Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-08-03

Kempax Research

Loop report · 2026-08-03

Global AI data-center and GPU capacity: evidence-qualified update

31 July 2026. This update separates what can be directly supported about data-center construction, electricity, GPU configuration and commercial GPU-cloud pricing. The review found no project-level comparison row for which one evidence chain establishes all of: construction or IT capacity, firm electricity status, a project GPU inventory specification, operator, investor backer, and a real commercial price. Accordingly, no incomplete project is presented as a complete, price-comparable offer.

The most concrete new electricity records are a 1.2-GW grid interconnection at the Crusoe–Lancium Abilene campus, a 960-MW planned contracted consumption arrangement for Cumulus, and 100 MW of energized critical IT load at Polaris Forge 1. These are not interchangeable and must not be totaled. Separately, Lambda publishes self-service B200, H100 and A100 GPU-cloud list rates in USD per GPU-hour; those rates are commercial offers, not campus, rack or fleet prices.

Complete project comparison

Qualifying projects Result Reason --- --- --- None No complete comparison is published The power ledger supplies firm-status evidence, while the pricing ledger supplies provider-specific public GPU-cloud rates. Neither ledger links a commercial price to a project that also has a cited GPU inventory specification and project backing.

This is a coverage limitation, not evidence that the excluded projects have no commercial terms. A project can enter the complete comparison only after a source establishes every required field at the same project or sellable-offer scope.

Electricity-qualified construction records

Project, company and status Construction IT capacity Electricity capacity and status GPU inventory specification Backing Commercial price Evidence --- --- --- --- --- --- --- Stargate Crusoe–Lancium Clean Campus, Abilene, Texas — Crusoe developer operator; Lancium energy-campus partner; Oracle OCI platform Planned: eight buildings, about 4m sq ft. Initial two-building phase: 980,000 sq ft and 200+ MW ; later source says the first two buildings were energized. Grid interconnection: 1.2 GW . This evidences site supply access, not energization of the full interconnection; energized MW is not disclosed. No numerical project inventory disclosed. Crusoe says Oracle began delivering NVIDIA GB200 racks in June 2025 and that the build can support hundreds of thousands of GPUs. No project-specific backing amount retained in this evidence set. Not disclosed for this project. Crusoe, 18 Mar 2025; Crusoe, 30 Sep 2025 Cumulus Data Center Campus, northeast Pennsylvania — AWS purchaser developer; Cumulus Digital Talen Energy seller and power provider Planned: 960 MW campus. Contracted development: agreement milestones permit development to 960 MW of power consumption ; Talen expected power sales under a PPA. This is not evidence that 960 MW is energized. Not disclosed in the reviewed source. Transaction parties are disclosed; no project commercial GPU-price or distinct investor amount is retained here. Not disclosed for this project. Talen, 4 Mar 2024; Talen, 26 Aug 2024 Polaris Forge 1, Ellendale, North Dakota — Applied Digital developer operator; CoreWeave long-term lessee Energized: Building 1 at 100 MW critical IT load . Contracted deployment: 400 MW for CoreWeave under long-term leases; later buildings remain future deployment. Energized: 100 MW . The 400 MW figure is contracted lease deployment, not utility supply. No project GPU count model disclosed in the reviewed source. CoreWeave is the named long-term lessee; no project-specific financing allocation is disclosed. Not disclosed for this project. Applied Digital, 24 Nov 2025

Other announced global construction capacity

These announced figures remain useful pipeline indicators but lack the firm power and or commercial-price evidence required above. They are not totals and are not comparable to energized load.

Project company Disclosed scope Status of electricity and price evidence Source --- --- --- --- Hyperion, Louisiana — Meta Blue Owl-managed funds Up to 5 GW announced campus capacity; US$27bn JV transaction No firm supply or commercial customer price in the disclosure. Meta, 21 Oct 2025 Frontier, Texas — Vantage Data Centers 1.4 GW announced campus capacity; 10 buildings, 3.7m sq ft; US$25bn planned investment No firm supply or commercial customer price in the disclosure. Vantage, 19 Aug 2025 SINES DC, Portugal — Start Campus 1.2 GW IT capacity IT capacity is not grid supply; commercial price not disclosed. Start Campus Business Wire, 8 Aug 2024 UAE–US AI Campus, Abu Dhabi — G42-led consortium 5 GW announced campus headline No firm supply or commercial customer price in the disclosure. G42, May 2025 Cyberjaya, Malaysia — Vantage Data Centers 256 MW announced campus; US$1.6bn GIC ADIA platform investment No firm supply; investment is not a commercial price. Vantage capacity announcement; Vantage funding announcement India expansion — NTT DATA NTT 400 MW additional capacity over three years; US$1.5bn programme investment No firm supply or commercial price in the disclosures. NTT; Fortune India

Disclosed GPU deployments and configurations

Operator deployment Inventory or configuration Specification Capacity price limitation Evidence --- --- --- --- --- xAI Colossus, Memphis 200,000 NVIDIA H100 GPUs xAI also reported an earlier 180,000-GPU milestone H100: 80 GB HBM3 , up to 3.35 TB s memory bandwidth Construction was reported in days, but facility MW, project backing allocation and commercial price are unavailable. xAI; NVIDIA H100 Oracle OCI Zettascale Up to 131,072 NVIDIA Blackwell GPUs ; announced maximum, not utilization B200: 192 GB HBM3e , 8 TB s bandwidth Facility MW, project backing and public numerical price unavailable. Oracle, 11 Sep 2024; NVIDIA Blackwell Tesla Cortex, Gigafactory Texas 50,000 H100 GPUs H100: 80 GB HBM3 , up to 3.35 TB s Facility MW, project backing and commercial price unavailable. Tesla, 23 Jul 2024; NVIDIA H100 Yotta Shakti Cloud, India 4,000 NVIDIA H100 GPUs H100: 80 GB HBM3 , up to 3.35 TB s Facility MW, project backing and validated service price unavailable. Yotta; NVIDIA H100

Verified commercial GPU-cloud list rates

The following are the only validated commercial prices in the corrective evidence set. They are Lambda public self-service per-GPU-hour list rates for the selected eight-GPU instance configuration, retrieved 31 July 2026; applicable sales tax may apply.

Provider configuration GPU specification Commercial list price Scope --- --- --- --- Lambda GPU Cloud, 8x B200 SXM6 8 GPUs; 180 GB VRAM per GPU ; 208 vCPUs; 2,900 GiB RAM; 22 TiB SSD US$6.69 per GPU-hour Provider offer, not a campus or fleet price. Lambda GPU Cloud, 8x H100 SXM 8 GPUs; 80 GB VRAM per GPU ; 208 vCPUs; 1,800 GiB RAM; 22 TiB SSD US$3.99 per GPU-hour Provider offer, not a campus or fleet price. Lambda GPU Cloud, 8x A100 SXM 8 GPUs; 80 GB VRAM per GPU ; 240 vCPUs; 1,800 GiB RAM; 19.5 TiB SSD US$2.79 per GPU-hour Provider offer, not a campus or fleet price.

Source for all three offers: Lambda pricing, retrieved 31 Jul 2026.

Interpretation and unresolved coverage

• Do not add grid interconnection, contracted consumption, energized IT load, announced campus capacity or GPU counts. Each has a different scope and maturity state.
• Funding, investment and transaction amounts are not commercial prices. No price is inferred by dividing a budget by MW or GPU quantity.
• The strict power-qualified sample is US-only. It strengthens evidence quality but does not establish globally balanced firm-supply coverage.
• No reviewed source ties a real commercial rate to a named project that also discloses all required capacity, power, GPU and backing fields. That is the remaining publication risk.

Report details

<sub Research date: 31 July 2026. Source dates appear beside dated disclosures; undated source pages are identified by retrieval date where applicable. Candidate report only; no publication action has been taken. Evidence ledgers: power capacity, commercial pricing, project register, and GPU register. Run: looprun-8b2a55d0-56b3-4cf8-971c-8d8268f47da8; session: loopsession-a035f8b0-933c-48cd-9db0-be2637876651.< sub