AI,
to Scale.
How much compute doesthe world actually have?
From a single chip to a planet of data centers. Explore the places, power, and extraordinary growth behind the intelligence on your screen.
The cloud has
an address.
A chatbot answer arrives in a small box on your screen. The machinery behind it can fill warehouses, require new substations, and take years to build. We made this feature to connect those two worlds.
Announcements arrive in megawatts, billions of dollars, and unfamiliar chip names. Those numbers become more useful when you can place the facilities on a map, line them up, and see the hardware inside.
Understanding AI’s physical scale helps us ask better questions about access, cost, energy, and who gets to build what comes next.
We assembled public facility estimates, chip-shipment records, building annotations, hardware documentation, and published training budgets. Every visualization has a defined scope. The data is an evolving picture of the largest projects, with uncertainty and gaps—not an exact inventory of every computer on Earth. See how we assembled it.
Intelligence lives somewhere.
Explore 94 documented major AI facilities, including Tesla Cortex. This is a curated atlas, not a complete worldwide inventory. Select a location, inspect the evidence, or move from the map into a globe you can turn.
Dot area scales approximately with estimated compute, with a minimum visible size.Data: Epoch AI + editorial supplement, CC BY 4.0. This atlas prioritizes major AI projects;
a region with few dots may have substantial untracked capacity.
How complete is this map?
Epoch AI prioritizes large facilities and has stronger coverage in the United States. Its live research estimates that its source collection covers roughly 43% of global deployed AI compute as of October 7, 2026, with a wide 90% interval of 23–81%. That uncertainty matters. We include all 93 Epoch records plus 1 separately sourced Tesla campus. The coverage estimate applies to Epoch’s collection; it has not been recalculated for this addition. An empty part of this map does not mean an absence of AI. Coverage and limitations →
Whose facility is it?
The building operator, chip owner, and AI customer can be different organizations. Our inspector reports the source’s chip owner and users, retaining “likely” and “speculative” labels. A project name or partnership announcement alone does not establish who owns its chips. SpaceXAI is displayed as SpaceX / xAI, with the original source name preserved in downloads. Acquisition announcement →
AI that learns to move.
Tesla Cortex at Gigafactory Texas broadens the picture to vehicle and humanoid robot autonomy. Inspect Cortex 1 and Cortex 2 separately within the campus. Tesla reports both in production in its latest update; installed capacity, workload utilization, and future ramping remain distinct. Company-defined H100e and compute MW disclosures are shown separately from our comparable totals. Tesla’s latest report →
A much bigger world,
one quarter at a time.
Drag through the timeline. Compare physical chip count with the computing capacity those chips represent.
2.5× the shipment-based capacity
from four quarters earlier
Source: Epoch AI chip shipments. Estimated theoretical capacity of shipped chips, not a count of currently online GPUs. Incomplete quarters are excluded. Missing updates retain each designer’s last published value; retirements are not subtracted. Per-designer uncertainty ranges are available in the downloaded source.
Our best answer to “how much?”
The latest available series represents about 31.2M H100-equivalents from six tracked chip designers, through Q2 2026. Amazon’s last published value is from Q4 2025. This shipment-based proxy includes chips awaiting installation and does not subtract retirements. It supports a useful estimate of scale, rather than an exact count of working GPUs. Read the definitions →
And “how fast?”
Epoch AI’s historical analysis estimates roughly 3.4× growth per year in the compute represented by the stock of AI chips since 2022—about a doubling every seven months. That is a fitted historical trend, rather than a promise about the next seven months. Faster chips and more shipments both contribute. Explore the analysis →
One H100-equivalent is a reference amount of computing performance. A newer chip can represent several. Millions of H100-equivalents do not mean millions of physical H100 GPUs.
Line up the giants.
Move along the skyline. Switch the measurement. Then compare the actual mapped roof outlines at the same scale.
Put the real building outlines
next to each other.
Compare the roofs at one shared scale. Switch to site layout to see how far apart the buildings sit.
Roofs are grouped for comparison; their shapes and sizes are unchanged. Spaces between them do not represent distances on site.
Footprints: Epoch AI’s latest building annotations, including unfinished buildings (dashed). Approximate roof area calculated from those polygons; this is neither total campus land area nor total floor area. Footprint dates vary by site and do not change with the capacity-date selector. Click a tower to add its facility to this comparison.
From a chip
to a small industrial world.
Turn the model, pull it apart, and move through six layers. Each layer solves a different problem.
The campus
MWthe unit of electrical capacityMultiple buildings, substations, cooling systems, and backup equipment make a campus. This illustration explains the parts, rather than depicting a particular site.
Explore the sourceEach stage is fitted to the viewing window. The transition is conceptual, not a single physical scale or a literal bill of materials. Server and rack stages use different documented chip architectures. Building layouts and heights are illustrative.
More than arithmetic
H100-equivalents compare theoretical peak dense 8-bit operations per second. Practical performance also depends on memory, bandwidth, software, precision, and how the chips communicate. A capacity ratio is not automatically a training-speed ratio. Why the reference is imperfect →
Heat is part of the story
Computers require power delivery and cooling around them. The IEA describes cooling shares ranging from about 7% at efficient hyperscale facilities to over 30% at less-efficient enterprise sites. Location, weather, equipment, and design all affect the result. The components of electricity demand →
What can all this compute do?
Think of compute as a resource budget. The useful result depends on the model, data, and task you give it.
The same broad infrastructure can help teach a model, answer users, or run experiments. A megawatt rating alone cannot tell us how many videos it will generate, how many questions it will answer, or which scientific problem it will solve.
Teach a model
Training / learning patterns from dataMany connected accelerators repeatedly process examples and adjust a model’s parameters. Meta’s published Llama 3.1 budgets make one real training effort tangible in the experiment below.
Make it useful to people
Inference / running the trained modelServe conversations, coding requests, image generation, and other tasks. The workload changes the number of requests a system can handle. Amazon documents Project Rainier’s use for Claude training and inference across multiple data centers.
Explore more possibilities
Research / many experiments in parallelRun alternative training recipes, evaluate model behavior, and investigate scientific applications. More compute can enable more experimentation; good data, valid methods, and validation determine whether an experiment produces a useful result.
What does a training
budget feel like?
Meta reported 30.8M H100 GPU-hours for its 405B model. One GPU-hour means one GPU working for one hour. Divide that budget across a hypothetical team of GPUs.
An arithmetic illustration using Meta’s reported GPU-hour budgets, not a prediction of training time. It assumes perfect division of the recorded work; networking, model memory, scheduling, failures, and changing efficiency limit real scaling. These are hypothetical physical H100s, not the atlas’s H100-equivalent units. The world’s facilities cannot all be combined into this configuration.
If you had access to all of it…
You could allocate capacity to many independent jobs: training runs, services for users, and experiments. Access to the chips would still leave substantial work in assembling datasets, building software, connecting hardware, and operating the systems.
…it would not become one computer.
Large training jobs depend on tightly coordinated, high-bandwidth clusters. You cannot add every GPU on the map and assume the same speedup. Meta’s training paper describes the communication and reliability engineering needed even inside purpose-built clusters.
Put the power
in everyday terms.
Change utilization and overhead to see how capacity becomes an annual energy comparison.
A megawatt is capacity.
A year of power is energy.
PUE is total facility power divided by IT power. 1.20 means adding 20% for cooling and other overhead. These sliders set assumptions; they are not measurements of the selected site.
6.96 TWh per year under these assumptions.
Comparison uses EIA’s 2022 average of 10,791 kWh purchased per U.S. residential customer. This is an annual energy equivalent, not a claim that the facility supplies homes, runs at this load, or has this measured PUE. Capacity: Epoch AI.
The IEA estimated that all data centers accounted for about 1.5% of global electricity consumption in 2024. This includes storage, ordinary cloud computing, and other non-AI work.
Its 2025 report projected more than a doubling by 2030 in the Base Case. Alternative scenarios change the outcome; hardware efficiency, adoption, and grid constraints matter.
Every facility in the atlas.
Search the complete snapshot. Each row opens its location in the atlas. Operating capacity is separate from expansion plans.
| Facility / location | Chip owner | Status | H100-equivalents | IT capacity |
|---|---|---|---|---|
| Colossus 2Memphis, United States | SpaceX / xAI | Operating | 1.1M | 946 MW |
| Microsoft Fairwater AtlantaFayetteville, United States | Microsoft | Operating | 768.8K | 636 MW |
| Anthropic-Amazon New CarlisleNew Carlisle, United States | Amazon | Operating | 685.9K | 910 MW |
| Google Pryor (North)Pryor, United States | Operating | 636.7K | 368 MW | |
| DayOne NusajayaJohor Bahru, Malaysia | Not identified | Operating | 583.8K | 473 MW |
| Meta PrometheusNew Albany, United States | Meta | Operating | 536.3K | 471 MW |
| OpenAI Stargate AbileneAbilene, United States | Oracle | Operating | 509.3K | 421 MW |
| DayOne KempasKempas, Malaysia | Not identified | Operating | 480.6K | 421.7 MW |
| Microsoft Fairwater WisconsinMount Pleasant, United States | Microsoft | Operating | 445.7K | 369 MW |
| Google ColumbusColumbus, United States | Operating | 332K | 303 MW | |
| Google New AlbanyNew Albany, United States | Operating | 323.9K | 333 MW | |
| Google Mesa, United States | Operating | 289.5K | 183 MW | |
| Colossus 1Memphis, United States | SpaceX / xAI | Operating | 275.8K | 340 MW |
| Google Council Bluffs (East)Council Bluffs, United States | Operating | 255.7K | 237 MW | |
| CoreWeave Denton TXDenton, United States | CoreWeave | Operating | 252.7K | 262 MW |
| QTS Richmond 1Sandston, United States | Not identified | Operating | 242.5K | 238 MW |
| Meta RosemountRosemount, United States | Meta | Operating | 215.7K | 178 MW |
| Amazon Madison Mega SiteCanton, United States | Amazon | Operating | 214.3K | 284 MW |
| Google BristowBristow, United States | Google (unknown) | Operating | 209.9K | 238 MW |
| Microsoft GoodyearGoodyear, United States | Microsoft | Operating | 205.2K | 202 MW |
| Meta JeffersonvilleJeffersonville, United States | Meta | Operating | 203.1K | 178 MW |
| Google Kansas City EastKansas City, United States | Operating | 202.6K | 105 MW | |
| Meta TempleTemple, United States | Meta | Operating | 202.5K | 178 MW |
| AWS New AlbanyNew Albany, United States | Amazon | Operating | 201.1K | 213 MW |
| Google OmahaOmaha, United States | Operating | 198.6K | 237 MW | |
| Google LincolnLincoln, United States | Operating | 197.6K | 141 MW | |
| QTS Richmond 2Sandston, United States | Not identified | Operating | 182.9K | 180 MW |
| Google The DallesThe Dalles, United States | Operating | 175.3K | 154 MW | |
| QTS Richmond 3Sandston, United States | Not identified | Operating | 174.8K | 144 MW |
| Meta MontgomeryMontgomery, United States | Meta | Operating | 174.5K | 153 MW |
| Meta KunaKuna, United States | Meta | Operating | 173.2K | 152 MW |
| Meta CheyenneCheyenne, United States | Meta | Operating | 172.8K | 152 MW |
| Amazon RidgelandRidgeland, United States | Amazon | Operating | 171.3K | 228 MW |
| Meta HuntsvilleHuntsville, United States | Meta | Operating | 166.8K | 146 MW |
| Coreweave HeliosAfton, United States | CoreWeave | Operating | 159.6K | 132 MW |
| Microsoft Project OsmiumCumming, United States | Microsoft | Operating | 155.6K | 190 MW |
| Meta Eagle MountainEagle Mountain, United States | Meta | Operating | 151.4K | 133 MW |
| Google MidlothianMidlothian, United States | Operating | 143.6K | 103 MW | |
| Google Storey CountyClark, United States | Operating | 133.9K | 161 MW | |
| Google LancasterLancaster, United States | Operating | 130.4K | 137 MW | |
| Meta-QTS Hillsboro 2Hillsboro, United States | Meta (likely) | Operating | 123.8K | 180 MW |
| Google Waltham CrossCheshunt, United Kingdom | Operating | 123.8K | 88 MW | |
| VNET Bayin UlanqabUlanqab, China | VNET | Operating | 122.8K | 221 MW |
| Google ArcolaArcola, United States | Operating | 108.1K | 77 MW | |
| Google Red OakRed Oak, United States | Operating | 108.1K | 77 MW | |
| Google PapillionPapillion , United States | Operating | 100.6K | 237 MW | |
| Meta Gallatin, United States | Meta | Operating | 99.5K | 87 MW |
| Meta Los LunasLos Lunas, United States | Meta | Operating | 99.5K | 87 MW |
| CoreWeave Ellendale NDEllendale, United States | CoreWeave | Operating | 93.5K | 68 MW |
| Meta SarpySpringfield, United States | Meta | Operating | 92K | 80 MW |
| Huawei HoringerHohhot, China | Huawei | Operating | 89.4K | 184.3 MW |
| CoreWeave Chester VAChester, United States | CoreWeave | Operating | 88.4K | 82 MW |
| Oracle BatamBatam, Indonesia | Oracle (likely) | Operating | 84.4K | 72 MW |
| AWS BerwickBerwick, United States | Amazon | Operating | 81.4K | 108 MW |
| CoreWeave Marble NCMarble, United States | CoreWeave | Operating | 78.8K | 65 MW |
| Meta AikenAiken, United States | Meta | Operating | 78.3K | 69 MW |
| Huawei WuhuWuhu, China | Huawei | Operating | 76.3K | 157 MW |
| Core42 Lake MarinerBarker, United States | Core42 | Operating | 71.8K | 58 MW |
| Microsoft SAT40San Antonio, United States | Microsoft | Operating | 64.7K | 75 MW |
| STACK Infrastructure NVA02Manassas, United States | Not identified | Operating | 61.1K | 85 MW |
| Anthropic Lake MarinerBarker, United States | AI XPV Platform | Operating | 58.6K | 42 MW |
| Microsoft-Nebius New JerseyVineland, United States | Nebius | Operating | 56.8K | 50 MW |
| CoreWeave NorwayØvrebø, Norway | CoreWeave | Operating | 50.8K | 42 MW |
| Nebius MantsalaMantsala, Finland | Nebius | Operating | 48.5K | 75 MW |
| Southgate MelbourneMelbourne, Australia | Firmus | Operating | 46.5K | 41 MW |
| Stream PhoenixGoodyear, United States | Not identified | Operating | 46.1K | 64 MW |
| Microsoft SAT14San Antonio, United States | Microsoft | Operating | 45.2K | 47 MW |
| Vantage TX1San Antonio, United States | Not identified | Operating | 38.7K | 32 MW |
| CoreWeave Dalton 1 & 2Dalton, United States | CoreWeave | Operating | 32.2K | 28 MW |
| Alibaba ZhangbeiZhangbei, China | Alibaba | Operating | 31.9K | 169 MW |
| Start Campus Sines Data CampusSines, Portugal | Nscale | Operating | 31.8K | 33 MW |
| xAI QTS AtlantaAtlanta, United States | SpaceX / xAI | Operating | 11.5K | 17.3 MW |
| Nscale KeflavikReykjanesbær, Iceland | Nscale | Operating | 5.8K | 5.1 MW |
| Anthropic Barber LakeColorado City, United States | Cipher Mining | Future site | Not operating | Not operating |
| CoreWeave Lancaster Greenfield siteLancaster, United States | CoreWeave | Future site | Not operating | Not operating |
| CoreWeave Muskogee OKMuskogee, United States | CoreWeave | Future site | Not operating | Not operating |
| Crusoe Abilene ExpansionAbilene, United States | Microsoft | Future site | Not operating | Not operating |
| EcoDataCenter 2Borlänge, Sweden | Mistral AI | Future site | Not operating | Not operating |
| GoodnightClaude, United States | Google (speculative) | Future site | Not operating | Not operating |
| Google Cedar RapidsCedar Rapids, United States | Future site | Not operating | Not operating | |
| Google Fort WayneFort Wayne, United States | Future site | Not operating | Not operating | |
| Meta Bowling GreenBowling Green, United States | Meta | Future site | Not operating | Not operating |
| Meta HyperionHolly Ridge, United States | Meta | Future site | Not operating | Not operating |
| Microsoft Narvik NorwayBjerkvik, Norway | Nscale (likely) | Future site | Not operating | Not operating |
| OpenAI Stargate LordstownLordstown, United States | Softbank | Future site | Not operating | Not operating |
| OpenAI Stargate MichiganBenton, United States | Oracle | Future site | Not operating | Not operating |
| OpenAI Stargate MilamBurlington, United States | Softbank | Future site | Not operating | Not operating |
| OpenAI Stargate New MexicoSanta Teresa, United States | Oracle | Future site | Not operating | Not operating |
| OpenAI Stargate ShackelfordAbilene, United States | Oracle | Future site | Not operating | Not operating |
| OpenAI Stargate UAEAbu Dhabi, United Arab Emirates | G42 (likely) | Future site | Not operating | Not operating |
| OpenAI Stargate WisconsinPort Washington, United States | Oracle | Future site | Not operating | Not operating |
| QTS Cedar RapidsCedar Rapids, United States | Not identified | Future site | Not operating | Not operating |
| QTS Eagle MountainEagle Mountain, United States | Meta (speculative) | Future site | Not operating | Not operating |
| Tesla Cortex · Gigafactory TexasAustin, United States | Tesla | Operating | Unknown / non-comparable | Unknown / non-comparable |
Public information.
A traceable picture.
This feature is an independent editorial visualization by Superpower Daily. The base facility and chip estimates are credited to Epoch AI. Tesla Cortex is a separately attributed editorial addition from Tesla’s investor reports, with additional explanations grounded in primary hardware documentation, research papers, and energy statistics.
Data retrieved Oct 7, 2026. Facility download dated October 5, 2026; chip download dated October 1, 2026. This is a reviewed snapshot, not a live feed.
- Scope: 94 documented major facilities: 93 records from Epoch’s source snapshot (including 20 future sites) plus one Tesla campus with two separately described clusters. They span 12 countries. This is not a complete worldwide inventory. A record may describe a campus or a cluster of nearby buildings; counts are not counts of individual buildings.
- Estimates: many capacity values are inferred from permits, cooling equipment, imagery, and public statements. Rounded numbers make the uncertainty more visible. Source uncertainty and calculation workbooks are linked in the inspector.
- Time: Epoch’s current view uses the source CSV’s snapshot values. Tesla cluster disclosures are filtered by publication date and never backfilled into earlier views. Historical and projected views use the last recorded timeline event at or before the selected date. We do not interpolate missing values or fill earlier dates with future capacity.
- Tesla capacity: company-defined installed H100e and compute MW are not assumed equivalent to Epoch’s operating peak-FP8 H100-equivalents or IT MW. Comparable figures remain unknown, are omitted from totals and rankings, and are not inferred from power or GPU nameplate ratings. The campus point is approximate; cluster locations and roof outlines are not supplied.
- Global growth: cumulative chip shipment estimates start in 2022. Incomplete quarterly rows are excluded. Where a designer has no new quarter, its latest published cumulative estimate is retained and dated. We do not calculate a total confidence interval by adding per-designer intervals.
- Physical size: footprint polygons are the source’s latest annotated building roofs, including unfinished buildings. Approximate area uses a local latitude-adjusted projection. They do not represent complete campus boundaries, floor area, or roof outlines at every historical date.
- Comparisons: tower heights encode one metric on a shared linear scale. The 3D hardware is schematic and does not reproduce an actual campus. Electricity and GPU-hour examples are calculations with visible assumptions, not measured site performance.
- Reproducibility: source ZIPs, derived data, and download checksums are available here. Links and confidence labels are preserved. Updated source releases may revise both historical estimates and future schedules.
Facility estimates, locations, annotated building outlines, and expansion timelines. CC BY 4.0; snapshot retrieved October 7, 2026.
How satellite imagery, permits, cooling equipment, and disclosures support estimates; definitions and uncertainty.
Six designers; cumulative shipments from 2022. Incomplete quarterly rows excluded; each designer’s last published value is carried forward and dated.
H100-equivalent definition, estimates, and the distinction between delivered and deployed chips.
Historical trend in compute represented by AI chips; a fitted trend is not a forecast.
Eight GPUs per DGX H100 system. Hardware explorer server stage uses this architecture.
72 Blackwell GPUs and 36 Grace CPUs per liquid-cooled rack. A different architecture from an H100 server.
Reported training budgets: 1.46M, 7M, and 30.84M H100 GPU-hours for the 8B, 70B, and 405B models.
Training infrastructure, parallelism, and the practical constraints on distributing a training job.
All data centers: estimated 415 TWh in 2024; 945 TWh in its 2030 Base Case. These totals include non-AI computing.
2022 U.S. average: 10,791 kWh purchased per residential utility customer annually; used as a dated comparison baseline.
Documented example of a multi-data-center cluster used by Anthropic for training and inference.
September 2025 announcement: planned project capacity is distinct from operating capacity.
Official explanation of the Wisconsin campus and its AI infrastructure.
Training, inference, networking, and plans for Prometheus and Hyperion.
Latest Cortex cluster disclosures: company-defined installed capacity and production status, not metered usage or comparable Epoch IT power.
Dated H100e disclosures and Cortex 2 training activity. Early-ramp capacity includes expected capacity; Tesla’s conversion basis is not assumed equivalent to Epoch’s.
February 2, 2026 announcement. Display label SpaceX / xAI preserves the source owner SpaceXAI in downloads.
Geographic boundaries. Political boundaries are cartographic context, not editorial positions.
A few useful answers.
Does the atlas include Tesla and SpaceX / xAI?
Yes. Tesla Cortex at Gigafactory Texas is a separately sourced campus with Cortex 1 and Cortex 2 cluster disclosures. It supports vehicle and humanoid robot autonomy development. Its company-defined installed capacities are not added to Epoch’s comparable operating-compute or IT-power totals. Colossus 1, Colossus 2, and xAI QTS Atlanta appear under SpaceX / xAI. This is a collection of documented major facilities, not a worldwide census.
How much AI compute does the world have?
Our latest available shipment series represents about 31.2M H100-equivalents across six tracked chip designers, through Q2 2026, with Amazon’s last published value from Q4 2025. This is an estimated shipment-based proxy, not an exact census of operating GPUs; it does not subtract retirements. Separately, 94 documented facilities are mapped; the 93 Epoch records with comparable estimates contain about 14.4M H100-equivalents as of Oct 7, 2026.
How fast is AI computing capacity growing?
Epoch AI’s historical analysis estimates that the compute represented by the stock of AI chips has grown about 3.4 times per year since 2022, doubling roughly every seven months. This describes a historical fitted trend. Our chart shows the underlying shipment estimates; continuing that trend is not guaranteed.
Which is the largest AI data center?
In this Oct 7, 2026 snapshot, Colossus 2 is the largest tracked operating facility by estimated compute, at about 1.1M H100-equivalents. Rankings change when you switch to IT power or include projected expansion. The dataset is not a census of every facility.
What is an H100-equivalent?
It is a comparison unit for theoretical peak dense 8-bit computing performance, using one NVIDIA H100 as the reference. A newer chip can count as multiple H100-equivalents. It is not the physical GPU count, a benchmark score, or a promise of real-world training speed.
How much electricity do AI data centers use?
The atlas reports estimated IT power capacity in megawatts, not metered electricity consumption. Actual consumption depends on utilization and cooling overhead. The IEA estimated 415 TWh for all data centers in 2024 and projected 945 TWh in its 2030 Base Case; those figures also include non-AI workloads.
Could we combine all the world’s AI compute into one computer?
The capacity can support many independent jobs, but it cannot simply be pooled into one giant training run. Chips need compatible software, sufficient memory, fast networking, and coordinated access. Data centers across continents do not have the same communication characteristics as a tightly connected training cluster.
Are the buildings and 3D models accurate?
Footprint comparisons use Epoch AI’s annotated building polygons at a shared map scale. They include unfinished annotated roofs and are not complete campus land area or floor area. The rotatable chip-to-campus models are original schematic illustrations; their layouts and building heights are illustrative.
Can I reuse the data or cite this page?
Yes. Epoch’s facility and chip datasets and our separately labeled editorial supplement are distributed under CC BY 4.0. Cited company documents retain their own terms. Download the source files and transformation manifest below, cite the source and this page’s retrieval date, and preserve the distinctions between estimates, operating capacity, and projections.
A clearer view of
the world we’re building.
Share a facility, cite the snapshot, or come back to the sources as the next generation of infrastructure takes shape.
Another question.
Another world to explore.
AI, at Work.
Your job is a world of tasks. Where does AI fit?
Explore a profession, remix a workday and put your judgment to work.
Explore the feature
AI, Behind the Shot.
How do you turn an idea into a film that holds together?
Move the camera. Choose your tools. Build your next sequence.
Explore the feature