Meta reportedly eyes Google AI chips for 2027

Meta is reportedly in talks to buy billions of dollars’ worth of Google’s AI chips for use in its data centers from 2027, according to The Information.

The discussions also include Meta potentially renting Google Cloud’s tensor processing units (TPUs) as early as next year.

This would be a shift for Google, which so far has limited the use of its TPUs to its own data centers.

If finalized, the deal could position Google as a direct competitor to Nvidia in the data center chip market.

Meta is one of Nvidia’s largest customers, with planned spending of up to US$72 billion this year.

Shares of Alphabet rose over 4% in premarket trading following the report, while Nvidia’s stock fell by 3.2%.

Other companies, including Anthropic, have also recently expanded their use of Google’s AI chips.

Neither Meta, Alphabet, nor Nvidia have commented on the reported talks.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Google’s TPU cost edge may not lead to broad Nvidia displacement

TPU v6e hits up to 4x performance per dollar over Nvidia H100 on some large language model training, and TPU v5e reached 2.7x over TPU v4 on GPT-J benchmarks 1 2. This edge holds in Google data centers with custom interconnects that run all-reduce, the gradient-aggregation step in distributed training, 10x faster than Ethernet GPU clusters 1.
Software maturity sets the pace. Nvidia CUDA, Compute Unified Device Architecture, is stable and widely supported. TPU support in PyTorch, an open-source deep learning framework, feels pretty brittle to developers 3. Moving needs PyTorch/XLA, Accelerated Linear Algebra, which raises migration work for enterprise AI teams 4 3 and kept many TPUs at Google for years 3. Meta may adopt TPUs 3.
Nvidia keeps a strong lead through its ecosystem and room to cut margins if needed 3. TPU growth hinges on Google extending its in-house cost edge to external use while breaking some CUDA lock-in 3.

Cross-accelerator complexity lifts demand for enterprise AI infrastructure

MLOps and migration tools see rising demand. Turning a PyTorch model to TPU means swapping CUDA device calls for torch_xla.device(), part of the PyTorch/XLA library for TPUs 5. Teams also wrap data loaders with MPDeviceLoader, a multi-process data-loading helper in torch_xla 5. They add optimization barriers, compiler fences that stop unsafe code motion, to avoid compiler issues 6. Anthropic, an AI startup, has expanded TPU use yet many companies still need specialists.
A window for services is open now. Snap, parent company of Snapchat, locked in 10,000 TPU v6e chips through capacity management 1. Snap and peers use reserved capacity planning with committed use discounts 1. Infrastructure firms like Introl, a cloud and AI infrastructure consultancy, guide enterprises on infrastructure deployment decisions 1. These choices balance performance with cost and operational complexity 1. Hybrid GPU TPU strategies are spreading as companies hedge accelerator risk 1.