Nebius, Lambda and CoreWeave Say ‘No’ to TPUs Amid Google’s Push
We’ve been tracking Google’s effort to expand the reach of its tensor processing units for quite a while, and two developments last week helped us understand where things are actually headed.
On one hand, Google said on its earnings call that it plans to sell its TPUs directly to customers for use in their own data centers—something we first reported last fall. The move marks a significant shift: For years, TPUs were largely confined to Google Cloud.
On the other hand, though, three senior executives from leading neocloud providers— Nebius, Lambda and CoreWeave—told us they aren’t planning to adopt TPUs anytime soon.
Taken together, the signals point to a complicated reality for Google. Even as it opens up the market for its own chips, the very companies that could be best positioned to distribute that hardware aren’t convinced it’s worth their investment.
“We bleed green at Lambda,” said Lambda Chief Financial Officer Chuck Fisher after I asked him about TPUs onstage at The Information’s Financing the AI Revolution event last week.
Marc Boroditsky, chief revenue officer of Nebius, told me 99% of the demand his firm sees is for Nvidia graphics processing units. In the rare cases that a prospective customer inquires about TPUs at Nebius, he said, they are often Google alumni who’ve used the chips before.
“Unfortunately, the diaspora [of] folks from Google does not make a market,” Boroditsky said.
At CoreWeave, the calculus is similar. Vice President of Corporate Development Nick Robbins pointed out the top TPU users—Google itself, Anthropic and Meta Platforms—are also among the largest buyers of GPUs. That makes him quite confident that renting Nvidia GPUs will always be a good business.
“For any incremental dollar we invest or megawatt we allocate, we are making a risk-adjusted bet on what is going to deliver the highest return for the longest,” Robbins said. “When 99% of the market wants one thing and you say, ‘OK, that could change, it could go down to 90%,’ it's still really hard to not focus on the 99% that goes to 90%.”
None of the executives’ answers was terribly shocking. Lambda, Nebius and CoreWeave are deeply tied to Nvidia, which is their No. 1 supplier, a key investor and in some cases a major GPU customer.
But their reluctance helps explain how Google’s TPU strategy is playing out and why it looks different from Nvidia’s.
Last year, Google explored working more closely with neocloud providers, including CoreWeave and Crusoe, to host TPUs alongside GPUs in their data centers. Those conversations largely went nowhere. Instead, Google struck a deal with software and data center startup Fluidstack to help deliver TPUs to Anthropic—and backed that effort by guaranteeing billions of dollars in leases and debt tied to those deployments.
Google went from trying to get its TPUs into multiple neoclouds to signing a deal with Fluidstack, a newcomer in the neocloud space, for Anthropic deployments. It shows that the market for TPUs outside Google Cloud is relatively concentrated at this point.
As CEO Sundar Pichai put it, Google plans to deliver TPUs to a “select group of customers,” particularly in areas like financial services and frontier AI. That’s instead of, say, trying to make them as ubiquitous as GPUs are.
So it makes sense that rather than trying to seed a broad third-party ecosystem right away, it is focusing on a smaller set of customers that are willing to go deep. This will reduce Google’s reliance on the neoclouds for distribution, at least in the near term.
Google has been busy working on how to make the financing happen. As we reported in February, it has signed an agreement with an unidentified large investment firm to fund a joint venture that would lease TPUs to other customers. Google has also been talking to potential finance partners to borrow money for special purpose vehicles that would buy TPUs and lease them to customers. (While Nvidia did help finance some of the early neoclouds, Google’s strategy is different because there’s not an existing market for TPUs, so it is playing a more active role off the bat.)
Longer term, even Boroditsky from Nebius said TPUs could gain a larger foothold. He knows several companies developing cloud software that—if it works—would make it easier for developers to use multiple types of chips, limiting hardware lock-in.
“It’s going to be difficult to entrench a single brand—there’s going to be a lot more diversity,” he said at the end of the panel. “It happened in CPUs, and I suspect the same movie is going to play out again.”