Meta has quietly signed a multibillion-dollar agreement to lease access to Google’s custom AI chips so it can train and run large models, instead of buying the hardware outright. The structure of the deal, which rents guaranteed compute throughput rather than racks of silicon, signals a shift toward treating machine intelligence like a utility.
As reported by The Information, the multi-year arrangement gives Meta access to Google’s Tensor Processing Units and is worth billions; the outlet reported that Meta may first rent capacity from Google Cloud before later deploying TPUs in its own data centers. The Information also reported that Google is lining up outside financing and joint-venture structures to lease TPUs to other customers, a sign the company wants recurring revenue from compute rather than one-and-done chip sales.
Leasing compute, not chips
WRAL TechWire has framed the transaction as a purchase of “compute throughput” - guaranteed streams of computational output - rather than ownership of physical accelerators, arguing that leasing reduces obsolescence and gives developers flexibility as software frameworks evolve. A 2017 technical paper documenting TPUs notes that Google began deploying the accelerators inside its datacenters in 2015, so opening that infrastructure externally represents a notable business pivot. According to WRAL TechWire, the economic logic is about selling continuous outputs, not one-off hardware. See the original TPU architecture paper on arXiv for background on early in-datacenter TPU use.
Why hyperscalers are hedging
Meta is diversifying its compute supply: the company has announced a long-term partnership with AMD to deploy up to 6 gigawatts of Instinct GPUs and said it will align hardware and systems roadmaps with the chipmaker in a company release. Meta's announcement complements existing agreements with Nvidia and Meta’s internal silicon work. At the same time, Anthropic has already committed to a large expansion of Google TPU capacity; a PR Newswire release said Anthropic will access up to one million TPUs, underscoring why major model builders prefer multi-vendor resilience.
What this means for the chip market
According to The Information, Google is exploring finance and joint-venture routes to make TPUs available as a metered service, a move that would let it collect continuous revenue as customers run workloads on its stack. Analysts say that shift could loosen Nvidia’s one-vendor dominance by giving hyperscalers more bargaining power and by creating a market for metered “intelligence” services rather than one-time hardware sales.
For Bay Area readers, the deal is another reminder that decisions by local cloud and platform companies can ripple into data center siting, energy demand and hiring across the region. Expect follow ups on delivery timelines, financing terms and any regulatory or enterprise pushback as the industry tests what it means to lease intelligence instead of owning silicon.
