The deal itself
In October 2025, Google agreed to supply Anthropic with up to 1 million of its tensor processing units, a commitment reported at tens of billions of dollars without a more precise figure disclosed. The deployment was expected to bring more than 1 gigawatt of new capacity online during 2026, and the arrangement was described at the time as one of the largest AI hardware commitments any lab had made. The deal deepened an already close relationship between Anthropic and Google Cloud, which has hosted Anthropic workloads for years.
Six months later, the relationship expanded again. In April 2026, Anthropic signed a further deal bringing Broadcom into the arrangement alongside Google, covering roughly 3.5 gigawatts of TPU capacity with deployment slated to begin in 2027. Broadcom separately disclosed in a regulatory filing that it holds a Long Term Agreement with Google to help develop future TPU generations, plus a Supply Assurance Agreement to provide networking and other rack components for Google's next-generation AI infrastructure through 2031.
Why Anthropic needed it
The scale of the commitment tracks the scale of Anthropic's growth. The company's run-rate revenue reached about 30 billion dollars by early 2026, up from roughly 9 billion dollars at the end of 2025, a more than threefold increase in a few months. That growth was partly financed by a 30 billion dollar Series G round closed in February 2026 at a 380 billion dollar post-money valuation, one of the largest private funding rounds completed by any AI company to date.
The TPU capacity also supports a specific public commitment. In November 2025, Anthropic pledged to invest 50 billion dollars in US compute infrastructure, and the company has said the Google and Broadcom capacity will count toward that pledge and will sit mostly within the United States. A Financial Times report from late March 2026 added that Google may provide construction financing, potentially up to 5 billion dollars, for a Texas data center operated by Nexus Data Centers specifically to serve Anthropic's workloads.
What TPUs actually change for buyers
For most of the current AI infrastructure build-out, Nvidia GPUs have been the default unit of capacity, and GPU scarcity has shaped pricing, allocation, and vendor negotiating leverage across the industry. A deal of this size, from a lab as significant as Anthropic, demonstrates that TPU-based capacity can scale to the same order of magnitude as the largest GPU commitments being signed anywhere, with multi-gigawatt capacity and multi-year supply assurance agreements of its own.
That matters because it establishes a credible second architecture rather than a niche alternative. Broadcom's involvement is particularly notable: the same company that helps build custom silicon for multiple hyperscalers is now contractually committed to Google's TPU roadmap through 2031, which signals a long enough planning horizon that enterprise buyers can reasonably treat TPU capacity as a durable option rather than a short-term workaround for GPU shortages. A six-year supply commitment from a component supplier is not the kind of agreement companies sign around a technology they expect to be a temporary stopgap, and that alone should reset how procurement teams weigh TPU-backed cloud offerings against GPU-only ones.
The portability problem this creates
The catch is that TPUs and GPUs are not drop-in substitutes. Code, frameworks, and model architectures optimized for Nvidia's CUDA ecosystem do not automatically run efficiently, or at all, on Google's TPU stack, and teams that have spent years building Nvidia-specific tooling face real engineering cost in making workloads portable. Anthropic has the scale and the engineering resources to manage that transition as a strategic choice. Most enterprises do not have that luxury and have not needed to think about it, because GPU-based cloud capacity has been the only practical option for mainstream AI workloads.
That is changing. As TPU capacity scales and more labs sign meaningful commitments, the gap between Nvidia-only tooling and a genuinely portable AI stack becomes a cost center rather than a theoretical risk. Enterprises relying entirely on one vendor's silicon architecture for AI workloads should treat this as an early warning rather than a future problem to revisit later, because the engineering cost of retrofitting portability tends to grow with the size of the codebase you have already built, not shrink with time.
What this means for your infrastructure roadmap
This does not mean every enterprise needs a TPU strategy today. Most organizations are still consuming AI capacity through managed APIs or standard cloud GPU instances, several layers removed from the underlying silicon choice, and that insulation is likely to continue for typical workloads. What it does mean is that the assumption underlying a lot of infrastructure planning, that GPU-based cloud capacity is the only serious path to scale, is no longer accurate at the frontier.
The practical move is to start testing portability on a small scale now, before it becomes urgent: evaluate whether your model training or inference pipelines could run on TPU-backed cloud instances without a full rewrite, and factor that optionality into contract negotiations with your primary cloud provider. A credible alternative architecture, even one you never use, is leverage in a GPU-scarce market, and leverage is worth building before you need it. Boards asking why AI infrastructure costs keep rising deserve a straight answer that includes this point: a single-architecture dependency is itself a cost, measured in whatever premium you pay for not having a second option on the table.
