Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-08-14

Kempax Research

Loop report · 2026-08-14

Global data-center and GPU capacity — daily update

Public evidence through 14 August 2026 shows a large, expanding, but non-additive AI infrastructure pipeline. It is sufficient to identify representative projects that are operating, under construction, permitted, contracted, or announced; it is not sufficient to produce a defensible global total of IT load, energized electricity, installed GPUs, or project cost. Source-defined MW may mean IT load, facility or campus capacity, grid interconnection, generation, PPA supply, or a program target. The current stance is Monitor .

The clearest evidence is split across nested scopes: Crusoe reports a 200 MW Abilene phase and a 1.2 GW campus plan; Meta describes 5 GW of Richland compute while separate utility reporting describes about 2.26 GW of planned generation; Nscale describes a 20 MW initial Norway phase inside a 230 MW build-out; and France reporting separates 13,800 GB300 GPUs from an earlier 18,000-system claim. These figures are retained as separate claims, not added. Crusoe Abilene, Meta Richland, Nscale Norway, Mistral financing

Executive view

Operating or reported live: Examples include the first reported-live Abilene buildings, xAI Colossus phase-one power service, AWS Project Rainier, the first Microsoft Fairwater building, ODATA QR03 energization, CoreWeave Barcelona’s inaugurated 15 MW facility, Nebius Mäntsälä’s expanded site, YTL’s first Johor AI facility, Nxera DC Tuas, BDx CGK4 phase one, KDDI Osaka Sakai, and Cassava’s reported Cape Town go-live. “Live” is a source status, not an audited current utilization or GPU stocktake.
Under construction or ready-for-service: Meta Richland, Meta Sturgeon County, Bell AI Fabric, QScale Q01 Building B, Scala Tamboré’s substation, SoftBank Tomakomai, and AirTrunk SYD3 have construction or ready-for-service evidence. Capacity remains phase-, campus-, or facility-scoped.
Contracted or permitted: Applied Digital’s 250 MW and 150 MW CoreWeave leases, Scala AI City’s 5 GW approved connection, BDx’s 788 MVA contracted grid power, Datagrid’s 140 MW supply arrangement, and the Talen AWS 960 MW development threshold are commitments or approvals, not energized IT load.
Announced or planned: Vantage Frontier, CloudHQ Querétaro, Nebius Lappeenranta, Pure DC SJK01, DataVolt NEOM, Stargate UAE, HUMAIN’s AI-factory program, NEXTDC S7, and the reported Xinjiang portfolio remain forward-looking or only partly verified.
Confidence: High means an official dated source clearly states the claim; medium means an official claim with an ambiguous capacity basis or a reported operating status; low means a secondary, search-lead, or unresolved scope claim. ND means not disclosed in the reviewed evidence, not zero.

Data-center pipeline

The entries below are representative and deliberately not additive. Status is shown as Operating reported live , Under construction ready-for-service , Contracted permitted , Announced planned , or Program portfolio . Capacity and power use the source’s own basis.

Americas

Operating reported live — Crusoe Stargate Abilene, Abilene, Texas, US. A 200 MW phase was reported live; the same campus has a 1.2 GW announced expansion. Neither source separates IT load from critical or campus load, and utility import MW is not disclosed. GB200 racks were reported, but count and current inventory are ND. Crusoe is developer operator; Oracle and OpenAI are workload partners; Blue Owl Capital and Primary Digital Infrastructure are named in the second-phase financing context. The reported US$15B is a program JV value, not a GPU price. Confidence: medium. 200 MW phase, live announcement, 1.2 GW campus, Crusoe JV
Operating expanding — xAI Colossus, Memphis, Tennessee, US. Phase-one service is reported at 150 MW; a separate additional 150 MW approval is excluded. IT load is ND. xAI’s public claims conflict between 180,000 and 200,000 H100s, while other reporting cites a 100,000-GPU build; no audited 2026 count is selected. xAI operates the site; MLGW and TVA are power counterparts; Valor Equity Partners, Andreessen Horowitz, and Sequoia are named company investors. The US$6B Series B is corporate financing, not site capex or GPU price. Confidence: medium. xAI Colossus, MLGW quick facts, Series B
Operating reported live — AWS Project Rainier, New Carlisle, Indiana, US. AWS reports nearly 500,000 Trainium2 chips; Trainium2 is a custom accelerator and is excluded from GPU totals. Site IT and utility MW are ND. AWS Amazon operates the cluster with Anthropic as an AI customer partner; the reported US$11B project figure is infrastructure scope, not accelerator price. Confidence: medium. About Amazon, AWS update
Operating expanding — Microsoft Fairwater, Mount Pleasant, Wisconsin, US. The first building was reported operational on 23 June 2026; site IT and interconnection MW are ND. Microsoft describes “hundreds of thousands” of NVIDIA GPUs without a count or model mix. The US$3.3B Wisconsin plan and later US$4B expansion are investment signals, not GPU prices. Confidence: medium. Microsoft, operating report
Operating reported energized — ODATA QR03, Querétaro, Mexico. ODATA announced a 300 MW IT-capacity campus plan and later energization, but online IT MW is ND. ODATA operates the facility with Aligned Data Centers platform backing; GPU inventory, financing, and price are ND. The full 300 MW is not treated as energized IT load. Confidence: medium. ODATA 300 MW, energization
Operating — Equinix SP6, São Paulo, Brazil. The AI-ready facility opened on 16 April 2026; IT and utility MW and site GPU inventory are ND. A reported R$570M investment is not a disclosed GPU or power price. Confidence: medium. Equinix opening, AI-ready facility
Under construction expanding — Meta Hyperion, Richland Parish, Louisiana, US. Meta’s July 2026 update describes construction and expansion around 5 GW of compute capacity; that is retained as compute wording, not relabeled IT load. Separate utility reporting describes about 2.26 GW of planned gas generation. Meta operates; Entergy Louisiana is the utility counterpart; Blue Owl-managed funds are the development partner. The current US$50B program report and earlier approximately US$27B development cost and US$7B Blue Owl contribution are nested or successive scopes. Confidence: medium. Meta update, Blue Owl JV, Entergy
Contracted ready-for-service — Applied Digital Polaris Forge 1, Ellendale, North Dakota, US. Two CoreWeave lease announcements cover 250 MW and a separate 150 MW addition. Together they describe a 400 MW lease campus scope, not confirmed live IT load. Applied Digital operates; CoreWeave is the tenant; Macquarie offers up to US$5B of funding. Lease revenue and funding are not capex or GPU prices, and site GPU inventory is ND. Confidence: medium. 250 MW lease, 150 MW lease, Macquarie funding
Under construction — Meta Sturgeon County, Alberta, Canada. Groundbreaking was reported on 8 July 2026 for a 1 GW data-center project. IT versus campus basis, separate utility import MW, GPU inventory, and outside investor allocation are ND. A reported C$13B project value remains a reported price signal pending primary reconciliation. Confidence: medium. Meta, Reuters
Under construction — Bell AI Fabric, RM of Sherwood, Saskatchewan, Canada. Bell BCE describes a 300 MW campus; IT critical and utility import splits are ND. CoreWeave and Cerebras are named customers, not GPU inventories. Bell BCE is developer operator, Bird Construction is a contractor, and the announced C$1.7B investment is not a GPU or tariff quote. Confidence: medium. BCE, Bird
Under construction — QScale Q01 Building B, Québec, Canada. Building B is described as 60 MW AI-ready infrastructure; QScale separately describes 142 MW of campus infrastructure. The bases are not defined, and they are not added. QScale operates; Goldman Sachs Alternatives’ acquisition and up to C$320M growth financing are capital signals; acquisition price and GPU inventory are ND. Confidence: medium. Building B, 142 MW campus, financing
Under construction — Scala Tamboré, São Paulo, Brazil. Scala reports a 600 MW master-plan campus and a 560 MW substation under construction; the substation is utility infrastructure, not IT load. Scala operates with DigitalBridge platform backing. The R$6.2B committed state investment is not allocated to one building or GPU procurement. Confidence: medium. Substation, investment
Permitted planned — Scala AI City, Eldorado do Sul, Brazil. Brazil approved a 5 GW power connection; a separate evidence packet reports 54 MW of initial IT capacity. The connection is electricity interconnection, not IT load or energized capacity. Scala operates with DigitalBridge platform backing; the US$500M initial investment is phase program scope, not a GPU price. Confidence: medium. Scala agreement, 5 GW approval
Announced planned — Vantage Frontier, Shackelford County, Texas, US. Vantage describes a 1.4 GW campus and US$25B program; IT, grid, schedule, and project financing are not separated. Vantage is a DigitalBridge-backed developer operator. Confidence: low to medium. Vantage, DigitalBridge
Announced planned — CloudHQ Querétaro, Mexico. A 900 MW campus and US$4.8B investment are reported, but IT versus utility basis, permitting, schedule, financing, and GPUs are ND. Confidence: low to medium. CloudHQ, reported investment
Contracted multi-site — Ascenty AI expansion, São Paulo region, Brazil. Ascenty reports 150 MW of contracted AI data-center capacity and about US$1.2B of AI contract investment signal; site allocation, IT basis, grid capacity, GPU inventory, and capex split are ND. Brookfield is the platform sponsor and Vertical Data is a disclosed partner. Confidence: medium. Ascenty contracts, Vertical Data partnership
Announced planned — Google Canelones, Uruguay. Construction is confirmed by Google and local authorities; service date, IT load, grid MW, GPU inventory, and Google’s investment allocation are ND. A secondary report cites US$850M, which is not treated as verified capex. Confidence: medium. Google, Uruguay XXI, Canelones

EMEA

Operating expanded — Nebius Mäntsälä, Finland. Nebius reported expansion to 75 MW and up to 60,000 GPUs; the MW basis is not separately labelled IT versus total electrical capacity, and the H100 H200 split is unknown. Nebius operates; site financing and GPU prices are ND. The reported more than US$1B annual revenue potential is revenue, not capex or price. Confidence: medium. Nebius 75 MW, Nebius update, H200
Operating reported site deployment — CoreWeave–MERLIN Edged Barcelona, Spain. The facility was inaugurated on 13 May 2025 with 10,224 H200s announced for a 15 MW facility; current 2026 stock is ND. CoreWeave is cloud operator; MERLIN Properties and Edged are facility counterparts; site financing and purchase price are ND. Confidence: medium. CoreWeave
Operating reported live — Sines Data Campus, Portugal. Start Campus lists a 37.5 MW SIN01 facility inside a 1.2 GW campus with national-grid access. The 1.2 GW is a campus headline, not energized MW; renewable claims are not a generation or PPA quantity. Start Campus operates; REN is the reported grid counterpart; the stated €8.5B is combined investment value, not tariff or GPU price. Confidence: medium. Start Campus
Operating reported live — Cassava AI Factory, Cape Town, South Africa. Cape Town was reported live in May 2026 with 3,000 NVIDIA GPUs for the South African first phase; the model split and site IT load are ND. A separate Johannesburg 20 MW target is not assigned the 3,000 GPUs. Cassava Technologies Africa Data Centres operates; NVIDIA and Stanlib strategic investments are reported without terms; the broader up-to-US$720M program signal is not site capex or GPU price. Confidence: medium. Cassava, ITWeb
Announced construction milestone — Stargate Norway, Kvandal near Narvik. Nscale and Aker describe a 20 MW initial phase, 230 MW planned build-out, and a further 290 MW expansion option; these are not a normalized IT-load total. Nscale Aker plan a 50 50 JV with OpenAI as infrastructure customer partner. The plan targets 100,000 NVIDIA GPUs by end-2026 and about US$1B of initial commitment including up to US$250M equity; no GPU price is disclosed. Confidence: medium. Nscale, OpenAI Norway
Contracted target construction — Mistral Compute, Bruyères-le-Châtel, France. The site is described as 40 MW in 2025 and 44 MW of powered capacity in 2026; IT load is not split. Public disclosures conflict between 13,800 individual GB300 GPUs and 18,000 Blackwell Blackwell Ultra systems. Eclairion owns runs, Fluidstack is the infrastructure partner, Mistral is the customer developer, and a seven-bank consortium provided US$830M debt. That debt supports infrastructure hardware and is not a GPU price. Confidence: medium. Fluidstack Eclairion, CNBC, NVIDIA France
Under construction power secured — Pure DC SJK01, Seinäjoki, Finland. Phase one is 110 MW IT, with a 550 MW-plus IT pathway and more than 700 MVA renewable-power access. MVA is retained as MVA and not converted to MW. Pure DC SDC Ventures is developer operator; the more than €1.5B phase-one figure and separate €1.3B financing are capital signals, not electricity prices. Confidence: medium. SJK01 launch, financing
Announced planned — Nebius Lappeenranta AI Factory, Finland. Nebius announced a 310 MW AI factory with first customer capacity expected in 2027. The MW is facility AI-factory capacity, not separately reported IT load; site GPU count, utility MW, and financing allocation are ND. A reported US$10B project framing is secondary and not independently confirmed company guidance. Confidence: medium. Nebius, Reuters
Construction progress planned — Stargate UAE, Abu Dhabi, UAE. The program describes a 1 GW AI cluster inside a 5 GW UAE–US campus; 200 MW was expected to come online in 2026, but current commissioning is not independently confirmed. G42 leads; Khazna develops; OpenAI, Oracle, NVIDIA, Cisco, and SoftBank are alliance partners. GPU count and project capex are ND. Confidence: medium. OpenAI UAE, G42 update
Announced planned program — HUMAIN–NVIDIA AI factories, Saudi Arabia. The Kingdom-wide program targets up to 500 MW over five years and names 18,000 GB300 units in the first phase, with several hundred thousand NVIDIA GPUs over five years. Site, utility interconnection, delivery status, financing, and whether “GB300” means individual GPUs or systems remain unresolved. HUMAIN is a Public Investment Fund subsidiary and NVIDIA is the technology supply partner. Confidence: medium. NVIDIA HUMAIN
Announced planned — DataVolt NEOM Oxagon, Saudi Arabia. The first phase is 300 MW and the wider campus 1.5 GW, with operations targeted for 2028. These are facility campus capacities, not IT load or energized power. DataVolt develops operates with NEOM as host development partner; a reported US$5B agreement investment is not a GPU price. An August 2026 construction headline could not be page-verified, so status remains planned announced. Confidence: low to medium. NEOM, Data Centre Magazine, AGBI lead

APAC and Oceania

Operating — Yotta NM1 Shakti Cloud, Navi Mumbai, India. Current operator material reports 52 MW IT power, while a November 2024 page reports 36 MW IT power; these are a dated scope conflict, not an additive range. Yotta’s initial cluster was more than 4,000 H100s and the advertised platform target reached 16,384 or more, but neither is a 2026 stocktake. Yotta Data Services Hiranandani Group operates with NVIDIA as partner; site finance and GPU price are ND. Confidence: medium. NM1 facility, H100 receipt, NVIDIA profile
Operating phased — YTL Green Data Center Park and AI Cloud, Johor, Malaysia. The first NVIDIA-powered facility was reported operational at 72 MW; the park is described in dated material as 500 MW and later up to 600 MW, with an earlier 500 MW on-site solar plan. These are different park scopes and not current energized IT totals. YTL Power YTL Data Centers develops and operates; NVIDIA is technology partner; a reported US$4.3B deal is a program signal, not 72 MW capex. GB200 NVL72 is named, but count is ND. Confidence: medium. YTL status, Malaysia location, GB200 announcement
Operating — Nxera DC Tuas, Singapore. Nxera opened the facility on 9 February 2026; source-defined AI-ready capacity is 58 MW with no clean IT facility or utility split. Singtel Nxera operates; a S$643M green loan is site-specific financing, while KKR’s 20% stake at a S$5.5B regional-platform valuation is not DC Tuas capex. H100 clusters are offered but count is ND. Confidence: medium. Nxera opening, green loan, KKR stake
Operating expanded — BDx CGK4, Jatiluhur, Indonesia. Phase one launched in 2024. Dated releases describe a 500 MW campus, a later 650 MW renewable-aligned scope, and 788 MVA contracted grid power within a 1.2 GW BDx portfolio commitment. These are separate scopes; MVA is not converted to MW and phase-one IT load is ND. BDx Indosat Ooredoo Hutchison Lintasarta operate the JV; a US$320M loan led by Bank Permata, BCA, and KB Bank supports portfolio expansion and refinancing. H100-class deployments are named, count ND. Confidence: medium. BDx launch, PLN power commitment, financing
Operating — KDDI Osaka Sakai, Osaka, Japan. KDDI began operations on 22 January 2026. A secondary report describes a 48 MW facility, retained as facility power rather than IT load. KDDI operates; Sharp sold the site and HPE Supermicro are technology partners; a reported ¥10B acquisition price is a site transaction, not AI fit-out or GPU price. GB200 NVL72 is installed; rack and GPU count are ND. Confidence: medium. KDDI, reported 48 MW and price
Under construction — SoftBank Tomakomai, Hokkaido, Japan. SoftBank reported more than 300 MW of expected power capacity and 300 MW secured; a secondary report describes a 50 MW first phase. IT load, grid import detail, and site GPU count are ND. SoftBank operates; METI selected the project for a ¥30B subsidy; about 4,000 Hopper GPUs are an operator-level Japan platform count, not a Tomakomai inventory. Confidence: medium. SoftBank construction, power notice, METI program
Under development — AirTrunk SYD3, Western Sydney, Australia. AirTrunk describes 320-plus MW of IT load in 2024 and 330-plus MW in a later platform page; these are dated planning updates, not additive or proof of live load. The site is planned around a 132 kV substation with no separate MW interconnection disclosed. AirTrunk is Blackstone-backed; Blackstone and CPP Investments agreed to acquire the platform for A$24B; a separate A$4.6B sustainability-linked loan covers multiple developments. GPU inventory is ND. Confidence: medium. SYD3, platform acquisition
Under construction last reported milestone — SK Group AWS Ulsan, South Korea. Groundbreaking was reported on 29 August 2025. A secondary index labels the project 103 MW, but official sources do not define that as IT or utility import; GPU inventory and financing are ND. A reported US$5.13B build has not been confirmed in official evidence. Do not allocate South Korea’s separate 260,000-chip national program to Ulsan. Confidence: low to medium. Korea Times, SK Group, secondary project record
Announced contracted target — Yotta D2, Greater Noida, India. Yotta committed 20,736 liquid-cooled B300s for an August 2026 target, while a later page claims 30,000; installed count is ND. The site is described as 60 MW scalable to 250 MW, with facility grid IT basis unresolved. Yotta operates; NVIDIA is the commercial partner; more than US$2B of deployment investment and a separate greater-than-US$1B engagement are not GPU prices. Confidence: medium. 20,736 B300 release, later 30,000 claim
Permitted planned — Datagrid Southland AI factory, Invercargill, New Zealand. Resource consent and a 280 MW planned campus are reported, with a separate 140 MW Mercury supply arrangement. Neither is verified live IT load; utility interconnection MW and strike price are ND. Datagrid develops; Mercury is the power partner; a reported NZ$53M investment is financing context, not total capex or power price. Confidence: low to medium. Datagrid campus, Mercury arrangement
Announced planning approval subject — NEXTDC S7, Eastern Creek, Australia. NEXTDC says the site could accommodate approximately 550 MW subject to development approval; this is potential facility capacity, not firm IT load or interconnection. NEXTDC is developer operator; OpenAI is an MoU partner, not a disclosed equity investor. The A$353M site purchase is land site acquisition, not total campus capex; third-party JV financing remained under evaluation. Confidence: medium. S7 acquisition, OpenAI MoU
Reported planned portfolio — Xinjiang, China. Secondary reporting describes approximately 115,000 NVIDIA chips across a multi-site desert portfolio, with different reports citing 36 or 39 sites. IT load, electricity, operator, financing, procurement status, and model mix are ND. This is a high-uncertainty diligence lead, not verified construction or operation. Confidence: low. Bloomberg, DatacenterDynamics
Program portfolio — South Korea national AI infrastructure. A 31 October 2025 announcement identifies a 260,000-GPU chip supply or deployment target involving NVIDIA, the government, Samsung, SK Group, and Hyundai. Model mix, sites, delivery, MW, ownership, financing, and price are ND. This program is not added to Ulsan or any other campus row. Confidence: low to medium. NVIDIA, Korea Times

Power, interconnection, and system constraints

Susquehanna, Pennsylvania, US: Talen describes a 960 MW AWS campus-development power-consumption threshold and a separate 1,920 MW nuclear PPA through 2042, ramping to full volume by 2032. The figures are not additive and neither is current IT load. Talen campus, Talen Amazon PPA
Crane Clean Energy Center, Pennsylvania, US: Microsoft’s PPA covers about 835 MW of future carbon-free nuclear generation for 20 years; it is generation PPA supply, not current data-center IT load. Restart and approvals remain material conditions. Constellation Microsoft filing
Ireland: CRU reports data centers at 22% of national electricity demand in 2024; EirGrid’s forecasted contracted demand rises from 9.4 TWh in 2025 to 14.6 TWh in 2034. These are system constraints, not project capacity. CRU policy, EirGrid policy
Singapore: IMDA’s 2024 roadmap targets at least 300 MW of additional data-center capacity and reports more than 1.4 GW existing capacity; a March 2026 response says there is no predetermined national ceiling and new capacity is primarily allocated through competitive applications. The 300 MW is a policy objective, not a project or energized-load total. IMDA roadmap, MDDI response

GPU inventories and specifications

There is no defensible public global installed-GPU total in this evidence set. The records below keep operating claims, historical deliveries, orders, commitments, future targets, company-level fleets, custom accelerators, and cloud topology ceilings separate.

xAI Colossus: 180,000 to 200,000 H100s are conflicting first-party claims; no point estimate is selected. H100 reference: 80 GB HBM3 and 3.35 TB s. xAI operates; investors named in company financing include Valor Equity Partners, Andreessen Horowitz, and Sequoia; GPU purchase price is ND. xAI, H100
CoreWeave Barcelona: 10,224 H200s were announced for the 15 MW Barcelona facility; current stock is ND. H200 reference: 141 GB HBM3e and 4.8 TB s. CoreWeave operates with MERLIN Edged facility roles; site price and OEM quote are ND. CoreWeave, H200
Israel-1: 2,048 H100s across 256 HGX H100 systems were reported operating; this is separate from the 1,000-B200 first-phase national program. Project financing and price are ND. Israel-1 report, Israel program
Yotta NM1: More than 4,000 H100s were disclosed in the March 2024 initial cluster, with historical targets reaching 16,384 or 32,768 depending on source and date. Current 2026 stock is ND. Yotta Hiranandani Group is the operator backer context and no site GPU quote was disclosed. Yotta H100 receipt, Yotta NVIDIA plan
Yotta D2: 20,736 liquid-cooled B300s were committed for an August 2026 target; a later Yotta page says 30,000. Installed count is ND. No OEM price is public. Yotta 20,736, Yotta later claim
DataSection Thailand: 4,696 B200s across 587 servers were reported ordered procured; delivery and commissioning are ND. B200 reference: 192 GB HBM3e and 8 TB s. DataSection operates; ASRock Rack is a supplier, not a disclosed investor; price is ND. DataSection filing, B200
Mistral France: 13,800 individual GB300s in 2026 financing coverage conflict with an earlier 18,000-system claim. GB300 NVL72 reference: 72 Blackwell Ultra GPUs and 36 Grace CPUs. Eclairion Fluidstack Mistral roles are disclosed; the US$830M bank debt is financing, not GPU price. CNBC, GB300
Cassava South Africa: 3,000 NVIDIA GPUs are attributed to the first South African phase; H200 SXM5 is reported in secondary coverage, but model split and installed count are ND. Cassava Africa Data Centres operates; NVIDIA and Stanlib investment terms are ND; the broader up-to-US$720M program value is not a GPU price. Cassava, DCD
Nscale Aker Norway: 100,000 NVIDIA GPUs are targeted by end-2026, with a separate more-than-30,000 Rubin tranche for Microsoft scheduled for 2027 on the same campus. Neither is treated as delivered, and the tranches are not added. Nscale Aker are planned JV owners; OpenAI and Microsoft are disclosed customers partners; the approximately US$1B initial commitment is not GPU price. Nscale, Microsoft tranche
Crusoe Lancium Abilene: First GB200 racks were reported delivered live, while separate reports cite more than 450,000 future GB200 GPUs, a 100,000-GPU fabric, and up to 50,000 systems racks per building. Installed count is ND and claims are not additive. Crusoe Lancium operates with Oracle OpenAI relationships; Blue Owl and Primary Digital provide capital context; the US$15B JV is not a GPU price. Crusoe live, GB200 report, GB200
HUMAIN NVIDIA Saudi Arabia: First phase names 18,000 GB300s and a longer-term several-hundred-thousand-GPU program, all forward-looking. HUMAIN is a PIF subsidiary and NVIDIA the technology partner; no site finance or OEM price is disclosed. Do not multiply 18,000 by the 72-GPU rack specification. NVIDIA HUMAIN, GB300
NVIDIA Noetra Japan: 27,500 Rubin GPUs plus 13,750 Vera CPUs are planned; delivery, site, and energized state are ND. NVIDIA and Noetra are named with Japanese government industrial partners; a planned up-to-US$1B NVIDIA investment is not a GPU price. Japan announcement, Vera Rubin
AWS Rainier: Nearly 500,000 Trainium2 chips are an operating AWS claim, not GPUs. Trainium2 has 16 chips in a Trn2 instance and 64 in an UltraServer; chip price and site power are ND. AWS Rainier, Trainium2
Google Ironwood: Google describes 256- and 9,216-chip configurations; the 9,216 figure is a configured pod ceiling and nearly 10 MW, not a global fleet. Ironwood reference: 192 GB HBM and 7.37 TB s. It is a TPU, not a GPU; purchase price is ND. Ironwood overview, Ironwood architecture
Oracle Blackwell cloud: Up to 131,072 NVIDIA Blackwell GPUs is an advertised topology availability ceiling, not a verified installed fleet. Site, investor, financing, and price are ND. Oracle

Possible price signals — kept in separate classes

GPU purchase proxies: A provider comparison gives H100 PCIe at about US$25,000–30,000-plus per card; secondary screening gives H100 PCIe US$25,000–30,000, H100 SXM5 US$35,000–40,000, DGX H100 US$300,000–460,000, DGX H200 US$400,000–500,000, and B200 US$30,000–50,000. These are secondary or provider estimates, not NVIDIA OEM list prices and not project quotes. RunPod H100 comparison, secondary market screening
GPU cloud rentals: Lambda lists about US$3.99 per H100 GPU-hour and US$6.69 per B200 GPU-hour; RunPod lists about US$2.89 H100 PCIe, US$2.99 H100 SXM, US$4.39 H200, US$5.89 B200, and US$7.39 B300 per GPU-hour; CoreWeave lists about US$50.44 per 8-GPU H200 instance-hour and US$68.80 per 8-GPU B200 instance-hour. These are volatile rentals, not purchase prices, leases, capex, or electricity tariffs. Lambda, RunPod, CoreWeave
Bundled cloud instance: AWS lists a P5.48xlarge at US$98.32 hour for eight H100s in the captured price snapshot, or a derived US$12.29 per GPU-hour. The instance includes CPU, networking, and system economics. AWS on-demand pricing
Critical-IT capex benchmark: TeraWulf Fluidstack cited US$8M–10M per MW of critical IT load for a project described as 168 MW critical IT and 240 MW gross. Digital Realty Blackstone and Equinix xScale provide broader arithmetic proxies of about US$14M MW and more than US$10M MW, respectively. These are not project-specific quotes for the rows above. TeraWulf Fluidstack, Digital Realty Blackstone, Equinix xScale
Electricity and colocation benchmarks: The May 2026 EIA observation gives US industrial electricity at 8.71 cents kWh and Texas at 6.33 cents kWh. Singapore’s regulated tariff for 1 July–30 September 2026 is 31.91 Singapore cents kWh before GST. CBRE reports average wholesale asking rates of US$196.25 kW-month in North America for 250–500 kW in H2 2025 and US$403 kW-month in Singapore in Q1 2026. None is a named-project tariff or PPA strike price. EIA, SP Group, CBRE North America, CBRE global
Project finance and valuation signals: Mistral’s US$830M debt, Yotta D2’s more-than-US$2B deployment investment, Nscale’s approximately US$1B initial commitment, Abilene’s US$15B JV, Meta Hyperion’s approximately US$27B earlier development cost US$50B current program report, DataVolt’s reported US$5B agreement, and AirTrunk’s A$24B platform acquisition are financing, capex, or valuation scopes. They must not be divided by GPU counts or MW to manufacture a GPU, power, or lease price.

Product specification anchors

The specifications below identify the named hardware; they are not additional inventory.

- NVIDIA H100: 80 GB HBM3, 3.35 TB s. - NVIDIA H200: 141 GB HBM3e, 4.8 TB s. - NVIDIA B200 DGX B200: 192 GB HBM3e, 8 TB s; DGX B200 is an eight-GPU system with about 14.3 kW maximum system power. - NVIDIA GB200 NVL72: 72 Blackwell GPUs and 36 Grace CPUs, about 13.4 TB HBM3e, 130 TB s rack NVLink. - NVIDIA GB300 NVL72: 72 Blackwell Ultra GPUs and 36 Grace CPUs, about 20 TB GPU memory, 37 TB fast memory, 130 TB s NVLink. - NVIDIA Vera Rubin NVL72: 72 Rubin GPUs and 36 Vera CPUs, about 20.7 TB HBM4 and 260 TB s rack NVLink; per-GPU reference is 288 GB HBM4 and 22 TB s. Vendor specifications are subject to change. - AMD Instinct MI350X MI355X: MI350X is cited at 288 GB HBM3E, 8 TB s, and 1,000 W typical board power; MI355X is cited at 288 GB HBM3E and 8 TB s. No site-level AMD inventory was found. - AWS Trainium2: Trn2 instances use 16 chips and UltraServers use 64; it is a custom accelerator, not a GPU. - Google Ironwood TPU: 192 GB HBM and 7.37 TB s; a configured 9,216-chip superpod is nearly 10 MW. It is a topology ceiling, not a global fleet. Product sources: H100, H200, DGX B200, GB200 NVL72, GB300 NVL72, Vera Rubin, AMD MI350X, AMD MI355X availability, Trainium2, Ironwood

Method, coverage, and confidence

The common evidence cutoff is 14 August 2026 UTC. The EMEA memo’s direct-source cutoff is 5 August 2026, with a 14 August refresh that preserved URLs and treated new search-result titles as leads rather than page-verified facts.
Only 1 GW = 1,000 MW was normalized. MVA, GVA, kV, TWh, facility MW, critical IT MW, grid interconnection MW, generation MW, PPA supply, and policy allocations remain separate. No power factor, PUE, or GPU-per-MW conversion was assumed.
“Operating” means the cited source describes operation, opening, energization, or live service; it is not an audited stocktake. “Contracted” means a lease, PPA, supply option, or power commitment; it is not energized IT load. “Announced” and “planned” remain forward-looking. Program and platform figures are not assigned to sites without explicit allocation.
High-confidence claims are official, dated, and scope-clear. Medium-confidence claims are official but basis-ambiguous or supported by a reported operating milestone. Low-confidence claims rely on secondary reporting, search leads, or unresolved project boundaries. ND is not zero.
Coverage is a bounded sample spanning the United States, Canada, Mexico, Brazil, Colombia, Chile, Uruguay, Europe, the Middle East, Africa, India, Japan, South Korea, Southeast Asia, Australia, New Zealand, China, and global cloud offerings. It is not a census and does not estimate market share.
The main gaps are current energized MW, utility and interconnection records, IT-versus-facility definitions, PPA strike prices and tariffs, delivered GPU bills of material, site-level ownership and financing allocations, OEM purchase quotes, and current status after target dates.

Unresolved risks

xAI’s 180,000 versus 200,000 H100 claims, Yotta D2’s 20,736 versus 30,000 B300 claims, Mistral’s 13,800 GPU versus 18,000 system claims, and Abilene’s multiple GB200 deployment claims cannot be reconciled from the retained public evidence.
HUMAIN’s 18,000 GB300 wording may refer to individual GPUs or systems; it must not be multiplied by the 72-GPU rack specification.
Norway’s 100,000-GPU target and separate 30,000-plus Rubin tranche, South Korea’s 260,000-chip program, and China’s reported 115,000-chip portfolio are program or portfolio claims, not installed site totals.
1.2 GW Abilene, 5 GW Richland compute, 5 GW Scala connection, 1 GW Stargate UAE cluster, 500 MW HUMAIN program power, 550 MW S7 potential, 280 MW Datagrid campus, 788 MVA BDx grid power, and 300 MW Yotta D2 scalability use different boundaries and must not be summed.
Named-project electricity tariffs, connection fees, PPA strike prices, GPU purchase quotes, and final financing allocations remain mostly undisclosed. Current price signals are rentals, secondary estimates, benchmark rates, or broader capital scopes.
Several current source pages could not be independently re-opened during the refresh; no status was upgraded solely from a search result. The unresolved items therefore remain visible rather than being presented as verified current facts.

The report is a public-evidence candidate, not a bankable estimate of global available power, installed GPUs, or total project cost.

Selected capacity scale is shown below only to make the different source-defined claims legible; the values are not additive.

Report details

<sub Coverage cutoff: 14 August 2026 UTC · direct-source EMEA cutoff: 5 August 2026 with refresh · owner: General Codex Agent · run: looprun-a6027d3d-d990-4595-ae7a-96b240f3fad1 · session: synthesize global report · publication state: candidate pending service-authored evaluation.< sub