Biggest AI compute deals

OpenAI contracted to buy an additional $250 billion of Azure services from Microsoft under the definitive agreement announced on 28 October 2025. The announcement gives no spending deadline, chip count or location for that purchase.

The companies amended their partnership on 27 April 2026, allowing OpenAI to serve all its products across cloud providers while keeping Microsoft as its primary cloud partner. That update states no replacement value for the Azure purchase.

Anthropic signed a cloud hosting agreement with Google in March 2026 and amended it between May and October. It commits Anthropic to spend at least about $121.3 billion between April 2026 and July 2033, whether or not it uses the capacity; any shortfall is paid to Google.

The filing does not split the amount by chip type. In October 2025 Anthropic had announced plans to use up to one million Google TPUs, worth tens of billions of dollars.

Anthropic announced a signed Amazon agreement on 20 April 2026 for up to 5 GW to train and run Claude. It initially disclosed more than $100 billion over ten years, spanning Graviton and Trainium2 through Trainium4, with inference expansion in Asia and Europe.

Its November filing specifies at least $110 billion from May 2026 to April 2036.

OpenAI and AWS announced a $100 billion expansion over eight years on 27 February 2026. OpenAI committed to consume approximately 2 GW of Trainium capacity through AWS for Stateful Runtime, Frontier and other AI workloads. The purchase expands their earlier $38 billion agreement. The release names no deployment sites.

Anthropic announced a $30 billion Azure compute purchase commitment on 18 November 2025, alongside plans to use Nvidia Grace Blackwell and Vera Rubin systems for Claude. The announcement gave no fixed contract length.

A November 2026 filing raises the minimum commitment to $44.9 billion between November 2026 and June 2033, following amendments from February through October.

Anthropic signed four Nscale GPU services agreements on 25 August 2026 for up to $44.6 billion, disclosed in Nscale’s 18 September S-1. They cover Nvidia Vera Rubin NVL72 systems at Monarch in West Virginia, with multi-year terms for each accepted tranche. Nscale had no binding financing commitments for the project when it filed.

Anthropic’s November filing confirms service terms extending into 2033.

OpenAI agreed to buy $38 billion of AWS compute over seven years, announced on 3 November 2025. The capacity includes hundreds of thousands of Nvidia GPUs, including GB200 and GB300 systems, for ChatGPT inference and training future models. AWS targeted deployment of all initial capacity before the end of 2026.

The companies expanded the agreement by $100 billion over eight years in February 2026. That expansion is a separate entry in this list.

Meta’s five-year Nebius agreement, announced on 16 March 2026, is worth up to approximately $27 billion. It includes $12 billion of dedicated Nvidia Vera Rubin capacity across multiple locations, with delivery starting in early 2027.

Meta also committed to buy up to $15 billion of capacity remaining after Nebius sells to other customers. That makes $27 billion the maximum contract value; actual purchases under the residual-capacity component depend on other sales.

Meta agreed to buy approximately $21 billion of AI cloud capacity from CoreWeave through December 2032. CoreWeave announced the expanded agreement on 9 April 2026. Dedicated capacity across multiple locations will include early Nvidia Vera Rubin deployments and support Meta’s AI inference workloads. This follows the companies’ 2025 capacity agreement.

OpenAI signed a multi-year agreement for 750 MW of Cerebras inference systems, announced on 14 January 2026. The capacity is intended to speed up responses to OpenAI customers, with deployment starting in 2026.

Cerebras’s May 2026 prospectus values the deal at more than $20 billion. It describes a December 2025 master agreement, deployment through 2028 and service terms of three or four years per tranche, extendable to five. OpenAI also has an option for another 1.25 GW through 2030; exercise of that option is required to reach 2 GW.

Anthropic signed a 20-year lease at TeraWulf’s Justified Data campus in Hawesville, Kentucky, announced on 6 July 2026. TeraWulf reports approximately 401 MW of critical IT capacity and $19 billion in contracted revenue over the initial term. Two five-year extension options could raise revenue to about $33 billion.

On 4 November 2026, TeraWulf announced Google credit support for financing and a definitive Fluor construction agreement for the project. The buyer named in the lease is Anthropic.

Microsoft signed a five-year agreement for dedicated Nebius GPU capacity in Vineland, New Jersey, on 8 September 2025, disclosed in a filing that day. The base contract value is about $17.4 billion through 2031, subject to deployment and availability. Additional services or capacity could raise it to about $19.4 billion. Deployment was scheduled in tranches during 2025 and 2026, with obligations dependent on Nebius confirming the required financing.

Meta signed a CoreWeave cloud capacity order on 25 September 2025, filed on 30 September. The initial commitment is up to approximately $14.2 billion through 14 December 2031, conditional on delivery and service availability, with an option to expand through 2032.

CoreWeave later announced a $21 billion expanded agreement on 9 April 2026. The original filing gives no chip count or site list.

OpenAI signed a contract worth up to $11.9 billion with CoreWeave, announced on 10 March 2025, for AI infrastructure to train and serve its models. OpenAI also received $350 million of CoreWeave stock as part of the deal; that equity is not counted here.

CoreWeave's first quarter 2026 filing gives October 2030 as the end of this agreement.

Anthropic committed $11.6 billion over seven years to Akamai Cloud, announced on 24 September 2026. The purchase covers distributed infrastructure and software for growing CPU workloads. Akamai disclosed no specific deployment locations or chip quantity. A further $9 billion is potential expansion subject to agreement; the announced equity warrant is separate from the cloud purchase.

Akamai’s 8-K says the two project plans were signed on 18 September, each with seven years from its service start date. Payments depend on delivery and service availability.

Microsoft signed a $9.7 billion agreement for IREN GPU services at Childress, Texas, on 2 November 2025, filed on 3 November. It has a five-year average service term through 2031, with Nvidia GB300 systems across four Horizon facilities totaling about 200 MW. Microsoft prepays 20% of each tranche’s value.

IREN’s August 2026 annual report says Microsoft accepted Horizon 1 that month. Horizon 2 to 4 deliveries were targeted for the fourth quarter, with contractual grace periods extending into the second quarter of 2027.

Fluidstack signed a $7 billion, 15-year lease for 245 MW of IT capacity at Hut 8’s River Bend campus in Louisiana, announced on 17 December 2025. Google backstops lease payments. Three five-year renewal options could raise the value to about $17.7 billion; a right of first offer covers up to another 1,000 MW, conditional on more power.

Hut 8’s 4 November 2026 results retain the River Bend expansion right in their notes. Fluidstack is the contractual buyer for this AI hosting capacity. The original release targets the first data hall for the second quarter of 2027.

OpenAI signed a third CoreWeave contract worth up to $6.5 billion, announced on 25 September 2025, to train its next models. CoreWeave's first quarter 2026 filing dates the end of this order to 31 May 2031. With the March and May contracts, OpenAI's commitments to CoreWeave come to about $22.4 billion.

Nvidia signed a CoreWeave order with an initial value of $6.3 billion on 9 September 2025, disclosed in a 15 September filing. Nvidia must buy capacity left unsold to CoreWeave’s other customers through 13 April 2032, subject to delivery, availability and termination provisions. The filing gives no fixed chip count or location for this residual-capacity purchase.

Jane Street committed approximately $6 billion to CoreWeave’s AI cloud on 15 April 2026. The agreement gives the trading firm compute across multiple facilities, including Nvidia Vera Rubin systems, plus storage, connectivity and software for large-scale machine learning research. CoreWeave disclosed no contract length. Jane Street’s separately announced $1 billion equity investment is excluded from the compute value.

In May 2025 OpenAI expanded its CoreWeave agreement by up to $4 billion, a second contract on top of the March deal. CoreWeave disclosed the figure when it announced the third contract in September; no separate end date has been published.

Nvidia bought $3.4 billion of managed GPU cloud services from IREN over five years, announced on 7 May 2026. IREN will use air-cooled Blackwell systems within approximately 60 MW at Childress, Texas. The services include orchestration and cluster management software with Mirantis, for Nvidia’s internal AI and research workloads.

Nebius announced a roughly $3 billion, five-year AI infrastructure agreement with Meta on 11 November 2025. Its filed shareholder letter said capacity would be deployed over the following three months; the amount was limited by available capacity. It did not identify a site or chip model for this contract. The companies announced a new agreement worth up to $27 billion in March 2026.

OpenAI and Oracle announced an agreement for 4.5 GW of additional Stargate data center capacity in the United States on 22 July 2025. OpenAI gave no dollar value or contract term. Its figure of more than 2 million chips covers the additional capacity together with Abilene, Texas.

In March 2026, Oracle said it had completed leases for the additional 4.5 GW, while disputing reports of problems with the Abilene project.

OpenAI agreed to deploy 6 GW of AMD Instinct GPUs across multiple generations under a definitive agreement announced on 6 October 2025. The first 1 GW of MI450 systems was scheduled to begin deployment in the second half of 2026. AMD described the term as multi-year and disclosed no fixed purchase price or deployment location.

In a 2026 update, AMD said OpenAI expected Helios systems to come online in the fourth quarter of 2026, with deployment accelerating in 2027.

Anthropic announced its SpaceX capacity agreement on 6 May 2026, initially describing Colossus 1 capacity for Claude users. SpaceX’s June prospectus disclosed about 325,000 Nvidia GPUs across Colossus and Colossus II at $1.25 billion a month through May 2029, with reduced ramp fees and cancellation rights.

Anthropic’s 2 November filing reports amended arrangements totaling approximately $3.1 billion a month after full delivery, expected in December 2026. The filed aggregate is up to $120.5 billion into 2029. Cancellation requires 90 days’ notice after initial periods ending in January and March 2027.

Google signed a SpaceX cloud services agreement on 5 June 2026 for approximately 110,000 Nvidia GPUs, with CPUs, memory and related hardware. SpaceX filed the agreement details that day. The filed price is $920 million a month from October 2026 through June 2029, following reduced fees during the ramp.

Google can terminate or accept fewer GPUs at a lower fee if delivery misses the September deadline and grace period. After 31 December 2026, either party can terminate on 90 days’ notice. The filing does not identify a specific Colossus campus.

Anthropic signed a Google and Broadcom agreement for multiple gigawatts of TPU capacity, announced on 6 April 2026. The next-generation capacity is expected to come online starting in 2027, with the vast majority in the United States, to train and run Claude. The announcement gives no dollar value or exact service term.

On 6 May, Anthropic described the arrangement as a 5 GW agreement. Broadcom’s 6 April filing specifies approximately 3.5 GW accessed through Broadcom as part of the wider commitment, dependent on Anthropic’s continued commercial success.

Meta and AMD announced a definitive agreement for up to 6 GW of Instinct GPUs on 24 February 2026. It spans multiple years and chip generations, beginning with a custom MI450-based GPU, Venice CPUs and Helios racks. Shipments for the first gigawatt were scheduled for the second half of 2026. No fixed dollar value or exact term was disclosed.

On 27 October 2026, AMD said initial Helios systems were installed and being tested at Meta. That update gives no revised capacity commitment.