Nvidia's stock closed down 2.6% today following reports that Meta may purchase Google’s TPU AI chips.
According to sources cited by The Information, the potential deal for Tensor Processing Units could be worth billions of dollars. Meta is reportedly considering leasing TPUs hosted by Google as early as next year, with plans to deploy the chips in its own data centers starting in 2027.
Google unveiled its latest TPU, codenamed Ironwood, in April. The chip—comprising two dies—features 192GB of high-speed HBM memory and six custom AI processing modules based on two distinct architectural designs.
Each Ironwood chip integrates four SparseCores optimized for handling large embeddings—mathematical structures used to store AI model data—and two TensorCores engineered to accelerate matrix multiplication, a core computation in AI workloads.
Google deploys Ironwood chips in liquid-cooled clusters capable of housing up to 9,216 chips. A single cluster delivers 42.5 exaflops of performance, with one exaflop equating to a quintillion (10^18) calculations per second.
It remains unclear whether Meta’s potential agreement with Google would involve Ironwood or a future TPU variant. Given Meta’s planned 2027 deployment timeline, the company might instead opt for Ironwood’s successor, expected to launch next year.
While Google could theoretically supply Meta with full TPU clusters similar to those powering its cloud infrastructure, the social media giant may choose to acquire only the chips and integrate them into its custom-designed servers and racks.
Meta previously developed its own inference chip, called MTIA. Reuters reported in February that a new version is slated for deployment by the end of 2025. The potential TPU deal suggests Meta might scale back its MTIA ambitions—or alternatively, use MTIA for inference tasks while running training workloads on Google’s TPUs.
Following the news, Google’s shares rose 1.6% at market close, while Broadcom—which collaborates with Google on TPU design—gained 1.9%.
The rumored partnership has drawn attention beyond Wall Street, including from Nvidia. “We’re happy for Google’s success—they’ve made tremendous progress in AI, and we continue to supply them,” the chipmaker stated on X. “NVIDIA is a full generation ahead in the industry—it’s the only platform that runs every AI model across all computing scenarios.”
Nvidia’s flagship Blackwell Ultra AI accelerator delivers 10 petaflops of performance when processing 4-bit floating-point operations, with one petaflop representing one quadrillion (10^15) calculations per second. The company’s upcoming Rubin GPU series, expected next year, is projected to offer even higher performance.
