Google began developing tensor processing units, or TPUs, in 2013, initially to run its own services. For years, that strategy kept the chips inside a tightly integrated cloud offering: Google sold customers the infrastructure, models and silicon together. But as Nvidia’s GPUs became the default hardware for the AI boom, Google’s closed approach limited its reach.¹

The pivot started in the second quarter, when Google opened TPUs to outside data centers. On Tuesday, it went further: Oracle said it would offer the chips to its cloud customers from next year, the first time TPUs will be available through another hyperscaler. The agreement gives Google a new route to customers that might not otherwise buy Google Cloud — and gives Oracle another weapon in its own cloud contest with Amazon and Microsoft.²

The stakes are larger than one partnership. Nvidia chief executive Jensen Huang had earlier dismissed the prospect of broad TPU growth, asking, “Without Anthropic, why would there be any TPU growth at all?” Yet Anthropic is not Google’s only opening: OpenAI is reportedly planning TPU purchases, while customers increasingly mix chips from Google, Nvidia, AMD and Amazon depending on workload and cost.

Google’s latest earnings commentary reinforced the direction of travel. Chief financial officer Anat Ashkenazi said TPU sales should see a “meaningful revenue ramp in 2027,” while cloud chief Thomas Kurian called them a meaningful contributor to third-quarter results. Analysts project a potentially huge business, though Nvidia remains vastly larger.

Still, expansion carries a price. Kurian has argued that Google’s own silicon pays back faster because the company can optimize the whole stack. External sales, by contrast, are estimated to generate lower margins than Google Cloud overall. Google is therefore making a calculated exchange: less control and efficiency per deployment, in return for scale — and a chance to make Nvidia’s dominance less absolute.