[Exclusive] Meta in Discussions to Tap Google Cloud’s TPUs Amid Billion-Dollar AI Push

Meta Platforms is reportedly in talks to access millions of dollars worth of Google TPUs (Tensor Processing Units), industry sources told The Left Shift. While no formal deal has been announced, the discussions signal Meta’s intent to strengthen its AI infrastructure as the race for high-performance compute intensifies.

The TPUs are expected to be used exclusively by Meta’s newly formed Super Intelligence team, which is focused on developing next-generation general intelligence systems.

Meta has previously partnered with Google Cloud on a six-year, $10 billion agreement to host its generative-AI workloads.

Sources suggest that the newest talks may focus on Meta leveraging Google’s custom TPUs — hardware optimised for large-scale model training and serving — rather than relying solely on mainstream GPU-based servers.

Recently, AI startup Anthropic announced a major expansion of its partnership with Google Cloud, revealing plans to scale its use of Google’s Tensor Processing Units (TPUs) to as many as one million units.

Last year, Meta and NVIDIA deepened their AI infrastructure partnership, with Meta reportedly securing some 350,000 NVIDIA H100 GPUs by the end of 2024 as part of its push toward advanced AI model development.

However, there have been signs that Meta is looking to diversify its GPU needs and not solely rely on NVIDIA. For instance, Meta is acquiring RISC-V chip startup Rivos to strengthen its in-house AI chip efforts. Rivos designs GPUs and AI accelerators on the RISC-V standard, while Meta’s custom MTIA chips, built with Broadcom and TSMC, are already being tested in its data centres.

For Meta, using Google’s TPUs could enable it to scale its AI operations more cost-effectively and competitively. Moreover, Meta has also been investing heavily in building data centres. Recently, it signed a $27 billion joint venture (JV) to develop a large-scale AI data centre campus in Louisiana.

There’s speculation that Meta may outfit one of its upcoming AI data centers with Google’s TPUs — potentially giving its newly formed Super Intelligence team access to massive compute power for advanced AI model training and research.

Google Cloud, meanwhile, has been aggressively scaling its TPU offerings. The technology previously powered internal initiatives like DeepMind’s protein-folding breakthroughs and is now being marketed broadly to enterprise customers.

Several leading organisations — including Apple, Cohere and startups such as Safe Superintelligence — now deploy TPUs for large-scale model development and inference, drawn by their high efficiency and Google-backed ecosystem.

Meta has not replied to The Left Shift's request for comment at the time of publication.