Anthropic has locked in a massive cloud deal with Google, giving the AI company access to 1 million Tensor Processing Units (TPUs) and 1 gigawatt of compute power by 2026.
The agreement is about compute; real, measurable, high-grade compute. Industry estimates peg the cost of building 1 gigawatt of AI data center capacity at around $50 billion, with $35 billion typically going just to the chips.
This project is central to Anthropic’s roadmap, and unlike OpenAI’s loud 33-gigawatt “Stargate” vision, this one is built for execution.
Claude expands while Anthropic multiplies its vendor stack
Founded by former OpenAI researchers, Anthropic isn’t out here performing. It’s calculating. It’s methodical. And it’s scaling the hell out of its infrastructure. Their Claude models — including the newer Claude Code — are split across Google’s TPUs, Amazon’s Trainium chips, and Nvidia GPUs. Each chip does different work. Training is done in one spot, inference in another, and research elsewhere.
That setup wasn’t just for show. Krishna Rao, Anthropic’s CFO, said, “Anthropic and Google have a longstanding partnership and this latest expansion will help us continue to grow the compute we need to define the frontier of AI.” Google backed that up by saying TPUs deliver “strong price-performance and efficiency.”
With this kind of architecture, Anthropic can stretch every dollar of compute further than the single-cloud players. Their pricing, performance, and power costs are all tuned per vendor, letting them avoid overpaying just because of exclusivity.
Thomas Kurian, CEO of Google Cloud, called the deal an expansion of an already successful relationship, noting, “Anthropic’s choice to significantly expand its usage of TPUs reflects the strong price-performance and efficiency its teams have seen with TPUs for several years.” Google’s new Ironwood TPUs are part of the package too.
Business is growing just as fast as the infrastructure. Anthropic’s annual revenue run rate is closing in on $7 billion, with over 300,000 businesses now running on Claude — up 300× in just two years. The number of large enterprise clients spending over $100,000 annually has gone up nearly sevenfold in the past year.
And then there’s Claude Code. It pulled in $500 million in annualized revenue just two months after launching. That number alone made Claude Code the fastest-growing product ever, according to Anthropic.
Amazon builds deeper ties while Google increases capital
While Google is loading up TPUs and stacking cash into the company, Amazon still holds the deeper seat. The retail giant has now invested $8 billion into Anthropic, compared to Google’s confirmed $3 billion in equity. That $8 billion isn’t just a number. It makes AWS the primary cloud provider, not just on paper but in the way the infrastructure is physically built.
Take Project Rainier for instance, the custom supercomputer built by Amazon specifically for Claude, which runs on Trainium 2 chips that skip the markup that comes with traditional options. That’s more compute per dollar, which is really the only thing that matters when you’re burning power by the gigawatt.
Wall Street already noticed. Alex Haissl from Rothschild & Co Redburn estimated that Anthropic added one to two percentage points to AWS’s growth during the last quarter of 2024 and the first quarter of 2025. He expects that number to hit over five points in the second half of 2025.
Scott Devitt from Wedbush told CNBC that once Claude becomes standard for enterprise devs, it directly pushes up AWS revenue. He said that dynamic could drive Amazon’s cloud growth for “many, many years.”
Back at Google, the investments keep stacking. After the first $2 billion and 10% equity stake, Google added another $1 billion in January. That came before the TPU expansion was announced.
And while AWS had an outage Monday, Claude kept running — unaffected — because of Anthropic’s multi-cloud approach.
Join a premium crypto trading community free for 30 days – normally $100/mo.
Source: https://www.cryptopolitan.com/anthropic-inks-deal-for-million-google-tpus/
