Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-08-07

Kempax Research

Loop report · 2026-08-07

Global data-center capacity, electricity, and GPU inventory — 7 August 2026

Report date: 2026-08-07 UTC. Evidence cutoff: 2026-08-06 UTC.

Executive summary

The evidence does not support one defensible global capacity or GPU total. This is a bounded, representative evidence set rather than a census, forecast, or audited operating inventory. The rows mix operational IT load, critical IT load, facility or campus capacity, utility interconnect capacity, generation, contracted power, cloud topology ceilings, installed claims, and future targets. Several rows are nested: for example, the 1.2 GW Abilene grid interconnection, its 200+ MW initial phase, and the separate 672 MW Microsoft critical-IT factory are on the same site; SINES’ 1.2 GW secured IT figure includes phased buildings; and ODATA’s 200 MW energized infrastructure is not its 300 MW IT maximum.

The largest disclosed signals in this sample are:

Announced or planned: DataVolt NEOM Oxagon describes a 1.5 GW factory with US$5bn phase-one investment, but the source does not define the 1.5 GW as IT load, facility load, grid interconnect, or generation. NEOM DataVolt source · EMEA APAC and power memos, workspace research data-centers apac-latam latest.md , workspace research power latest.md
Secured contracted: Start Campus reports 1.2 GW of secured IT capacity with grid power at SINES, while Crusoe Lancium reports 1.2 GW of total power capacity and grid interconnection at Abilene. These are different capacity bases and are not additive. SINES source · Abilene source
Under construction or expanding: Meta describes Hyperion Richland as expanding from an earlier over-2-GW compute-capacity statement to 5 GW of compute capacity , without identifying the figure as IT or facility load. Meta Blue Owl disclosed approximately US$27bn of development costs. Meta project source · JV source
Operational energized examples: Applied Digital reports 100 MW critical IT load operational at Polaris Forge 1 Building 1; STT reports 24 MW IT load for Jakarta 2; ODATA reports 200 MW energized in its first Querétaro energy phase. These values measure different things and are not a regional subtotal. Polaris source · STT source · ODATA source
GPU quantity evidence: xAI’s Colossus page contains both 200,000 and 180,000 NVIDIA H100 claims; AWS reports nearly half a million Trainium2 chips at Project Rainier, which are not GPUs; Yotta reports more than 4,000 H100s in its first NM1 cluster; Stargate Norway targets 100,000 NVIDIA GPUs by end-2026; and NVIDIA announces a Japan Noetra plan for 27,500 Rubin GPUs . Installed, delivered, target, and cloud-configuration states are kept separate below. GPU inventory memo
Pricing: No reliable public OEM list price or signed GPU purchase price was found in the bounded primary evidence. The report therefore separates project financing, cloud rental quotes, and secondary-market estimates; none is a GPU procurement price unless explicitly labeled.

Evidence state and confidence

Capacity state uses the upstream power memo’s taxonomy: R = source-reported operational or energized value; A = announced, allocated, secured, contracted, or financed value; M = modeled, forecast, target, or future milestone; I = direct unit normalization only, such as 1 GW to 1,000 MW. A state label does not independently verify a developer claim.

Confidence is an editorial evidence-quality judgment:

High: direct, dated operator, utility, government, or financing disclosure with a relatively explicit unit or stage.
Medium: direct disclosure but future-dated, self-reported, ambiguous on capacity basis, or partly platform-level.
Low: material source conflict, secondary consideration, undated historical status, or a planned figure with little corroboration.

The same project can have high confidence for a product specification and medium or low confidence for current construction status. ND means not disclosed in the reviewed evidence, not zero.

Global project and capacity ledger

Each row is one selected project or campus. Figures preserve the source’s unit, currency, date, and capacity basis. No rows are summed.

North America

Project geography Last disclosed stage and capacity Electricity interconnect generation Company, operator, investors, and price signal GPU AI signal Confidence and evidence --- --- --- --- --- --- Crusoe Lancium Stargate, Abilene, Texas, US First two-building phase reported live 2025-09-30; six-building expansion announced under construction 2025-03-18, current completion not reverified. A 1.2 GW total power capacity and 1.2 GW grid interconnection ; initial phase 200+ MW , with a separate two-building 206 MW disclosure. Neither is labeled IT load. Crusoe expansion · live update Behind-the-meter battery solar, nearby wind, and gas backup are described; generation ratings are ND. Crusoe Lancium developer operator; Oracle OCI customer for OpenAI workloads. Initial phase described as US$3.4bn fully funded with Blue Owl Real Estate and Primary Digital Infrastructure JV sponsors. A later US$15bn second-phase JV was disclosed; scope overlap is unresolved. JV sources · second-phase JV Each building designed for up to 50,000 NVIDIA GB200 NVL72 systems racks and initial materials say up to 100,000 GPUs on one fabric; design early-deployment claims, not installed count. Medium. North America memo · power memo Crusoe Microsoft AI factory, Abilene, Texas, US Site clearing preparation announced 2026-03-27; first building expected to energize mid-2027. Two buildings: A 672 MW critical IT load total, each 336 MW. The same-site projection is 2.1 GW , including the Stargate campus above. Crusoe Microsoft source 900 MW behind-the-meter power plant and battery storage; generation site power, not IT load. Crusoe developer operator; Microsoft supporting customer. Project capex, lender, and equity investor ND. Next-generation AI architectures mentioned; GPU model count ND. Medium. North America memo Applied Digital Polaris Forge 1, Ellendale, North Dakota, US Building 1 reported operational after two 50 MW RFS phases; Buildings 2 and 3 were scheduled for mid-2026 and 2027, not reverified. R 100 MW critical IT load ; A 400 MW contracted critical IT load across three leases; campus engineered to 1 GW and has 1+ GW in load studies. RFS source Final utility interconnect, facility load, generation, PUE, and tariff ND in the reviewed release. Applied Digital owner operator; CoreWeave lessee. Macquarie preferred-equity facility up to US$5bn , with US$112.5m initial draw for the 400 MW build-out. Approximately US$11bn anticipated lease revenue is commercial revenue, not capex. leases · financing GPU model count ND. Medium. North America memo · power memo Meta Hyperion, Richland Parish, Louisiana, US Under construction and expanding. Earlier source: over 2 GW compute capacity ; July 2026 update: 5 GW compute capacity . The two statements are not additive and neither defines IT versus facility load. Richland update Entergy agreement covers over 5,200 MW of new gas combined-cycle plants, transmission, storage, nuclear uprates, and up to 2,500 MW renewables ; regional power package, not a direct Hyperion interconnect or IT load. Meta and Blue Owl funds 80 20 JV; Meta construction property manager. Approximately US$27bn development costs; approximately US$7bn Blue Owl cash contribution; PIMCO and other bond investors debt-funded part of Blue Owl capital. JV · Entergy agreement GPU model count ND. Medium. North America memo · GPU memo Meta El Paso, Texas, US Under construction; expected online 2028. A 1 GW compute capacity , basis unspecified. Meta announcement Direct electrical interconnect, generation, and utility capacity ND. Meta initial occupant operator; BlackRock-related funds, Global Infrastructure Partners, and HPS are disclosed partners. Approximately US$14bn total development cost; Meta investment above US$10bn; approximately US$4.9bn BlackRock cash contribution and US$12.5bn debt financing are transaction figures, not an additive capex total. GPU model count ND. Medium. North America memo Meta Sturgeon County, Alberta, Canada Meta reported ground broken 2026-07-08; Alberta major-project database still listed the project as Proposed at access, so administrative status conflicts. A 1 GW AI-optimized data center , basis unspecified. Meta · Alberta status Meta says it will fund new generation grid infrastructure and match energy use with clean renewable energy; Alberta page describes a gas plant. No separate interconnect MW. Meta operator investor; Greenlight LP, AltaLink, Capital Power, AESO, AUC, Clark Builders and PCL named as energy grid construction partners. CA$13bn+ investment and approximately CA$60m local infrastructure. GPU model count ND. Low–medium due source-level status energy conflict. North America memo Talen Cumulus AWS, Berwick, Pennsylvania, US Campus sold to AWS; buildout remains a development ceiling. A 960 MW power consumption development capacity , not current IT load. Talen campus sale Co-located Susquehanna nuclear supply; later Amazon PPA is 1,920 MW through 2042, ramping by 2032, and is not added to the 960 MW campus figure. PPL transmission reconfiguration noted. Talen Cumulus and AWS. Secondary regional source reports US$650m purchase price, US$350m at close plus US$300m escrow; Talen officially confirms the escrow release but not the total, so consideration is low-confidence. official escrow · secondary price GPU model count ND. Medium for 960 MW; low for US$650m. North America memo · power memo xAI Colossus, Memphis, Tennessee, US Operating facility with expansion utility plan; current energized load not independently reverified. xAI page claims 200,000 H100s in headline and 180,000 H100s in its “by the numbers” section; 1m GPUs by 2026 is a target. xAI Colossus · Memphis page City MLGW deck dated 2024-07-09: 8 MW existing electric capacity , availability rising to 50 MW , developer-funded 150 MW substation , MLGW capital cost US$760,000 , and approximately US$24m margin allowance. These are service substation figures, not IT load. MLGW deck xAI operator; project capex, lender, and investor ND. H100 quantity is unresolved; no single installed count is accepted. Low–medium due internal count conflict and historical utility plan. North America memo · GPU memo Microsoft Fairwater, Mount Pleasant, Wisconsin, US First facility reported fully operational 2026-06-23; second in foundation steel underground work, scheduled for 2028. Microsoft completion Data-center IT facility MW ND. 250 MW solar project is planned generation supply, not data-center load. Microsoft Azure operator and investor; National Grid Renewables energy partner. US$3.3bn Mount Pleasant investment through 2026 and US$4.7bn local hyperscale construction 2024–28 are differently scoped. investment Massive NVIDIA GB200 cluster described; each rack has 72 Blackwell GPUs ; total racks GPUs ND. technical overview High for stage rack spec; low for capacity inventory. North America memo QScale Q01, Lévis, Québec, Canada Operating campus; Building B under construction. R 142 MW electrical campus capacity , R 114 MW protected client IT load ; Building B adds 60 MW IT . QScale Building B · campus specs Owned 266 MVA 2N 120 kV substation; MVA not converted to MW. Current page quotes approximately US$0.045 kWh , not necessarily a contract tariff. QScale operator; Infrastructure at Goldman Sachs Alternatives portfolio company; HPE AI-cloud partner. Building B $700m build, currency not stated; older project disclosure $867m total including Québec loan equity and Desjardins financing, date-scoped and non-additive. financing Building B designed for NVIDIA GB300-class systems and liquid-cooled cabinets above 600 kW ; installed GPU count ND. Medium. North America memo

Europe, Middle East, and Africa

Project geography Last disclosed stage and capacity Electricity interconnect generation Company, operator, investors, and price signal GPU AI signal Confidence and evidence --- --- --- --- --- --- SINES DC, Portugal — Start Campus SIN01 operational by 2025-04; SIN02–06 phased plan with full campus targeted by 2030; later status not reverified. A 1.2 GW secured IT capacity with grid power ; SIN01 14 MW expanding to 26 MW; later buildings up to 240 MW IT each. Start Campus 1.2 GW · SIN01 REN grid access secured; 100% renewable target; PUE 1.1 target in power memo. No tariff PPA strike price. Start Campus; Davidson Kempner Capital Management and Pioneer Point Partners shareholders; U.S. bank unnamed. €8.5bn construction investment value disclosure. AI HPC and liquid-cooling ready; GPU count model ND. Medium–high for reported capacity; lower for later-stage current status. EMEA memo · power memo Stargate Norway, Narvik Kvandal — Nscale Aker OpenAI Announced 2025-07-31; preparatory work construction mobilization described, not operational in source. M A 230 MW planned , M A 290 MW expansion ambition , initial 20 MW ; expansion is not added. Nscale · OpenAI Renewable hydropower; limited local transmission capacity noted, no interconnect MW. Nscale Aker planned 50 50 JV; OpenAI initial offtaker. Approximately US$1bn initial 20 MW phase, including up to US$250m equity on a 100% basis; no GPU price. M 100,000 NVIDIA GPUs targeted by end-2026; model variant and delivered quantity ND. Medium. EMEA memo · power memo Glomfjord, Norway — Nscale Current facility page describes 30 MW operational capacity , expandable to 60 MW ; page is undated. 100% renewable hydropower; interconnect MW ND. Nscale owner operator; investors and price ND. Commercial page says access to “thousands of GPUs”; service statement, not installed inventory, and models ND. facility page Medium–low due undated page and access inventory distinction. EMEA memo AWS Europe Spain Aragón region — Amazon Web Services Region operational since 2022; new Zaragoza, Huesca, and Teruel facilities announced 2026, site status MW ND. 100% renewable matching stated; interconnect MW ND. AWS Amazon. €33.7bn is a multi-year Spain cloud AI data-center and supply-chain program, not site-level capex. AI ML server manufacturing and repair facility planned; GPU model count ND. Amazon Spain Medium. EMEA memo Yondr London campus, Slough, UK First building 20 MW RFS 2024-07; second building 10 MW of 30 MW live 2025-04; third 40 MW broke ground 2025-02 and expected mid-2026, not reverified. Campus 100+ MW ; capacity type not stated. Yondr first · second No interconnect MW; second building PUE 1.21 and HVO solar features reported. Yondr developer owner operator; investor, lender, and price ND. GPU model count ND. Medium. EMEA memo Yondr Bischofsheim Frankfurt, Germany 40 MW site fully operational reported 2025-07; first 20 MW RFS 2024-12. Capacity type not stated. Yondr operational 100% renewable electricity in prior year; interconnect MW ND. Yondr owner operator; investors, lenders, and price ND. GPU model count ND. High for direct stage capacity; medium for load basis. EMEA memo STACK Liederbach Frankfurt, Germany Planned announced 2022; current permitting construction operation not re-confirmed. A 80 MW campus across four facilities; load type ND. STACK Utility interconnect MW ND; excess-heat rainwater design proposed. STACK developer operator; site acquired from Coca-Cola Europacific Partners; finance, capex, and price ND. GPU model count ND. Low–medium. EMEA memo Stargate UAE, Abu Dhabi, UAE — G42 OpenAI Oracle Announced 2025-05-22; 200 MW expected live in 2026 , construction funding close not evidenced in cited release. A M 5 GW UAE–US AI campus , 1 GW compute cluster , 200 MW first cluster; load basis unspecified. G42 · OpenAI Planned mix nuclear, solar, gas; grid interconnect MW ND. G42 builds; OpenAI and Oracle operate; NVIDIA, Cisco, SoftBank partners. Capex debt equity ND. NVIDIA Grace Blackwell GB300 systems named; count ND. Medium–low due future stage and undefined basis. EMEA memo · GPU memo DataVolt NEOM AI factory, Oxagon, Saudi Arabia Agreement announced 2025-02; phase one expected operational 2028; construction not evidenced in cited release. A 1.5 GW factory campus target , basis unspecified. NEOM Entirely renewable ambition; interconnect generation storage substation ratings ND. DataVolt developer investor operator; NEOM Oxagon land infrastructure partner. Phase one US$5bn . High-density energy-efficient architectures; GPU model count ND. Medium–low. EMEA memo · power memo Teraco JB7, Isando campus, Johannesburg, South Africa Construction commenced 2024-11; completion scheduled 2026, not reverified. A 40 MW critical power load ; Isando campus up to 110 MW critical power load . Teraco 68 MVA utility supply ; MVA not converted to MW. Separate 120 MW utility-scale solar PV feeds Eskom wheels power; generation, not JB7 IT load. Teraco Digital Realty; R8bn syndicated loan through Absa and other institutions, not stated as JB7-only capex. GPU model count ND; high-density AI readiness only. Medium–high for disclosed units; medium for current completion. EMEA memo Raxio UG1, Namanve near Kampala, Uganda Operational since 2021. R 1.5 MW IT power , 2 MW cooling, 400 racks, 1,000 m² white space. Raxio Uganda 33 kV utility supply directly from nearest substation; 2N redundancy; interconnect MW ND. Raxio operator; platform-level Roha Meridiam ownership and IFC financing, not site-specific. No site price. GPU model count ND. High for direct site figures. EMEA memo Raxio ET1, Addis Ababa, Ethiopia Operational since 2023. R 3 MW IT power , 3 MW cooling, up to 800 racks, 2,000 m² white space. Raxio Ethiopia 33 kV utility supply directly from nearest substation, N+N; interconnect MW ND. Raxio operator; platform-level ownership finance only. No site price. GPU model count ND. High for direct site figures. EMEA memo SJK01, Seinäjoki, Finland — Pure Data Centres SDC Ventures Phase 1 announced 2026-07-14; permissions and power requirements described as secured; first data-hall substation live. A 110 MW phase 1 ; M A 550 MW+ IT pathway subject to permissions contracting. Pure DC More than 700 MVA renewable power access; preserved as MVA, not converted. Pure Data Centres with SDC Ventures. €1.5bn+ phase-one and €7.5bn+ full-campus investment disclosures; not electricity prices. Modular AI-ready direct-liquid cooling; GPU count model ND. Medium. power memo Aurora Energy Precinct AI campus, South Australia Planned target. A M 17 MW starter; A 200 MW anchor after 275 kV connection; M A up to 1 GW longer-term pathway. 1414 Degrees Starter uses existing 33 kV; anchor depends on 275 kV line. 900 MW solar potential and 140 MW 280 MWh battery are separate energy assets, not data-center load. 1414 Degrees and unnamed Australian operator; operator to bring capital technical expertise; project price and tariff ND. AI infrastructure only; GPU count model ND. Medium–low due target unnamed operator. power memo STT Jakarta campus, Indonesia Jakarta 2 commenced operations with R 24 MW IT load ; A 360 MW AI-ready IT pipeline; Jakarta 5 and 6 each 40 MW planned. STT GDC On-site high-voltage substation energized; MW MVA rating ND. Renewable strategy with PLN; tariff ND. ST Telemedia Global Data Centres; green financing mentioned, amount investor lease price ND. Designed for NVIDIA Vera Rubin Vera Rubin Ultra; no inventory count. Medium–high for 24 MW; medium for pipeline. power memo

Asia-Pacific and Latin America

Project geography Last disclosed stage and capacity Electricity interconnect generation Company, operator, investors, and price signal GPU AI signal Confidence and evidence --- --- --- --- --- --- STACK Johor Bahru campus, Iskandar Puteri, Malaysia Planned; initial delivery scheduled Q4 2026, construction not stated. A 220 MW campus , basis unspecified. STACK TNB 275 kV infrastructure and on-site substation; whether 220 MW is IT or utility capacity ND. STACK developer operator; project investor, lender, capex, lease, and power price ND. Cloud AI ML design; GPU model count ND. Medium–low. APAC LatAm memo Empyrion MY1, Johor, Malaysia Regulatory approved planned 2025-12; earliest energization Sept 2026 and first-phase RFS Q4 2026. A 200+ MW campus, five 40 MW buildings; basis unspecified. Empyrion 145 MW initial power allocation from TNB; full-campus scalability and green PPAs described. Empyrion Digital; largest investment to date, amount financing partner ND. AI cloud HPC-ready; GPU model count ND. Medium. APAC LatAm memo PDG JC3, Greater Jakarta, Indonesia Ground broken 2025-11-19; under construction; first phase targeted RFS Q4 2026. A 120 MW AI-ready hyperscale campus , basis unspecified. JC3 Dual-grid PLN feeds; full power capacity reported contracted; voltage interconnect MW ND. Princeton Digital Group; US$1bn campus investment. Platform equity context: Mubadala led 2022 round with US$350m , Warburg Pincus and Ontario Teachers’ also invested; not a JC3-specific financing commitment. PDG equity Direct-to-chip cooling and AI-ready design; GPU model count ND. Medium–high for project disclosure; medium for platform investors. APAC LatAm memo Project Minato, Osaka, Japan — SC Capital Partners Construction commenced after 2025-12-03 groundbreaking; operations scheduled early 2028. A 100 MW allocated power ; IT load ND. SCCP Direct access to robust power infrastructure; utility interconnect name, voltage, and MW ND. SC Capital Partners SC Zeus Data Centers developer manager operator. First phase approximately US$600m , reported fully secured; consortium includes a wholly owned ADIA subsidiary, Japanese real-estate developer, and SCCP RECAP funds. Hyperscale AI-era design; GPU model count ND. Medium–high. APAC LatAm memo AirTrunk OSK2, Osaka, Japan Planned announced; delivery timeline not disclosed. A up to 100 MW total campus IT capacity ; OSK1’s 20 MW is separate context and not added. AirTrunk OSK2 interconnect ND. AirTrunk; acquisition consortium led by Blackstone and CPP Investments with founder Robin Khuda. Approximately US$8bn across existing and planned Japan projects, not OSK2-specific. investors Cloud AI workload design; GPU model count ND. Medium–low due OSK2-specific finance power gaps. APAC LatAm memo Yotta NM1, Panvel Navi Mumbai, India Operational since 2020; first H100 cluster delivered 2024-03. A R 50 MW power , 7,200 racks; five-building park 250 MW power 30,000 racks at full build-out. Power basis is not labeled IT load. NM1 Redundant on-site power; 2020 source cited captive solar and gas co-generation. Current generation interconnect status not reverified. Yotta Infrastructure Yotta Data Services; Hiranandani Group subsidiary context and Nidar Infrastructure ownership in 2025 investor presentation. Project capex, lender, tariff, and GPU purchase price ND. More than 4,000 NVIDIA H100s in first cluster; 32,768 by end-2025 was a target, not verified. Variant not disclosed; do not assign SXM PCIe or convert to MW. H100 release · presentation Medium–high for first-cluster claim; medium for current inventory. APAC LatAm memo · power memo ODATA DC QR03, Querétaro, Mexico First facility launched available in 2025; campus related facilities under construction in cited releases, later status not reverified. A up to 300 MW IT across five buildings; first building 72 MW IT . ODATA launch R 200 MW energized Feb 2025; path toward 400 MW infrastructure, 400 kV switching-substation expansion, 6.2 miles transmission, two substations. Energized electricity infrastructure is not 300 MW IT. ODATA energy ODATA, an Aligned Data Centers company; Ammper renewable-energy partner. Over US$3bn launch disclosure and US$3.3bn energization disclosure; scope date conflict, not a GPU price. Cloud AI cooling density; GPU model count ND. Medium–high. APAC LatAm memo · power memo Ascenty Sumaré Vinhedo São Paulo, Brazil Vinhedo operational expanding; Sumaré 3 construction began 2026-03, delivery targeted Q3 2027. A 150 MW contracted capacity across new facilities; Sumaré 3 90 MW initial plus 90 MW expansion, Vinhedo 2 50→80 MW, Vinhedo 3 90 MW AI, new São Paulo 20 MW. Mixed bases, not summed. DCD report Sumaré 3 has 32 km transmission line from CPFL substation; ONS access permit; interconnect MW tariff ND. Ascenty JV of Digital Realty and Brookfield Infrastructure. US$1.2bn announced investment; no project lender. Sumaré 3 liquid-cooled, up to 1 MW rack; GPU count model ND. Medium. APAC LatAm memo Scala Campus Tamboré, Barueri São Paulo, Brazil Phase 2 inaugurated 2024-08; six data centers operational and four under construction at that date. R 158 MW IT Phase 2; 200 MW operating construction activity then; M A 600 MW full-campus maximum across 17 buildings. Scala Three substations planned for full campus; utility interconnect MW voltage ND; 100% renewable stated. Scala; DigitalBridge-founded sponsored platform. R$6.2bn committed Phase 2 and R$10bn total invested committed including Phase 1; not GPU price. AI ML-ready liquid cooling, 24–100 kW rack; GPU model count ND. Medium. APAC LatAm memo

Company, investor, and financing map

This map links disclosed roles and capital to the project or platform scope stated by the source. Platform investors, lenders, program amounts, and corporate transactions are not assigned to a site by inference.

Project company Developer, operator, customer Investor, lender, utility, or strategic partner Disclosed financing or price category Evidence --- --- --- --- --- Crusoe Lancium Stargate Abilene, US Crusoe Lancium developer operator; Oracle OCI customer for OpenAI workloads. Blue Owl Real Estate and Primary Digital Infrastructure JV sponsors; SoftBank, OpenAI, Oracle, and MGX named in the broader Stargate context. US$3.4bn initial phase fully funded; later US$15bn JV; scope overlap unresolved. These are project JV figures, not GPU or electricity prices. Initial JV · Second JV Abilene capacity Applied Digital Polaris Forge 1, US Applied Digital owner operator; CoreWeave lessee. Macquarie Asset Management preferred-equity provider. Facility up to US$5bn; US$112.5m initial draw; about US$11bn prospective lease revenue. Financing and revenue are not capex, GPU price, or tariff. Macquarie · Lease Operational IT load Meta Hyperion, Louisiana, US Meta developer property manager; Meta Blue Owl funds 80 20 JV. Blue Owl funds; PIMCO and other bond investors debt-funded part of Blue Owl capital; Entergy Louisiana power partner. Approximately US$27bn development cost and approximately US$7bn Blue Owl cash contribution; JV-scoped figures, not GPU price. JV · Power agreement Meta project update Start Campus SINES, Portugal Start Campus developer operator; AI HPC customer allocations remain site-specific. Davidson Kempner Capital Management and Pioneer Point Partners shareholders; REN grid partner. €8.5bn campus construction investment value disclosure; no tariff or PPA strike price disclosed. Campus capacity SIN01 status Pure DC SJK01, Finland Pure Data Centres with SDC Ventures; AI infrastructure campus developer operator. Power and financing partners are not allocated to a specific building in the reviewed release. €1.5bn+ phase one, €7.5bn+ full-campus pathway, and a separate €1.3bn financing disclosure; none is an electricity or GPU price. Launch · Financing Power evidence PDG JC3, Greater Jakarta, Indonesia Princeton Digital Group developer operator. Mubadala, Warburg Pincus, and Ontario Teachers’ Pension Plan are platform-level investors; PLN provides dual-grid context. US$1bn campus investment; platform equity is not JC3-specific financing. JC3 · Platform equity Campus construction Teraco JB7, South Africa Teraco Digital Realty developer operator; AI-ready facility. Absa and other institutions provided a syndicated loan; utility and solar partners are separate from IT load. R8bn syndicated loan across JB7 and other construction; not stated as JB7-only capex or GPU price. Teraco financing JB7 capacity DataVolt NEOM Oxagon, Saudi Arabia DataVolt developer investor operator; NEOM Oxagon host and infrastructure partner. NEOM Oxagon; renewable-power ambition is not a quantified delivered supply. US$5bn phase-one investment; 1.5 GW factory campus basis is unspecified and not a GPU or tariff. NEOM DataVolt Project announcement Raxio platform, Africa Raxio operator; UG1 and ET1 are site rows, while ownership and finance are platform-level. Roha and Meridiam equity; IFC, Proparco, and EAAIF debt context. More than US$380m group committed capital and portfolio financing; not Uganda Ethiopia site capex or GPU price. Capital update IFC platform financing

Electricity, interconnect, and system context

These records are context rather than extra project capacity. They must not be added to the project ledger.

System or benchmark Disclosed fact and date Interpretation confidence --- --- --- Ireland connection policy CRU says data centers were 22% of national electricity demand in 2024 ; EirGrid forecast sector demand from 9.4 TWh in 2025 to 14.6 TWh in 2034 . New connections must provide generation storage to match requested maximum import demand and meet at least 80% of annual demand with additional Irish renewable projects. CRU System policy and forecast, not project delivered MW or IT load. High for policy statement. Singapore Green Data Centre Roadmap IMDA roadmap reports more than 1.4 GW existing data-center capacity , aims to provide at least 300 MW additional capacity near term, and supports facilities with PUE 1.3 or lower . IMDA Policy objective threshold, not a project interconnect or national total. High for dated policy fact. Singapore current policy posture MDDI said 2026-03-03 there is no predetermined upper limit to total data-center capacity; growth is calibrated within resource constraints and new capacity is primarily allocated through competitive Calls for Application. MDDI Do not treat the 300 MW roadmap figure as a fixed cap or automatic queue. High. Singapore regulated tariff benchmark SP Group lists S$0.3191 kWh before GST and S$0.3478 kWh including GST for 2026-07-01 through 2026-09-30. SP tariff · Q3 release Regulated retail benchmark, not a hyperscale contract PPA price and not applicable to other regions. High for tariff date; not project-specific . QScale public electricity signal QScale Q01 page quotes approximately US$0.045 kWh , date not exposed. Public page estimate benchmark, not necessarily a signed tariff; preserve as a price signal only. Medium–low. QScale

GPU inventory specification matrix

The matrix combines the requested specification, quantity state, company cloud OEM role, and possible price category. “ND” means not disclosed or not verified in the reviewed evidence, not zero. Quantities are never added across projects or converted from systems chips to discrete GPUs without an explicit source basis.

Model system Technical specification anchor Quantity status Company, cloud, or OEM Possible price category Evidence --- --- --- --- --- --- NVIDIA H100 80 GB HBM3; up to 3.35 TB s memory bandwidth. xAI page contains 200,000 headline and 180,000 internal H100 claims; Yotta reports more than 4,000 H100s in its first NM1 cluster and later targets. No single audited current total. NVIDIA OEM; xAI and Yotta Shakti Cloud operators. OEM purchase price ND. Cloud observations: US$2.99–3.99 GPU-hour; secondary screening: US$25k–40k card, indicative only. Lambda · RunPod · secondary screen H100 specification · xAI · Yotta NVIDIA H200 141 GB HBM3e; 4.8 TB s bandwidth. AWS advertises P5e P5en scaling within a 20,000 H100 H200 GPU ceiling; this is a cloud topology ceiling, not installed fleet inventory. xAI Kao allocations ND. NVIDIA OEM; AWS cloud; xAI Kao named project contexts. OEM price ND. RunPod observation: US$4.39 GPU-hour; CoreWeave HGX example is per 8-GPU instance-hour. RunPod · CoreWeave H200 · AWS P5 NVIDIA B200 DGX B200 B200 anchor: 192 GB HBM3e and 8 TB s; DGX B200: eight GPUs, 1,440 GB aggregate GPU memory, and about 14.3 kW maximum system power. DataSection reports 4,696 B200 GPUs across 587 servers as a planned deployment record; Israel’s national program reports 1,000 B200 accelerators available allocated. Neither is a live global stocktake. NVIDIA OEM; DataSection ASRock Rack and Israel Innovation Authority allocation contexts. OEM price ND. Cloud observations: US$5.89–6.69 GPU-hour; secondary B200 screening US$30k–50k card, indicative. Lambda · RunPod · secondary screen B200 · DGX B200 · DataSection · Israel program NVIDIA GB200 NVL72 72 Blackwell GPUs + 36 Grace CPUs; 13.4 TB HBM3E; 130 TB s rack NVLink. Abilene describes up to 50,000 systems racks per building and up to 100,000 GPUs on one fabric; source unit and installed quantity remain unresolved. NVIDIA OEM; Crusoe Lancium, Oracle, and OpenAI workload context; YTL also names GB200 NVL72. OEM rack price ND. Abilene US$3.4bn US$15bn figures are project JV finance, not hardware prices. NVIDIA · Crusoe NVIDIA GB300 NVL72 72 Blackwell Ultra GPUs + 36 Grace CPUs; 20 TB GPU memory; 37 TB fast memory; 130 TB s NVLink. Stargate UAE names GB300 systems with count ND. HUMAIN NVIDIA wording includes 18,000 GB300 and other Blackwell quantities, but unit status ambiguity remains. NVIDIA OEM; G42 OpenAI Oracle Cisco SoftBank in UAE; HUMAIN PIF NVIDIA in Saudi Arabia. OEM NVL72 price ND. B300 cloud rental at US$7.39 hour is a standalone proxy, not NVL72 pricing. RunPod GB300 · G42 · HUMAIN NVIDIA NVIDIA Vera Rubin NVL72 Preliminary: 72 Rubin GPUs + 36 Vera CPUs; 20.7 TB HBM4; 260 TB s rack NVLink; specifications subject to change. Japan Noetra announces 27,500 Rubin GPUs and 13,750 Vera CPUs planned; not installed or site-allocated. NVIDIA Noetra and Japanese government industry program; product OEM price ND. OEM and cloud price ND. NVIDIA up to US$1bn and Brookfield NAVER funding are program project amounts, not GPU prices. Rubin program · Rubin spec AMD Instinct MI350X MI355X MI350X: 288 GB HBM3E; 8 TB s; 1,000 W typical board power. MI355X operator listing: 288 GB HBM3E and 8 TB s. MI350X installed quantity orders ND. MI355X is listed on Crusoe Cloud, but physical quantity and shipment status ND. AMD OEM; Crusoe Cloud channel for MI355X. OEM, rental, and verified resale price ND in reviewed evidence. AMD MI350X · Crusoe MI355X AWS Trainium2 not a GPU Trn2 uses 16 chips; Trn2 UltraServer uses 64 chips linked by NeuronLink. AWS reports nearly half a million Trainium2 chips at Project Rainier; this is an operational chip claim, not an NVIDIA GPU count. AWS Amazon custom accelerator; Anthropic customer compute partner. Public chip purchase list price and comparable rental price ND. Rainier · Trn2 Google Ironwood TPU not a GPU 4,614 TFLOPS chip; 192 GB HBM; 7.37 TB s bandwidth. A 9,216-chip pod is 42.5 exaflops and nearly 10 MW. 256-chip and 9,216-chip cloud configurations are available; deployed pod count ND. Google silicon cloud; outside investor and project finance ND. Cloud rate and purchase price ND. The nearly 10 MW value is one pod, not a fleet. Ironwood · architecture power Intel Gaudi 3 not a GPU 128 GB HBM2E; 3.7 TB s; 1,678 TFLOP s FP8 BF16 matrix peak; 24×200G Ethernet; 900 W OAM TDP. PCIe card shipping; quantity and installed inventory ND. Intel OEM; IBM Cloud, Denvr Dataworks, and AWS EC2 channels named. OEM list, cloud rental, and verified resale price ND. Intel · Gaudi white paper

GPU family and system specification ledger

Specifications are product or system facts, not evidence that a given project owns or has installed the quantity. Product availability is not inventory. The evidence cutoff is 2026-08-06 UTC ; individual price observations retain their own capture dates, and preliminary specifications remain subject to change.

GPU accelerator family or system Capacity-planning specification Form factor power Quantity, operator, and price state Confidence and source --- --- --- --- --- NVIDIA GB200 NVL72 36 Grace CPUs + 72 Blackwell GPUs; 13.4 TB HBM3E; 576 TB s aggregate GPU-memory bandwidth; 130 TB s NVLink; NVFP4 1,440 PFLOPS sparse and FP16 BF16 360 PFLOPS sparse. Liquid-cooled rack-scale system; rack power TDP ND on product page. NVIDIA product available; installed quantity, orders, operator allocation, and OEM price ND. High for spec. NVIDIA · GPU memo NVIDIA GB300 NVL72 72 Blackwell Ultra GPUs + 36 Grace CPUs; 20 TB GPU memory and 37 TB fast memory; up to 576 TB s memory bandwidth; 130 TB s NVLink; FP4 1,440 PFLOPS sparse 1,080 PFLOPS dense; 800 Gb s network per GPU. Fully liquid-cooled rack; rack power TDP ND. NVIDIA marks “Available Now”; installed quantity order volume OEM price ND. High for current published spec. NVIDIA · GPU memo NVIDIA DGX B200 Eight Blackwell GPUs; 1,440 GB total GPU memory; 64 TB s HBM3E bandwidth; FP4 144 PFLOPS sparse; 14.4 TB s aggregate NVLink. 10 RU; approximately 14.3 kW maximum system power . Product listed; installed fleet, MSRP, and signed purchase price ND. High for spec power. NVIDIA · GPU memo NVIDIA Vera Rubin NVL72 72 Rubin GPUs + 36 Vera CPUs; 20.7 TB HBM4; 1,580 TB s aggregate bandwidth; 260 TB s rack NVLink; 288 GB HBM4 and 22 TB s per Rubin GPU; FP4 3,600 PFLOPS inference 2,520 PFLOPS training. Liquid-cooled rack; rack TDP ND; specifications preliminary subject to change. NVIDIA says ramping full production and systems shipping; installed fleet MSRP ND. Japan Noetra program separately announces 27,500 Rubin GPUs. Medium–high; preliminary. NVIDIA · GPU memo AMD Instinct MI350X 288 GB HBM3E; 8 TB s bandwidth; 8,192-bit interface; 9.2 PFLOPS MXFP4 MXFP6; 2.3 PFLOPS FP16; 144.2 TFLOPS FP32; 4.6 POPS INT8. OAM; PCIe 5.0 x16 + seven Infinity Fabric links; 1,000 W typical board power . Launched 2025-06-12; installed quantity, orders, MSRP, and project allocation ND. High for spec. AMD · GPU memo AMD Instinct MI355X Crusoe page lists 288 GB HBM3E and 8 TB s bandwidth; complete stable throughput table not used. Form factor TDP ND in cited availability evidence. Available on Crusoe Cloud; quantity, orders, and purchase price ND. Medium. Crusoe · GPU memo AWS Trainium2 Trn2 UltraServer Trn2 instance has 16 Trainium2 chips; UltraServer has four Trn2 instances 64 chips linked by NeuronLink. Per-chip memory, throughput, and TDP ND in cited page. AWS system form factor; power ND. Trn2 UltraServer available; Project Rainier reports nearly half a million operational Trainium2 chips. Not a GPU. AWS price product rental varies by instance and is not a purchase price. High for topology; medium for current quantity. AWS Trn2 · Rainier Google Ironwood TPU 4,614 TFLOPS peak per chip; 192 GB HBM; 7.37 TB s HBM bandwidth; 1.2 TB s bidirectional ICI. 9,216-chip superpod: 42.5 exaflops, 1.77 PB shared HBM, 9.6 Tb s ICI. Liquid-cooled; 9,216-chip pod nearly 10 MW ; per-chip TDP ND. Google Cloud customer configurations 256 or 9,216 chips; deployed pod count global inventory ND; no outside investor or purchase price. High for published system facts. Google · architecture power Intel Gaudi 3 OAM reference: 128 GB HBM2E; 3.7 TB s HBM bandwidth; 96 MB SRAM; FP8 BF16 1,678 TFLOP s; 24×200G Ethernet; PCIe Gen5 x16. OAM 900 W TDP ; PCIe exact board power not used. PCIe card shipping; IBM Cloud, Denvr Dataworks, and AWS EC2 channels named; quantity, transaction price, and installed inventory ND. High for published spec. Intel · white paper NVIDIA H100 H200 through AWS P5 P5.4xlarge: one H100 with 80 GB HBM3; p5.48xlarge: eight H100s 640 GB aggregate HBM3, 3,200 Gbps EFA and 900 GB s NVSwitch. P5e P5en: eight H200s 1,128 GB aggregate HBM3e. Per-GPU TDP ND on AWS page. AWS advertises P5 P5e P5en scaling to 20,000 H100 H200 GPUs and up to 20 exaflops; cloud capacity ceiling, not installed inventory. High for product cloud ceiling. AWS P5 · GPU memo

Disclosed deployments and inventory status

This table is intentionally separate from the product ledger. It distinguishes installed or operational claims, targets, system configurations, and cloud availability.

Site program Quantity or configuration Status, unit, and interpretation Company operator and finance Evidence confidence --- --- --- --- --- xAI Colossus, Memphis 200,000 H100 headline and 180,000 H100 internal “by the numbers”; roadmap to 1m GPUs by 2026 . First-party operating claim with internal conflict; roadmap is a target. No single verified installed total. xAI; project investor financing ND. xAI · GPU memo · Low–medium AWS Project Rainier Nearly half a million Trainium2 chips operational; a dated expectation of more than 1m chips for Claude is not substituted. Operational chip claim, not GPUs; site allocation and current post-forecast count ND. AWS Amazon host; Anthropic customer compute partner; project financier ND. AWS · Medium–high Stargate Norway 100,000 NVIDIA GPUs targeted by end-2026. Future target; model variant, delivery, and installed count ND. Nscale Aker with OpenAI offtake; approximately US$1bn initial phase. Nscale · Medium Yotta NM1, India More than 4,000 H100s in first cluster; 32,768 end-2025 scale target. First-party cluster statement; target fulfillment not independently verified; H100 variant ND. Yotta Shakti Cloud; purchase price ND. Yotta H100 release · Medium–high for first cluster Crusoe Lancium Abilene Up to 50,000 GB200 NVL72 systems racks per building ; initial release says up to 100,000 GPUs on one fabric. Design early-deployment wording; source unit ambiguity prevents conversion to discrete GPUs. Crusoe Lancium Oracle; project JV financing as above. Crusoe · Medium–low for quantity Microsoft Fairwater, Wisconsin Each GB200 rack packs 72 Blackwell GPUs ; total racks GPUs ND. Rack specification, not inventory. Microsoft Azure; price is data-center investment, not GPU price. Microsoft technical overview · High for rack spec QScale Q01, Québec Building B designed for GB300-class systems and liquid-cooled cabinets above 600 kW . Design readiness; no GPU count. QScale; price capex is building project finance, not GPU price. QScale · Medium Stargate UAE NVIDIA Grace Blackwell GB300 systems named; count ND. Hardware family, not installed quantity. G42 OpenAI Oracle with NVIDIA, Cisco, SoftBank; project financing ND. G42 · Medium–low NVIDIA Noetra Japan national AI infrastructure 27,500 Rubin GPUs and 13,750 Vera CPUs announced. Planned program quantity; delivery schedule, site, power, and commitment status ND. NVIDIA and Noetra; finance partner capital allocation ND. NVIDIA Vera Rubin · Medium–low Google Cloud Ironwood 256-chip and 9,216-chip customer configurations; nearly 10 MW for 9,216-chip pod. System topology pod maximum; not global installed inventory. Google Cloud; outside project finance ND. Google Ironwood · High for configuration AWS P5 P5e P5en Up to 20,000 H100 H200 GPUs and 20 exaflops advertised. Orderable architectural cloud ceiling; not verified owned-fleet count. AWS; project finance ND. AWS P5 · High for advertised ceiling; low for inventory inference

Price and financing signals

Project and infrastructure finance

The following are disclosed capital, financing, purchase, or investment signals. They are not converted into $ MW, $ GPU, or electricity prices.

Project program Disclosed amount and currency Date unit interpretation Confidence and source --- ---: --- --- Crusoe Lancium Abilene US$3.4bn initial phase fully funded; later US$15bn second-phase JV 2024-10-15 and 2025-05-21 disclosures; scope overlap unresolved. Medium. initial JV · second JV Applied Digital Polaris Forge 1 Macquarie preferred-equity facility up to US$5bn ; US$112.5m initial draw; approximately US$11bn contracted lease revenue Financing draw and lease revenue, not capex or GPU price. 2025 disclosures. Medium. Macquarie Meta Hyperion Approximately US$27bn development costs; approximately US$7bn Blue Owl cash contribution JV development cost and equity contribution; PIMCO other bond investor debt part of financing. 2025. High for disclosure; not GPU price. Meta Blue Owl Meta El Paso Approximately US$14bn total development cost; approximately US$4.9bn BlackRock cash contribution; US$12.5bn debt financing Transaction-scoped figures, not additive capex. 2026-07-28. Medium. Meta BlackRock Meta Sturgeon County CA$13bn+ investment; approximately CA$60m local infrastructure Project developer disclosure 2026-07-08. Medium–low due stage conflict. Meta Talen Cumulus AWS US$650m secondary-reported purchase consideration; US$350m close + US$300m escrow Transaction consideration; official Talen source confirms escrow but not full US$650m. 2024. Low. secondary report QScale Q01 $700m Building B, currency not stated; older $867m full project, including Québec government loan equity and Desjardins financing Date-scoped project building figures; not current valuation or GPU price. 2021 2026. Medium–low. Building B · financing Start Campus SINES €8.5bn construction investment value 2024 2025 developer disclosures; campus-wide and phased. Medium–high. Start Campus Stargate Norway Approximately US$1bn initial 20 MW phase; up to US$250m equity on 100% basis 2025-07-31 planned JV capital; not electricity price. Medium. Nscale AWS Spain €33.7bn planned Spain investment Multi-year cloud AI data-center and supply-chain program; not site-level capex. 2026-03-02. Medium. Amazon DataVolt NEOM Oxagon US$5bn phase-one investment 2025-02 agreement; phase-one expected operational 2028. Medium. NEOM Teraco JB7 R8bn syndicated loan 2024-11; loan through Absa and other institutions, not stated as JB7-only capex. Medium. Teraco Raxio platform US$100m Raxio-announced IFC financing; IFC disclosure shows US$293m estimated total project cost and a US$70m approved-investment line Ten-DC 13.5 MW portfolio context; figures require reconciliation and are not UG1 ET1 site prices. 2024 2025. Low–medium. Raxio · IFC SJK01 Finland €1.5bn+ phase 1; €7.5bn+ full campus Announced 2026-07-14; not electricity prices. Medium. Pure DC Project Minato Osaka Approximately US$600m first phase Reported fully secured; operations scheduled early 2028. Medium–high. SCCP PDG JC3 US$1bn campus investment Project-specific 2025 announcement. Medium–high. PDG ODATA QR03 Over US$3bn launch disclosure and US$3.3bn energization disclosure Two releases, scope date difference unresolved. Medium–high. launch · energization Ascenty Brazil US$1.2bn announced investment Mixed campus expansion program, not one building or GPU purchase. 2026-05-29. Medium. DCD Scala Tamboré R$6.2bn Phase 2; R$10bn campus including Phase 1 Committed invested figures, not utility cost or GPU price. 2024-08-15. Medium. Scala

GPU and cloud pricing

Category Public signal Unit, date, and interpretation Confidence source --- ---: --- --- OEM product list prices ND for H100, H200, B200 GB300, Rubin, MI350X, MI355X, Trainium2, Ironwood, and Gaudi 3 in the reviewed primary pages ND is not zero; vendors direct buyers to partners or publish specifications without MSRP. Retrieved 2026-08-05. High for “not disclosed in evidence.” NVIDIA GB300 · AMD MI350X · Intel Gaudi CoreWeave HGX H100 H200 B200 US$49.24 US$50.44 US$68.80 USD per 8-GPU HGX instance-hour in North America; cloud rental, not hardware purchase. 2026-08-05 observation. Medium; dynamic provider quote. CoreWeave pricing Lambda B200 H100 US$6.69 US$3.99 USD per GPU-hour on the 8-GPU instance plan; other commitment tiers differ. 2026-08-05 observation. Medium; dynamic cloud quote. Lambda pricing RunPod B300 H200 B200 H100 SXM US$7.39 US$4.39 US$5.89 US$2.99 Provider-listed USD hour model rates; page updated 2026-07-27 and retrieved 2026-08-05. Terms geography may vary. Medium; dynamic cloud quote. RunPod pricing Secondary-market screening estimates H100 PCIe US$25k–30k ; H100 SXM5 US$35k–40k ; DGX H100 8-GPU US$300k–460k ; DGX H200 8-GPU US$400k–500k ; B200 broad US$30k–50k Analyst-style estimates published 2026-07-20; not OEM list prices or verified transactions. Low–medium; supplier quotes required. IntuitionLabs

Do not divide a data-center capex, lease, JV, or financing figure by a GPU target. Do not convert instance-hour prices to ownership cost without utilization, commitment, networking, storage, tax, and lifecycle assumptions.

Methodology

Read the current upstream evidence and review artifacts: data-center pipeline, GPU inventory specifications, ownership financing, and evidence review, with the regional and power registers linked in the source list.
Preserved publisher URLs, publication dates where exposed, original currency, units, and the access as-of date. Primary developer, operator, utility, government, regulator, investor, lender, and manufacturer disclosures were preferred; secondary evidence is explicitly marked.
Merged duplicate projects across regional and power memos rather than creating additive rows. Kept nested phases, campus maxima, grid access, generation, PPA power, facility power, and IT load in separate fields.
Used only direct unit normalization for GW-to-MW readability where needed. MVA was not converted to MW without a disclosed power factor. GPU system counts were not converted to MW or vice versa.
Separated installed or operational claims from orders, targets, cloud availability, product availability, rack system design, and customer topology. Product specifications are not assigned to a project unless the source does so.
Assigned confidence as an editorial evidence-quality label based on source directness, date, explicitness, and conflict; it is not a statistical probability.

Unresolved risks and data gaps

The evidence is a bounded sample, not a global census. No global total, regional total, or comparable $ MW estimate is calculated.
Web-index coverage may be incomplete or stale. Several project pages are undated or historical; construction, energization, financial close, and operating status require refresh before an investment, procurement, or publication decision.
Most values are self-reported. Independent meters, signed interconnection studies load letters, current permits, utility connection offers, PPA schedules strike prices, and project-specific tariffs were generally unavailable.
“Compute capacity,” “critical power,” “power consumption,” “electrical capacity,” “secured IT capacity,” “allocated power,” and “campus capacity” are not interchangeable. MVA is not treated as MW; PPA or generation capacity is not treated as IT load.
GPU evidence mixes discrete GPUs, rack-scale systems, chip accelerators, customer configurations, cloud ceilings, and future targets. xAI’s 200,000 180,000 H100 discrepancy remains unresolved; AWS Rainier’s Trainium2 count is not a GPU count; Noetra’s Rubin count is planned, not installed.
Investor and financing disclosures can be platform-level rather than project-level. Raxio’s US$100m US$70m US$293m figures and Cumulus’s US$650m secondary purchase report require reconciliation.
Current OEM prices and signed GPU purchase prices are not public in the reviewed evidence. Cloud rates and secondary estimates are volatile and use different units.

Next steps and ownership

Human reviewer — current-status pass: Recheck Abilene, Polaris Forge 1, Hyperion, Sturgeon County, Cumulus, SINES later buildings, Stargate Norway UAE, DataVolt Oxagon, Teraco JB7, ODATA QR03, PDG JC3, Osaka projects, and Yotta against current utility, permit, lender, and ready-for-service records.
Power and data reviewer — definition pass: Add explicit fields for IT load, facility load, campus maximum, interconnect allocated power, generation, PPA, PUE, source date, and evidence state; reject any total that crosses definitions.
GPU and procurement reviewer — inventory and price pass: Resolve the xAI count discrepancy, refresh delivered counts and cloud ceilings, confirm H100 variant at Yotta, and obtain supplier OEM quotes if procurement-grade prices are needed.
Finance reviewer — reconciliation pass: Reconcile overlapping project JV figures for Abilene, Meta El Paso Hyperion, QScale, Raxio, and Cumulus before using any amount as capex, equity, debt, or purchase consideration.
Publication authority: Treat this file as a draft candidate. Route it to the service-authored current-run evaluation; no publication or closure is claimed here.

Source list and evidence register

Canonical upstream evidence read for this report:

Data-center pipeline and capacity ledger
GPU inventory and specification ledger
Ownership, investor, and financing map
Evidence quality review

The regional, power, and product memo links below are retained as supporting registers. Direct publisher, operator, utility, government, investor, lender, manufacturer, and pricing URLs are preserved beside the material claims throughout the tables.

North America evidence memo
EMEA evidence memo
APAC and Latin America evidence memo
Global power and capacity evidence memo
GPU inventory specification evidence memo

Report details

<sub Report date: 2026-08-07 UTC. Evidence cutoff: 2026-08-06 UTC. Workspace: ws-01. Loop: loop-0c93c098-2728-4106-a8bf-b79ea6a93bdc. Run: looprun-e03ee008-8278-4da4-997a-0c773bd11cb0. Session: loopsession-43397353-f837-4396-99dc-438b29f1c8d0. Owner: General Codex Agent. Status: candidate ready for current-run evaluation; publication and closure remain pending. Startup files were not modified. No external communication, schedule change, or production-data change was performed.< sub