For years, Google has promoted to customers the advantage of acquiring its chips, models, and cloud services on a single platform to obtain a complete technology stack, a model that Amazon and Microsoft have also sought to replicate.
Today, the company is expanding its equipment beyond its own ecosystem to reach more customers and build a larger business around AI chips—a market currently dominated by NVIDIA.
The day before its third-quarter earnings report, Alphabet said on Tuesday that Oracle will offer Google’s Tensor Processing Units (TPUs)—custom chips the internet giant has been developing since 2013. This marks the first time customers will be able to use TPUs outside of Google’s infrastructure, highlighting efforts by cloud business head Thomas Kurian to find growth elsewhere. Kurian previously worked at Oracle for more than 20 years.
Chris Ballard, Managing Director at Check Capital Management, commented on Google’s position in the chip industry: “They are a very strong competitor and could ultimately take market share away from NVIDIA.”
Kurian, who joined Google at the end of 2018, is currently in a relatively favorable position. Analysts expect Google Cloud's revenue to grow by approximately 90% year-over-year, following an 82% increase in the second quarter, with TPU contributing to the growth momentum.
Uday Cheruvu, a portfolio manager at Harding Loevner who holds Alphabet stock, said: “Google used to be great at doing its own thing, but it never truly served its customers well. That mindset has changed.”
Before announcing its partnership with Oracle this week, Google opened its TPUs to external data centers in the second quarter.
Estimates for TPU business vary widely, but all point to significant growth. TD Cowen forecasts TPU sales of $9.4 billion in the second half of this year and $62.5 billion in 2027. Wells Fargo expects third-quarter sales of approximately $9 billion and $99 billion next year.
Chief Financial Officer Anat Ashkenazi stated on the last earnings call in July that only a small portion of TPU revenue from existing agreements will be recognized this year, with the "vast majority" to be recognized in 2027. She also noted that, during the company’s expansion of its own infrastructure, Google plans to utilize more external capacity in the third quarter.
Google's cloud revenue for the second quarter was $25 billion, trailing behind Amazon Web Services' $42 billion and Microsoft Azure's $29 billion.
But Google is growing the fastest.
Ballard said, "It's almost as if the bigger the cake gets, the faster they cut into it." As of June 30, his company held approximately $236 million in Alphabet stock.
Optimize the entire technology stack
Although Kurian hopes to bring more TPUs to market through external partners, he must balance the benefits of expanding the customer base with the company’s own technological needs.
At a Goldman Sachs conference in September, Kurian stated that Google’s AI servers can recoup their costs in less than two years, while servers running Google’s custom chips can break even in roughly half that time.
Kurian said, "This is because we are able to optimize the entire technology stack."
Industry research firm SemiAnalysis estimates that the ownership and operational costs of Google's latest Ironwood TPU are approximately 44% lower than those of NVIDIA's GB200 servers.
In addition, selling TPU externally puts pressure on profit margins. TD Cowen estimates that the operating profit margin for TPU sales is around 20%, compared to an overall operating profit margin of approximately 35% for Google Cloud last quarter. Cheruvu stated that he believes it would be difficult for Google Cloud’s operating profit margin to significantly exceed the mid-30% range.
Google declined to comment.
One of Google's largest TPU customers is also one of its main competitors in the AI field.
Anthropic, which competes with Gemini using its Claude models, agreed last October to use up to one million TPUs in a deal worth tens of billions of dollars. Mizuho stated that this agreement will bring nearly one gigawatt of online capacity this year and described the relationship as Google “arming itself against its closest model competitor.”
CNBC confirmed that, following Tuesday's initial report by The Information, OpenAI also plans to purchase TPUs from Google.
Anthropic is securing external funding to purchase Google chips. This week, Broadcom, which collaborates with Google on TPUs, disclosed a $63 billion financing facility backed by Blackstone and three major banks. The facility will purchase TPUs and networking equipment from Broadcom and lease them to Anthropic for a five-year term. Analysts at William Blair stated that this financing covers approximately half of Anthropic’s projected demand through 2027, suggesting that additional funding rounds are likely to follow.
In its IPO filing submitted on Monday, Anthropic stated that the company uses chips from Google, Amazon, NVIDIA, and Advanced Micro Devices to "select chips based on the optimal cost-performance and efficiency characteristics for each workload."
Regarding Google and Broadcom, Anthropic stated that these partnerships "enable us to access compute capacity powered by Google TPUs across multiple generations of the TPU platform."