Google spent a decade building its own AI chips, known as tensor processing units or TPUs, mostly for its own use. This week showed how much that has changed. On Oct 27, Oracle said it will become the first major cloud provider outside Google to offer TPUs to its customers, starting in 2027 and ramping up over several years [1]. The same day, OpenAI, long one of Nvidia’s biggest buyers, was reported to be planning to buy Google’s chips too [2].
The demand is real and large. Anthropic’s IPO filing shows a $125.2 billion, five-year commitment to lease TPU computing capacity, with Broadcom, which co-designs the chips with Google, lending up to $42 billion to help fund it [3]. For investors, the question is whether Google has turned an internal tool into a business that can chip away at Nvidia’s dominance.
#TPU Sales Are Starting to Show Up in Google Cloud
The early evidence sits inside Google Cloud. Alphabet said cloud revenue more than doubled to $30.7 billion in the third quarter, up 103%, and named TPU systems sales as a meaningful contributor [4]. That growth rate has climbed every quarter for a year.
Year-over-year figures are the honest read here, since they compare each quarter with the same period a year earlier. Analysts expect much more to come. TD Cowen estimates TPUs could bring in $9.4 billion in the second half of 2026 and become a $138 billion business by 2028, about a third of Google Cloud’s expected revenue that year [2]. Google’s finance chief has said most revenue from its outside TPU deals will be booked from next year.
#Why Customers Want an Alternative to Nvidia
The appeal is cost and supply. Nvidia’s chips are powerful but expensive and often hard to get. Google says its newest training chip delivers nearly three times the computing power per pod of the previous generation [1]. Oracle’s backing matters because it lets companies use TPUs without moving the rest of their computing over to Google.
Nvidia is not standing still. This week it agreed to invest $3 billion in Meta’s huge share sale, with Meta expanding its use of Nvidia hardware, even though Meta has also rented TPUs from Google. Nvidia still has the deepest software ecosystem in AI, and that keeps many developers loyal.
A sharper way to see the markets in just 5 minutes.
Same news, different lens. We cut through the noise and hand you the overlooked ideas and the deeper read the crowd misses. Join 38,000+ investors seeing the markets differently.
#The Risks Behind the Headline Numbers
Much of the TPU story rests on a few very large customers, led by Anthropic. Anthropic’s own filing flags that Broadcom’s dual role as supplier and lender could create conflicts of interest. If one or two big buyers slowed their spending, the forecasts would look very different.
Building this capacity is also costly for Google. Alphabet spent $51.4 billion on capital projects last quarter, pushing free cash flow negative, and its shares fell about 10% after results despite the cloud surge [5]. Investors want proof that TPU revenue will grow fast enough to justify the bill.
The next markers to watch are Alphabet’s fourth quarter cloud figures, the timing of Oracle’s TPU rollout, and whether OpenAI’s reported purchase becomes a signed deal. Google now has a credible alternative to Nvidia. Whether its TPU business can generate attractive returns on the enormous investment required is still an open question.