For years Google pitched a single bundle: its accelerators, its models and its cloud, sold as one stack that Amazon and Microsoft kept trying to copy. This week it began renting pieces out. Alphabet said a day before its third-quarter results that Oracle’s cloud will carry Google’s tensor processing units, the custom silicon the company has refined since 2013.
That marks the first time buyers can reach TPUs without running workloads inside Google facilities. The move belongs to cloud chief Thomas Kurian, who logged more than two decades at Oracle before moving to Google in 2018 and is now hunting for growth further afield. Wall Street looks for roughly 90 percent growth in Google Cloud revenue, up from 82 percent in the second quarter, with chips feeding the number.
Projections for the chip line diverge sharply but all point upward. Numbers from two banks bracket the range. TD Cowen puts second-half TPU revenue at $9.4 billion and 2027 at $62.5 billion, while Wells Fargo looks for about $9 billion in the third quarter and $99 billion across next year. Google first let outside data centers run TPUs in the second quarter, and chief financial officer Anat Ashkenazi has said most money from signed deals lands in 2027.
Economics are less generous than for cloud. Selling chips yields maybe a fifth of revenue as operating profit, against the mid-thirties the overall cloud unit posted in the prior quarter, per TD Cowen. On the cost side, the research firm SemiAnalysis puts Ironwood, Google’s newest accelerator, at roughly 44 percent lower total cost than an Nvidia GB200 server.
The biggest buyer is a competitor. Anthropic agreed last October to use up to one million TPUs, and OpenAI also intends to purchase them, CNBC confirmed.
