Oracle to Sell Google's AI Chips, Breaking Google's Closed-Stack Model

For years, Google sold its artificial intelligence hardware only to customers who also bought its cloud services, a bundled strategy that kept its custom chips inside its own data centers. That approach changes this week. Alphabet announced on Tuesday that Oracle will begin selling Google's tensor processing units, or TPUs, to its own customers.

This marks the first time buyers can access TPUs without renting space in Google's infrastructure. The move signals a shift in strategy for Thomas Kurian, Google Cloud's chief executive, who spent more than 20 years at Oracle before joining Google in late 2018. By opening its hardware to external partners, Google aims to expand its reach in a market long dominated by Nvidia.

The decision comes on the eve of Google's third-quarter earnings report. Analysts expect Google Cloud revenue to surge approximately 90% compared to the same period last year, following an 82% increase in the second quarter. Uday Cheruvu, a portfolio manager at Harding Loevner, noted that Google has historically struggled to serve customers outside its own ecosystem. He stated that the company's mindset regarding external sales has changed.

Google began allowing outside data centers to host TPUs in the second quarter of this year. The Oracle partnership extends that access to a major cloud provider. Estimates for the TPU business vary by firm. TD Cowen projects $9.4 billion in TPU sales for the second half of this year and $62.5 billion for 2027. Wells Fargo anticipates approximately $9 billion in third-quarter sales and $99 billion in revenue for next year.

CFO Anat Ashkenazi told investors in July that only a small portion of TPU revenue from current agreements will be recognized in this fiscal year. She said the vast majority of that revenue will arrive in 2027. She also noted that Google planned to utilize more outside capacity in the third quarter while it continues to build its own infrastructure.

Despite trailing Amazon Web Services, which posted $42 billion in second-quarter revenue, and Microsoft Azure, which earned $29 billion, Google is growing faster than its rivals. Chris Ballard, managing director of Check Capital Management, observed that Google accelerates as the market expands. His firm held about $236 million in Alphabet stock as of June 30.

Kurian must balance external sales with internal needs. At a Goldman Sachs conference in September, he said Google's AI servers recoup their costs in under two years. Servers running Google's own silicon achieve payback in roughly half that time. He attributed the speed to the company's ability to optimize the entire software and hardware stack.

Industry research from SemiAnalysis suggests Google's latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server. However, external sales carry lower margins. TD Cowen estimates a 20% operating margin on TPU sales, compared to Google's overall cloud margin of about 35% in the previous quarter. Cheruvu does not expect Google's cloud operating margin to widen much beyond the mid-30s range.

Anthropic, a major AI competitor to Google, is the company's largest TPU customer. In October of last year, Anthropic agreed to use up to 1 million TPUs in a deal valued at tens of billions of dollars. Mizuho analysts said that agreement brings nearly a gigawatt of capacity online this year. They described the relationship as Google arming its closest model competitor. OpenAI is also planning to purchase TPUs from Google, a fact CNBC confirmed after The Information reported the intentions on Tuesday.

Anthropic is using external financing to fund its Google silicon. This week, Broadcom detailed a $63 billion vehicle backed by Blackstone and three major banks. The entity will buy TPUs and networking gear from Broadcom and lease them to Anthropic for five years. William Blair analysts said the financing covers about half of Anthropic's 2027 needs. In its IPO filing on Monday, Anthropic stated it uses chips from Google, Amazon, Nvidia, and Advanced Micro Devices to select the best options for each workload.