Google LLC said this week it has struck a deal with the artificial intelligence startup Anthropic PBC that could be worth tens of billions of dollars.

The deal will give Anthropic access to up to one million of Google Cloud’s tensor processing units or TPUs, which it wants to use to train and run large language models such as Claude. The company, which is perhaps the closest competitor to OpenAI, will also use Google’s cloud services, it said.

Anthropic said in an announcement that it has chosen Google’s TPUs because of their price performance and efficiency, and also due to its previous positive experience with using the chips. The companies’ partnership dates back to 2023, and the new deal will help the AI company to “continue to grow the compute we need to define the frontier of AI,” said its Chief Financial Officer Krishna Rao.

The companies didn’t talk about numbers, but it’s likely that the deal will bring Google tens of billions in revenue over the next few years. Anthropic said it will gain access to well over a gigawatt of compute capacity by 2026, and experts peg the average cost of that much computing power at around $50 billion per year.

AI companies have been scrambling to secure access to the data center resources they need to power their LLMs. This year, OpenAI has struck multiple deals with companies including Nvidia Corp., Advanced Micro Devices Inc. and Broadcom Inc. to secure access to their chips. It’s also partnering with Oracle Corp. and SoftBank Group Corp. on the ambitious Stargate initiative, which aims to bring an incredible 33-gigawatts online in the next few years.

While Anthropic’s deal with Google is smaller, it’s a strong validation of the TPUs, which are a unique asset for the cloud provider. It’s said that Goole’s TPUs are more power efficient because they use a vertical power delivery system, whereas Nvidia’s graphics processing units rely on lateral power delivery. Google works with Broadcom to design and manufacture the TPUs, but there is talk of a deal with Taiwan-based MediaTek Inc., which could further reduce the costs of producing them.

Google doesn’t just sell TPU capacity to cloud customers. It also uses them extensively to power its own AI services, and it’s believed to be a huge business. In a recent analysis, D.A. Davidson suggested that if Google were to spin off the TPU unit and its AI research organization DeepMind, those two entities could be worth as much as $900 billion.

Anthropic’s multicloud bet

Anthropic’s deal with Google does raise some interesting questions about the model maker’s ongoing relationship with Amazon Web Services Inc. Like Google, Amazon is a major investor in Anthropic and provides the company with the vast majority of its AI processors, but in recent months some analysts have questioned the competitiveness of its chips.

In a note to clients last week, Jeffries analyst Blayne Curtis suggested that Anthropic’s growing use of Google’s TPUs doesn’t reflect well on Amazon’s specialized Trainium processors. This week, UBS analyst Karl Keirstead wrote that there are signs Anthropic could shift towards Google as its primary compute provider.

Still, Amazon retains a much larger stake in Anthropic. While Google has invested $3 billion in the AI startup to date, Amazon has funded it to the tune of $8 billion, and is still considered to be its primary source of cloud computing infrastructure. Anthropic has previously said that Project Rainier, a custom-built supercomputer designed to power its Claude LLMs, runs exclusively on Amazon’s Trainium 2 chips.

For its part, Anthropic insisted that it remains “committed to our partnership with Amazon”. It noted that AWS remains its primary training partner and cloud provider, and will continue to work with it on Project Rainier.

If anything, the deal simply shows that Anthropic is pursuing a sensible multicloud strategy, using a combination of Amazon’s Trainium chips, Google’s TPUs and Nvidia’s GPUs. By not relying on a single provider, the company’s services are much more resilient, as we saw earlier this week when Claude remained online, unaffected by the recent AWS cloud outage that took out many other service providers.

Anthropic’s strategy is further justified by its explosive growth this year. The company now has an annual revenue run rate approaching $7 billion, and Claude is used by more than 300,000 businesses globally, up 300-times in the last two years. Moreover, large customers that deliver at least $100,000 per year in revenue have increased almost sevenfold in the past year.

The company’s agentic coding assistant Claude Code is also doing very well, generating more than $500 million in annual revenue within two months of its launch.

Photo: Google Cloud