Meta Signs Multibillion-Dollar Deal to Lease Google TPUs as Big Tech Races to Diversify Beyond Nvidia
Meta has signed a multibillion-dollar deal to lease Google Ironwood TPUs as it diversifies its AI chip procurement beyond Nvidia, while Google targets 10 percent of Nvidia's data center revenue.
Meta Platforms has signed a multibillion-dollar agreement to lease Google’s custom Tensor Processing Units for training and running its next-generation large language models, according to a report by The Information confirmed by multiple outlets in late February. The deal represents the largest known instance of one hyperscaler renting another’s proprietary AI silicon, and it sharpens Google’s emerging rivalry with Nvidia in the data center chip market.
Under the arrangement, Meta will initially rent TPU capacity through Google Cloud before potentially purchasing millions of the processors outright for deployment in its own data centers starting in 2027. The chips in question are Google’s Ironwood TPUs, the company’s seventh-generation accelerators launched in November 2025, which can scale to 9,216 units in a single server pod with 9.6 terabits per second of interconnect bandwidth and 1.77 petabytes of shared high-bandwidth memory.
A Four-Way Silicon Strategy
The Google TPU lease is one component of what has become the most aggressive multi-vendor chip procurement strategy in the industry. Meta’s 2026 capital expenditure for AI infrastructure is projected at $115 billion to $135 billion, nearly doubling the $72 billion it spent in 2025.
That spending is spread across at least four hardware tracks. Meta expanded its Nvidia relationship in February with a deal for millions of Vera Rubin GPUs expected later in 2026, including Nvidia’s standalone CPUs for the first time. A separate arrangement with AMD covers Instinct MI400 series accelerators in a deal reportedly worth billions over five years beginning in the second half of 2026. Meta is also developing its own MTIA chips internally, though the program has faced technical delays that have pushed the next-generation design further into the year.
The diversification reflects a calculated effort to match specific silicon to specific workloads while using competitive pressure to negotiate pricing. By maintaining relationships with Nvidia, AMD, and Google simultaneously, and developing in-house alternatives, Meta avoids the single-supplier dependency that has constrained other AI companies.
Google’s Bid to Crack Nvidia’s Grip
For Google, the Meta deal validates a decade-long investment in custom silicon that began with the original TPU in 2015. Google Cloud executives believe the arrangement could generate revenue equivalent to as much as 10 percent of Nvidia’s annual data center business within a few years, a target that would represent billions of dollars given Nvidia’s dominance of the AI training chip market.
Google has formed a joint venture with an unidentified large investment firm to lease TPUs to external clients, with Meta as an early customer. The structure, which rents guaranteed compute throughput rather than physical racks of hardware, signals a shift toward treating AI compute as a metered utility rather than a capital expenditure.
The Meta agreement builds on a broader commercialization push. Google previously secured a six-year cloud deal with Meta worth over $10 billion, and its largest TPU commitment to date is with Anthropic, which contracted for over a gigawatt of AI compute capacity on Google Cloud.
Nvidia’s Position
Nvidia, which holds approximately 80 percent of the AI training chip market and a market capitalization exceeding $3 trillion, has responded by emphasizing its platform advantages. The company has argued that its CUDA software ecosystem and full-stack integration remain a generation ahead of alternatives, and that its platform is the only one capable of running every major AI model.
The market response has been muted so far. Analysts note that while TPU leasing introduces meaningful competition, Nvidia’s installed base and software lock-in give it structural advantages that will take years to erode. The more immediate effect may be on pricing, as Meta and other large buyers gain leverage from having credible alternatives.
The deal also raises questions about the evolving relationship between hyperscalers. Google and Meta are direct competitors in advertising, social media, and AI research, yet the economics of AI infrastructure have created new forms of commercial interdependence. Whether the arrangement represents a durable shift in how AI compute is procured or a transitional measure while Meta scales its own chip program remains to be seen.