AI compute deals. When will the power arrive?

Google campus under construction since June 2025. Connected to Goodnight 1 wind farm (265.5 MW, operational) and Goodnight 2 (265.5 MW, under construction). Wind generation ratings are not campus load or energized IT capacity. Campus MW, agreement value and first-capacity date are undisclosed. Separate from Crusoe’s Abilene campuses.

Project Plans 2 and 3 signed September 18 under the May 5 master agreement. Approximately $11.6B is subject to delivery, availability and termination provisions. Potential additional $9B, equity warrants and supplier capital spending are excluded. Site, MW and dated first service are undisclosed.

Hardware: CPU cloud infrastructure

Dedicated managed inference for Thinking Machines Lab. The disclosed amount is annual pricing, not a lifetime contract value. Site, MW allocation and dated first-service milestone are undisclosed. Customer commentary describes production use but does not establish a delivery or energization date.

Hardware: NVIDIA HGX B200 / Quantum-2 InfiniBand

The consortium signed lease documentation with an Australian Anthropic subsidiary for the first stage of Western Downs Digital Park, subject to customary approvals. The disclosure does not specify leased MW, contract value or delivery date. Consortium funding obligations also remain subject to approvals. Campus estimates are not an Anthropic allocation.

Perplexity contracted Crusoe Cloud for model training and managed inference. Contract value, GPU count, MW allocation, sites and first-service date are undisclosed. Crusoe also plans to adopt Perplexity software for its employees; that reciprocal software purchase is not additional compute capacity.

Hardware: NVIDIA GB300 NVL72 clusters / InfiniBand

Preferred sovereign infrastructure partnership and planned integration for Palantir customers. No dedicated megawatt allocation or compute purchase value is disclosed. Excluded from investment and capacity totals.

Initial compute commitment for Figure’s Helix models. Intent to expand above $6B is not an additional signed commitment. Nscale’s separate equity investment is undisclosed and excluded. GPU count is a potential upper limit; customer MW and energization date are undisclosed. Deployment target does not establish operational delivery.

Hardware: Potential deployment of up to 100,000 NVIDIA Vera Rubin GPUs

An unnamed neocloud operator contracted 14 MW within the 37 MW LON-East campus. Company-reported facility delivery does not establish customer occupancy or a utility energization date. Price and term are undisclosed. The campus total is not the customer allocation.

Binding term sheet with an unnamed sovereign AI company based in India for Blackwell cloud GPUs and allied services. Customer identity is undisclosed; do not attribute the order to Sarvam. Site, MW, GPU count and first-service date are undisclosed. Original INR value is retained without currency conversion.

Hardware: NVIDIA Blackwell cloud GPUs; count undisclosed

Program ambition for up to 1 GW by 2030, supported by enterprise compute commitments. Individual sites and incremental capacity beyond earlier Mistral projects are not allocated in this announcement. Excluded from physical capacity totals.

Dedicated IBM Cloud cluster for Together AI open-source inference. Site, MW and GPU count are undisclosed. Expected availability is a target, not delivery evidence. Together AI’s separate financing is excluded.

Hardware: NVIDIA HGX B300 / Spectrum-X

Three Indonesian facilities with 360 MW of contracted IT power, expected online in 2028. This is the regional deployment allocation, not CoreWeave’s global contracted-power total.

Delta Forge 1: 15-year take-or-pay lease with an unnamed US hyperscaler. Date is the results disclosure, not the undisclosed signing date. Each campus is a separate lease; customer identity is not inferred. Hosting revenue is not included in compute purchase values.

Delta Forge 2: 15-year take-or-pay lease with an unnamed US hyperscaler. Date is the results disclosure, not the undisclosed signing date. Each campus is a separate lease; customer identity is not inferred. Hosting revenue is not included in compute purchase values.

Polaris Forge 3: 15-year take-or-pay lease with an unnamed US hyperscaler. Date is the results disclosure, not the undisclosed signing date. Each campus is a separate lease; customer identity is not inferred. Hosting revenue is not included in compute purchase values.

Crusoe and Lancium announced a 1 GW campus for an unnamed hyperscale customer. Construction is expected to start in Q3 2026. This is a separate project from IREN’s Childress campus.

Joint operations company for the existing Narvik development. Its planned 2026 start refers to the operations company, not energized customer capacity. No incremental power allocation or contract value is disclosed.

Colocation agreement across two campuses. CoreWeave stated that initial capacity at Stockholm 4 South was already online. No deal-specific power allocation or contract value is disclosed. The company-wide active and contracted power figures are not assigned to these campuses.

The May 6 announcement covers more than 300 MW at Colossus 1. SpaceX’s June 5 prospectus describes later agreements across Colossus 1 and 2: approximately 325,000 GPUs and $1.25B per month through May 2029, with reduced fees during the May–June ramp. Either party can terminate on 90 days’ notice after the initial three months. No guaranteed total or incremental MW is inferred. Orbital compute remains an expression of interest.

Hardware: 220,000+ NVIDIA GPUs

Expansion separate from the earlier 5,120-GPU programme. April disclosure valued this project at approximately $2.8B; July 21 described Yotta 2 at approximately $2.5B without reconciling the difference. September 8 reports equipment deliveries across the wider Yotta programme, not operational acceptance of this tranche. Project value remains outside compute-purchase totals. No MW conversion is inferred from GPU count.

Hardware: 20,736 NVIDIA B300 GPUs

Executed agreement with an unnamed US AI service provider specialising in large-scale model training. Committed value covers GPU cloud computing and allied services for six months. Customer identity, site, GPU count, MW and service commencement are undisclosed.

Anthropic committed more than $100B over ten years to AWS technologies for up to 5 GW of new Trainium and related capacity. Nearly 1 GW is expected before the end of 2026. Amazon also announced a $5B equity investment, with up to $20B more later; those equity figures are omitted from the dollar column.

Hardware: Trainium2–4

Expanded Microsoft agreement adds more than 30,000 Rubin GPUs at Narvik in 2027. The 230 MW figure describes the campus, not the power allocated to this expansion.

Hardware: 30,000+ NVIDIA Rubin GPUs

Multi-year agreement to support Claude development and deployment, with capacity starting later in 2026. No exact dollar or megawatt figure is disclosed.

Expanded agreement for approximately $21B through December 2032. Includes new capacity and exercise of an option under a previous order. Read with the earlier agreement rather than treating both as independent projects.

The April 6 announcement is for multiple gigawatts of next-generation TPU capacity starting in 2027. Anthropic’s May 6 recap states a 5 GW agreement with Google and Broadcom. No public dollar figure appears on either Anthropic page. Builds on a prior October 2025 TPU-capacity increase.

Hardware: Google TPUs

Site securing and development agreement for the Sievari industrial area. Fortum supports grid design and permitting; preliminary land-sale work is underway. Power capacity remains subject to design and permitting. No first-capacity date or contract value is disclosed.

New 900 MW campus serving Microsoft, next to the existing Abilene development. Includes two buildings and onsite generation. The combined 2.1 GW Abilene figure includes the existing campus.

Includes $12B of dedicated capacity and up to $15B of additional capacity that remains after sales to other customers. The $27B ceiling is not all dedicated capacity.

Hardware: NVIDIA Vera Rubin

Letter of intent for up to 1.35 GW at Monarch. The campus’s potential power expansion above 8 GW is not the Microsoft allocation. A letter of intent is not a completed supply contract.

Hardware: NVIDIA Vera Rubin NVL72

An additional $100B over eight years expands the earlier $38B AWS agreement. Approximately 2 GW of Trainium capacity is included. Amazon’s equity investment is separate.

Hardware: Trainium3 / Trainium4

Four-year NVIDIA DGX Cloud engagement within Yotta's Blackwell Ultra supercluster. April says NVIDIA is expected to account for roughly half the tranche offtake. The full 20,736 GPUs, host facility MW and over $2B build budget are not NVIDIA allocations or additional contract value. The April deployment completion target is September 30, 2026. No customer service acceptance was verified.

Long-term €1.2B infrastructure partnership at Borlänge. EcoDataCenter will build and operate the facility. The disclosed euro amount is retained without a dollar conversion.

Hardware: NVIDIA Vera Rubin

OpenAI announced 750 MW of inference compute to be introduced in phases through 2028. The announcement does not disclose a contract price or individual sites.

Hardware: Cerebras wafer-scale systems

15 MW deployment at Verne’s Icelandic campus. The power basis is not specified as critical IT load, so this figure remains outside the IT power subtotal.

Hardware: Approximately 4,600 NVIDIA Blackwell Ultra GPUs

Anthropic committed to purchase $30B of Azure compute capacity and to contract additional capacity up to 1 GW, initially on NVIDIA Grace Blackwell and Vera Rubin systems. NVIDIA and Microsoft also committed to invest up to $10B and $5B in Anthropic; those equity figures are omitted from the dollar column. Amazon remains Anthropic’s stated primary cloud and training partner.

Hardware: Grace Blackwell / Vera Rubin

Anthropic announced a $50B investment in U.S. computing infrastructure, building data centers with Fluidstack in Texas and New York. The announcement says Fluidstack can deliver gigawatts of power but does not state a megawatt figure. This is an infrastructure build, not a cloud offtake.

The results recap a $6.3B NVIDIA arrangement; the September 9 order gives NVIDIA access to residual unsold capacity. This backstop can overlap capacity sold to other customers and is excluded from aggregate purchase values. The date identifies the cited results disclosure.

Approximately $9.7B over five years, with 20% prepayment. Horizon 1–4 support 200 MW of critical IT load within the 750 MW Childress campus; the full campus is not assigned to this contract.

Hardware: NVIDIA GB300

Multibillion-dollar agreement for tens of thousands of NVIDIA GPUs. No exact contract price or megawatt allocation is disclosed in the announcement.

Hardware: NVIDIA GPUs, including GB300 NVL72

Microsoft disclosed an incremental $250B Azure services purchase by OpenAI. The equity ownership and prior investment are separate from this services commitment.

Filed September 30 for a September 25 order. Up to approximately $14.2B through December 2031, subject to delivery and availability terms. An option to expand through 2032 is separate.

September expansion only. CoreWeave also disclosed a $4B May expansion and approximately $22.4B in cumulative OpenAI contracts. Do not add the cumulative figure again.

September disclosure of the July Oracle agreement: over $300B across five years and up to 4.5 GW of additional Stargate capacity. This is part of the Stargate program, not capacity additional to its headline total.

Approximately $17.4B for dedicated GPU capacity, subject to deployment and availability. Additional services could increase the value to approximately $19.4B. Microsoft can terminate delayed tranches under the filing’s conditions.

Third lease dated August 28 adds 150 MW to the initial 250 MW, bringing CoreWeave’s campus allocation to 400 MW. The record contains only the incremental allocation. Reported campus delivery remains attached to the original record to avoid counting it twice.

Two ten-year hosting agreements. Approximately $3.7B initial-term revenue; $8.7B includes optional extensions. Google’s $1.8B backstop and warrants are financing support, not another compute purchase. The 160 MW expansion is separate. The August 2026 filing reports partial lease commencement by June 30, with later phases extending into early 2027. Its 102 MW energized figure covers Lake Mariner across customers, not a Fluidstack allocation.

Initial leases cover 250 MW for approximately 15 years, with approximately $7B expected revenue. A separate 150 MW option was subsequently exercised, taking the campus to 400 MW under contract. The initial allocation is not the whole campus. Hosting revenue is excluded from compute purchase totals.

Initial compute agreement. Later expansions are separate records; the September cumulative figure is not another contract.

Stargate is a new company that intends to invest $500B over four years in U.S. AI infrastructure for OpenAI, beginning with $100B. Initial equity funders: SoftBank, OpenAI, Oracle, and MGX. Technology partners: Arm, Microsoft, NVIDIA, Oracle, and OpenAI. This is a project-company build, not a bilateral cloud offtake. The $100B figure is the initial deployment; the $500B figure is the stated intention.

xAI reported a 100,000-Hopper-GPU Colossus training system in use. The $6B funding round is not a compute contract price. This record describes a self-build milestone.

Hardware: 100,000 NVIDIA Hopper GPUs

Government announcement of support for a Canadian AI data center serving Cohere. The up-to-C$240M government contribution is not the total project cost or a compute purchase price. Builders are not named in this source.

Yotta reports Sarvam 1 was trained on Shakti Cloud in India. Uses the October 24, 2024 announcement date in the body; the page header says November 24. Price, reserved GPU count and MW are undisclosed. Sarvam released 30B and 105B models on March 6, 2026, trained in India using IndiaAI Mission compute. That later release does not identify Yotta as their provider and is not a new compute contract.