Google and Meta Team Up on AI Chips
In a major challenge to Nvidia's dominance, Google has signed a multibillion-dollar deal to supply Meta with custom AI chips. The partnership signals a strategic shift by tech giants to develop their own silicon, aiming to reduce costs and dependence on a single supplier for the hardware powering the AI arms race.
The deal centers on Google's custom-built Tensor Processing Units (TPUs), with Meta leasing them for training and running its next-generation AI models. This multi-year agreement is valued at several billion dollars and is seen as a move to diversify Meta's AI hardware suppliers beyond Nvidia. This partnership is structured in two potential phases: an initial period of leasing Google's TPU capacity, followed by the possibility of Meta directly purchasing the chips for its own data centers as early as 2027. This signals a deeper collaboration and a significant step for Google in commercializing what started as internal infrastructure. Meta's decision to partner with Google comes after reported "technical challenges" and delays with its own in-house "MTIA" chip development. The move to lease TPUs provides Meta with immediate access to powerful AI hardware while it re-evaluates its internal silicon strategy. The agreement is part of a broader, aggressive hardware strategy by Meta, which includes a multi-year deal for millions of Nvidia's Blackwell and Rubin GPUs and a pact with AMD valued at up to $100 billion. Meta's capital expenditure for 2026 is projected to be between $115 billion and $135 billion, largely for AI infrastructure. This collaboration is a direct challenge to Nvidia's market dominance, which currently controls over 80% of the market for GPUs used in AI. While Nvidia's recent quarterly revenue soared to $68.1 billion, this and similar deals indicate a growing market for viable alternatives. Google is actively positioning its TPUs as a compelling, lower-cost alternative to Nvidia's GPUs. The latest "Ironwood" TPUs are said to offer four times better performance for training and inference than the previous generation, a key factor in attracting large-scale customers like Anthropic and now Meta.
