Google's AI Chips Are Leaving the Building
TPUs, the custom chips Google has built since 2013, are now available through Oracle for the first time outside Google's own data centres. OpenAI wants in too, and Anthropic already has a deal worth tens of billions.
Key points
- For the first time in the technology's 13-year history, Google's TPU chips are available through a third-party cloud provider, Oracle, as of this week.
- TD Cowen forecasts $62.5 billion in TPU sales for 2027; Wells Fargo puts that number at $99 billion.
- Anthropic, whose Claude AI models compete directly with Google's own Gemini, agreed in October 2025 to use up to one million TPUs in a deal worth tens of billions of dollars.
- OpenAI, maker of ChatGPT, is planning to buy TPUs from Google, confirmed by CNBC after The Information reported it first.
- A $63 billion financing vehicle backed by Blackstone and three large banks will buy TPUs and lease them to Anthropic, covering roughly half of Anthropic's projected 2027 compute needs.
Google built its TPU, or tensor processing unit, a custom chip designed to run AI calculations, for more than a decade as a private competitive weapon. This week that changed.
Alphabet confirmed on Tuesday that Oracle will offer Google TPUs through its own cloud platform, making Oracle the first company to sell access to TPUs outside Google's infrastructure. CNBC reported the move a day before Alphabet's third-quarter earnings.
Why is Google opening up now?
Google Cloud boss Thomas Kurian is chasing a bigger slice of the AI chip market that Nvidia still dominates. More partners means more TPUs in more hands, building the customer base faster than Google can fill data centres alone.
The economics make a reasonable case. Research firm SemiAnalysis estimates Google's latest Ironwood TPU costs about 44% less to own and operate than a comparable Nvidia GB200 server. At a Goldman Sachs conference in September, Kurian said AI servers running Google's own silicon pay for themselves in roughly one year, versus about two years for external silicon, because Google can tune every layer of the stack together.
There's a cost to selling outside, though. TD Cowen estimates TPU sales carry a 20% operating margin, well below Google Cloud's overall margin of roughly 35% last quarter. That gap will widen as external volume grows, and it's the number worth watching.
Who is already buying?
The biggest TPU customer is also one of Google's sharpest AI rivals.
Anthropic, the company behind the Claude family of AI models, locked in a deal last October for up to one million TPUs. Mizuho's analysts put that agreement at nearly a gigawatt of computing capacity online this year. As we reported on 17 September, Oracle's AI cloud business just had its best quarter ever, taking on over $100 billion in debt to build the kind of infrastructure it'll now use to sell Google's chips, which makes Tuesday's announcement a logical next step. The supplier-competitor dynamic is unusually pointed: Google is, in Mizuho's words, arming its closest model competitor.
This week, chip designer Broadcom detailed a $63 billion financing vehicle backed by Blackstone and three large banks that will purchase TPUs and networking gear, then lease them to Anthropic for five years. William Blair's analysts say the arrangement covers about half of Anthropic's projected 2027 compute needs, which implies more financing deals are coming. In its IPO filing published Monday, Anthropic said it sources chips from Google, Amazon, Nvidia and AMD, selecting hardware based on price, performance and efficiency for each specific task.
OpenAI is also planning to buy Google TPUs, CNBC confirmed Tuesday. That's a notable pivot: our story from 17 September found that OpenAI built its first custom AI chip in under 20 months, promising response times up to 3.6 times faster than Nvidia hardware. Buying Google silicon alongside its own suggests OpenAI isn't betting on a single supply chain.
With $25 billion in second-quarter cloud revenue, Google still trails Amazon Web Services at $42 billion and Microsoft Azure at $29 billion. But it's growing fastest, and analysts expect cloud revenue to have jumped roughly 90% year-on-year when third-quarter numbers land.
For anyone using Claude or ChatGPT, the practical upshot is simple: the chips running some of the most-used AI tools increasingly come from Google, whether those companies admit it or not.
Common questions
What is a TPU and why does it matter?
A TPU, tensor processing unit, is a chip Google designed specifically to run AI software, as opposed to general-purpose chips or Nvidia's GPUs. Built for one job, it does that job faster and more cheaply than most alternatives.
Does this affect what AI tools I can use?
Not directly. The deal is between Google and other businesses. End users of ChatGPT or Claude won't change how they log in or pay, but the underlying hardware running those services may now be Google silicon.
What does Anthropic's IPO filing say about chip choices?
Anthropic said in its Monday filing that it uses chips from Google, Amazon, Nvidia and AMD, selecting whichever option offers the best combination of price, performance and efficiency for a given task.
