Anthropic has announced a new agreement with Google and Broadcom to supply several gigawatts of next-generation computing power. This includes specialized chips–TPUs, which Google develops for artificial intelligence tasks. The new capacity is planned to become operational starting in 2027.
To put this into perspective, «several gigawatts» is not just a big number. Modern large data centers consume hundreds of megawatts. A gigawatt is on the level of an entire city district. So we're talking about an infrastructure so massive its scale will be noticeable even among the most ambitious projects in the industry.
Why Does Anthropic Even Need So Much Power?
The short answer is: demand is growing faster than the company can keep up.
Anthropic's annual revenue has exceeded $30 billion–and this is considering that at the end of 2025, it was around $9 billion. A roughly threefold increase in a few months isn't gradual scaling; it's a dramatic leap.
Another figure is also telling. When the company announced a new funding round in February, it was mentioned that over 500 corporate clients were spending more than a million dollars a year each on Claude. Now, there are more than a thousand such clients. That's a doubling in less than two months.
Simply put, Claude has evolved from an experimental tool into a core business infrastructure. And that means the load on its servers is growing not by the day, but by the hour.
Not the First Step, But a Continuation of a Grand Strategy
The new agreement isn't a pivot in strategy, but its logical continuation. Back in November 2025, Anthropic announced its intention to invest $50 billion in developing American computing infrastructure. The current deal with Google and Broadcom is the largest concrete commitment under this plan. The vast majority of the new capacity will be located in the United States.
The partnership with Google is not starting from scratch: back in October, the companies agreed to increase access to TPU capacity. The current agreement expands these arrangements to a fundamentally new level.
A Bet on Hardware Diversity
An interesting detail in this whole story is how Anthropic is building its infrastructure. The company is intentionally not betting on a single chip supplier or a single cloud platform.
Claude is trained and runs on several types of hardware: chips from Amazon, Google, and NVIDIA. The logic here is pragmatic–different tasks are better solved with different equipment, and this flexibility allows for selecting the optimal option in each specific case. Furthermore, it reduces reliance on a single supplier and increases the overall stability of the system.
Meanwhile, Amazon remains the primary cloud partner: the joint work with AWS continues as part of a project called Rainier.
As for customer availability, Claude remains the only «frontier» model (meaning one of the most powerful in the industry) that can be used across all three of the world's largest cloud platforms: Amazon Web Services, Google Cloud, and Microsoft Azure.
What This Means in Practice
From the perspective of a user or a developer working with Claude, this news means that in the coming years, the company plans to secure a reserve of computing power sufficient for future growth without any degradation in service quality.
Expanding infrastructure on this scale is usually a response to a specific problem: when demand starts to outpace the system's capabilities, latency increases, and reliability drops. Anthropic, it seems, is trying to solve this problem proactively, rather than playing catch-up.
For the industry as a whole, this is another signal that the race for computing resources hasn't gone away–and, it seems, is only gaining momentum. Companies developing large language models are forced to think not in terms of server racks, but in terms of gigawatts and multi-year infrastructure commitments.
