Google's push to commercialize its TPU chips has encountered significant obstacles at the distribution stage. According to a recent report from The Information, during the same week Google announced plans to sell TPU chips directly to customers, executives from three major U.S. cloud service providers—Nebius, Lambda, and CoreWeave—publicly stated they have no immediate plans to adopt the technology.
During an earnings call, Google CEO Sundar Pichai indicated the company intends to sell TPUs to a "select group of customers," with a focus on the financial services and cutting-edge AI sectors. This statement itself signals a strategic shift for Google—away from pursuing widespread adoption akin to GPUs and toward targeting a limited number of clients willing to engage in deep collaboration.
However, the collective rejection from these three cloud providers highlights the practical challenges Google faces in building a TPU ecosystem.
Lambda CFO Chuck Fisher responded succinctly when asked about TPUs: "We bleed green," a clear reference to NVIDIA's brand color. Nebius Chief Revenue Officer Marc Boroditsky provided more direct data, noting that 99% of his company's customer demand is for NVIDIA GPUs. Occasional inquiries about TPUs, he mentioned, typically come from former Google employees familiar with the chips. "Unfortunately, the circle of ex-Google employees isn't large enough to constitute a market," Boroditsky remarked.
Nick Robbins, Vice President of Corporate Development at CoreWeave, explained the decision from an investment return perspective. He pointed out that the primary users of TPUs—Google itself, Anthropic, and Meta Platforms, Inc.—are also major purchasers of GPUs. This means that even if TPU market share grows, demand for GPUs will persist. "For every dollar we invest, for every megawatt we allocate, we are making a risk-adjusted bet on what will deliver the highest, longest-term return," Robbins stated. "When 99% of the market wants the same thing, even if that drops to 90%, it's difficult not to focus on that 90%."
The stance of these three companies is not surprising. Nebius, Lambda, and CoreWeave all have deep ties to NVIDIA, which serves as their primary supplier, a key investor, and, in some cases, a significant GPU customer.
Google has not been idle in its efforts to collaborate with mainstream cloud providers. Reports indicate that last year, Google explored closer partnerships with providers like CoreWeave and Crusoe, hoping to co-locate TPUs alongside GPUs in their data centers. However, these negotiations largely stalled.
Ultimately, Google pivoted to a less prominent partner—the software and data center startup Fluidstack—providing billions in lease and debt guarantees to support its TPU deployments for Anthropic. This shift, from attempting to enter established cloud provider networks to signing with emerging players like Fluidstack, indicates that the market for TPUs outside of Google Cloud remains highly concentrated.
On the financing front, Google is actively making moves. A February report from The Information revealed that Google has signed an agreement with an undisclosed major investment firm to establish a joint venture for leasing TPUs to other customers. Google is also in talks with potential financing partners about creating special purpose vehicles to borrow funds for purchasing TPUs and leasing them to clients.
Unlike NVIDIA's strategy of supporting early cloud providers, Google finds itself needing to take a more active role in financing—primarily because a ready-made market for TPUs does not yet exist.
Despite their clear short-term stance, Nebius's Boroditsky acknowledged that TPUs could gain more market share in the future. He noted that several companies are developing cloud software which, if successfully deployed, could make it easier for developers to work across different chip types, thereby reducing reliance on any single hardware vendor. "It's difficult for a single brand to maintain a monopoly—there will be more diversity in the future," he said. "This happened in the CPU space, and I expect the same story to play out again."
This perspective holds significance for the broader AI chip market. While NVIDIA currently dominates both AI training and inference chip sectors, the maturation of multi-chip compatible software could be a key variable in altering this landscape.
Comments