Google Takes Its TPU Fight With Nvidia Beyond Its Own Cloud
Google Takes Its TPU Fight With Nvidia Beyond Its Own Cloud
Google’s TPU push began as an internal effort to improve Search and other services. Since 2013, the company has built its own tensor processing units while selling customers a tightly integrated package of chips, models and cloud services. That strategy is now changing: Google opened TPUs to outside data centers in the second quarter, broadening its reach beyond its own infrastructure.
On Tuesday, Alphabet and Oracle announced that Oracle Cloud customers will be able to access Google TPUs from next year—the first time the chips will be offered through another hyperscaler. The partnership gives Google a new route to customers in a market Nvidia has long dominated. Check Capital’s Chris Ballard said Google is “a very strong competitor” that could eventually take Nvidia market share.
The case for Google rests on a widening customer base and the economics of its silicon. Cloud chief Thomas Kurian has said servers powered by Google’s own chips can pay for themselves in roughly half the time of other AI servers because the company can optimize the whole stack. SemiAnalysis estimates Google’s latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server.
But the expansion also exposes tensions. Google remains a major Nvidia customer, and Nvidia’s $96.2 billion in quarterly revenue still dwarfs Google’s TPU business. External sales may carry only a 20% operating margin, below Google Cloud’s roughly 35% margin, according to TD Cowen estimates.
On Wednesday, CFO Anat Ashkenazi said TPU sales should see a “meaningful revenue ramp in 2027,” while Kurian called them a “meaningful contributor” to third-quarter results. Demand is not limited to friendly customers: Anthropic uses Google chips alongside Amazon, AMD and Nvidia hardware, while OpenAI is reportedly preparing to buy TPUs. For Nvidia, the more consequential threat may not be one rival chip, but customers increasingly spreading workloads across several of them.
Write a comment