How to read this census
A GPU count is only meaningful with its class and its date. Providers blend three kinds of numbers in their marketing, and the differences decide real capacity. Equally important is whose GPUs they are: some of the largest "clouds" below operate chips that belong to their customers.
Deployed
Racked, networked and revenue-ready. The only number a buyer can actually consume this quarter. CoreWeave's 250,000 (Dec 2024) and Yotta's 16,384 H100s are deployed figures.
Ordered / contracted
A signed purchase order or lease with a delivery window — real money, not yet compute. IREN's 50,000+ B300s (delivering H2 2026) and Nscale's ~200,000 GB300s for Microsoft sit here.
Planned
An announced target with no confirmed hardware behind it. Lambda's "1M GPUs and 3 GW" ambition, Neysa's 20,000-GPU plan, TensorWave's 2 GW of secured capacity. Ambition, not capacity.
The ownership trap
Crusoe's Abilene campus runs Oracle's and OpenAI's GB200 racks; Fluidstack operates clusters owned by Anthropic (largely Google TPUs) and is building a 1 GW French site whose anchor customer is Mistral. Counting those chips as "Crusoe's fleet" or "Fluidstack's fleet" double-counts the same silicon. For landlord-operators this census records power and halls, and counts GPUs only where the provider owns the chips. Marketplace listings (Vast.ai's 1,400+ hosts, Shadeform, Akash) are other people's hardware and are never counted as a fleet.
The big fleets
Providers whose disclosed fleet, signed pipeline or operating footprint clears roughly 10,000 GPUs or 100 MW. ClusterMAX tier shown where rated (v2.0, November 2025).
| Provider | ClusterMAX | GPUs (verified) | Silicon | Footprint | Source · as of |
|---|---|---|---|---|---|
| CoreWeaveNASDAQ: CRWV | Platinum | 250,000+last disclosed count, Dec 2024 — not updated since; 2026 capex of $31–35B implies far more | H100 → H200 → GB200/GB300 NVL72, RTX PRO 6000 | 43 data centers · 850 MW active · 2.2 GW+ contracted | S-1 / 10-Kmar '25 / q4 '25 |
| NscaleUK · vertical hyperscaler | Unrated '25 | 12,600 live+ ~200,000 GB300 contracted to Microsoft (Texas 104k, Norway 52k, UK 23k, Portugal 12.6k); + 96,000 Rubin GPUs for 2027 | GB300 NVL72 → Vera Rubin NVL72 | Texas 240 MW → 1.2 GW · Sines PT · Loughton UK 50–90 MW · Narvik & Glomfjord NO | nscale.comoct '25 / may '26 |
| IRENNASDAQ: IREN | Underperf. '25 | 150,000company-stated fleet after 50,000+ B300 order finishes deploying H2 2026; $9.7B Microsoft GB300 deal signed earlier | H100, H200, B200 → GB300, B300 | Childress TX · Mackenzie & Canal Flats BC · Prince George BC — air-cooled, renewable-rich | iren.com / SECmar '26 |
| Oracle (OCI)NASDAQ: ORCL | Gold | undisclosedsingle OCI Superclusters scale to 131,072 Blackwell GPUs; also 50,000 AMD MI355X (Oct 2025) | H100, H200, B200, GB200/B300 + AMD MI355X | 162 cloud regions built or underway · largest single site 800 MW · gigawatt nuclear site planned | oracle.comsep '24 / oct '25 |
| NebiusNASDAQ: NBIS | Gold | undisclosedper-site ceilings: Mäntsälä FI to 60,000 · Kansas City to 35,000 · UK 4,000–9,000 B300 | H100, H200, B200, GB200, GB300 | 220 MW connected end-'25 → 1 GW+ secured end-'26 · Finland, Vineland NJ, UK, Iceland, Israel, Paris | nebius.comoct '24 – oct '25 |
| Crusoeprivate | Gold | undisclosedown cloud fleet; buildings designed for 100,000 GPUs on one fabric — Abilene racks belong to Oracle/OpenAI | H100, H200, GB200 NVL72 | Abilene 200 MW live · 1.2 GW campus · 4.9 GW contracted across 5 campuses | dcdjul '24 / jul '26 |
| Fluidstackprivate · NY | Gold | operates others'deploys & runs Anthropic's fleet (largely Google TPUs, $50B program) and Mistral-anchored French capacity | NVIDIA H100/H200/GB200 + Google TPU | ~1 GW named sites: River Bend LA 245 MW · Lake Mariner NY 360 MW · Abernathy & Barber Lake TX · France 1 GW | anthropic.comnov '25 |
| Lambdaprivate | Silver | tens of thousandsundisclosed exact; NVIDIA leases back 18,000 for $1.5B; Microsoft deal for tens of thousands GB300 NVL72; stated goal 1M+ GPUs / 3 GW | H100, H200, GH200, B200, GB300 NVL72 | 15 US data centers · first GB300 NVL72 live Sep '25 (Mountain View) | lambda.ainov '25 |
| Lightning AImerged Voltage Park, jan '26 | Bronze* | 36,000+owned & operated — the largest clean private disclosure after CoreWeave; debt-free fleet; $500M+ ARR | H100, B200, GB300 | 60 MW active · 6 facilities · 4 US states | voltagepark.commar '26 |
| Firmusaustralia · pre-IPO | Silver | 18,500 liveGB300s ordered & online from Apr '26 at Southgate Melbourne; 54,000 GB300 across Southgate by mid-'26; Tasmania zone to ~36,800 | H200, L40S → GB300 NVL72 | Tasmania 90 MW stage 1 (→ 1.6 GW national plan '28) · Melbourne · Singapore | firmus.cooct '25 / apr '26 |
| TensorWaveUS · all-AMD | Silver | 10,000AMD-only fleet; single 8,192 × MI325X liquid-cooled cluster is the core; MI355X ramping; 2 GW capacity secured | MI325X, MI355X (ROCm; CUDA via SCALE) | Arizona · Florida · Pennsylvania (+20 MW TecFusions, Jan '26) | dcd / futuriomjun '26 |
| Microsoft AzureNASDAQ: MSFT | Gold | undisclosed100,000+ GB300 access leased via neoclouds (Nscale, CoreWeave, Nebius, Lambda, IREN); first 4,608-GPU GB300 NVL72 cluster live Oct '25 | Full NVIDIA stack + Maia silicon | 400+ data centers · Fairwater AI campuses (WI, GA) | press reportsoct '25 |
| AWSNASDAQ: AMZN | Silver | undisclosednever discloses GPU counts; P5e/P6e-B200/GB300 families + Project Rainier (Trainium2) for Anthropic | NVIDIA H100/H200/B200/GB300 + Trainium | Global region footprint | —no disclosure |
| Google CloudNASDAQ: GOOGL | Silver | undisclosednever discloses GPU counts; A3/A4/A4X (GB300) + TPU v5p–v7 fleets | NVIDIA H100/H200/B200/GB300 + TPU | Global region footprint | —no disclosure |
| Applied Digital → ChronoScaleNASDAQ: CHRN | — | undisclosedcloud unit (~$75.2M TTM revenue, Aug '25) spun out as ChronoScale, May '26; parent is landlord for 400 MW CoreWeave campus | H100-class clusters | Ellendale ND (Polaris Forge 1–2, 600 MW leased) + colo sites | applied digital IRmay '26 |
* Voltage Park held Silver in ClusterMAX 2.0 (Nov 2025) and Lightning AI held Bronze; the merged company (Jan 2026) has not been re-rated. IREN landed in "Underperforming" in Nov 2025 — before its Microsoft GB300 deal and 150,000-GPU expansion, a reminder that ratings lag balance sheets.
Sovereign & regional fleets
National champions and regional clouds, mostly subsidy-backed. Fleets are smaller, but disclosure is often better — governments want receipts for their money.
| Provider | ClusterMAX | GPUs (verified) | Silicon | Footprint | Source · as of |
|---|---|---|---|---|---|
| Yottaindia · shakti cloud | Unavail. '25 | 16,384 liveH100s at NM1 Navi Mumbai; second 16,384 cluster planned at D1 Greater Noida for 32,768 total; B200s on the way | H100, L40S → +B200 | NM1 Navi Mumbai (Tier IV) · D1 Greater Noida · 700 MW campus | nvidia case studyjul '26 |
| Naver Cloudkorea | Unavail. '25 | 60,000allocated to public cloud within Korea's 260,000-GPU national compute project | Blackwell-class | Korea (sovereign) | semianalysisnov '25 |
| Neysaindia · velocis | Bronze | ~2,000 livescaling to 20,000+ on $1.2B Blackstone-led financing (Feb '26) — would be ~⅓ of India's AI GPU capacity | H100, H200, L40S, L4, MI300X | Mumbai · Chennai colos · Telangana MoU with NTT DATA | weka / pressfeb '26 |
| GMI Cloudus / taiwan | Bronze | 7,000Blackwell GPUs at the $500M Taiwan AI DC opened Mar '26; US fleet undisclosed | H100, H200, GB200 NVL72/NVL4, B300 | Taiwan (own DC) + US colos | taiwanplusmar '26 |
| SoftBankjapan | — | undisclosedworld's first DGX B200 SuperPOD; GB200 NVL72 system; lead on Noetra national project — 27,500 Rubin GPUs with 43 partners | B200, GB200 NVL72 → Rubin | Japan | meti / pressjul '26 |
| Sakura Internetjapan | Unavail. '25 | 2,000 → ~4,000Hopper GPUs; ~10,800 total planned incl. B200 at Ishikari; largest METI subsidy recipient (~¥50.1B) | H100, H200 → B200 | Ishikari, Hokkaido (renewable by '27) | nvidia newsnov '24 |
| GMO GPU Cloudjapan | Silver | undisclosedTOP500/Green500-listed H200 cluster; +256 H200 added Q4 '25; HGX B300 deployed '26; METI-subsidized (~¥1.9B) | H200, B300 | Tokyo | gmo.jpmay '25 |
| NeevCloudindia | Bronze | 8,000 ordered40,000 planned by end '26 (company statements via ClusterMAX 2.0) | H100-class | India | semianalysisnov '25 |
| Scalewayfrance · iliad | Silver | ~5,000"top-tier" GPUs per parent iliad (Feb '25); multiple 1,016-GPU H100 SuperPODs; €3B group AI plan | H100, H200, GB200 | Paris DC5 (OpCore) · Warsaw | iliadfeb '25 |
| Vultrus · private, 33 regions | Silver | undisclosedfleet size never published; GB300 NVL72 deployments via HPE announced Jun '26 | NVIDIA + AMD; GB300 NVL72 | 33 cloud regions worldwide | hpe / businesswirejun '26 |
| RunPodus · dev cloud | Bronze | undisclosed"thousands of GPUs weekly" added; Q1 '26 deployments = 3× the entire prior-year fleet | H100, H200, B200, B300, RTX PRO 6000 | 31 global regions | runpod.iojun '26 |
The medallion long tail
Rated providers with no verified fleet disclosure. Listed so the census is complete — absence of a number is itself information.
| Provider | ClusterMAX | What's public | Silicon offered |
|---|---|---|---|
| Together AI | Silver | Fleet undisclosed; $305M Series B at $3.3B (Feb '25) | H100, H200, B200, GB200 |
| Cirrascale | Silver | Fleet undisclosed; long-running HPC cloud, GPU-backed debt facilities | H100, B200, MI300X |
| Gcore | Silver | Fleet undisclosed; AI Cloud Stack software runs on customer estates of 20,000+ GPUs | H100, H200, B200, MI300X |
| Hyperstack | Bronze | Fleet undisclosed; NexGen Cloud arm, EU renewable DCs | H100, A100, L40 |
| Shadeform | Bronze | Broker — owns no GPUs; aggregated marketplace across providers | multi-cloud catalog |
| STN | Bronze | Fleet undisclosed; US-based clusters | H100, H200, B200 |
| Atlas Cloud | Bronze | Fleet undisclosed | H100, H200 |
| Prime Intellect | Bronze | Aggregator — owns no fleet; decentralized training focus | marketplace catalog |
| CUDO Compute | Bronze | Fleet undisclosed; UK, partner-DC model | H100, H200, MI300X |
| Qubrid | Bronze | Fleet undisclosed | H100, H200 |
| Latitude.sh | Bronze | Fleet undisclosed; bare-metal specialist, global edge sites | H100, H200, L40S |
| Verda (ex-DataCrunch) | Bronze | Fleet undisclosed; Finland, EU-sovereign pricing floor | H100, H200, B200 |
| Denvr Data | Bronze | Fleet undisclosed; Canada/US; also Intel Gaudi 3 | GB200, B200, MI300X, Gaudi 3 |
| IBM Cloud | Bronze | Fleet undisclosed | H100, H200, MI300X |
| DigitalOcean | Bronze | Fleet undisclosed; GPU Droplets + Paperspace lineage | H100, H200, MI300X, RTX 6000 |
| Hot Aisle | Bronze | Fleet undisclosed; Michigan, AMD-focused, bare metal | MI300X, MI355X |
| Buzz HPC | Bronze | Fleet undisclosed; Canada, HIVE lineage | H100, H200 |
| Vast.ai | Bronze | Owns zero GPUs — 1,400+ independent hosts list hardware; sells them the operating system | marketplace: RTX → B200 |
The census totals
Summing only company-stated figures across neoclouds — CoreWeave, IREN, Lightning AI, Firmus, Yotta, Nscale, TensorWave, GMI, Scaleway, Neysa. Lambda, Nebius, Crusoe, Vultr and RunPod are large but undisclosed, so the true neocloud total is well past this.
Signed for delivery through 2027: Nscale's ~200k GB300 + 96k Rubin, IREN's 50k+ B300, Neysa's 18k expansion, Yotta's second 16.4k, Sakura's path to ~10.8k, Japan's 27.5k Rubin national order.
SemiAnalysis's ClusterMAX 2.0 market view (Nov 2025) counts 209 GPU clouds, brokers and sovereign projects — 37 earned a medallion. The long tail explains why no census can stay finished.
Two calibrations for these numbers. First, the hyperscalers: AWS, Azure and Google never disclose GPU counts — though Microsoft alone admitted to 100,000+ GB300 access just through neocloud leases, and Google sells TPUs to neoclouds. Second, the single-site record isn't a cloud at all: xAI's Colossus campus in Memphis passed half a million GPUs under one roof, more than every disclosed neocloud fleet on this page combined.
Frequently asked
Is there an official list of GPU counts per cloud?
No. SemiAnalysis tracks fleet size inside its paid models and scores it under ClusterMAX's availability criterion, but publishes no per-provider table. Public figures are scattered across SEC filings, earnings calls, company newsrooms and government subsidy announcements. This page assembles them, dated and sourced.
Which neocloud has the most GPUs?
CoreWeave — last disclosed count 250,000 (Dec 2024), certainly higher now with 43 DCs, 850 MW active and $31–35B of 2026 capex. IREN states 150,000 once its B300s finish deploying in late 2026. Nscale holds the largest signed private pipeline: ~200,000 GB300 for Microsoft plus 96,000 Rubin for 2027.
Why do Crusoe and Fluidstack show no GPU count?
Much of what they build isn't their silicon: Crusoe's Abilene racks belong to Oracle/OpenAI; Fluidstack operates Anthropic's and Mistral's clusters. A single "fleet size" would double-count customer chips. The census records their power and halls instead.
What's the difference between deployed, ordered and planned?
Deployed = racked and revenue-ready. Ordered/contracted = signed purchase with a delivery window. Planned = announced target, no confirmed hardware. Every figure here carries its class and date — a 2027 plan is not a rival's 2026 capacity.
Why are ratings shown at all?
Because fleet size and fleet quality move together in buyers' minds, and ClusterMAX (v2.0, Nov 2025) is the market's de-facto quality score. But note IREN: rated "Underperforming" in Nov 2025, a 150,000-GPU Microsoft-backed fleet by Mar 2026. Ratings lag balance sheets.
A number is stale or wrong — what then?
Send it through the corrections form with a primary source. Corrections ship within a week and the stamp date moves. Private companies: a dated, verifiable fleet figure is always welcome — on the record.
How the census is built
Primary sources only: SEC filings (S-1, 10-K, 8-K), earnings materials, company newsrooms, named-customer press releases, and government announcements (METI subsidy awards, IndiaAI Mission, Korea's national compute project). Each row shows the figure, its class (deployed / ordered / planned), the silicon, the footprint, and the source with its date. We don't count landlords' buildings as GPU fleets, don't count marketplace listings as owned hardware, and don't repeat figures we can't trace. Verified 2 August 2026 · by Rahul Patwardhan · part of the AI Infra Ecosystem Map.