Anthropic shields itself with several gigawatts of Google TPU after de facto admitting that its infrastructure can no longer cope
Anthropic has announced a new agreement with Google and Broadcom to secure several gigawatts of next-generation capacity based on TPUs starting in 2027. The company presents the move as its largest computing commitment to date, a key piece to continue scaling Claude and respond to a demand that, according to its own data, has skyrocketed in 2026. Beyond the headline, the news leaves a fairly clear reading: Anthropic needs to shield its infrastructure now because current growth is beginning to strain its real service capacity.
The company ensures that its run-rate revenue already exceeds 30,000 million dollarscompared to about 9,000 million at the end of 2025, and that the number of business customers who spend more than one million annually has doubled in less than two months, now exceeding one thousand. These are very striking figures, but they also serve to remind us of something less striking: in artificial intelligence it is not enough to sell a lot of software if there is then a lack of chips, energy and deployment capacity to sustain that growth.
In fact, the announcement fits almost as a natural continuation of the episode we saw just a few days ago, when Anthropic began charging extra to the most intensive users of Claude Code due to the pressure that external agent tools were generating on its infrastructure. So the company was talking about operational limitations and usage patterns that its plans were not designed for. Now comes this new step: ensuring massive power in the future because commercial success is already colliding with the physical limit of the system.
More business alone does not solve the AI bottleneck
Anthropic insists that it maintains a diversified hardware strategy, combining AWS Trainium, Google TPU and NVIDIA GPUand remember that Amazon remains its primary cloud and training provider. But this agreement with Google and Broadcom shows that TPUs are gaining weight as part of the structure that Claude needs to continue growing. It also reinforces, albeit in passing, Google’s position as an increasingly important strategic partner within the Anthropic ecosystem.
The bottom line is that the AI market is no longer decided just by who has the most powerful model. It is also a question of who manages to deploy more capacity, faster and with less operating cost per token. The software scales, but the infrastructure does not scale for free. You can raise prices, tighten limits, or attract more enterprise customers, but if the demand for agents, scheduling, and automation continues to rise, the bottleneck ends up appearing in the data centers.
That’s where this news connects with a broader trend in the sector. We recently saw how Google introduced TurboQuant to compress AI memory up to six times without penalizing performance, precisely with the aim of reducing one of the great costs of modern inference. Those kinds of advances matter a lot because the industry no longer needs only more capable models, but also much more efficient systems.
The next big battle will be to maintain intelligence without increasing the cost
That doesn’t mean efficiency is going to solve everything. In fact, more efficient AI can also absorb more demand and further expand the total use of the system. But even with that paradox, the direction seems inevitable: efficiency without loss of intelligence will become increasingly important. Not only to make the service cheaper, but to prevent commercial growth from becoming unsustainable in terms of infrastructure.
Anthropic’s announcement goes right along those lines. In the short term, the company needs more capacity because Claude is outgrowing the available infrastructure. In the medium term, it also needs more efficient models and inference techniques so that this growth does not become a spiral of cost and scarcity. More gigawatts help, but not enough on their own. The AI that is coming will need much more hardware, yes, but also much more intelligence when using it.
