Anthropic is putting its eggs in Google and Broadcom’s basket. It has signed a deal for several gigawatts of cloud capacity for its AI models on Google hardware.
AI company Anthropic announced on Easter Monday that it has signed a new deal with Google. The deal secures Anthropic several gigawatts of additional capacity in Google’s data centers and also includes the use of Google’s in-house chips. Broadcom is joining the partnership as a third party, as it assists Google in the development of those chips.
Broadcom formally announced the news on the same day via a document filed with the US market watchdog SEC. In the document, Broadcom announces that it will continue to develop and supply specialized Tensor Processing Units (TPUs) to Google until at least 2031. It also mentions that starting in 2027, Anthropic will purchase 3.5 gigawatts of capacity on Google’s chips.
Not without risk
However, Broadcom’s document contains a striking clause, as noted by The Register. “Whether Anthropic actually uses this expanded AI computing capacity depends on the company’s continued commercial success,” it literally reads. This implies that Broadcom still has some doubts about Anthropic’s viability. The recent fuss with the Pentagon will have contributed to those doubts.
In its own announcement, Anthropic tries to wave away those doubts by mentioning that revenue has risen to thirty billion dollars, compared to 9 billion dollars in 2025. Yet Anthropic, like OpenAI, is currently not profitable. This is due to the high investment costs for training and running LLMs.
Fifty billion per gigawatt
Financial Times estimates the total cost of one gigawatt of AI capacity at thirty to fifty billion dollars. The newly concluded deal between Anthropic and Google has a potential value of hundreds of billions of dollars. The revenue Anthropic generates is therefore far from sufficient to cover the costs.
The company relies on generous investors to pay the bills. The latest investment round of competitor OpenAI shows that investors are still happy to open their wallets to keep the AI hype alive, and care little about their return on investment. But no company can continue to run at a loss forever, including Anthropic and OpenAI.
Loyal to AWS
One party will be less pleased with the triangular relationship between OpenAI, Google, and Broadcom. AWS has already pumped several billion dollars into Anthropic in exchange for the purchase of chip infrastructure and cloud services. To avoid biting the hand that feeds it, Anthropic confirms that AWS remains ‘the primary cloud provider and training partner’.
However, the deal with Google should ensure that it can better distribute its capacity. Anthropic is struggling with the necessary growing pains to continue supporting the popularity of Claude. Users complain about limited usage caps, and measures taken to better distribute capacity appear to make no difference for now.
“We train and run Claude on various types of AI hardware from AWS, Google, and Nvidia, allowing us to allocate workloads to the chips most suitable for them. This diversity of platforms translates into better performance and greater resilience for customers who rely on Claude for critical tasks,” Anthropic writes.