AI Labs Now Lease Compute From Direct Competitors

The New Infrastructure Playbook

Frontier AI laboratories are fundamentally restructuring how they acquire computing capacity, increasingly turning to competitors and unconventional providers rather than relying solely on traditional cloud platforms. Anthropic has emerged as the most visible example of this shift, reportedly proposing to purchase up to $10 billion in computing capacity from Meta over two years—a deal Meta was still evaluating as of mid-July 2026, according to Forbes.

The arrangement would position Meta, which develops the competing Llama model family, as an infrastructure supplier to Anthropic's Claude. This represents a striking departure from conventional competitive boundaries in the AI sector.

Why It Matters

This infrastructure fragmentation introduces new operational and contractual risks for enterprises deploying AI at scale. When model providers depend on competitors, bitcoin miners, and startup data center operators for compute capacity, the reliability of AI services becomes tied to complex, often undisclosed contractual terms and the financial health of multiple counterparties. Enterprises need visibility into these dependencies before committing to long-term AI deployments.

Anthropic's Diversified Compute Strategy

Anthropic maintains substantial workloads on Amazon Web Services and Google Cloud, but has layered on several alternative arrangements. The company reportedly pays xAI $1.25 billion monthly for access to the Colossus 1 facility in Memphis through May 2029, according to a SpaceX offering filing from May 2026. Either party can terminate this agreement with 90 days' notice.

In a more traditional arrangement, Anthropic signed a 20-year lease with TeraWulf on July 6, 2026 for approximately 401 megawatts at a Kentucky campus. TeraWulf disclosed the deal should generate roughly $19 billion in contracted revenue, with initial capacity expected in the second half of 2027.

Separately, Anthropic committed $50 billion in November 2025 to custom-built U.S. infrastructure developed with Fluidstack in Texas and New York, though ownership and financing structures were not disclosed.

The Hidden Costs of Outsourcing

These arrangements do not eliminate construction, financing, or permitting risk—they redistribute it across model developers, infrastructure operators, lenders, and utilities. Oracle's Project Jupiter in New Mexico illustrates the challenges. The campus, intended for OpenAI workloads and authorized for up to $165 billion in industrial revenue bonds, is awaiting an air quality permit from state regulators. On July 2, 2026, Oracle revised its application to replace gas-fired generation with Bloom Energy fuel cells after residents complained about the approval process.

Oracle reported capital expenditure of $55.7 billion in fiscal 2026, up from $21.2 billion the prior year, with guidance pointing to roughly $70 billion in fiscal 2027.

The trend has created opportunities for previously unknown suppliers. Sharon AI, an Australian provider, disclosed a $1.32 billion five-year agreement with an unnamed global AI lab on July 16, 2026, covering New Zealand infrastructure with revenue expected to begin in 2027.

Questions for Enterprise Buyers

Enterprises purchasing inference services at scale should ask three questions of any provider, Forbes reports. First, which specific facilities and upstream capacity providers serve their workload. Second, what migration and backup provisions exist if arrangements terminate or sites face delays. Third, whether any material facility operates on permits not yet granted.

The compute market now rewards speed in assembling power, chips, interconnection, and regulatory approval—capabilities increasingly found outside traditional hyperscalers. For organizations signing multi-year AI commitments, understanding these contractual dependencies matters more than the ownership headlines suggest.

These details were first reported by Janakiram MSV in Forbes.