Meta Platforms (META) has signed a multi-billion-dollar deal to rent artificial intelligence chips from Google (GOOGL), according to a Thursday report in The Information, citing anonymous sources.
The chips in question are Google’s Tensor Processing Units (TPUs), which will be used by Meta to train and run its next-generation large language models (LLMs). The report comes as tech firms continue to pour billions into chips and data center infrastructure to meet the growing demands of AI compute.
Meta Reportedly Agrees to Buy Billions of Dollars Worth of Google’s Tensor Processing Units for Its Data Centers
In December, Reuters reported that Google was pushing to make its TPUs a viable alternative to Nvidia’s (NVDA) market-leading GPUs, which power the vast majority of AI applications in the world today.
TPU sales have become a crucial revenue growth engine for Google’s cloud platform as they are positioned as a compelling and lower-cost alternative to Nvidia GPUs. The company believes there is a real opportunity to increase its market share and prove to investors that its AI investments are generating returns.
In November, Google announced “Ironwood,” its most advanced Tensor Processing Unit. Customers can scale a single AI server pod to up to 9,216 Ironwood TPUs that will be connected by high-speed interconnects, providing up to 9.6 terabits per second (TB/s) of bandwidth. The chips can also be connected to a colossal 1.77 petabytes (PB) of high-bandwidth memory (HBM).
According to Google, the Ironwood series can deliver more than 118 times the FP8 ExaFLOPS of its nearest competitor and demonstrated four-times better performance in training and inference than Trillium, the company’s previous generation TPU.
The first company to partner with Google for TPUs was Anthropic, which touted the massive price and performance gains they provide. The chips reportedly enabled the company to serve massive Claude models at scale. Google previously announced that it had struck a deal with Anthropic worth “tens of billions of dollars,” giving the AI firm access to 1 million TPUs.
The Information reported that Meta is also in talks with Google to buy the TPUs for its data centers as early as 2027, though the status of those discussions could not be determined. However, that would be a separate deal from today’s agreement, which only relates to cloud access. Meanwhile, Google also has an agreement with an unidentified investment firm to fund a joint venture that would lease TPUs to other customers.
Meta Announces $60 Billion Chip Deal with AMD for 6GW of AI Compute Power
Earlier this week, Advanced Micro Devices (AMD) said it would sell up to $60 billion in AI chips to Meta over the next five years. The chipmaker will provide up to 6 GW of its Instinct series GPUs to power the social media giant’s AI training and inference workloads globally.
The first deployment will use the Instinct GPU on the MI450 platform, the AMD EPYC GPUs, and the Helios rack-scale AI server architecture, which are co-engineered by Meta and optimized for its AI workloads.
As part of the deal, AMD granted Meta warrants entitling it to acquire up to 10% of the AMD stock, structured to vest over the course of the deal. This effectively ties Meta’s computing demand to AMD’s equity upside, further deepening the strategic partnership.
Meta also has plans to design and develop its own AI chips. It has reportedly been working with Taiwan Semiconductor Manufacturing Co. (TSMC) on an updated model for its Meta Training and Inference Accelerator (MITA) chips, which is expected to launch later this year.
