Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-07-29

Kempax Research

Loop report · 2026-07-29

Global data-center and GPU capacity — announced-project snapshot

As of 29 July 2026, the reviewed evidence identifies four large announced US campus or programme entries. Three have directly reported electrical capacity: Meta Hyperion up to 5 GW , an OpenAI–Oracle Stargate increment 4.5 GW , and Vantage Frontier 1.4 GW ; a further five-site Stargate announcement has no captured per-site capacity. These figures are planning or campus-power disclosures, not installed IT load, and must not be added together because Stargate announcements may overlap. Public project-specific GPU inventories and procurement prices remain sparse; the only quantified deployment in this evidence set is a third-party report of 100,000 NVIDIA H100 GPUs at xAI Colossus.

Announced data-center projects

Project Location and status Facility IT capacity Electric capacity Company and capital attribution Investment or price basis Evidence --- --- ---: ---: --- --- --- Hyperion Richland Parish, Louisiana, US — development announced IT load not disclosed Up to 5 GW campus capacity Meta; Blue Owl Capital is the financing development partner US$27B joint venture; transaction value, not a capacity price Meta announcement Stargate Oracle incremental capacity US — planned expansion; site allocation not captured IT load not disclosed 4.5 GW additional capacity OpenAI and Oracle; founding partners include SoftBank and MGX Not disclosed for this increment OpenAI: Stargate advances with Oracle Stargate sites US — five further sites announced Not disclosed Not disclosed OpenAI Stargate Not disclosed OpenAI: Five new Stargate sites Frontier Shackelford County, Texas, US — planned; first delivery previously stated as H2 2026 10 buildings; 3.7m sq ft ; IT load not disclosed 1.4 GW total campus capacity Vantage Data Centers US$25B planned investment; not a capacity price Vantage announcement; Frontier data sheet

GPU inventory, specifications, and commercial position

Operator or reference Status Accelerator and specification Quantity capacity Company and investor attribution Price as of 29 Jul 2026 basis Evidence --- --- --- ---: --- --- --- xAI Colossus, Memphis Reported deployment — estimate NVIDIA H100: 80 GB HBM3, 3.35 TB s memory bandwidth 100,000 H100 GPUs reported; not yet corroborated by xAI or NVIDIA xAI; no project-level lender investor list established Not disclosed; no vendor MSRP verified DatacenterDynamics report; NVIDIA H100 Stargate Commitment project — not installed inventory GPU model and system configuration not disclosed Not disclosed SoftBank has financial responsibility; OpenAI operational responsibility; Oracle and MGX are named initial equity funders; Arm, Microsoft, NVIDIA, Oracle and OpenAI are named initial technology partners Hardware price not disclosed. The up-to- US$500B four-year project commitment US$100B stated as immediately deploying is not a GPU price or inventory count OpenAI: Announcing The Stargate Project CoreWeave GPU cloud Commercial fleet financing; model mix unverified Specific model-by-model fleet mix not established in the reviewed evidence Not disclosed CoreWeave announced an US$8.5B GPU-backed financing facility GPU purchase price not disclosed; financing principal is not divided by a fleet estimate CoreWeave investor overview; financing announcement NVIDIA GB200 NVL72 Product reference, not an operator inventory 72 Blackwell GPUs and 36 Grace CPUs per NVL72 system 72 GPUs per rack; no installed-rack count implied NVIDIA supplier platform; deployers not specified Price unavailable; NVIDIA product page does not establish MSRP NVIDIA GB200 NVL72; technical overview

Model reference

Accelerator Public specification Commercial treatment --- --- --- NVIDIA H100 80 GB HBM3; 3.35 TB s Do not combine PCIe and SXM units unless the deployment specifies form factor; verified supplier MSRP unavailable. Source NVIDIA H200 141 GB HBM3e; 4.8 TB s A Hopper refresh; not interchangeable with H100 by unit count. No validated price captured. Source NVIDIA B200 192 GB HBM3e; 8 TB s Count physical GPUs; for GB200 NVL72, retain the separate 72-GPU rack mapping. No validated price captured. Source

Pricing interpretation

No date-, region-, and term-specific supplier hardware price or provider quote was established for the listed inventories. The report therefore uses not disclosed rather than turning announced financing, project investment, cloud-rental comparisons, or secondary-market figures into a per-GPU procurement price. The US$27B Hyperion joint venture, US$25B Frontier plan, Stargate up-to-US$500B commitment, and CoreWeave US$8.5B financing are capital or financing figures only.

Methodology and assumptions

The scope is a conservative sample of individually announced projects, not a global construction census. The cited evidence primarily covers US projects; absence of a region or operator is not evidence of no capacity.
“Electric capacity” is reported only where a source gives campus or additional power capacity. It is not treated as IT load, deployed capacity, or annual electricity consumption.
“Facility IT capacity” remains not disclosed unless separately stated. Building count and floor area are physical-facility descriptors, not MW of IT load.
A public GPU quantity is classified as deployment, commitment, product reference, or estimate. Financing and investment values are never converted to GPUs or unit prices.
No aggregate MW GW total is calculated: the Stargate Oracle increment, five-site announcement, and any individual campus may overlap. GB200 NVL72 is excluded from fleet totals because it is a product specification.

Unresolved evidence gaps

Exact primary-page publication dates and complete source text were not retained for several capacity announcements.
Project-level IT load, construction milestones, and GPU allocations are undisclosed for Hyperion, the reviewed Stargate entries, and Frontier.
The xAI 100,000-H100 claim needs a primary xAI or NVIDIA confirmation before financial-model use.
CoreWeave needs a dated filing or issuer disclosure for fleet count and model mix.
No validated procurement quote supports a hardware-price comparison; pricing requires dated, region- and term-specific supplier or provider evidence.
Non-US project coverage has not been validated in this research set.

Report details

<sub Research basis: 29 July 2026. Source registers: data-center capacity research and commercial GPU inventory research. All figures are reported announcements or explicitly labelled estimates; missing values are intentionally not inferred.< sub