Google Starts Selling Its AI Chips Through Outside Partners, Taking Aim at Nvidia

Google has spent years selling a simple pitch: buy our chips, our models and our cloud all in one place, and you get a stack no competitor can match. That pitch is changing.

Alphabet announced Tuesday, a day before its third-quarter earnings report, that Oracle will offer Google's tensor processing units, or TPUs, the custom AI chips Google has been building since 2013. It's the first time outside customers will be able to access TPUs without going through Google's own infrastructure, marking a real shift in how the company competes in the AI chip market that Nvidia currently dominates.

The move is a notable pivot for a company that has long guarded its silicon as a reason to pick Google Cloud over Amazon Web Services or Microsoft Azure. Now it's pushing that same hardware out to rivals' turf through partners, betting that reaching more customers matters more than keeping the technology exclusive.

Cloud chief Thomas Kurian, who spent more than two decades at Oracle before joining Google in late 2018, is the one steering that bet. "They're a very strong competitor, and it's very possible that they could eventually take market share from Nvidia," said Chris Ballard, managing director of Check Capital Management, speaking about Google's chip position, according to CNBC.

Google had already opened TPUs to outside data centers in the second quarter, so the Oracle deal extends a strategy already in motion rather than starting one from scratch.

A business growing fast, even if it's still smaller

Kurian isn't making this move out of weakness. Analysts expect Google Cloud revenue to climb roughly 90% year over year in the third quarter, building on 82% growth in the second quarter, with TPUs cited as a key driver of that momentum.

"Google was great at doing its own thing, but it was never able to serve customers," said Uday Cheruvu, a portfolio manager at Harding Loevner, which holds Alphabet shares. "That mindset has changed."

Estimates for the TPU business itself vary, but all point in the same direction. TD Cowen projects $9.4 billion in TPU sales for the second half of this year, rising to $62.5 billion in 2027. Wells Fargo is even more bullish, forecasting about $9 billion in third-quarter sales alone and $99 billion next year.

Alphabet CFO Anat Ashkenazi cautioned on the company's July earnings call that only a small slice of TPU revenue from current agreements will show up this year, with the "vast majority" landing in 2027. She also said Google would lean on more outside capacity in the third quarter while it continues building its own infrastructure.

Even with the growth, Google Cloud is still the smallest of the major hyperscalers in raw terms. It posted $25 billion in second-quarter cloud revenue, behind AWS at $42 billion and Azure at $29 billion. But it's expanding faster than either. "It's almost like the bigger the pie is getting, the faster they're accelerating into it," Ballard said. His firm held about $236 million in Alphabet stock as of June 30, according to CNBC.

The tradeoff Kurian has to manage

Opening TPUs to the outside world isn't free. Kurian has to balance growing the customer base against keeping enough chip supply for Google's own AI workloads, and the company's internal numbers suggest why it wants to protect some of that advantage.

Speaking at a Goldman Sachs conference in September, Kurian said Google's AI servers pay for themselves in under two years, while servers running the company's own silicon reach that payback point in roughly half the time. "It's because we can optimize that whole stack," he said.

That cost edge shows up in outside research too. SemiAnalysis estimates Google's latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server, a gap that helps explain why Google is eager to put more of these chips into the market even as it keeps a chunk of the advantage for itself.

There's also a margin question. Selling TPUs externally comes with thinner profits than running its own cloud: TD Cowen estimates a 20% operating margin on TPU sales, well below the roughly 35% overall cloud margin Google posted last quarter. Cheruvu said he doesn't expect Google's cloud operating margin to climb much beyond the mid-30s given that mix. Google declined to comment for the CNBC report.

Its biggest TPU customer is also a Gemini rival

The clearest sign of how far Google is willing to go is its relationship with Anthropic, maker of the Claude models that compete directly with Google's own Gemini. The two companies struck a deal last October giving Anthropic access to up to 1 million TPUs, worth tens of billions of dollars. Mizuho estimates that agreement brings nearly a gigawatt of computing capacity online this year, and describes the arrangement as Google effectively arming its closest model competitor.

OpenAI is reportedly moving in the same direction. The ChatGPT maker is planning to buy TPUs from Google, a detail CNBC confirmed after The Information first reported it on Tuesday.

Anthropic is also getting outside financial help to pay for all that Google hardware. This week, Broadcom, which partners with Google on TPU development, outlined a $63 billion financing vehicle backed by Blackstone and three major banks. That vehicle will buy TPUs and networking equipment from Broadcom and lease them to Anthropic over five years. Analysts at William Blair estimate the financing covers about half of Anthropic's projected 2027 compute needs, suggesting more deals of this kind are likely on the way.

Anthropic's own IPO filing, released Monday, confirmed it sources chips from four suppliers: Google, Amazon, Nvidia and Advanced Micro Devices. The company said this lets it "select chips based on the best price-performance and efficiency characteristics for each workload." On the Google and Broadcom arrangements specifically, Anthropic said the partnerships "provide us with access to compute capacity powered by Google TPUs across multiple generations of the TPU platform."

Taken together, the Oracle deal, the OpenAI talks and the Anthropic financing all point to the same shift: Google is no longer trying to keep its chip advantage locked inside its own walls. It's betting that selling TPUs more broadly, even to competitors, builds a bigger business than hoarding the technology ever could.