Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-07-30

Kempax Research

Loop report · 2026-07-30

Global data-center and GPU-capacity report

30 July 2026 — Public disclosures point to a sizeable pipeline of AI-oriented campuses across the United States, UAE, Portugal, and India, but the figures are not a comparable installed-capacity total. The source set contains announced campus or programme power claims of up to 5 GW Meta Hyperion and UAE–US AI Campus , 4.5 GW an Oracle-related Stargate increment , 1.4 GW Vantage Frontier , and separately a 1.2-GW IT-capacity claim for SINES DC. These categories must remain separate; no aggregate is calculated because projects and programmes may overlap and most are not operating-capacity disclosures.

Public GPU disclosures are more limited. xAI reports 100,000 H100 GPUs at Colossus; CoreWeave reported more than 250,000 GPUs across its fleet as of 31 December 2024; Oracle announced a planned ceiling of up to 131,072 Blackwell GPUs; and Meta reported 6,080 A100 GPUs live in RSC’s initial phase. None supplies a current, auditable global GPU inventory or a defensible hardware unit price.

Data-center projects and construction

Project Location and stage Capacity disclosure Electricity power disclosure Company and investor backing Possible price or investment not capacity price Source --- --- --- --- --- --- --- Hyperion Richland Parish, Louisiana, US; development JV announced, not treated as operational IT load: unknown Up to 5 GW campus capacity announced claim; definition needs confirmation Meta; JV with funds managed by Blue Owl Capital US$27bn JV reported; not a power or GPU price Meta; Reuters Frontier Shackelford County, Texas, US; planned mega-campus, initial delivery expected H2 2026 10 buildings 3.7m sq ft; IT-load MW unknown 1.4 GW total campus capacity Vantage Data Centers; DigitalBridge identified as backer US$25bn planned investment; not a capacity price Vantage; DigitalBridge Stargate Oracle incremental buildout US; planned programme increment; site allocation not established IT load: unknown 4.5 GW additional capacity stated for the partnership; do not add to individual Stargate sites OpenAI and Oracle; initial Stargate equity funders: SoftBank, OpenAI, Oracle, MGX Stargate intends up to US$500bn over four years; no price for this increment OpenAI—Oracle partnership; Stargate announcement UAE–US AI Campus Stargate UAE Abu Dhabi, UAE; phase 1 unveiled in May 2025; not treated as operational IT load: unknown 5 GW headline campus capacity; portion that is IT load or firm delivered utility supply is unknown G42-led consortium with US partners Not disclosed G42; US Embassy SINES DC Sines, Portugal; multi-building campus; SIN01 inaugurated, but no phase MW in service disclosed 1.2 GW IT capacity for the announced campus Power supply: unknown Start Campus Not disclosed Start Campus; Business Wire; SIN01 notice TCS AI data-centre buildout India; announced national buildout, not a site-specific construction disclosure 1 GW planned buildout Power supply: unknown Tata Consultancy Services Tata Group US$6.5bn plan reported; not a capacity price DatacenterDynamics; Hindustan Times

Public GPU inventory and specification

Operator scope GPU inventory and specification Quantity and status Power context Possible price Source --- --- --- --- --- --- xAI Colossus, Memphis NVIDIA H100 Hopper ; memory form-factor variant unknown 100,000 H100 GPUs , operator operating claim 2024 Site power unknown ; do not infer from GPU count Hardware price not disclosed xAI Colossus CoreWeave fleet 32 data centers NVIDIA mix including H100 H200 and other generations; count by model unknown More than 250,000 GPUs as of 2024-12-31; fleet total, not a site count More than 360 MW active power as of 2024-12-31; facility context, not GPU-only draw Unit price not disclosed CoreWeave S-1 Oracle Cloud Infrastructure zettascale cluster NVIDIA Blackwell GPUs in GB200 NVL72 systems; B200 count separately unknown Up to 131,072 Blackwell GPUs , planned configuration; not installed-inventory evidence Electrical capacity unknown Purchase price not disclosed Oracle Meta AI Research SuperCluster NVIDIA DGX A100 systems A100 GPUs; exact memory SKU unknown 6,080 A100 GPUs live in initial phase; 16,000 was a prospective build target Electrical capacity unknown Purchase price not disclosed Meta RSC

Investors, financing, and possible pricing

No reviewed disclosure establishes a dated per-GPU, per-rack, per-MW, or customer capacity price. The following figures are financing or investment context only and must not be divided by a GPU or MW count.

Company project Disclosed parties Financial figure and classification Capacity-price conclusion Source --- --- --- --- --- Stargate SoftBank, OpenAI, Oracle and MGX are named initial equity funders; SoftBank holds financial responsibility and OpenAI operational responsibility Up to US$500bn over four years; US$100bn stated as deploying immediately — announced investment-program ceiling Not disclosed OpenAI CoreWeave Full lender holder roster unknown in the reviewed material US$8.5bn GPU-backed financing facility — debt financing, not a procurement price Not disclosed CoreWeave Nebius Investor list requires primary-release extraction US$700m strategic equity financing; separately US$1bn aggregate-principal convertible notes Not disclosed ; amounts must remain separate Equity release; notes release Crusoe Brookfield US$750m facility , Upper90 US$225m facility ; JV with Blue Owl Capital and Primary Digital Infrastructure Credit facilities JV backing; not equipment capex or a customer contract Not disclosed Brookfield facility; Upper90 facility; JV

Gaps and methodology

This is a conservative register of publicly announced projects and fleet disclosures, not a global census. “Unknown” means the reviewed evidence did not establish the field.
IT capacity, campus electrical capacity, utility supply, active facility power, and annual energy consumption are different measures. They are never substituted for one another.
No project row is counted as operating capacity. SINES has an inaugurated phase, but its in-service MW is not disclosed. Oracle’s configuration and Meta’s 16,000-GPU figure are clearly labelled planned.
No total MW GW or GPU count is calculated. The Stargate Oracle programme may overlap other Stargate projects; fleet totals are not site inventories; announced capacity may not be delivered.
Before financial modelling or publication review, obtain dated primary confirmation of: capacity definitions and commissioning milestones; project-level GPU allocations; post-2024 CoreWeave fleet and power; Oracle commissioning; and a dated supplier quote or contract for any capacity-price claim.

Report details

<sub Research date: 30 July 2026. Publication candidate only; no publication action has been taken. Source inputs: data-center project register, GPU inventory, and investor pricing research. Run: looprun-9f650df7-0ef7-4f55-9c94-f9906d825d55; session: loopsession-67ed78c1-0370-49fb-b212-a51a605531f9; owner: General Codex Agent.< sub