Meta Platforms Strikes Multibillion-Dollar Deal to Rent Google TPUs

Meta Platforms has reportedly agreed to a multibillion-dollar deal to rent custom artificial intelligence chips from Google Cloud, according to a report by The Information. The agreement will give Meta access to Google’s tensor processing units (TPUs) to train and run its next-generation large language models.

In October, The Left Shift published an exclusive report on Meta's talks with Google about a potential multi-billion-dollar deal for Google's TPUs.

While NVIDIA continues to dominate the market with its GPUs, Google’s TPUs are emerging as a lower-cost alternative. Last year, AI startup Anthropic announced a major expansion of its partnership with Google Cloud, revealing plans to scale its use of Google’s TPUs to as many as one million units.

Google’s latest TPU generation, known as Ironwood, allows customers to scale up to 9,216 chips in a single server pod, offering high-speed interconnects and massive shared high-bandwidth memory. The company claims Ironwood delivers significant performance gains over prior generations.

Interestingly, Meta remains one of Nvidia’s largest customers and has also committed billions toward AI chips from Advanced Micro Devices (AMD), including its latest Instinct GPUs.

Diversifying suppliers allows Meta to optimise workloads across different chip architectures while strengthening its negotiating leverage.

The report also noted discussions between Meta and Google about potentially purchasing TPUs for Meta’s own data centres, though no agreement has been finalised.

Meanwhile, Meta continues developing its in-house MTIA chips, though recent reports suggest technical delays in their rollout.