Google strikes multibillion-dollar AI chip supply deal with Anthropic
Anthropic has secured access to up to one million Google TPUs in a deal it has valued in the tens of billions, with well over a gigawatt of capacity due online in 2026 as the partners deepen their cloud and AI hardware ties.
Google has agreed to supply Anthropic with access to up to one million of its Tensor Processing Units in a deal that Anthropic has valued in the tens of billions of dollars, significantly expanding compute for the Claude model family.
The companies have said the additional capacity is due to come online in 2026, with Anthropic adding the computing power will be capable of delivering well over a gigawatt of capacity as data centres are commissioned in 2026.
The move has deepened an existing partnership between the firms: Anthropic has trained and served Claude on Google Cloud for several years and has cited TPU price‑performance and efficiency as key reasons for expanding.
Scale and energy footprint
The companies have indicated the deployment will bring “well over a gigawatt” online next year; a gigawatt is roughly comparable to the power needed for about 350,000 U.S. homes and is a key parameter in measuring the industrial scale of current AI infrastructure.
Anthropic has maintained a diversified compute strategy that includes Google’s TPUs, Amazon’s Trainium and NVIDIA GPUs; Amazon has remained a key investor and Anthropic’s primary cloud provider even as the startup has expanded its Google relationship.
Funding and market position
Founded in 2021, Anthropic has recently reported a valuation of about $183 billion after raising a further $13 billion, reflecting investor demand for frontier‑model development and enterprise AI services.
Recent competitor and ecosystem developments
OpenAI has said it has no current plan to use Google’s in‑house TPUs and has been targeting a first custom chip “tape‑out” this year to reduce reliance on third‑party suppliers.
NVIDIA has unveiled its Blackwell Ultra platform, including the GB300 NVL72 rack‑scale system for training and test‑time scaling, with availability slated for the second half of 2025.
Amazon Web Services has detailed Project Rainier, a U.S. compute cluster designed to interconnect hundreds of thousands of Trainium2 chips for large‑scale AI training.