Google Sells Custom AI Chips

Google's cloud revenue is projected to surge 90% year over year, a jump that analyst Chris Ballard of Check Capital Management attributes to the growing strength of the company's custom silicon. This financial momentum drives a strategic pivot announced Tuesday, just before Alphabet's third-quarter earnings report. The company will allow Oracle to host its tensor processing units, or TPUs, for external customers. It is the first time Google has made these chips, developed since 2013, available outside its own data centers.

The move signals a departure from the full-stack approach Google has long defended. Previously, the company argued that buying chips, models, and cloud services from a single provider offered the best performance. Cloud chief Thomas Kurian, who joined Google from Oracle in 2018, is now pursuing growth through third-party partners. This shift follows an earlier step in the second quarter when Google opened TPUs to outside data centers. Uday Cheruvu, a portfolio manager at Harding Loevner, noted that while Google excelled at internal development, it previously struggled to serve external customers, a mindset he says has changed.

Financial forecasts for the TPU business vary but consistently indicate rapid expansion. TD Cowen models $9.4 billion in TPU sales for the second half of 2026 and $62.5 billion for 2027. Wells Fargo projects approximately $9 billion in sales for the current quarter and $99 billion next year. Chief Financial Officer Anat Ashkenazi stated in July that only a small portion of TPU revenue from existing agreements will be recognized in 2026, with the vast majority arriving in 2027.

Selling chips externally introduces margin pressures. TD Cowen estimates a 20% operating margin on TPU sales, compared to Google's overall cloud margin of about 35% last quarter. Cheruvu expects Google's cloud operating margin to remain in the mid-30s. Despite this, Google remains smaller than its rivals in total cloud revenue. Its second-quarter figure of $25 billion trails Amazon Web Services at $42 billion and Microsoft Azure at $29 billion, though Google is growing faster.

Kurian must balance external sales against internal efficiency. He noted at a September Goldman Sachs conference that Google's AI servers pay for themselves in under two years, while those running its own silicon reach payback in about half that time. SemiAnalysis estimates the latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server.

Anthropic, a direct competitor in the model space, is Google's largest TPU customer. The company agreed in October to use up to 1 million TPUs in a deal worth tens of billions of dollars. CNBC has also confirmed that OpenAI plans to purchase TPUs from Google. Anthropic's IPO filing on Monday stated it uses silicon from Google, Amazon, Nvidia, and Advanced Micro Devices to optimize price-performance for specific workloads.

Financing for these massive hardware purchases is becoming a distinct market. Broadcom detailed a $63 billion vehicle backed by Blackstone and three large banks to buy TPUs and networking gear for lease to Anthropic. William Blair analysts say this plan covers about half of Anthropic's projected 2027 needs, indicating further financing rounds are likely.