Anthropic to Expand Compute With Up to 1 Million Google TPUs

Anthropic has announced plans to dramatically expand its artificial intelligence infrastructure on Google Cloud, gaining access to up to one million TPUs, the chips Google designed to train and run AI models. The deal, valued at tens of billions of dollars, is expected to bring more than one gigawatt of computing capacity online in 2026.

The figure puts the deal among the largest computing contracts in the race to build language models. For Anthropic, the company behind Claude, this is about more than training more capable systems: it also needs to serve a rapidly growing business customer base and have the resources to test its models before deployment.

One million chips doesn’t mean one million home computers

A TPU, short for Tensor Processing Unit, is a processor specialized for the mathematical operations that underpin neural networks. Unlike a more general-purpose GPU, such as those Nvidia dominates, it is specifically designed to efficiently move and multiply enormous matrices of numbers.

Google has used these units in its own products and models for years and sells them through its cloud business. The seventh generation, Ironwood, is part of that offering. Anthropic’s plan to use up to one million of these accelerators is a significant validation of the TPU as an option for frontier workloads—those that require entire data centers rather than a handful of servers.

The figure of more than one gigawatt helps put the scale in perspective. A gigawatt is a unit of power, not energy: it indicates the electrical demand the infrastructure can reach while operating. Bringing that capacity into production requires more than chips; it also requires buildings, networks, cooling systems, and power supplies. At this stage of AI development, the bottleneck is no longer just designing a model—it is also building the physical infrastructure needed to train and serve it.

Google sells capacity to a company competing in models

The deal has another industrial dimension. Google is developing Gemini and selling its own AI services, while Anthropic is competing for enterprise customers with Claude. Even so, Google Cloud gains a high-volume customer for its infrastructure and shows that its platform can attract labs outside Google.

“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,” said Thomas Kurian, CEO at Google Cloud. “We are continuing to innovate and drive further efficiencies and increased capacity of our TPUs, building on our already mature AI accelerator portfolio, including our seventh generation TPU, Ironwood.”

That balance between cost and performance matters as much as raw power: training and operating an advanced model involves processing enormous amounts of data and responding to millions of requests, so small efficiency gains translate into very different costs at scale.

For Google, the contract also strengthens a way to compete with Nvidia that does not involve selling chips to third parties, but rather renting out its own infrastructure. Nvidia remains central to the AI accelerator market, but major technology companies are developing their own silicon to reduce dependence, control costs, and secure supply.

A three-platform strategy

Anthropic will not concentrate all its capacity at Google. The company describes a diversified strategy built around three chip families: Google’s TPUs, Amazon’s Trainium, and Nvidia’s GPUs.

Trainium is the AI accelerator designed by Amazon Web Services. Anthropic continues to count Amazon as its primary training partner and cloud provider, and it is still working with the company on Project Rainier, a massive cluster distributed across multiple U.S. data centers that will bring together hundreds of thousands of AI chips. Nvidia’s GPUs round out the mix.

This combination has a practical advantage: it keeps Claude’s future from depending on a single architecture or provider. It also requires Anthropic to adapt its software, tools, and training processes to different platforms—a considerable technical undertaking. In return, however, it gains more flexibility to choose where it is most efficient to train, test, or serve each model.

Enterprise demand explains the investment

Anthropic says it now serves more than 300,000 business customers. Its number of large accounts—customers that each represent more than $100,000 in run-rate revenue—has grown nearly sevenfold over the past year.

That growth explains why the investment is not aimed exclusively at the next model. The additional capacity will help meet rising demand and enable more extensive testing, alignment research, and responsible deployment at scale. Alignment refers to the set of techniques used to make a system follow instructions and behave within limits defined by its creators.

The expansion is planned for 2026, so there is still work to do before the announcement becomes operational data-center capacity. But it makes one reality of the industry clear: companies seeking to remain at the frontier of language models need to secure chips, energy, and cloud capacity years in advance. Anthropic has chosen to do so without tying itself to a single class of processor, while Google has secured a much larger role for its TPUs in that plan.