Google Will Sell TPU Chips to Select Customers Under Multi-Gigawatt Anthropic, Meta Pacts
Google will sell TPU chips for on-premises installation to select customers, shifting from a rental model into AI labs and HPC deployments. TPU 8t for training and TPU 8i for inference are backed by multi-gigawatt pacts with Anthropic and a multibillion-chip deal with Meta, pitting Google directly against Nvidia.
1. Shift in Distribution Model
Google is moving from renting TPU capacity via its data centers to selling physical TPU hardware to a select set of customers for installation in their own facilities. This pivot aims to expand its addressable market and offer tailored solutions for AI labs and high-performance computing operators.
2. New TPU 8t and 8i Launches
The company introduced TPU 8t, optimized for AI training workloads, and TPU 8i, tailored for inference tasks. These next-generation chips promise higher throughput and improved power efficiency over prior models.
3. Major Customer Agreements
Google secured multi-gigawatt hardware purchase agreements with Anthropic for deployment starting in 2027 and finalized a multibillion-chip contract with Meta. These deals underscore enterprise demand for on-premises AI acceleration.
4. Competitive Impact on Nvidia
This on-premises sales strategy intensifies competition in the AI accelerator market, directly challenging Nvidia’s GPU dominance. Enterprises now have an alternative source for AI hardware, which could influence Nvidia’s data center sales and pricing dynamics.