On February 26, 2026, The Information reported — and Reuters picked up — that Meta Platforms has signed a multibillion-dollar, multiyear deal to rent Google Cloud’s custom TPUs to train and run its next-generation large language models. Neither the exact dollar figure nor the duration is public. The direction is what matters: the company famous for building its own data centers and buying chips by the millions is now renting accelerator capacity from one of its biggest rivals.
The same reporting adds a second thread: Meta is separately negotiating to buy millions of TPUs and deploy them inside its own facilities. That follow-on deal is not finalized. Together, the two negotiations mark a structural change in the AI compute market — Google’s TPU is turning from an internal asset into a merchant product.
The Deal: Billions to Rent Google Cloud TPUs
According to the reports, the rented TPUs are consumed through Google Cloud, not installed in Meta’s own buildings, and are earmarked for training and serving next-generation LLMs. Sources described the agreement only as “multibillion-dollar” and multiyear, with no precise figures disclosed.
The interesting half is the negotiation running alongside it: Meta discussing the outright purchase of millions of TPUs for its own data centers. If that closes, it would be the first time Google sells its homegrown accelerator at scale into someone else’s facility — a decisive step toward being a merchant silicon supplier rather than a cloud-only one.
Why a Self-Build Empire Rents
Meta’s infrastructure story for a decade has been “build it ourselves”: planet-scale data centers, private networks, in-house accelerators. The reported rationale for renting now comes down to three things: avoid depending on a single chip supplier, match different processors to different workloads, and use competition among vendors to negotiate better prices.
Lay Meta’s February chip moves side by side and the diversification strategy is hard to miss:
- Mid-February: a multibillion-dollar Nvidia purchase covering millions of next-generation Vera Rubin GPUs, with availability later this year
- February 24: a multibillion-dollar AMD deal spanning Instinct MI400-series GPUs, plus an option to acquire a 10% equity stake in AMD if performance milestones are met
- In-house MTIA: a training-optimized successor built with TSMC was expected this year but is reportedly delayed by technical challenges
Ironwood and Google’s Merchant-Chip Ambition
Google’s current flagship is the Ironwood TPU, launched in November 2025. A single pod scales to 9,216 chips with 9.6 Tbps of bandwidth and up to 1.77 petabytes of shared HBM; Google claims more than 118 times the FP8 ExaFLOPS of its nearest competitor and 4x the performance of the previous Trillium generation.
The specs are half the story. Google reportedly believes it can capture up to 10% of Nvidia’s data center revenue within a few years, and it has started selling chips directly into private data centers. There is precedent at hyperscale: the earlier Google-Anthropic agreement was described as worth tens of billions of dollars for access to as many as one million TPUs.
Where This Leaves Nvidia
Losing one customer is not a crisis for Nvidia. The signal is bigger than the sum: the largest compute buyers in the world are now seriously building second supply lines. Nvidia reported its FY2026 fourth-quarter results the same week (see our NVIDIA Q4 FY2026 earnings breakdown), and data center demand remains the whole story — but when Google pushes TPUs beyond its own cloud into the open market, Nvidia’s moat faces pressure on two fronts at once: raw silicon and cloud bundling.
For buyers, that is exactly the point. Vera Rubin’s delivery schedule, MI400’s real-world performance, and Ironwood’s rental pricing are now mutual leverage in every negotiation.
Takeaways for Infrastructure Teams
- Buy like a portfolio: Meta now runs rental (Google Cloud TPUs), purchase (Nvidia Vera Rubin, AMD MI400), and in-house (MTIA) in parallel. Treat compute sourcing as portfolio management, not a one-off vendor decision.
- TPUs are now a real alternative: with Google willing to rent — and possibly sell — TPUs, teams outside the Google ecosystem have their first credible non-GPU option. Bargaining power comes from credible alternatives, and those now exist.
- Keep an exit path in the stack: training and inference pipelines hard-wired to a single accelerator ecosystem are a single point of failure in the 2026 supply chain. Cross-hardware capability in your frameworks and schedulers is appreciating in value.
Sources
- Google signs multibillion-dollar AI chip deal with Meta — Reuters
- Google and Meta reportedly strike new, multibillion-dollar AI chip deal — SiliconANGLE
- Meta signs multi-billion-dollar deal to rent Google AI chips — Yahoo Finance
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.