Meta Eyes Google’s TPUs in Multi-Billion Dollar Shift Away From Nvidia

Facebook parent Meta Platforms is engaged in ongoing discussions with Alphabet Inc.'s Google to procure and rent billions of dollars worth of Google's tensor processing units (TPUs) for its AI data centers, marking a potential shift from its heavy reliance on Nvidia's graphics processors.

The negotiations, first reported by The Information and corroborated by multiple outlets, center on Meta integrating Google's TPUs into its own infrastructure starting in 2027. Additionally, the talks include provisions for Meta to rent TPUs from Google Cloud as early as 2026. This arrangement would represent Google's first major foray into selling or leasing its custom AI chips directly to external customers for on-premises use, expanding beyond its traditional cloud-only model.

Meta, one of the world's largest AI infrastructure spenders, currently allocates up to $72 billion annually on Nvidia chips. The proposed deal with Google aims to diversify suppliers amid surging demand for AI hardware alternatives to Nvidia's supply-constrained and high-cost GPUs. Google Cloud executives have internally projected that broader TPU adoption could capture up to 10% of Nvidia's annual revenue, equivalent to billions of dollars.

The discussions come as Meta ramps up its AI investments, including a previously announced $600 billion commitment over the next three years for U.S. infrastructure and jobs focused on AI data centers. Google's TPUs, optimized for machine learning workloads, have gained traction with other clients; for instance, AI startup Anthropic recently expanded its Google agreement to access up to one million TPUs, valued at tens of billions.

Neither Meta, Google, nor Nvidia responded to requests for comment on the talks. The report has already rippled through markets: Alphabet shares rose more than 3% in premarket trading on Monday, while Nvidia's stock dipped 3.2%. Google's cloud unit, a key revenue driver, has seen recent boosts from investments like Warren Buffett's Berkshire Hathaway and positive reception for its Gemini 3 AI model.

This development underscores intensifying competition in the AI chip sector, where over 4 million developers depend on Nvidia's proprietary CUDA software ecosystem. For Meta, the move aligns with broader industry efforts to mitigate risks from single-supplier dependency in powering next-generation AI applications.

Sources close to the matter emphasized that the talks remain preliminary, with no final agreement reached. Further details could emerge as negotiations progress into 2026.