China AI Infrastructure Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell, AI & Machine Learning | Last updated: September 6, 2026 | License: CC BY 4.0
China’s installed intelligent computing capacity reached 2,185 EFLOPS at FP16 precision on June 30, 2026, up 177% year over year, while the nationwide utilization rate of computing facilities stood at 71.4%. Applying that rate to the installed base leaves roughly 625 EFLOPS unused, a stranded block larger than China’s entire national capacity twelve months earlier.
Quick Answer
China’s Ministry of Industry and Information Technology reported 2,185 EFLOPS (FP16) of intelligent computing capacity as of June 30, 2026, with computing facilities running at 71.4% utilization. According to Axis Intelligence Research, that yields a China Effective Compute Index (CECI™) reading of 1,560.1 EFLOPS, equivalent to approximately 1.58 million NVIDIA H100 accelerators at dense FP16 throughput, with 624.9 EFLOPS sitting idle.
Key Findings
- According to China’s Ministry of Industry and Information Technology, national intelligent computing capacity reached 2,185 EFLOPS (FP16) on June 30, 2026, a 177% increase year over year.
- According to Axis Intelligence Research, the CECI™ reading for June 30, 2026 is 1,560.1 EFLOPS of effective deployed compute, calculated as 2,185 EFLOPS multiplied by MIIT’s reported 71.4% facility utilization rate.
- According to Axis Intelligence Research, China’s 624.9 EFLOPS of idle intelligent computing capacity at mid-2026 equals 79.3% of the country’s entire installed base of 788 EFLOPS one year earlier.
- According to MIIT, eastern China held 55.9% of national intelligent computing capacity at end-June 2026, western China 32.6%, central China 10.6% and the northeast 0.9%.
- According to Axis Intelligence Research, quarterly net additions to China’s intelligent computing base fell from roughly 401 EFLOPS per quarter in the second half of 2025 to 292 EFLOPS in Q1 2026 and 303 EFLOPS in Q2 2026, meaning the headline 177% annual growth rate conceals a sequential plateau.
How Much AI Computing Power Does China Have in 2026?
The single authoritative number is 2,185 EFLOPS of intelligent computing capacity as of June 30, 2026, disclosed by MIIT spokesperson Xie Cun at a State Council Information Office press conference on July 20, 2026 and published by Xinhua and the State Council Information Office.
Four qualifiers travel with that number, and anyone citing it without them will be citing it wrong.
It is measured at FP16 precision, the half-precision format used by AI accelerator cards. That is a substantially more permissive basis than the double-precision LINPACK benchmark behind the TOP500 supercomputer list, and it is not interchangeable with general-purpose computing capacity. It is nameplate capacity, not delivered work. It is nationwide, spanning telecom operator facilities, hyperscaler data centres, provincial intelligent computing centres and university clusters. And it sits on the far side of a definitional break in MIIT’s own published series, discussed below.
China Intelligent Computing Capacity by Reporting Date
| As-of date | Intelligent computing capacity (EFLOPS, FP16) | Net addition vs prior reading | Announced by | Source |
|---|---|---|---|---|
| 2025-06-30 | 788 | not applicable | Xiong Jijun, MIIT deputy minister | People’s Daily via Henan provincial government portal |
| 2025-12-31 | 1,590 | +802 | Zhang Yunming, MIIT vice minister | State Council Information Office, Jan 22, 2026 |
| 2026-03-31 | 1,882 | +292 | Zhang Yunming, MIIT vice minister | SCIO press conference, Apr 21, 2026 |
| 2026-06-30 | 2,185 | +303 | Xie Cun, MIIT spokesperson | SCIO / Xinhua, Jul 20, 2026 |
Source: Ministry of Industry and Information Technology disclosures at State Council Information Office press conferences. Net additions calculated by Axis Intelligence Research.
The end-2025 figure of 1,590 EFLOPS was announced alongside the disclosure that China hosted more than 6,000 AI enterprises and a core AI industry expected to exceed 1.2 trillion yuan for the year. Our China AI statistics page carries the market-size and adoption layer; this page carries the physical compute layer.
Sarah Mitchell: The 177% headline is real and it is also the least interesting thing in the release. Run the increments and the picture inverts. China added roughly 401 EFLOPS per quarter through the second half of 2025, then 292 in Q1 2026 and 303 in Q2. Quarter-on-quarter growth went 42.0%, then 18.4%, then 16.1%. Capacity is still climbing, but the second derivative turned over two quarters ago. When a build-out decelerates while the annual comparison still looks vertical, the annual comparison is measuring last year.
What Is the China Effective Compute Index (CECI™)?
CECI™ stands for the China Effective Compute Index. It is an Axis Intelligence Research metric that converts China’s reported nameplate intelligent computing capacity into an estimate of the compute actually in productive use, by applying MIIT’s own published facility utilization rate to MIIT’s own published capacity figure.
Nameplate exaflops are a construction statistic. They tell you what was installed, not what is running. CECI™ exists because those two quantities have diverged in China, and because the divergence is measurable from official disclosures without any modelling assumption of ours.
Formula and Inputs
CECI™ = Installed intelligent computing capacity (EFLOPS, FP16) × Nationwide computing facility utilization rate
CECI™ (2026-06-30) = 2,185 EFLOPS × 0.714 = 1,560.1 EFLOPS
Idle overhang = 2,185 − 1,560.1 = 624.9 EFLOPS
Both inputs come from a single MIIT disclosure dated July 20, 2026, which is what makes the calculation internally consistent: capacity and utilization are measured by the same authority, on the same date, across the same facility population.
CECI™ Reading, June 30, 2026
| Component | Value | Unit | Basis |
|---|---|---|---|
| Installed intelligent computing capacity | 2,185 | EFLOPS (FP16) | MIIT, reported |
| Nationwide facility utilization rate | 71.4 | percent | MIIT, reported |
| CECI™ effective compute | 1,560.1 | EFLOPS (FP16) | Axis calculated |
| Idle overhang | 624.9 | EFLOPS (FP16) | Axis calculated |
| Idle overhang as share of June 2025 installed base | 79.3 | percent | Axis calculated |
Source: MIIT via State Council Information Office, July 20, 2026; Axis Intelligence Research calculation.
Scope of the metric
MIIT publishes the 71.4% utilization figure for computing facilities nationwide, a population that includes general-purpose and storage-oriented capacity as well as AI accelerators. CECI™ therefore applies an all-facility utilization rate to the intelligent-compute subset. If AI accelerator utilization runs above the national average, CECI™ understates effective compute; if intelligent computing centres built ahead of demand drag below the average, CECI™ overstates it. MIIT does not publish a utilization rate specific to intelligent computing, and Axis Intelligence Research does not estimate one. The reading is published as a single, reproducible arithmetic operation on two official numbers, and every future reading will use the same two series.
This is a baseline reading. CECI™ is being published for the first time with the June 30, 2026 data point, and carries no history against which it can be described as high or low.
How Many H100-Equivalents Is China’s AI Compute Worth?
Exaflops are hard to reason about. Accelerator counts are not. Because MIIT reports at FP16, its figures can be converted directly into equivalents of the most widely benchmarked AI accelerator in the world.
According to NVIDIA’s H100 Tensor Core GPU datasheet, the H100 SXM delivers 1,979 TFLOPS of FP16 Tensor Core throughput with structured sparsity, which corresponds to 989.5 TFLOPS dense. Axis Intelligence Research uses the dense figure, because dense throughput is the conservative and more comparable basis for a fleet-level conversion.
H100-equivalents = EFLOPS × 1,000 ÷ 0.9895 PFLOPS per H100 (dense FP16)
China’s Intelligent Compute in H100-Equivalents, June 30, 2026
| Measure | EFLOPS (FP16) | H100-equivalents (dense FP16) |
|---|---|---|
| Installed nameplate capacity | 2,185 | 2,208,186 |
| CECI™ effective compute | 1,560.1 | 1,576,645 |
| Idle overhang | 624.9 | 631,541 |
Source: MIIT capacity and utilization data (July 20, 2026); NVIDIA H100 datasheet specifications; Axis Intelligence Research calculation.
According to Axis Intelligence Research, China’s installed intelligent computing base at mid-2026 is arithmetically equivalent to approximately 2.21 million NVIDIA H100 accelerators running dense FP16, of which roughly 631,541 H100-equivalents are not in productive use.
The conversion is an arithmetic equivalence, not a hardware census. China’s fleet is a heterogeneous mix of domestic accelerators and legacy imported silicon, and no accelerator-count disclosure exists at national level. What the conversion does is translate an unfamiliar unit into a familiar one at a fixed, published exchange rate that any reader can recompute. Vendor share within that fleet is covered on our AI chip market share page.
Sarah Mitchell: Six hundred and thirty thousand H100-equivalents of idle silicon is not a rounding error. It is more accelerator capacity than most countries will install this decade, powered up and waiting for workloads. The instinct is to read this as waste. The more useful reading is that China has decoupled its compute build from its compute demand deliberately: capacity first, absorption later, on the theory that a chip you cannot buy in 2028 is worth more idle in 2026 than efficient in a spreadsheet. Whether that theory pays depends entirely on whether the idle fraction shrinks. It is the number to watch each quarter, and it is the reason CECI™ exists.
Where Is China’s AI Computing Power Located?
At the July 2026 press conference, Xie Cun gave the regional split of the 2,185 EFLOPS: eastern China 55.9%, western China 32.6%, central China 10.6% and the northeast 0.9%. Axis Intelligence Research has converted those shares into absolute capacity and into CECI™ effective compute at the national utilization rate.
China Intelligent Computing Capacity by Region, June 30, 2026
| Region | Share of national total (%) | Installed capacity (EFLOPS, FP16) | CECI™ effective compute (EFLOPS) |
|---|---|---|---|
| Eastern | 55.9 | 1,221.4 | 872.1 |
| Western | 32.6 | 712.3 | 508.6 |
| Central | 10.6 | 231.6 | 165.4 |
| Northeastern | 0.9 | 19.7 | 14.0 |
| National | 100.0 | 2,185.0 | 1,560.1 |
Source: MIIT regional shares via State Council Information Office, July 20, 2026; Axis Intelligence Research calculation. Regional CECI™ applies the national 71.4% utilization rate uniformly; MIIT does not publish utilization by region.
Here is the finding that neither Chinese nor Western coverage has drawn out. The Eastern Data, Western Computing programme, launched in February 2022, exists to move computational load from the coastal provinces to the renewable-rich west. Four years in, the west holds 32.6% of national intelligent computing capacity.
Now compare that against physical hub build-out. Peer-reviewed analysis published in Frontiers in Energy Research in April 2026, drawing on National Data Administration figures, tabulates IT load capacity across the eight national computing hubs at 6,750 MW eastern and 3,250 MW western as of June 2024. The western share of hub IT load capacity is 32.5%.
According to Axis Intelligence Research, the western share of national intelligent computing capacity (32.6%, mid-2026) and the western share of national-hub IT load capacity (32.5%, mid-2024) are effectively identical. Two independent measurements, two years apart, of two different quantities, land within a tenth of a percentage point. The programme has moved roughly a third of the load west and held there.
National Computing Hub Performance, East Versus West
| Metric | Eastern hubs | Western hubs | Source |
|---|---|---|---|
| IT load capacity (MW) | 6,750 | 3,250 | Frontiers in Energy Research, Apr 2026 (NDA data, Jun 2024) |
| Average PUE | 1.35 | 1.11 | Frontiers in Energy Research, Apr 2026 |
| Renewable integration rate (%) | 28 | 48 | Frontiers in Energy Research, Apr 2026 |
| Net carbon intensity (kg CO2/kWh) | 0.62 | 0.47 | Frontiers in Energy Research, Apr 2026 |
| Five-year TCO per kW (USD) | 2,880 | 1,865 | Frontiers in Energy Research, Apr 2026 |
Source: Wang et al., Frontiers in Energy Research 14:1774740 (April 29, 2026), Table 2 and Section 4.4, drawing on National Data Administration reporting.
The economics point west and the workloads stay east. The same peer-reviewed analysis identifies why: inter-hub network latency runs 15 to 22 milliseconds, adequate for batch training and cold storage but not for interactive inference or transaction processing requiring sub-5-millisecond response. The authors estimate delay-tolerant work at 40% to 50% of total data centre workload. That ceiling, not construction capacity or electricity price, is what has pinned the western share near one third.
Sarah Mitchell: A 35% total-cost-of-ownership advantage and a 0.24-point PUE advantage should have pulled far more than a third of the load west by now. They did not, because latency is physics and cost is policy, and physics does not respond to a five-year plan. The interesting consequence is architectural: if western hubs are structurally limited to training and batch work while eastern hubs carry inference, then China’s two compute regions are not substitutes converging on a national average. They are specialised tiers, and reporting them as one blended national number, which is exactly what the 2,185 EFLOPS figure does, hides the split that actually determines what the capacity can be used for.
Why Do China’s AI Compute Statistics Contradict Each Other?
Three widely quoted figures describe China’s 2024 intelligent computing capacity, and they differ by a factor of eight. All three are accurate. They measure different things.
China 2024 Intelligent Computing Capacity: Three Reported Bases
| Reported value | Measurement basis | Issuing organization | What it counts |
|---|---|---|---|
| 90 EFLOPS | Share of total national computing capacity of 280 EFLOPS | National Data Administration (Apr 2025) | Intelligent compute as ~32% of an all-computing aggregate |
| 725.3 EFLOPS | FP16 intelligent accelerator card performance | IDC and Inspur Information, 2025 assessment report | Nameplate accelerator throughput at half precision |
| 1,590 EFLOPS | FP16 intelligent computing scale, end-2025 | MIIT (Jan 2026) | Nameplate accelerator throughput, one year later |
Source: National Data Administration figures reported via People’s Daily, August 2025; IDC and Inspur Information 2025 China AI computing power assessment report, cited in Xinhua, January 2026; MIIT via State Council Information Office, January 22, 2026.
The 90 EFLOPS and 725.3 EFLOPS figures are not in conflict, because the first is a subset of an aggregate computed on one basis and the second is a nameplate accelerator throughput computed on another. Reporting that treats them as a contradiction, or that averages them, produces a number describing nothing.
Axis Intelligence Research declines to publish a growth rate spanning the end-2024 and mid-2025 MIIT readings. The 90 EFLOPS end-2024 figure and the 788 EFLOPS June 2025 figure appeared in the same People’s Daily report, attributed to the same ministry, six months apart. An 8.8-fold increase in two quarters is not a physical construction rate; it is a change of measurement basis. Any growth rate computed across that break would be an artefact. This calculation is published as a retracted row in the accompanying dataset with the reasoning attached, rather than silently omitted.
According to Axis Intelligence Research, the earliest date from which China’s intelligent computing series can be treated as internally consistent is June 30, 2025. Every rate in this analysis begins there.
How Fast Is China Adding AI Compute Capacity?
Quarterly Net Additions to China’s Intelligent Computing Base
| Period | Net addition (EFLOPS, FP16) | Per-quarter average | Quarter-on-quarter growth (%) |
|---|---|---|---|
| H2 2025 (Jul-Dec) | +802 | 401.0 | 42.0 (implied quarterly) |
| Q1 2026 (Jan-Mar) | +292 | 292.0 | 18.4 |
| Q2 2026 (Apr-Jun) | +303 | 303.0 | 16.1 |
Source: MIIT capacity disclosures. Net additions and growth rates calculated by Axis Intelligence Research. The H2 2025 quarterly figure is the geometric implied rate across two quarters, as MIIT did not publish a Q3 2025 reading.
According to Axis Intelligence Research, China’s intelligent computing build-out has settled into a run rate between 292 and 303 EFLOPS of net additions per quarter, down from an implied 401 EFLOPS per quarter in the second half of 2025. Absolute additions in Q1 and Q2 2026 were close to identical, which is what a supply-constrained plateau looks like rather than a demand-driven slowdown.
That plateau also puts China well ahead of the forecast most widely used to model it. IDC and Inspur Information projected 1,460.3 EFLOPS for full-year 2026. Actual capacity at the halfway point was 2,185 EFLOPS, 49.6% above the full-year forecast, and 78.5% of the way to the same forecast’s 2028 figure of 2,781.9 EFLOPS. Forecasts built on the 2024 trajectory have been overtaken by roughly two and a half years.
Cluster and Facility Build-Out
| Metric | Value | As-of date | Source |
|---|---|---|---|
| Intelligent computing facilities with 10,000+ AI accelerators | 42 | 2025-12-31 | MIIT via Xinhua |
| Intelligent computing facilities with 10,000+ AI accelerators | 52 | 2026-06-30 | MIIT via CGTN |
| Standard racks in in-use computing centres (millions) | 14.45 | 2026-03-31 | MIIT |
| Major computing transmission channels around national hubs | 70+ | 2026-06-30 | MIIT |
| Network performance improvement between clusters (%) | ~10 | 2026-06-30 | MIIT |
Source: CGTN reporting of the July 20, 2026 State Council Information Office press conference; MIIT Q1 2026 press conference, April 21, 2026.
According to Axis Intelligence Research, China added 10 new ten-thousand-accelerator clusters in the first half of 2026, taking the national count from 42 to 52. Those facilities are the training-capable tier. A ten-thousand-accelerator cluster is the minimum practical unit for frontier pretraining runs, which makes the cluster count a better proxy for frontier capability than the aggregate exaflops number, and it is growing at 24% per half-year rather than 177% per year.
What Does China’s AI Infrastructure Cost in Electricity?
Chinese data centres consumed an estimated 260 TWh of electricity in 2024, up from 161 TWh in 2018, representing 2.67% of national electricity consumption. The International Energy Agency’s Energy and AI analysis projects Chinese data centre consumption reaching 277 TWh by 2030 under conservative efficiency assumptions, while the Frontiers in Energy Research baseline scenario projects 500 TWh and its high-growth scenario 600 TWh.
China Data Centre Electricity Consumption and Intensity
| Year | Data centre consumption (TWh) | Share of national electricity (%) | Average PUE |
|---|---|---|---|
| 2018 | 161 | 2.35 | 1.70 |
| 2020 | 190 | 2.53 | 1.60 |
| 2022 | 220 | 2.55 | 1.50 |
| 2024 (est.) | 260 | 2.67 | 1.40 |
| 2026 (proj.) | 340 | 3.12 | 1.32 |
| 2030 (baseline proj.) | 500 | 3.85 | 1.25 |
| 2030 (IEA conservative proj.) | 277 | 2.31 | 1.20 |
Source: Wang et al., Frontiers in Energy Research 14:1774740 (April 29, 2026), Table 1, compiled from National Bureau of Statistics, China Electricity Council, Greenpeace East Asia and IEA data.
The share of national electricity has stayed remarkably flat, between 2.35% and 2.67% across six years, and the reason is not efficiency. The paper’s decomposition attributes the stability to two large offsetting forces: workload growth pushed the ratio up 2.15 percentage points while China’s overall electricity system expansion pulled it down 1.62 points, with PUE improvement contributing only 0.21 points. The denominator did most of the work. Comparative figures for other markets sit on our AI data center energy consumption statistics page and in the data center electricity demand tracker.
Sarah Mitchell: Read the 625 EFLOPS idle figure next to the 1.40 average PUE and a cost picture emerges that neither number shows alone. Idle accelerators still draw power, still need cooling, still amortise. Whatever fraction of China’s stranded compute is powered and provisioned rather than dark is being paid for at a facility overhead of 40 cents on every dollar of IT load. That is the real price of building ahead of demand, and it does not appear in any capacity statistic.
Is China’s Computing Power Network Actually Connected?
The stated ambition of the national integrated computing network is that compute can be dispatched across provinces the way electricity is dispatched across a grid. Progress toward that is now partially quantified.
At the World Artificial Intelligence Conference in July 2026, academician Gao Wen reported that the first phase of the national computing network, resource aggregation, is roughly 70% complete, with about 70% of national compute resources registered in a continuously updated unified database made available to the National Data Administration. In February 2026 MIIT set out a “1+M+N” node architecture of one national computing internet service node, M regional nodes and N industry nodes. In April 2026 the computing power network was formally added to China’s list of strategic infrastructure networks.
According to Axis Intelligence Research, the gap between 70% resource visibility and genuine cross-provincial dispatch is the most consequential unmeasured variable in Chinese AI infrastructure. Visibility into a facility is not the ability to schedule work on it. Until MIIT or the National Data Administration publishes a cross-provincial dispatch volume, the 71.4% utilization rate and the 624.9 EFLOPS overhang should be read as a national aggregate of largely locally scheduled facilities, not as evidence about the performance of a functioning compute grid.
Methodology
Collection. Every figure in this analysis was retrieved from a primary or peer-reviewed source during production on September 5 and 6, 2026, and logged with its URL and retrieval date in the accompanying CSV. Capacity, utilization, regional distribution, cluster counts and network figures come from Ministry of Industry and Information Technology disclosures made at State Council Information Office press conferences and published through Xinhua, the State Council Information Office English portal and CGTN. Hub-level capacity, PUE, renewable integration, carbon intensity and total cost of ownership figures come from Wang, Lei, Zhang, Jia and Guo, “Analysis of China’s power development and the impact of data center energy consumption,” Frontiers in Energy Research 14:1774740, published April 29, 2026 under CC BY 4.0. Accelerator throughput specifications come from NVIDIA’s H100 Tensor Core GPU datasheet. No figure in this article was taken from a secondary aggregator, and no aggregator is linked.
CECI™ formula. CECI™ = installed intelligent computing capacity (EFLOPS, FP16) × nationwide computing facility utilization rate. Both inputs are drawn from the same MIIT disclosure dated July 20, 2026. Idle overhang is the arithmetic residual. Regional CECI™ values apply the national utilization rate uniformly across regions, because MIIT does not publish regional utilization. Scope is stated in full in the CECI™ section above.
H100-equivalent conversion. EFLOPS × 1,000 ÷ 0.9895, where 0.9895 PFLOPS is the dense FP16 Tensor Core throughput of the NVIDIA H100 SXM, derived from the datasheet figure of 1,979 TFLOPS with structured sparsity. The dense basis is used because it is the conservative choice and because MIIT’s series is a nameplate throughput measure rather than a sparsity-optimised one. The conversion is an arithmetic equivalence and is not an estimate of installed accelerator units.
Quarterly additions. Net additions are simple differences between consecutive published MIIT readings. The H2 2025 per-quarter figure is the geometric implied rate across two quarters, because MIIT published no Q3 2025 reading. Quarter-on-quarter percentages are computed on consecutive readings only.
Declined calculations. Axis Intelligence Research declines to publish any growth rate spanning the end-2024 and mid-2025 MIIT readings, on the grounds that the 90 EFLOPS and 788 EFLOPS figures rest on different measurement bases. This declined calculation is published as a retracted row in the dataset with its reasoning, rather than omitted.
Scope of the underlying data. MIIT reports at FP16 nameplate precision, which produces higher values than double-precision benchmarks and is not comparable to TOP500 rankings. Facility utilization is reported for all computing facilities rather than for intelligent computing specifically. No official accelerator unit count exists at national level. Cross-provincial compute dispatch volume is not published.
About This Dataset
Contents. The accompanying CSV covers China’s intelligent computing capacity, utilization, regional distribution, cluster and rack counts, national computing hub performance and data centre electricity consumption, spanning 2018 through mid-2026, plus every Axis-calculated figure appearing in this article with its formula reference.
Coverage. Temporal: 2018 to 2026-06-30. Spatial: China, with eastern, central, western and northeastern regional breakdowns and eight national computing hub locations.
Structure. One row per observation. Every row carries metric, value, unit, as-of date, source organization, source document, source URL, retrieval date, primary-source flag, Axis-calculated flag and a method note for calculated rows.
License. Creative Commons Attribution 4.0 International (CC BY 4.0).
Citation. Axis Intelligence Research, China AI Infrastructure Statistics 2026, 2026.
How to Cite This Research
APA: Axis Intelligence Research, & Mitchell, S. (2026, September 6). China AI infrastructure statistics 2026: 2,185 EFLOPS installed, 625 EFLOPS idle. Axis Intelligence Research. https://axis-intelligence.com/china-ai-infrastructure-statistics/
MLA: Axis Intelligence Research, and Sarah Mitchell. “China AI Infrastructure Statistics 2026: 2,185 EFLOPS Installed, 625 EFLOPS Idle.” Axis Intelligence Research, 6 Sept. 2026, axis-intelligence.com/china-ai-infrastructure-statistics/.
Chicago: Axis Intelligence Research, and Sarah Mitchell. “China AI Infrastructure Statistics 2026: 2,185 EFLOPS Installed, 625 EFLOPS Idle.” Axis Intelligence Research, September 6, 2026. https://axis-intelligence.com/china-ai-infrastructure-statistics/.
CECI™ attribution: “Axis Intelligence Research, China Effective Compute Index (CECI™), 2026, axis-intelligence.com/china-ai-infrastructure-statistics/.”
Frequently Asked Questions
Is China’s 2,185 EFLOPS figure comparable to US AI compute totals?
No, and the incompatibility is structural rather than a matter of adjustment. MIIT publishes nameplate FP16 accelerator throughput aggregated across all facility types nationwide. The United States publishes no equivalent national aggregate; US compute is estimated bottom-up from accelerator shipments, hyperscaler capital expenditure disclosures and data centre capacity, each on its own basis. Any direct comparison silently assumes that a Chinese state statistical aggregate and a Western analyst reconstruction measure the same quantity. They do not. Comparisons of frontier training capability are better anchored on the count of ten-thousand-accelerator clusters, which China reports at 52 as of June 30, 2026.
Does FP16 reporting inflate China’s computing power numbers?
It raises them substantially relative to double-precision benchmarks, which is the correct choice for measuring AI capacity and the wrong one for comparing against supercomputer rankings. FP16 is the working precision for AI training and inference, so an AI capacity statistic reported at FP16 is measuring the right thing. The error is comparing that number to TOP500 LINPACK results, which are FP64. Whether the sparsity-adjusted or dense FP16 basis underlies MIIT’s series is not disclosed, which is why Axis Intelligence Research uses the conservative dense figure for accelerator conversions.
Why does so much of China’s AI computing capacity sit unused?
MIIT reports 71.4% utilization across computing facilities nationwide as of June 30, 2026, leaving an estimated 624.9 EFLOPS idle by the CECI™ calculation. Three mechanics are documented in the source material. Provincial and municipal intelligent computing centres were commissioned on policy timelines rather than demand signals. Cross-provincial dispatch is incomplete, with roughly 70% of national resources visible in the unified database but no published dispatch volume, so capacity in one province cannot readily serve demand in another. And western hubs, which hold 32.6% of capacity, carry 15 to 22 millisecond latency to eastern demand centres, which excludes them from interactive workloads regardless of availability.
Has Eastern Data, Western Computing actually shifted workloads west?
Partially, and it has plateaued near one third. Western China held 32.6% of national intelligent computing capacity at end-June 2026. Western hubs held 32.5% of national-hub IT load capacity in mid-2024. Two years apart, on two different measures, the western share is the same. The binding constraint is latency rather than construction or cost: western hubs run at 1.11 average PUE against 1.35 in the east and roughly 35% lower five-year total cost of ownership, yet remain restricted to delay-tolerant work estimated at 40% to 50% of total workload.
How many ten-thousand-accelerator AI clusters does China operate?
Fifty-two intelligent computing facilities equipped with more than 10,000 AI accelerators each were in operation as of June 30, 2026, up from 42 at the end of 2025. That is the tier capable of frontier pretraining runs, and its 24% growth over the half-year is a materially different trajectory from the 177% annual growth in aggregate exaflops.
What is the CECI™ metric and how is it calculated?
The China Effective Compute Index (CECI™) is an Axis Intelligence Research metric equal to installed intelligent computing capacity in EFLOPS at FP16 multiplied by the nationwide computing facility utilization rate, both as published by MIIT for the same date. The June 30, 2026 baseline reading is 1,560.1 EFLOPS, from 2,185 EFLOPS at 71.4% utilization. It exists because nameplate capacity is a construction statistic and the gap between installed and used compute in China is now large enough to change the conclusions drawn from the headline number.
Which single figure should a journalist cite for China’s AI computing capacity?
For installed capacity, 2,185 EFLOPS at FP16 as of June 30, 2026, attributed to China’s Ministry of Industry and Information Technology, with the FP16 basis stated. For capacity actually in use, 1,560.1 EFLOPS, attributed to Axis Intelligence Research’s China Effective Compute Index. Citing the first without the FP16 qualifier invites a false comparison against supercomputer benchmarks; citing it without the utilization rate overstates compute actually in use by roughly 40%.
Is China’s AI infrastructure build-out slowing down?
Sequentially, yes; annually, no. Net additions ran an implied 401 EFLOPS per quarter through the second half of 2025, then 292 EFLOPS in Q1 2026 and 303 EFLOPS in Q2 2026. Quarter-on-quarter growth fell from 42.0% to 18.4% to 16.1%. The year-over-year figure remains 177% because it is still measuring against the pre-plateau base. Two consecutive quarters of near-identical absolute additions is the pattern of a supply-constrained run rate rather than a demand collapse.
How much electricity does China’s AI infrastructure consume?
Chinese data centres consumed an estimated 260 TWh in 2024, or 2.67% of national electricity, at an average PUE of 1.40. Projections for 2030 range from 277 TWh under the IEA’s conservative efficiency case to 500 TWh in the baseline case and 600 TWh under rapid AI adoption. Note that this is total data centre consumption, not AI-specific: no official series separates intelligent computing electricity draw from general data centre load in China.
Related Axis Intelligence Research: China AI statistics · AI data center statistics · AI chip market share · China AI Capex Statistics · DeepSeek statistics · China AI Chip Statistics · China Semiconductor Statistics · China Memory Chip Production ·