There are roughly 12,000+ data centers on Earth, half of them in the United States. Four companies — Amazon, Microsoft, Google, and Meta — are on track to spend $370 billion on them this year. Below is a full map, a country-by-country ranking, a state-by-state breakdown, an LLM-by-LLM tie-up, and the water, power, and square-footage reality of the AI build-out.
Snapshots, not forecasts. Every number on this page links to its source at the bottom.
Hover any country for its data-center count. The choropleth uses log-scaled color so a 5-facility country doesn't vanish next to the U.S.
If you stack the next nine countries against the US, you still don't reach the US count.
Hyperscale facilities (≥10,000 sq ft, >1 MW) account for 44% of global capacity, up from 20% in 2017. Projected 61% by 2030.
The US, Europe, and China together control 77% of all installed IT capacity on Earth.
Hyperscaler self-builds + colocation + sovereign-cloud build-outs. The 8-state top-8 holds 51% of US facilities.
Northern Virginia — known as "Data Center Alley" — alone holds ~1,020 facilities within a 20-mile radius, more than every country on Earth except the US itself.
Equinix and Digital Realty are pure colocation. Google, Meta, Microsoft, and AWS each self-build hyperscale campuses and additionally lease capacity from those two. The hyperscalers are scaling both.
| Project | Owner / Anchor tenant | Location | Scale | Capex |
|---|---|---|---|---|
| Stargate Initiative | OpenAI + Oracle + SoftBank | Multiple US sites | 5+ GW planned | $500B |
| Meta Hyperion | Meta | Richland Parish, LA | 5 GW · 2,250 acres | $10B+ |
| Anthropic-Amazon New Carlisle | Amazon (anchor: Anthropic) | New Carlisle, IN | ~2.2 GW · 1.5M sq ft | $11B (Project Rainier) |
| xAI Colossus 2 | xAI | Memphis, TN (+ MS) | 1.2+ GW · 1,112k H100-eq | ~$10B est. |
| Microsoft Fairwater Atlanta | Microsoft (OpenAI as user) | Fayetteville, GA | 769k H100-eq | undisclosed |
| Meta Prometheus | Meta | New Albany, OH | 763k H100-eq | undisclosed |
| OpenAI Stargate Abilene | Crusoe + Oracle + OpenAI | Abilene, TX | 1.2 GW · 509k H100-eq | ~$40B |
| Microsoft Fairwater Wisconsin | Microsoft | Mount Pleasant, WI | 446k H100-eq | ~$3.3B |
| Google Columbus | Columbus, OH | 409k H100-eq | undisclosed | |
| AWS Georgia expansion | Amazon | Georgia, USA | 11+ data centers | $35B |
| AWS Mississippi complex | Amazon (anchor: Anthropic) | Mississippi, USA | 2 sites · Project Rainier | $10B |
| Aligned Data Centers | BlackRock consortium | 50+ US campuses | — | $40B acquisition |
Compute capacity is given in H100-equivalent GPUs as estimated by Epoch AI from satellite imagery, permits, and public documents. Real performance is typically 20–50% of theoretical peak.
The open-weight Chinese models land on Hugging Face, but where they were trained is mostly a black box. Here is the most honest picture I can draw from public sources.
After the U.S. banned the H100 in 2022 and the H20 in 2024, China's frontier training runs on a constrained mix. DeepSeek trained V3 on 2,048 Nvidia H800s (a sanctioned China-export chip); newer work is shifting to Huawei Ascend 910B/910C. Domestic supply is the bottleneck, not power or land.
| Model | Lab | How they get compute | Where | Best public evidence |
|---|---|---|---|---|
| DeepSeek V3 / R1 | DeepSeek (Hangzhou) | Pattern B · parent cluster | Hangzhou, Zhejiang | 2,048 H800s · 671B MoE · V3 paper cites Fire-Flyer lineage |
| Qwen3 / Qwen3-Max | Alibaba (Qwen team) | Pattern A · self-hosted | Alibaba Hangzhou-Yuhang + Beijing-Tongzhou | 235B-A22B MoE · 100+ open-weight releases · 40M downloads |
| Kimi K2 / K2.5 | Moonshot AI (Beijing) | Pattern A · colocation | Beijing (racks in Alibaba + smaller landlords) | 1T params · 32B active · trained on 15.5T tokens · MoE |
| GLM-4.5 / 4.6 | Zhipu AI (Beijing) | Pattern A · colocation | Beijing colocation · undisclosed | Open-weight MoE · competitive with Claude 4 Sonnet on some benchmarks |
| Doubao 1.5 Pro | ByteDance Volcano Engine | Pattern C · self-build | Inner Mongolia & Guizhou (Doubao + TikTok) | Closed-weight · estimated 7–10 GW self-build by 2027 |
| ERNIE 4.5 / 4.5 Turbo | Baidu | Pattern C · self-build | Yangquan (Shanxi) + Baoding (Hebei) | Closed-weight · sovereign enterprise cloud focus · Kunlun 2 silicon |
| Hunyuan Turbo / T1 | Tencent Cloud | Pattern A · self-hosted | Shenzhen + Tianjin + Shanghai | Closed-weight · Mamba-style hybrid architecture |
| Pangu / MindSpore | Huawei Cloud | Pattern C · self-build | Guizhou + Dongguan + Inner Mongolia | Ascend 910B/910C silicon · only credible Nvidia competitor at scale |
| MiniMax-M2 / Yi-Lightning | 01.AI (Beijing) | Pattern A · colocation | Beijing (Alibaba + smaller) | Yi-Lightning closed-weight · M2 open-weight · founder Kai-Fu Lee |
Disclosure rule of thumb: Alibaba, ByteDance, Tencent, Baidu, Huawei control a known facility (Pattern A or C). The frontier research labs — DeepSeek, Moonshot, Zhipu, 01.AI — are overwhelmingly Pattern A or B: they rent racks in hyperscaler buildings and don't own the addresses.
If you're tracking where AI training happens in China, watch three provincial names: Inner Mongolia (Hohhot), Guizhou (Guiyang), and Ningxia (Zhongwei). These are China's "national data center clusters" — designated by MIIT since 2020 specifically to absorb hyperscale compute.
Most frontier AI labs don't own their data centers. They sign exclusive compute deals with the four big hyperscalers — and increasingly, with Oracle, Crusoe, and CoreWeave.
$8B in Anthropic + Trainium silicon deal. Operates Project Rainier (New Carlisle, IN) and Mississippi campus for Anthropic workloads. 1.5M+ sq ft.
Originally exclusive OpenAI host. Fairwater Atlanta (769k H100-eq) and Fairwater Wisconsin (446k H100-eq) anchor the AI footprint. FY26 capex ~$120B+ including Stargate co-funding.
Builds its own TPUs. Hosts Anthropic on TPU clusters via a multi-billion-dollar deal. Also hosts smaller AI startups (Lovable, Windsurf) as "primary computing partners."
Zuckerberg committed $600B US infra through 2028. Hyperion (LA, 5 GW, nuclear-adjacent) and Prometheus (OH, 763k H100-eq) are the two new flagships. Also $10B/yr with Google Cloud.
Rose to the top tier after the Stargate announcement. Operates Abilene site for OpenAI with Crusoe. Sept 2025: 5-year $300B deal with OpenAI for compute starting 2027.
First true gigawatt-scale data center in execution (SemiAnalysis). Uses behind-the-meter gas turbines to bypass 5–7 yr grid interconnection queues. NAACP lawsuit over unpermitted turbines in MS; DOJ intervened citing national security.
Anchor customer on AWS Project Rainier (Indiana + Mississippi). Also buys TPU capacity from Google. Trains Claude on a Trainium + Nvidia hybrid — co-designed kernels for cost efficiency.
No longer Microsoft-exclusive. Distributing compute across Microsoft (Fairwater), Oracle (Abilene + 5 more Stargate sites), Nvidia, and Crusoe-built sites. The most distributed of any AI lab.
Built on crypto-mining foundations, became the first major "neo-cloud" pure-play GPU provider. Signed a $15B compute contract with Microsoft. Now adding Bloom Energy fuel cells at the Volo, IL site for behind-the-meter power.
The split that's reshaping the entire industry. AI facilities are not just "data centers but bigger" — they're a fundamentally different build.
| Dimension | Traditional cloud / colocation | AI / "AI factory" |
|---|---|---|
| Primary workload | Multi-tenant SaaS, web, email, databases, video streaming | Single-customer training runs or inference; one model, one company (sometimes one nation) |
| Hardware | CPU racks · 8–32 kW per rack | GPU racks · 60–130+ kW per rack (Nvidia Blackwell direct-to-chip liquid) |
| Site footprint | ~70% IT room, 30% support · dense racks | ~30% IT room, 70% support · power & cooling dominate |
| Power density | ~1.0 kW / sq ft | ~2.5–6.0 kW / sq ft (newer sites target 8+) |
| Cooling | Air / chilled water loop | Direct-to-chip liquid + rear-door heat exchangers + evaporative cooling towers (~19,000 L/min each) |
| Typical PUE | 1.4 – 1.6 | 1.1 – 1.3 (Google claims 1.10 industry-leading) |
| Water use | ~0.3 L / kWh (mild) | ~1.0–2.0 L / kWh for evaporative (cooling-tower) sites |
| Tenant model | Hundreds-to-thousands of customers sharing racks | "AI factory" — often one tenant; Jensen Huang's term |
| Latency priority | Distributed across metro regions | Co-locate as many GPUs as possible on one campus to minimize inter-chip latency |
| Demand signal | Proven — "proven model with proven returns" | Speculative — "build it and they will come" |
| Capex per facility | $200M – $1B | $5B – $40B+ per campus |
Air cooling used to be the default. AI racks now run liquid-to-chip — cold plates bolted directly to the GPU die. Microsoft shifted the bulk of new builds to water cooling in 2024.
Grid queues in the US are 5–7+ years. Hyperscalers increasingly bring their own power: gas turbines (xAI, Crusoe), fuel cells (Oracle/CoreWeave), even co-located nuclear (Meta Hyperion).
Power Usage Effectiveness = total facility power / IT power. AI sites have lower PUE because cooling is denser — but absolute power and water per site are dramatically higher.
The three numbers that decide whether a data center actually gets built.
| Region | 2024 (TWh) | 2030 projected | Growth |
|---|---|---|---|
| United States | 183 | 426 | +133% |
| China | 65 | 175 | +170% |
| Europe | 64 | 109 | +70% |
| Japan | 19 | 34 | +80% |
| Rest of world | 84 | 201 | +140% |
| Global | 415 | 945 | +128% |
Built with Pexels stock footage and HyperFrames. Pexels free-to-use license; video is CC-BY-style shareable with credit.
Clips: Pexels / Mikael Blomkvist, Tom Fisk, Dima Krivoy, MrColo, Pressmaster, Pavel Danilyuk, Kuiyibo Campos. Data: Synergy Research · IEA · LBNL · Epoch AI · US Data Map · FT · TechCrunch · company filings.
If a number on this page matters to you, click through to the source.