direct-supply project in Ningxia met it at commissioning.
How much does data center cooling cost in China?
Cold plate liquid cooling runs 4,000 to 5,000 yuan per kilowatt, about $550 to $690, against more than $1,380 in North America. Air cooling runs 2,000 to 3,000 yuan per kilowatt.
How many racks does China have?
Approaching 1 million units nationally by the end of 2026, per the China Industrial Research Institute.
Is China's data center build-out slowing because of chip export controls?
The market data does not show a slowdown: 13.9% overall growth, 30% in the AI segment, and 177% growth in intelligent compute capacity. Export controls constrained access to advanced Nvidia silicon, and Nvidia's share of China's AI chip market has fallen to single digits, but the capacity figures are growing anyway. Roughly 65% of compute available to Chinese entities is still projected to come from US-designed hardware in 2027, and about a third through remote cloud access.
Sources and update log
- China Industrial Research Institute via AskCI, January 2026: total market and AIDC forecasts, rack count.
- CAICT, Q1 2026: AI compute demand and supply growth rates.
- Ministry of Industry and Information Technology via Economic Daily, July 2026: intelligent compute capacity and utilization.
- Energy Institute, 2026 Statistical Review of World Energy, August 2026: data center electricity by region.
- Gartner, 2026: global data center electricity forecast.
- Institute for Progress, October 2026: H100-equivalent compute estimates and remote access share.
- TrendForce, August 2026: top nine cloud provider capex.
- Hansen Fluid, 2026: cooling capital cost per kilowatt.
- 15th Five-Year Plan, March 2026: national integrated computing network and treasury bond allocation.
- Alibaba and Tencent financial disclosures, 2026.
- PJM capacity auction results and independent market monitor attribution, 2024 to 2026.
- Company and state media reporting on the Ulanqab, Zhongwei, Gui'an and Lingang projects, 2026.
Update log. First published November 8, 2026. Figures will be revised when the next annual market report and the next Statistical Review are published. Percentages from third-party analysts are forecasts and change with market conditions.
The information provided on this page is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this page. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. Figures are cited as reported by the named source on the named date and may be revised upstream.
Limitations: forecasts are not results; single-firm estimates are labeled as such; currency conversions use the rate at the source's publication date; no figure on this page is derived from operational access to any facility.