Meta Platforms has signed an agreement to rent Google’s tensor processing units through Google Cloud to develop new AI models. The deal marks a major expansion of Google’s TPU commercialization strategy, which previously restricted the chips to its own internal use and select cloud customers. Meta is also in discussions to purchase TPUs outright for installation in its own data centers as soon as next year.
The agreement allows Meta to diversify its compute supply beyond Nvidia and AMD, supporting a projected $135 billion in AI infrastructure spending by 2026. Google is forming a joint venture to lease TPUs to other AI customers, underscoring a broader push to challenge Nvidia’s dominance in the AI hardware market. Some Google Cloud executives estimate expanding TPU sales could capture as much as 10% of Nvidia’s annual revenue.
On February 24, Meta and AMD announced a multiyear deal for up to six gigawatts of AMD’s Instinct GPUs. The deal is valued at up to $60 billion over five years and includes an equity warrant that could give Meta as much as 10% of AMD’s stock. A week earlier, Meta expanded its partnership with Nvidia to deploy millions of Blackwell and next-generation Vera Rubin processors.
Meta CEO Mark Zuckerberg stated the AMD deal is an important step for Meta to diversify its compute and deliver personal superintelligence. Morningstar analysts described Meta’s approach as a “multipronged silicon strategy.” The strategy leverages Nvidia for frontier model training, AMD for inference needs, Google’s TPUs for possible Llama workloads, and Meta’s in-house MTIA chips for core recommendation algorithms.
Google first developed TPUs more than a decade ago for internal AI workloads. In October 2025, Anthropic signed a deal worth tens of billions of dollars for access to up to one million TPUs. In December, Google collaborated with Meta on a project called “TorchTPU” to ensure full compatibility between TPUs and the PyTorch software framework.
The Wall Street Journal reported that Google has been expanding financial support for data center partners and exploring an investment in cloud startup Fluidstack to broaden TPU demand. The Guardian reported this week that Meta’s multi-vendor approach reflects a scale of operations that demands several alternatives. Analyst Nguyen, quoted in the report, said Meta has already grown large enough to require several alternatives.
Meta has plans for 30 data centers, with 26 located in the United States. The company’s AI infrastructure spending is projected to reach as much as $135 billion in 2026.
