7 segments · 99 companies · as of Sep 18, 2026
Neoclouds & GPU Clouds
Specialist operators renting accelerated compute and contracted AI hosting
Where it sits
Fed by: Servers & Systems · Data Center InfrastructureFeeds: Models & Platforms
Specialist clouds rent accelerated servers, while hosting developers supply powered buildings to the companies that own those servers. The distinction matters: a megawatt of colocation is not a megawatt of owned GPUs, and a signed multi-year lease is not current operating revenue.
Why now
Blackwell deliveries, Rubin reservations and financing are driving 2026 contracts. New leases at former mining sites are substantial, but commissioning remains uneven. September updates expand CoreWeave commitments and Crusoe platform value while several smaller pivots remain unproven.
Listed accelerated cloud operators
Listed operators provide rented accelerators, ranging from large frontier-lab suppliers to newly commercial businesses. Ranks reflect delivered scale and contract quality. Small initial leases and unnamed customers carry less weight than mature operating fleets.
OpenAI/Meta/Anthropic GPU cloud; $104.2B Q2 backlog plus >$25B Q3 commitments; 4.2 GW contracted (est.)
leaderTargets 5 GW contracted power by YE2026 versus 800 MW-1 GW connected (est.); these are targets, not live capacity
majorMicrosoft $9.7B/200 MW contract; separate Nvidia $3.4B five-year cloud contract and conditional $2.1B equity rights
majorMerchant Colossus 2: Google ~$920M/month, Reflection ~$150M/month; Anthropic ~$1.25B/month spans Colossus 1/2
majorOpenAI wafer-scale cloud; Q2 RPO $25.4B and >600 MW live/contracted through 2027 (est.), not all operating
challengerNvidia anchor; August 6 TCV $8.8B and secured power 212 MW (est.), largely a 2026-27 delivery pipeline
challengerGPU cloud plus Nscale NC-1 hosting: $865M/40 MW over 10 years (est.); first half handed over in August
challengerVia Northern Data/Quake AI, rents Nvidia GPUs; June acquisition closed; public customer TCV remains undisclosed
challengerApplied Digital-controlled cloud after May spinout; owns rented GPU infrastructure; customer TCV/MW undisclosed
nicheAI-native GPU customers; 155 MW total committed site capacity (est.) includes future builds, not only GPUs
nicheAugust adds an unnamed $600M AI-cloud customer contract (est.); expands GPU purchases into 2027
challengerALPHA-01: 504 B200 GPUs and $32.2M two-year unnamed lab lease (est.); live since May
emergingHGX B200 service earns revenue under an August multi-year AI customer contract; buyer and contract value undisclosed
emerging
Sources: investors.coreweave.com, assets.nebius.com, sec.gov, reuters.com, globenewswire.com, srgresearch.com
Large private GPU and TPU clouds
Private providers combine compute operations with large construction and financing programs. Their contract totals often include campus services as well as rented accelerators. Proposed IPOs, fundraising discussions and planned power are explicitly distinct from completed events.
Reported Jane Street cloud contract ~$13B over five years; disclosed >$140B platform TCV also includes campuses
leaderReported TCV $103B (est.), including Anthropic; Microsoft anchors Kvandal; proposed IPO has not closed
majorAnthropic announces $50B US build with its GPU/TPU cloud (est.); a development program, not delivered MW
majorReported Anthropic $35B/350 MW supply pact (est.) at Hut 8 Beacon Point; future delivery, IPO not completed
majorProject Southgate Nvidia cloud; August private round >$10.5B valuation (est.); client TCV/live MW undisclosed
majorDedicated GPU clusters and inference; Q2 annualized bookings $1.15B (est.) are not contracted backlog
majorAMD MI355X Ohio build targets 50 MW (est.); LuminArx/AMD-backed cloud, named anchor and backlog undisclosed
challengerAMD Instinct specialist; ~14 MW live versus 500 MW leased pipeline (est.); customer TCV undisclosed
challengerVoltage Park merger completed; Cursor/Reflection customers; >36k owned Nvidia GPUs (est.), dollar TCV undisclosed
challengerAnnounced Taiwan Blackwell factory: 16 MW/~7k GPUs (est.); March 2026 commissioning target remains unverified
challengerHyperstack GPU cloud; USD.AI finances a $34M B200 deployment (est.); customer backlog/MW undisclosed
nicheCanadian-owned Denvr AI Cloud offers H200/H100 bare metal and dedicated model endpoints across Canada and the US
nicheReported February Ori merger forms Brookfield's Radiant AI cloud; GPU services continue under the enlarged platform
niche
Sources: crusoe.ai, reuters.com, bloomberg.com, denvr.com
Mining sites with signed AI business
Power-rich mining operators are becoming AI landlords or GPU service providers. Only signed hosting or commercial compute businesses appear here. Options, grid reservations and uncontracted development pipelines are excluded from contracted capacity.
CoreWeave plus AMD leases: 1.1 GW/>$24B base value (est.); 437 MW billing mid-July, later AMD delivery
leaderCoreWeave/two unnamed hyperscalers: 1,410 MW/$36B base leases (est.); 175 MW live at July update
majorRiver Bend/Beacon Point: 949 MW IT/$26.6B base leases (est.); first data-hall deliveries in 2027
majorAnthropic signs 401 MW/$19B lease (est.); Lake Mariner 102 MW live (est.); Abernathy stake sold in July
majorAWS $5.5B/15-year lease (est.), first rent in August; Fluidstack/Google-backed Barber Lake still ramping
majorCoreWeave leases 526 MW IT at Helios (est.); first 133 MW IT delivered (est.); further phases due 2027
majorChildress Microsoft lease: 200 MW IT over five years (est.); first 50 MW tranche delivered in August (est.)
majorMeta-guaranteed Sandersville: 175 MW/$6.6B over 20 years (est.); initial delivery Q4 2027
challengerRockdale: AMD 50 MW plus unnamed lab 191 MW/$9.1B base (est.); 25 MW live, lab delivery starts Dec 2027
challengerNC-1 Nscale lease: 40 MW/$865M base term (est.); 20 MW billing in August (est.), balance targeted thereafter
challengerBUZZ signs $220M Bell/Cohere and $350M enterprise GPU contracts (est.); new clusters due late 2026-27
challengerOwns Tydal: signed 121 MW IT Volta lease, ~$4.7B/16 years; phases due Dec 2026/Mar 2027, with tenant exit at year 10
challengerCerebras Alabama hosting: 40 MW/$1.1B over 10 years (est.); deliveries scheduled December to Q1 2027
emergingMichigan neocloud lease: 20 MW critical load/10-year base term (est.); $1.2B includes renewal options (est.)
emergingEcoHash starts commercial GPU service in September; 3 MW AI-ready section (est.) is part of a 50 MW Georgia mining site
emerging
Sources: investors.corescientific.com, ir.applieddigital.com, hut8.com, investors.terawulf.com, cangoonline.com, stocktitan.net
GPU marketplaces and inference clouds
Marketplaces aggregate hardware supply, while inference clouds sell model execution rather than entire buildings. Fleet counts can include third-party GPUs. Undisclosed power and backlog are left unestimated, especially where customers buy short-duration usage.
Sells open-model inference and dedicated Nvidia clusters; contracted customer backlog and live MW undisclosed
majorServerless and dedicated GPU cloud plus Community Cloud; June funding $100M (est.); MW/backlog undisclosed
majorIndependent LPU inference cloud; August footprint 54 MW (est.); 200+ MW target for 2027 (est.), not backlog
majorSells wafer-scale inference; OpenAI anchors 750 MW phased supply (est.); not a Nvidia GPU reseller
challengerCombines AI development tools with Voltage Park-owned GPUs; marketplace-to-owned-cloud integration completed
challengerMatches developers with third-party GPU owners; marketplace listings are not owned fleet or committed backlog
nicheAggregates GPU rentals and inference, including Qumulus supply; own customer TCV and active MW undisclosed
nicheTrades short-duration GPU clusters for AI developers; exchange availability is not long-term contracted MW
nicheSigns Canopy Wave inference-cloud merger; closing targets end-2026, with SAIH still the current listed issuer
emergingManaged open-model inference and reinforcement-learning infrastructure; July Series D finances additional serving capacity
majorRuns dedicated model inference and post-training deployments; June Series F funds infrastructure for production workloads
majorManaged model endpoints use a multi-cloud GPU fleet; checkpointing reduces cold starts for intermittently used inference
challengerHosts open-model inference APIs and dedicated GPU deployments; May Series B funds expanded model-serving capacity
challengerRoutes open-model inference across distributed GPU providers; latency and utilization management are its service layer
nicheServerless GPU inference serves image, video and audio models; December Series D funds global capacity expansion
challengerOwns Replicate since December 2025; model catalog and deployment API supply inference alongside Workers AI
major
Sources: deepinfra.com, blog.fal.ai, blog.cloudflare.com, ir.saiheat.com, fireworks.ai, baseten.co
Asian and African GPU clouds
Regional suppliers serve national programs and local enterprises that need domestic GPU access. Rankings emphasize verifiable deployments and contracts. Announced chip orders, financing plans and government allocations are not counted as commissioned capacity.
Shakti Cloud's 20,736 B300 rollout (est.) includes Nvidia's four-year service commitment; full acceptance unverified
majorKoukaryoku cloud serves Japanese enterprises/researchers; ~1,100 Blackwell GPUs live by February (est.)
majorNational GPUaaS supplier; June allocation 4,120 B300/Rubin GPUs (est.), phased through H1 2027
majorVia NTT DATA, offers Japanese DGX B200 GPU cloud; enterprise GPU backlog and active GPU MW undisclosed
majorOsaka Sakai GB200 cloud serves Japanese enterprises; GPU customer backlog and active GPU MW undisclosed
majorVia KT Cloud AI Train, rents H100/A100 clusters; named anchor, GPU-only backlog and active MW undisclosed
nicheNational research customer on live B300/H100 GPUaaS; KRW15.3B service project (est.), July 2026-March 2027
majorRents Gwangju B200 capacity commercially to Korean enterprises; paid GPU backlog and utilization undisclosed
majorVietnam/Japan H100/H200 cloud serves enterprises and developers; B300 announced, customer TCV/MW undisclosed
majorAugust fleet: 96 H200 and 25 B300 servers (est.); further deliveries pending; customer TCV/MW undisclosed
challengerUnnamed Indian sovereign AI customer signs INR10B term sheet (est.) through June 2029 for deployed Blackwell
challengerAICA and national research GPU supplier; H100/B200 service contract May-December 2026; TCV/MW undisclosed
challengerJohor GB200 NVL72 GPU cloud is operating; 600 MW park potential (est.) is not commissioned GPU capacity
challengerVayu rents H100/H200/L40S to enterprise users; named GPU anchor, contracted backlog and active MW undisclosed
nicheReported IndiaAI allocation of ~4,096 B200 GPUs (est.); full commissioning and utilization remain unverified
challengerOperates Japanese B200/A100 GPU cloud and expands domestic sites; customer TCV and power not disclosed
nicheOwn-GPU contracts in Japan/Australia survive August cancellation of upstream rerented compute; Thailand remains an MOU
emergingCassava AI announces South African Nvidia GPUaaS and model APIs; broader African factory rollout remains phased
emergingAI Data Center GPU Cloud uses Infrinia AI Cloud OS and domestic GB200s; beta live, commercial launch targets October
emergingThai Nvidia Cloud Partner rents H100/H200 clusters; DDN confirms an initial 1,000-GPU deployment (est.)
challengerBlackstone-backed AI cloud plans 20k+ GPUs (est.); $600M equity plus proposed matching debt (est.) is not delivered capacity
challengerIndian cloud rents H100/A100 VMs, bare-metal GPUs and Kubernetes AI Pods; frontier-model ambitions no longer define the fleet
nicheH200 GPU instances and distributed inference pair with regional networking; AI Gateway provides unified model access
nicheGovernment-backed B200 allocation supplies Korean startups and researchers through Kakao Cloud, distinct from KOACC ownership
niche
Sources: sakura.ad.jp, internet.gmo, datasection.co.jp, business-standard.com, ddn.com, meti.go.jp
European sovereign GPU specialists
European specialists compete on local jurisdiction, power access and accelerator availability. Some operate live clouds while building much larger campuses. Announced sovereign projects are checked for pauses, changes of customer and unfinished construction.
European GPU cloud; 2026 revenue guide $3.0-3.4B and exit ARR $7-9B (est.) depend on phased capacity delivery
majorGPU DDTL funds Norway/Portugal/Iceland/UK clusters; reported Stargate UK pause is separate from Microsoft Kvandal
majorMistral Compute builds French Blackwell cloud; 13,800 GPUs financed for Paris cluster (est.); Sweden due 2027
challengerEuropean GPU instances and managed inference serve enterprise users; public GPU-only backlog/MW undisclosed
majorVia Scaleway, supplies European GPUs and trusted cloud for Airbus; GPU-specific backlog/MW undisclosed
majorFormer DataCrunch runs Nordic GPU cloud; 2026 funding supports expansion; customer TCV/live MW undisclosed
challengerHyperstack expands Nordic Blackwell clusters; $34M USD.AI facility (est.) funds Sweden B200 deployment
challengerExaion majority acquisition closed Q1 2026; French sovereign GPU cloud operates, with TCV/MW undisclosed
nicheEuropean sovereign GPU servers and AI inference; neither GPU customer backlog nor active GPU MW disclosed
nicheDedicated RTX PRO GPU servers for developers; no disclosed long-term GPU customer backlog or facility MW
nicheCommercial cloud and Blackwell services are live; proposed Stockholm and France campuses remain expansion projects
emergingReported Radiant/Ori combination adds European GPU operations to Brookfield's AI program; site-level delivered MW undisclosed
nicheManaged European Blackwell clusters offer B200/B300 HGX nodes and NVL72 racks; availability is product-specific
nicheFrench AI factories and on-demand H100 cloud operate alongside planned SoftBank Bosquel 1 GW campus (est.)
challengerUK sovereign AI platform offers B200 GPUs and CivoStack/FlexCore private deployments for regulated workloads
niche
Sources: assets.nebius.com, nscale.com, reuters.com, corporate.ovhcloud.com, ori.co, cudocompute.com
Chinese compute leasing providers
Domestic listed suppliers rent computing infrastructure, often alongside equipment sales or telecom work. Revenue from resale is separated from recurring leasing. Unnamed customer contracts, rented capacity and export restrictions limit comparisons with US neoclouds.
Xiechuang intelligent-compute business scales in H1 2026; GPU lease customers, MW and backlog not separately given
majorH1 compute leasing revenue RMB614M (est.); owned and re-rented compute mixed; no comparable contracted MW
majorIndependent GPU cloud; September says GPU utilization >90%; named anchor, contract backlog and MW undisclosed
challengerGPU bare-metal cloud expands Chinese sites; proposed RMB1.59B financing (est.) is not customer backlog
challengerMarch unnamed customer compute contract RMB428.5M over three years (est.); no disclosed MW or chip model
challengerSmall compute-leasing/integration business; H1 AI revenue RMB51.7M (est.), sharply lower; TCV/MW undisclosed
nicheSenseCore serves four external foundation-model providers; 2.4T daily tokens in July (est.) is throughput, not leased MW
majorCompute Cloud and edge AI gateway grow in H1; provides model access and scheduling alongside hosted compute
nicheH1 filing discloses compute leasing and equipment sales through Aofei Intelligent Computing; AI capacity is not all IDC MW
challengerWuxiang Yungu supports its operating Token Factory; H1 compute-network revenue includes service and infrastructure activities
challengerReported Beijing/Tianjin intelligent-compute services have RMB100M+ annual contracts (est.); distinct from rack colocation
nicheSongjiang phase two completes delivery and enters operation in H1; serves Shanghai public AI-compute infrastructure
challenger
Sources: money.finance.sina.com.cn, stockmc.xueqiu.com, disc.static.szse.cn, sensetime.com
Glossary
- Critical IT load
- Power available to computing equipment; gross site power also includes cooling and other facility loads.
- TCV
- Total contract value across a contract term; it may include options or services absent from accounting backlog.
- RPO
- Remaining performance obligations under accounting rules; not directly comparable with every company's TCV.
- GPUaaS
- GPU capacity rented as a cloud service; the operator or a financing partner supplies the hardware.
- Take-or-pay
- A customer owes contracted payments once delivery conditions are met, even if it uses less capacity.
Heat, spend, exposure and shares are editorial estimates. How to read the map ↗