Frontier AI Labs Turn to Competitors for Computing Power: Infrastructure Strategy Restructuring

Introduction

In July 2026, media reported that Anthropic proposed renting up to $10 billion in computing capacity from competitor Meta, while signing a dedicated computing agreement with xAI for $1.25 billion per month, and a 20-year, $19 billion lease with Bitcoin miner TeraWulf. These deals signal that frontier AI labs are fundamentally changing their infrastructure strategy: decoupling model development from infrastructure ownership, and distributing construction, financing, and licensing risks across multiple partners. Even Oracle's Project Jupiter for OpenAI has been hindered by air permitting issues. This trend suggests that the winners in the computing market are no longer purely cloud giants, but companies that can secure power, chips, and regulatory approvals fastest.

Technical Analysis: From Private Cloud to Hybrid Infrastructure Portfolio

Frontier AI labs are no longer relying solely on a single cloud provider. They are signing four types of agreements:

1. Public cloud contracts: Anthropic still makes heavy use of Amazon Web Services and Google Cloud. 2. Dedicated computing agreements: For example, Anthropic pays xAI $1.25 billion per month to use the Colossus 1 facility, with a contract through 2029, and either party can terminate with 90 days' notice. 3. Traditional real estate leases: TeraWulf's campus in Hawesville, Kentucky provides Anthropic with approximately 401 megawatts of capacity. The 20-year lease is projected to generate $19 billion in revenue, with initial capacity delivered in the second half of 2027. 4. Custom facilities: Anthropic has committed $50 billion to build dedicated infrastructure with Fluidstack in Texas and New York; ownership and financing details are not disclosed.

This hybrid strategy does not eliminate construction risks but redistributes them: model developers, infrastructure operators, lenders, and utilities share the burden. The most critical contract terms (such as take-or-pay, prepayments, and minimum revenue guarantees) are typically not publicly disclosed.

Enterprise Impact Analysis: Costs, Risks, and Decisions

Cost Impact - CAPEX: Labs no longer bear the massive capital expenditures of data center construction (e.g., Oracle's FY2026 CapEx of $55.7 billion, projected $90-95 billion in 2027). Instead, they shift the capital burden to operators through leases. - OPEX: Long-term leases provide predictable operating costs, but termination clauses (e.g., 90-day notice) may allow labs to adjust flexibly if demand declines.### Deployment and Operations Impact - Dedicated compute agreements and custom facilities ensure resource availability for training workloads, but depend on third-party facility licensing approvals and construction timelines. - Oracle’s Project Jupiter was delayed due to a hearing on air permit issues, highlighting that regulatory risks cannot be outsourced.

Security and Compliance - Renting compute resources from competitors (e.g., Anthropic renting xAI’s Colossus 1) raises concerns about data isolation and intellectual property protection. Security clauses in contracts are not publicly disclosed, requiring enterprises to assess attack surfaces. - Sovereign clouds and cross-border data flows: Some facilities are located in New Zealand (Sharon AI agreement) and other regions, involving compliance requirements.

Enterprise Takeaways: When purchasing large-scale inference services, three key questions should be asked: 1. Materiality: Which upstream facilities truly support the workload? 2. Redundancy: If a contract is terminated or construction is delayed, what migration and backup plans exist? 3. Regulatory Dependency: Are any key facilities still awaiting necessary permits?

Market Competition Analysis: Reshaping the Landscape

Cloud Provider Competition - AWS, Azure, Google Cloud still hold core positions, but labs are shifting some high-density demand to specialized providers, weakening traditional cloud vendor lock-in. - Oracle has grown through custom projects like Project Jupiter, but permit issues have exposed its regulatory challenges. Oracle’s FY2026 capital expenditure increased 163% year-over-year, signaling its aggressive bet on AI infrastructure.

Emerging Providers Benefit - Bitcoin miners (e.g., TeraWulf) are repurposing existing power capacity for AI computing, securing stable revenue through long-term leases. - Neoclouds (e.g., xAI, Fluidstack) are rapidly rising, gaining faster access to power and chips. - Smaller providers like Sharon AI enter the market through undisclosed agreements worth $1.32 billion, indicating increasing fragmentation.

Competitive Pressure - Meta faces a dilemma: accepting Anthropic’s $10 billion lease would directly support a competitor to its Llama models, potentially undermining its own AI model competitiveness. - xAI rents idle compute capacity to Anthropic, but if its own Grok demand grows, the economics of the deal will change.

Industry Trend Observation: Separating, Not Eliminating, RiskFrontier labs are pushing an extreme form of "Infrastructure as a Service": - Separation of Ownership and Operations: Labs focus on model R&D, outsourcing heavy-asset risks such as electricity, chips, land, and permits. - Formation of a Contract Ecosystem: Clauses such as take-or-pay, upfront payments, parent guarantees, and termination fees become core, but transparency is insufficient. - Increased Fragmentation: The market winners are entities that can most quickly integrate electricity (e.g., TeraWulf's 401 MW), chips (e.g., NVIDIA H100/B200), and approvals (e.g., New Mexico air permits).

In the long term, this may give rise to an "AI infrastructure brokerage" model—third parties coordinating suppliers, financing, and regulatory compliance to provide one-stop capacity guarantees for labs and enterprises. But as TeraWulf's $19 billion revenue spread over 20 years and reliant on yet-to-be-built capacity illustrates, risk is not fundamentally eliminated; it is merely dispersed across a longer supply chain.