Since 2018, a company that wanted to rent Google’s tensor processing units has almost always done it through Google Cloud. Starting in 2027, Oracle Cloud Infrastructure will offer the chips from its own data centers, and OpenAI, one of Nvidia’s most valuable customers, reportedly plans to buy TPUs as well.
Oracle and Google announced the OCI arrangement Tuesday at Oracle AI World in Las Vegas, calling OCI the first hyperscale cloud outside Google to offer TPUs. Capacity arrives in calendar 2027 and ramps over several years. Oracle wouldn’t say how big the first deployment will be, but Karan Batta, a senior vice president at OCI, told DCD it is a “multi-year, multi-gigawatt partnership” and that “when Google launches a new TPU, we will have it on day one.”
Hours later, on Tuesday evening, The Information reported that OpenAI plans to buy Google’s TPUs, citing a person familiar with the matter. Financial terms and other details couldn’t be learned, the outlet said. Bloomberg reported the same plan, also citing a person familiar with the matter, and CNBC said it had confirmed it. OpenAI and Google declined to comment to Bloomberg. Nvidia, for its part, has invested tens of billions of dollars in OpenAI, The Information noted.
How the chips got out of Google’s data centers
Google built TPUs for its own use more than a decade ago and began renting them to cloud customers in 2018. Selling the hardware for other people’s data centers is much newer. Google opened TPUs to outside facilities in the second quarter, CNBC reported this morning, and The Information names Fluidstack as one of the companies running them. Google is also working with Blackstone on Crux AI, a U.S.-based TPU cloud, and DCD said Oracle could be first to bring TPUs online outside Google’s platform depending on which one launches sooner.
Google’s biggest TPU customer is Anthropic, whose Claude models compete with Gemini. Anthropic agreed last October to use up to 1 million TPUs, and its IPO filing on Monday lists a planned $114.5 billion in spending with Google Cloud, The Information reported, without saying how much of that goes to TPU compute. In the same filing, quoted by CNBC, Anthropic said that using chips from Google, Amazon, Nvidia and AMD lets it “select chips based on the best price-performance and efficiency characteristics for each workload.” WebProNews covered the rest of that S-1 on Monday.
What OCI customers would be renting
The press release points to Google’s eighth generation, which is split into two chips. TPU 8t handles training, with nearly three times the compute per pod of the previous generation, and TPU 8i is built for inference and agent workloads. In a Google Cloud blog post co-written by Batta and Google senior director of product management Gurmeet Goindi, the companies put the eighth-generation TPUs at up to 12.6 petaflops of FP4 compute per chip. The TPUs will talk to each other over Google’s own interconnect and reach the rest of OCI through Oracle’s Acceleron networking.
Google’s pitch to developers leans on software. The company says TorchTPU, its PyTorch stack that it open-sourced last week, works with tools such as vLLM, SGLang and Ray, so a model can move between TPUs on Google Cloud and on OCI without rewriting its code. Oracle is also pitching proximity, since companies whose systems of record already live in Oracle databases could fine-tune and serve models next to that data.
“Different AI workloads have different compute needs,” Mahesh Thiagarajan, OCI’s president, said in the release. Oracle isn’t dropping Nvidia: the same day, it made Nvidia’s Vera Rubin NVL72 systems generally available on OCI, The Register reported.
For Google, outside sales carry a cost. TD Cowen estimates TPU sales earn a 20% operating margin, against Google Cloud’s overall margin of about 35% last quarter, CNBC said. The firm models $62.5 billion in TPU sales for 2027, and Wells Fargo forecasts $99 billion. Chief Financial Officer Anat Ashkenazi said in July that the “vast majority” of revenue from existing TPU agreements would show up next year.
Alphabet reports third-quarter results today and holds its earnings call at 4:30 p.m. ET. Analysts expect Google Cloud revenue to grow about 90% from a year earlier, CNBC reported.
from WebProNews https://ift.tt/R8CDBho
No comments:
Post a Comment