Anthropic is currently negotiating a potential $10,000 million, two-year deal to lease AI computing infrastructure from Meta. The agreement, if finalized, would mark a significant shift for Meta, transitioning from an internal AI power user to a commercial provider of AI computing services for external clients.
Proposed Infrastructure Deal Between Anthropic and Meta
The AI startup Anthropic, known for developing the Claude model, has initiated preliminary discussions with Meta Platforms to lease large-scale AI computing capacity. As reported by the outlet covering the negotiations, the proposed arrangement could reach a valuation of $10,000 million over a two-year duration. The terms currently under discussion involve a monthly service fee structure, with the option for Anthropic to terminate the contract before the full two-year term expires.
For Anthropic, securing additional compute resources is essential to sustain the growing demands of its model development and service deployment. The company already maintains infrastructure partnerships with major tech players, including Amazon, Google, and SpaceX. According to the same reporting, the initial proposal was submitted by Anthropic to Meta in June.
Meta’s Strategic Shift Toward AI Compute as a Service
If the deal proceeds, it would represent a fundamental change in how Meta manages its data center resources. While the company has historically invested in massive infrastructure to support its own internal AI research and product deployment, this agreement would signal an entry into the commercial market for AI compute services. This is a competitive landscape currently dominated by established cloud service providers.
Meta CEO Mark Zuckerberg has previously signaled that the company is open to monetizing its excess computing capacity if external demand proves sufficient. This strategy aligns with the firm’s aggressive capital expenditure plans, which include a projected investment of between $125,000 million and $145,000 million in AI infrastructure throughout 2026 to scale its data center capabilities.
Advancements in Proprietary MTIA Chip Production
Parallel to these negotiations, Meta is moving to reduce its reliance on third-party hardware by advancing its own custom silicon. The company’s Meta Training and Inference Accelerator (MTIA) chip is slated to enter official production in September 2026, according to recent reports on the project’s progress. This development follows a rapid six-week testing phase that confirmed the chip’s functional viability.
To execute this hardware strategy, Meta has leveraged a wide network of industry expertise. The design process involved collaboration with Broadcom, while technical contributions were sourced from partners including Samsung, Sandisk, and Sumitomo Electric. The final manufacturing is handled by the semiconductor foundry TSMC.
Looking ahead, Meta plans to accelerate its hardware release cycle significantly compared to industry norms. The company intends to announce updated versions of its MTIA chips every six months. This rapid iteration cycle is designed to provide greater architectural flexibility, allowing the firm to adapt its data centers more effectively to evolving AI workloads while simultaneously curbing long-term dependency on a limited number of external hardware suppliers.
Keep reading