Anthropic Commits to Spending $200 Billion on Google’s Cloud and Chips
The Takeaway
- Anthropic committed to spend $200 billion with Google Cloud over five years as part of their recent agreement.
- Contracts involving Anthropic and OpenAI are now worth over half of the $2 trillion in backlogs at major cloud providers.
- The arrangements show how much the AI boom still hinges on the plans of two cash-burning startups.
When Google last month said it would supply Anthropic with an astonishing five gigawatts of server capacity, the companies didn’t put a dollar figure on that commitment.
But as part of the deal, which begins next year, Anthropic plans to spend about $200 billion with Google over five years, according to a person with knowledge of it. The commitment means Anthropic represents more than 40% of the “revenue backlog” Google disclosed to investors last week, reflecting contractual commitments from its cloud customers.
While that sounds like a high percentage, revenue backlogs reported by Google’s cloud rivals are even more tied to Anthropic and its archrival OpenAI.
Together, contracts involving Anthropic and OpenAI now make up around half of the $2 trillion revenue backlog across Amazon, Microsoft, Google and Oracle, the four biggest U.S. cloud providers.
It’s a reflection of just how much the AI boom still hinges on the plans of two cash-burning startups.
Backlog figures don’t include revenue the cloud providers have already generated from the AI firms. Companies report them to give investors an indication of future revenue, based on long-term deals. Those backlogs have ballooned as Anthropic and OpenAI have grown.
The cloud providers have anticipated such deals for years. Google, Microsoft and Amazon have each invested billions of dollars in the two AI leaders in the hopes they will become large renters of cloud servers, and ultimately deliver the cloud providers revenue that is many times the size of their investments.
OpenAI and Anthropic are already among the largest, if not the largest, such renters of cloud servers. OpenAI is projecting to spend around $45 billion on servers this year, up from about $17 billion last year, The Information has reported. Much of OpenAI’s spending this year is likely to happen with Microsoft, an early backer that has invested more than $13 billion in the ChatGPT maker.
Anthropic has projected it would spend upwards of $20 billion renting servers this year, or about triple what it spent last year. That projection, however, occurred in mid-December, before Anthropic’s revenue exploded in the first quarter, so the final figures could be materially higher.
Google Cloud runs the application programming interface that makes up a majority of Anthropic’s business. That helps explain Google Cloud’s remarkable 63% revenue growth in the first quarter. (Revenue growth at the much bigger Amazon Web Services, which also helps power Anthropic’s API, accelerated nearly five percentage points to 28% in the first quarter compared to its growth in the prior quarter.)
In addition to reporting Google Cloud’s revenue acceleration, Google also last week said its revenue backlog roughly doubled to more than $460 billion in the first three months of this year; the new deal with Anthropic would account for much of that increase.
Similarly, Amazon’s revenue backlog, which it calls performance obligations, rose 49% to $364 billion in the first quarter. OpenAI increased its spending commitment with AWS by $100 billion in that period, Amazon said, representing more than 80% of the increase. Amazon in April also struck a server rental deal with Anthropic worth $100 billion for “up to” five gigawatts over the next 10 years. That capacity would consume more than two Hoover Dams’ worth of electricity.
In total, OpenAI and Anthropic spending commitments make up about half of the revenue backlog AWS has announced, including the April deal.
In aggregate, Anthropic projected late last year that it would pay Amazon, Google and Microsoft about $200 billion to run its Claude AI on their cloud servers through 2029, according to the startup’s most optimistic forecasts from mid-December. OpenAI, meanwhile, projected last quarter it would spend about $180 billion on servers in 2029 alone.
Some investors may be skeptical about OpenAI and Anthropic’s ability to spend so much money. The two companies’ projections assumed that by 2029, they will have grown their revenues 20 to 30 times compared to what they generated in 2025.
Oracle’s shares have fallen 45% since September, when it effectively announced a $300 billion cloud computing deal with OpenAI that’s set to run through 2032, as investors worry that the spending may not materialize or might happen on other clouds instead.
How Oracle is Different
Competitors do have at least one advantage over Oracle: they make money from Anthropic and OpenAI in other ways besides renting out servers, namely by selling the two firms’ AI to cloud customers. As part of an updated partnership agreement between Microsoft and OpenAI, for instance, Microsoft will keep 100% of sales from selling OpenAI models to Azure cloud customers, which could total billions of dollars this year. Previously, Microsoft gave OpenAI a 20% cut of such sales.
Similarly, Amazon and Google keep a significant share of sales of Anthropic models to their cloud customers. At the end of last year, Anthropic expected cloud providers to keep nearly $2 billion from reselling Anthropic models this year, and another $6 billion next year. Oracle doesn’t host and resell models from OpenAI and Anthropic yet.
OpenAI and Anthropic plan to go public long before 2029, making it easier for them to raise capital. They also could raise money from Nvidia, which has already invested billions of dollars in given each of them, and has plenty of reasons to ensure they can afford to keep spending heavily on servers powered by Nvidia chips. Nvidia has long acted as a kind of central bank to dozens of companies that buy or rent its server chips.
One advantage for Google is it uses in-house AI server chips to power computing for Anthropic, which likely produces better profit margins than when it rents out Nvidia’s expensive AI servers to cloud customers. (Google pays its chip partner Broadcom for some services related to the AI chip.)
Amazon also has invested in in-house AI chips to power services such as Bedrock, which Amazon Web Services customers use to access Anthropic and other models.
Amazon is poised to improve its financial position further when OpenAI starts using a lot of Trainium chips from AWS rather than just renting Nvidia chips. When Amazon announced it would invest $50 billion in OpenAI and expand the two companies’ cloud computing deal, Amazon also said in a securities filing that OpenAI use of Trainium would consume 2 GW of computing capacity.
Microsoft’s in-house AI server chip is nascent, while Oracle hasn’t produced such a chip.