OpenAI Switches to Google TPUs Amid Nvidia Shift

Tasia

Get ready to hear Google brag about TPUs on Wednesday. When Alphabet, Google’s parent company, reports third-quarter earnings tomorrow, the company’s in-house AI chips, tensor processing units, are sure to be on the lips of Google executives and Wall Street analysts. That’s because we got a bunch of new signs this week of the swelling customer demand for TPUs. Earlier today, Google said that Oracle will become the first major cloud provider outside Google itself to begin offering TPUs to customers through its own data centers. And a short while ago, Ryan Barwick published a scoop for us that says OpenAI—a major customer for Nvidia’s AI chips—plans to buy Google’s TPUs. But perhaps the juiciest details on TPU demand came on Monday courtesy of Anthropic’s IPO prospectus. In that filing, Anthropic chronicled the enormity of the financial commitment it has made to the Google chips, including a plan to spend $114.5 billion with Google Cloud (it didn’t specify how much of that will be for TPU-based compute). Anthropic also estimated it could spend an additional $125 billion to $150 billion on TPUs with Google and other providers. And the AI company will pony up $161.2 billion for TPU compute with Broadcom, which co-designs TPUs with Google. TPUs, you’ll recall, began as chips that Google used only internally starting over a decade ago. The company began allowing outside customers to use them via Google Cloud in 2018. More recently, Google has begun striking agreements with companies like Fluidstack and now Oracle to run TPUs in their own data centers. Google only began recognizing revenue from the sale of TPUs to outside data centers last quarter, though Chief Financial Officer Anat Ashkenazi said on an earnings call then that only a small portion of the revenues from its TPU agreements were reflected during that period. The “vast majority of revenues” from those agreements will kick in next year, Ashkenazi said. Analysts are predicting something of a windfall for Google when that happens. In a report earlier this month, TD Cowen’s John Blackledge estimated that TPUs will generate $9.4 billion in sales during the second half of this year, about 13% of the roughly $70 billion in revenue he was forecasting for Google Cloud during that period. By 2028, Blackledge thinks TPUs could be a $138 billion business, accounting for 34% of Google Cloud’s estimated revenue that year. While big, that figure would still be a fraction of the $692 billion in revenue that analysts, on average, are expecting Nvidia, the dominant provider of AI chips, to report in its fiscal year ending January 2028. Still, Google’s TPU growth has not gone unnoticed at Nvidia. Today, we got further evidence of Nvidia’s efforts to keep marquee companies in it s tent. The chipmaker announced an expansion of its partnership with Meta, which said it would increase its deployment of Nvidia chips. As part of the agreement, Nvidia will invest $3 billion into the nearly $100 billion in stock offerings that Meta announced this week. We reported earlier this year that Meta struck a deal to rent TPUs from Google. Nvidia isn’t going to let Google invade its turf without a fight. cloud.google.com/blog/pro..

cloud.google.com

Google and Oracle are partnering to bring Google TPUs to Oracle Cloud Infrastructure | Google Cloud Blog

Google and Oracle are partnering to bring Google TPUs to Oracle Cloud Infrastructure, making OCI the first hyperscaler outside Google to offer TPUs.