Anthropic 鎖定最多 100 萬顆 Google TPU，逾 1 吉瓦容量將於 2026 年上線

Anthropic has dramatically expanded its use of Google Cloud in a deal worth tens of billions of dollars that covers up to one million TPUs and brings well over a gigawatt of capacity online in 2026, the largest compute expansion in the company history and among the biggest chip commitments any AI lab has ever made with a single cloud provider.

The Deal
Announced on Thursday, October 22, the agreement turns one of the largest line items in the Anthropic draft S-1 filing into named silicon with a delivery date. The deal covers up to one million of Google tensor processing units and more than 1 gigawatt of capacity that will come online during 2026. Google Cloud CEO Thomas Kurian said the Anthropic choice reflects the strong price-performance and efficiency its teams have seen with TPUs for years, and pointed to the seventh-generation Ironwood chip in the Google accelerator lineup. Anthropic CFO Krishna Rao said the expanded capacity exists to meet exponentially growing demand while keeping models at the cutting edge.

Why the Demand Is There
The commercial numbers behind the deal are striking. Anthropic now serves more than 300,000 business customers, and its count of large accounts, each worth more than 100,000 dollars in run-rate revenue, has grown nearly 7x in the past year. That enterprise traction is what turns a chip order into a strategic necessity: every new coding, agent and analysis workload runs on capacity that has to be reserved years in advance.

The commitment also sits against the backdrop of the Anthropic IPO preparations. The S-1 draft that surfaced this week listed hundreds of billions of dollars in future cloud commitments, including roughly 450 billion dollars of mostly noncancelable cloud contracts with Google, AWS, Microsoft and neoclouds, more than 160 billion dollars for Google TPUs sourced through Broadcom, and up to 150 billion dollars more still in talks, according to Bloomberg. One deal of this size does not replace that stack, but it shows the largest pieces are converting from plans into signed capacity.

Multi-Platform by Design
Just as important as what changed is what did not. Anthropic restated its multi-platform compute strategy across Google TPUs, Amazon Trainium and NVIDIA GPUs, and reaffirmed Amazon as its primary training partner through Project Rainier, the cluster spanning multiple US data centers. The company is not betting on a single silicon vendor; it is using its scale to run a portfolio, and the Google Cloud expansion adds a gigawatt-class option rather than replacing an existing one.

The Bigger Picture
Gigawatt-scale compute deals are becoming the defining transactions of the AI race. Training runs for frontier models now require power and silicon on a scale that only a handful of providers can deliver, and the labs with the strongest revenue growth are locking in capacity before competitors can. For Anthropic, the deal is both a demand signal and a risk statement: it commits tens of billions of dollars to infrastructure that must be financed, built and powered on schedule, even as investors debate how long the AI capex cycle can run at its current intensity.

Sources: Anthropic Newsroom, Reuters, Bloomberg, DEV Community AI Daily Digest, threerouter