Frontier AI Labs Are Renting Compute From Their Competitors

AI frontier labs like Anthropic are radically reshaping their compute infrastructure strategies, moving beyond traditional cloud providers. Anthropic is diversifying with a proposed $10 billion deal with rival Meta, a $1.25 billion monthly agreement with xAI, a 20-year, $19 billion lease with TeraWulf, and a $50 billion commitment for custom facilities with Fluidstack. This trend involves separating model development from infrastructure ownership, reallocating significant construction, financing, and permitting risks across various partners, including neoclouds and even competitors. Oracle's Project Jupiter for OpenAI exemplifies these challenges, facing permitting hurdles. The market now rewards those who can rapidly secure power, chips, and regulatory approvals, leading to a fragmented landscape where contractual dependencies are paramount for enterprises.

Anthropic reportedly proposed buying up to $10 billion in computing capacity from Meta over two years, and Meta was still weighing the offer as of July 17, 2026. Meta develops Llama and competes directly with Claude, yet it is considering becoming an infrastructure supplier to its rival.

Frontier labs are separating model development from infrastructure ownership. They are signing a mixture of public cloud contracts, dedicated compute agreements, long-term data center leases and custom facilities built by partners. That mixture does not eliminate construction, financing or permitting risk. It reallocates the risk across model developers, infrastructure operators, lenders and utilities, often through contracts whose most important protections are never disclosed.

Anthropic still runs substantial workloads on Amazon and Google Cloud, so nothing here replaces the public cloud. Three recently disclosed arrangements illustrate the shift.

The first is a dedicated compute agreement. Anthropic reportedly pays xAI $1.25 billion per month for the Colossus 1 facility in Memphis through May 2029, a figure disclosed in SpaceX’s May 2026 offering filing. Either party can terminate on 90 days' notice, which cuts in both directions.

The second is a conventional real estate lease. TeraWulf announced on July 6, 2026 a 20-year agreement with Anthropic covering roughly 401 megawatts at a campus in Hawesville, Kentucky. The company told investors the lease should produce about $19 billion of contracted revenue and said it expects the arrangement to be supported by investment-grade credit. Initial capacity is due in the second half of 2027.

The third is the Meta proposal, which remains preliminary and may never be signed.

A fourth model sits alongside those three. On November 12, 2025 Anthropic committed $50 billion to custom-built United States infrastructure developed with Fluidstack in Texas and New York. The company said the facilities are built for its own workloads. That announcement did not specify the ownership, financing or land arrangements, so a dedicated facility should not be read as a directly owned one.

Oracle describes itself as the tenant of Project Jupiter in Doña Ana County, New Mexico, a campus intended to support OpenAI workloads and authorized for as much as $165 billion in industrial revenue bonds. Authorization is not issued debt, and the amount actually raised has not been made public. The campus is waiting on an air quality permit. That proceeding is with the New Mexico Environment Department, a state agency, whose secretary ordered a public hearing after residents complained that their names appeared in supportive comments without their consent.

On July 2, 2026, Oracle replaced its proposed gas-fired generation plan with a Bloom Energy fuel cell system and asked the department to approve the revised application. A lab can outsource the permitting work to an infrastructure partner, but it cannot outsource its dependence on the permit being granted and kept valid.

The spending behind that patience keeps climbing. Oracle reported capital expenditure of $55.7 billion in fiscal 2026, up from $21.2 billion the prior year. Chief financial officer Hilary Maxson guided to roughly $70 billion of Oracle's own spending in fiscal 2027, with total outlays of $90 billion to $95 billion once customer-funded and reimbursable amounts are included.

The same conditions are producing suppliers nobody had heard of in 2024. Sharon AI, an Australian provider listed on Nasdaq, disclosed a $1.32 billion five-year cloud computing service agreement with an unnamed global AI lab on July 16, 2026. The company said the deal covers infrastructure in New Zealand, with revenue expected to begin in 2027. A contract of that size may help support project financing, although the company has not said the agreement alone finances the build.

Contracted revenue is a long way from delivered revenue. TeraWulf's $19 billion arrives over 20 years and depends on capacity that does not yet exist, and Sharon AI's first dollar lands in 2027.

The risk allocation is harder to read than the headline numbers suggest. Take-or-pay commitments, prepayments, minimum revenue guarantees, parent guarantees, termination payments and equipment ownership all move the exposure around, and none of these terms is public for every agreement. To the extent these projects are financed against long-term customer commitments, investors should examine whether tenant termination rights run shorter than the financing tenor.

Availability also depends on someone else’s demand curve. SpaceX’s filing described the Colossus arrangement as a way to monetize unused compute capacity, and reporting has linked that spare capacity to weaker Grok usage. If xAI's own compute demand rises, the economics underpinning the deal will change.

In summary, the labs have not eliminated the construction problem. They have divided it among a larger set of counterparties, including bitcoin-mining operators, neoclouds, established cloud providers and, in a few cases, rival model companies.

Enterprises buying inference at scale should put three questions to any provider. The first is materiality, which means asking which facilities and upstream capacity providers actually serve the workload. The second is redundancy, which means asking what migration and backup provisions protect the service if either arrangement ends or a site is delayed. The third is regulatory dependence, which means asking whether any material facility is operating on a permit that has not yet been granted.

The compute market now rewards whoever can assemble power, chips, interconnection and regulatory approval fastest, and that ability sits outside the traditional hyperscalers more often than it did a year ago. For enterprises, developers, and operators now signing 20-year commitments, the market is becoming more fragmented, and its contractual dependencies are more consequential than the ownership question that gets the headlines.