COMPUTE INDEX
SOURCES.
The evidence behind the chips and clusters.
This page documents the source for each data point in the State of AI Report Compute Index charts. Data cutoff: 1 October 2026. GPU-count charts show NVIDIA data-center GPUs (GH200/GB200/GB300 counted as GPU dies; one NVL72 rack = 72 GPUs), split into Deployed / Installing / Announced. Figures are charted by owner/operator; renters are listed separately under the demand-side chart and are not double-counted. National-HPC counts are exact (published); large company, neocloud and sovereign figures are best-available estimates or announced commitments. Where a primary source did not disclose a GPU count, none is invented.
Research paper citations
View chart →A100 GPU count (Ampere)
View chart →Meta Research SuperCluster - 16,000
Source: Meta AI blog
Leonardo Booster (CINECA) - 13,824
Source: CINECA / Top500
Perlmutter (NERSC) - 7,168
Source: NERSC architecture docs
Polaris (Argonne) - 2,240
Source: ALCF
SoftBank - ~2,000 Ampere (DGX SuperPOD, operating since Sep 2023; SoftBank says 'approximately 2,000 NVIDIA Ampere GPUs')
Source: SoftBank
Berzelius (Linkoping) - 752
Source: Prior State of AI Compute Index
Alps (CSCS) - 576 A100 (144 nodes x 4)
Source: CSCS
Delta (NCSA) - 440
Source: NCSA Delta docs
Jean Zay (IDRIS) - 416
Source: IDRIS
NSCC ASPIRE 2A (Singapore) - 352
Source: NSCC
Gadi (NCI) - 16
Source: NCI
Hopper GPU count (H100 / H200 / GH200)
View chart →SpaceXAI Colossus 1 - 200,000 (150k H100 + 50k H200)
Tesla Cortex - ~66,000 (H100-equivalent)
Source: Tesla shareholder deck / 8-K
Meta GenAI clusters - 49,152 (2x 24,576 H100)
Source: Meta Engineering blog
CoreWeave (one cluster) - 42,000 H200
Source: Epoch AI; The Next Platform
Voltage Park - 24,000 H100
Source: Data Center Dynamics
JUPITER Booster (Julich) - 23,536 GH200
Source: EuroHPC JU; Top500 (June 2026)
Microsoft Eagle - 14,400 H100
Source: Top500
Alps (CSCS) - 10,752 GH200
Source: CSCS; Top500
ChronoScale (Applied Digital) - 6,144 H100 (deployed; third-party colos in CO/MN/UT, finance-leased, operator not owner)
AIST ABCI 3.0 - 6,128 H200
Source: NVIDIA / AIST
Isambard-AI (Bristol) - 5,448 GH200
NVIDIA Eos - 4,608 H100
Source: NVIDIA
MareNostrum 5 ACC (BSC) - 4,480 H100
Source: Barcelona Supercomputing Center
KAUST Shaheen III - 2,800 GH200
Source: KAUST; Top500
Venado (LANL) - 2,560 GH200
Source: Top500
Israel-1 (NVIDIA) - 2,048 H100
Source: NVIDIA
NSCC ASPIRE 2B (Singapore) - 1,536 H200
Source: NSCC
Gefion (DCAI, Denmark) - 1,528 H100
Source: NVIDIA; Novo Nordisk Foundation (DCAI is owner: NNF 85% / EIFO 15%)
Arrhenius (NAISS, Sweden) - 1,528 GH200 (deployed; entered production early June 2026, Top500 #42)
Source: EuroHPC JU supercomputer register (publishes the 1,528 total); NAISS (production date)
Jean Zay (IDRIS) - 1,456 H100
Source: IDRIS
Leonardo LISA upgrade (CINECA) - 1,328 H100, 166 8-way servers (installing; inaugurated 11 Jun 2026, user availability expected summer 2026)
Source: EuroHPC JU
Yotta Shakti Cloud (India) - 8,192 H100 (deployed; Nidar Infrastructure 424B3, Nov 2025: 'currently all 8,192 H100s are operational'. A further 2,048 H200 and 8,192 B200 were on order but the registration was withdrawn in Jan 2026 and neither is confirmed installed)
Source: Nidar Infrastructure SEC 424B3
NSCC ASPIRE 2A+ (Singapore) - 320 H100 (40 DGX H100 nodes)
Source: NSCC
SoftBank - ~4,000 Hopper (deployed since Oct 2024; SoftBank says 'approximately 4,000 NVIDIA Hopper GPUs')
Source: SoftBank
NCHC Nano4 (Taiwan) - 1,760 H200 (220 nodes x 8; in formal operation from 1 Jul 2026)
Source: NCHC; CNA
DeltaAI (NCSA) - 608 GH200
Source: NCSA Delta documentation
Blackwell GPU count (B200 / B300)
View chart →Deutsche Telekom Industrial AI Cloud (Munich) - up to 10,000 B200
Source: Deutsche Telekom
SoftBank DGX SuperPOD - >4,000 B200
Source: SoftBank
Naver Cloud - 4,000 B200 (deployed)
Source: Data Center Dynamics
E2E Networks (India) - 1,024 B200
Source: E2E Networks
SK Telecom Haein - >1,000 B200
Source: Data Center Dynamics
IREN - 50,000+ B300 (announced; purchase agreements, phased delivery through H2 2026)
Source: IREN
Yotta D2 (Greater Noida, India) - 20,736 HGX B300 (announced; go-live targeted Aug 2026)
Source: Yotta
Roche - 2,176 Blackwell on premises across the US and Europe (deployed; the separate cloud tranche taking Roche 'over 3,500' is excluded)
Source: Roche media release, 16 Mar 2026
WhiteFiber (Bit Digital) - 2,096 B300 (announced: 1,392 BaseTen Nov 2026 + 576 Iceland Dec 2026 + 128 ordered Jul 2026)
Source: WhiteFiber 10-Q
Eli Lilly (LillyPod, Indianapolis) - 1,016 B300 (DGX SuperPOD, deployed)
Source: NVIDIA
HIVE / BUZZ HPC (Bell AI Fabric, Manitoba) - 504 B200 (deployed; went live May 2026)
Source: HIVE Digital FY2026 results
NxtGen (India) - over 4,000 Blackwell in Dell PowerEdge XE9685L servers (announced; B200 vs B300 unstated, no operating confirmation)
Source: CIOL quoting Dell, 28 Jan 2026; Vertiv release, 18 Feb 2026
WhiteFiber (Bit Digital) - 384 B200 (deployed; service order with 'Deployment and revenue generation commenced in January 2026'. A minimum: total B200 fleet undisclosed)
Source: WhiteFiber 10-Q
NHN Cloud (Korea) - 7,656 B200 (deployed; two NIPA clusters of 6,120 procured under the MSIT program and operated by NHN, plus ~1,536 for NHN's own use. Part of the South Korea aggregate on the Grace-Blackwell chart)
Source: Asia Economy, 7 Aug 2026; Top500
Firebird DC-1 (Armenia) - 6,144 B200 (deployed; operational debut 8 Aug 2026)
Source: Firebird
Datasection - 5,080 B300 installing (Inzai, Japan) + 10,000 B300 announced (Sydney) + 4,696 B200 announced (Bangkok)
Source: Datasection releases
Nebius Israel (Modiin) - 4,000 B200 (deployed; Nebius: 'In October we launched in Israel with a deployment of NVIDIA B200s'; the 4,000 is from the launch release)
Source: Nebius
Kakao (Korea) - 2,040 B200 (deployed; government-procured, Kakao-operated, 84% of its 2,424 allocation)
Source: Asia Economy
Cloud GA instances (Google A4, AWS P6, Azure, Lambda, RunPod, DataCrunch)
Source: Vendor product pages; counts not disclosed
Grace-Blackwell GPU count (GB200 / GB300)
View chart →Stargate Abilene (Oracle/Crusoe) - at least 131,000 installing / 450,000 target (installing: Oracle 'delivered 131 thousand GPUs there in Q1' FY2027, Jun-Aug 2026, and 6 of 8 buildings 'delivered to the customer'. A one-quarter delivery figure, so a floor, chip not named on the call; delivery is not counted as operation. Oracle owns the silicon and leases to OpenAI; Crusoe is builder/landlord)
Source: Oracle Q1 FY2027 earnings call, 10 Sep 2026; Data Center Dynamics (450,000 target)
HUMAIN (Saudi) - 18,000 GB300 phase 1 / up to 600,000
South Korea (national + industry) - 260,000 (announced; a country-wide aggregate. The sovereign slice is MSIT 'up to 50,000' from an 'initial deployment of 13,000'; the rest is Samsung, SK Group, Hyundai and NAVER Cloud, and the mix includes RTX PRO 6000 Blackwell Server Edition as well as GB-series)
Source: NVIDIA newsroom
Nscale to Microsoft - 200,000 GB300 (announced)
Source: Nscale
SpaceXAI Colossus 2 - 550,000 deployed (110,000 GB200 + 440,000 GB300), 440,000 GB300 announced (220,000 targeted for the week of September 28 and 220,000 for November). The post does not establish that the first future tranche is racked or energizing, and its operational target is not treated as completion. Sourced to a CEO post, not a filing: 'Colossus 2 is 110k GB200 and 440k GB300.' The June prospectus (424B4 p.76) had 220,000 brought online. A further conditional December tranche is excluded. Google, Anthropic and Reflection are renters and carry no bar
Source: Elon Musk on X, 25 Sep 2026; SpaceX 424B4 (12 Jun 2026)
Argonne Solstice + Equinox (DOE) - 100,000 + 10,000 Blackwell (both announced; Argonne: 'Both systems are in the planning stages'. Announced as Blackwell, not explicitly Grace-Blackwell)
Together / Hypertec - 36,000 GB200
Source: Together AI
Not charted: Aolani Cloud (Malaysia, ByteDance as renter) - the WSJ-reported 36,000 GB200 plan was retired from the bars in October 2026. Aolani's own release (21 Sep 2026) says 22,000 new GPUs in early 2027 'will bring Aolani's total deployment to over 48,000 GPUs', which leaves no room for a 36,000 Blackwell cluster today
Stargate UAE / G42 - 35,000 GB300 (installing)
Source: OpenAI
SpaceXAI Colossus 1 - 30,000 GB200
Source: SemiAnalysis
AWS Project Ceiba - 20,736 GB200 (288 NVL72 racks)
Source: AWS; Data Center Dynamics
UK Nebius/Nscale - 14,000 Blackwell (announced)
Source: NVIDIA newsroom
CoreWeave - 8,192 GB300 NVL72 (deployed; MLPerf Training v6.0)
Source: CoreWeave
Mistral (Eclairion, Bruyeres-le-Chatel, France) - 13,800 GB300 (installing; CTO in Mar 2026: 'on stabilise actuellement la premiere tranche'. No operating confirmation after the end-June 2026 target)
Source: silicon.fr; Data Center Dynamics
Sines / Start Campus (Nscale/Microsoft, PT) - over 12,600 GB300 (deployed; Nscale's S-1 lists the Portugal site as Active, defined as online and revenue generating. Microsoft is the offtaker and carries no bar)
Source: Nscale S-1, 18 Sep 2026
Lambda - 10,000 GB300 NVL72 (deployed; Lambda: 'in production, on a 10,000-GPU NVIDIA GB300 NVL72 cluster'. Lambda owns and operates it; Microsoft and other customers are renters)
Source: Lambda blog, 31 Mar 2026
Nscale Glomfjord (Norway) - 2,304 GB200 (32 NVL72; announced: Nscale owns the GPUs, service was due to start 30 Aug 2025 with no confirmation it did. Renter is Spring (SG) Pte. Ltd.)
Source: Nscale S-1, exhibit 10.12
MareNostrum 5 AI Factory upgrade (BSC) - 1,920 (336 nodes x 4 GB200 NVL4 + 144 nodes x 4 GB300; announced, H1 2026 install target lapsed unconfirmed)
Source: Barcelona Supercomputing Center
Sferical AI (Linkoping) - 1,152 GB300 (16 GB300 NVL72 racks; announced, owned by Sferical AB, no operating confirmation)
Source: Sferical AI
NCHC Nano4 GB200 (Taiwan) - 144 (2 GB200 NVL72; deployed, opened to users as a free trial on 1 Sep 2026)
Source: NCHC
Taiwan Foxconn 'Big Innovation' - 10,000 GB300 (announced)
Source: NVIDIA newsroom
IT4LIA AI Factory (Bologna) - 8,000+ GB200 NVL4 (announced)
Source: EuroHPC JU
Microsoft Azure NDv6 GB300 - 4,608 (for OpenAI)
Source: NVIDIA newsroom
Jane Street - 4,032 GB300 (56x GB300 NVL72, owned Texas facility; deployed)
Source: NVIDIA AI Infrastructure
Microsoft Azure NDv6 GB200 - ~4,000
Source: Microsoft Azure blog
Nscale Keflavik (Verne campus, Iceland) - 4,608 GB300 (deployed; 64 GB300 NVL72 racks, benchmark suite run across all 64)
Source: Nscale, 16 Jul 2026
HIVE / BUZZ HPC (Merritt BC, Bell/Cohere) - 2,304 GB200 (announced; contracted not live, 'go live in late 2026 to early 2027')
Source: HIVE Digital Q1 FY2027
HIVE / BUZZ HPC (Merritt BC, GB300) - 2,016 GB300 NVL72 (announced; separate cluster and customer, delivery calendar Q4 2026, HIVE retains ownership)
Source: HIVE Digital 8-K, 19 Aug 2026
Blue Swan / BavariaAI (FAU Erlangen + LRZ) - 1,024 GPUs via GB200 NVL4 (announced; FAU: 'a further 1,024 AI chips from... NVIDIA', contract signed 23 Apr 2026, access targeted autumn 2026. NVIDIA rounded to 1,000)
HammerHAI (HLRS Stuttgart) - 860 GPUs (215 GB200 NVL4 nodes; announced, in service H2 2026)
Source: HLRS
Mimer (EuroHPC JU, hosted at Linkoping) - 400 GPUs (100 GB200 NVL4 systems; announced)
NERSC Cech (LBNL) - 144 Blackwell with 72 Grace CPUs (installing; early-access system for Doudna)
Source: NERSC
WhiteFiber (Bit Digital) - 72 GB200 (deployed; revenue generation began 7 Mar 2026. A further 216 GB200 remain unconfirmed)
Source: WhiteFiber 10-Q
Indosat (Indonesia) - 2,304 GB200
Source: Tom's Hardware
SoftBank - 1,224 GB200
Source: SoftBank
Frontier-lab compute agreements
View chart →OpenAI - NVIDIA 10 GW / up to $100B
Source: NVIDIA newsroom
OpenAI - Oracle $300B (~4.5 GW, Stargate)
Source: OpenAI; CNBC
OpenAI - Microsoft Azure $250B; AWS $38B; CoreWeave $22.4B
Source: CNBC ($1T AI deals guide)
OpenAI - AMD 6 GW / MI450 (non-NVIDIA)
Source: OpenAI; TechCrunch
OpenAI - Broadcom 10 GW of OpenAI-designed accelerators across multiple generations to 2029 (non-NVIDIA; term sheet, Oct 2025). 'Jalapeno' is the first of those generations and carries no GW figure of its own
Source: OpenAI / Broadcom
OpenAI - AWS ~2 GW Trainium3/Trainium4 (non-NVIDIA; 27 Feb 2026, inside a $100B/8yr expansion of the $38B AWS agreement)
Source: Amazon press release
OpenAI - Cerebras 750 MW wafer-scale (non-NVIDIA; signed 14 Jan 2026)
Source: Cerebras
Anthropic - Google + Broadcom 5 GW of next-generation TPU (non-NVIDIA; Anthropic's own May 2026 restatement. Broadcom supplies ~3.5 GW of that total). Builds on the Oct 2025 'well over a gigawatt' Google Cloud TPU expansion, which neither party restates as a separate addend
Source: Anthropic
Anthropic - AWS Project Rainier ~500k-1M Trainium2 (non-NVIDIA)
Source: Anthropic
Anthropic - Microsoft Azure + NVIDIA: $30B of Azure compute plus additional capacity up to 1 GW (the 1 GW is additional to the $30B, not a GW-denomination of it)
Source: Anthropic
Anthropic - AMD up to 2 GW / MI450 (non-NVIDIA; 22 Jul 2026, AMD equity up to $5B, deploys from H1 2027)
Source: AMD investor relations
Anthropic - SpaceXAI (Colossus 1) lease: 'more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs)'. The $1.25B/mo price is press-reported, not in either party's announcement
Source: Anthropic
Microsoft (offtaker) - Nscale 200k GB300; IREN 200 MW/$9.7B; Lambda; Nebius; Crusoe 900 MW
Source: Operator press releases
Meta (offtaker) - CoreWeave $35.2B; Nebius up to $27B
Source: The Globe and Mail; press
Startups (Cohere/Runway->CoreWeave; Thinking Machines/World Labs->Google Cloud; Reka/Suno->Oracle; Sarvam->Yotta 4,096 H100; Upstage->SKT ~500 B200; Inflection->CoreWeave 22k legacy)
Source: Respective cloud case studies / press
Compiled October 2026. Citation-chart data: Zeta Alpha (last refreshed 1 September 2026). Compute-cluster data: operator disclosures, EuroHPC, Top500 (June 2026), SemiAnalysis, The Next Platform, Data Center Dynamics, company earnings and SEC filings, and NVIDIA newsroom. Estimates and announced commitments are labelled as such in the underlying dataset.