Google’s TPU push hits wall as neoclouds stay loyal to Nvidia

Google made a notable strategic shift recently, announcing plans to sell its tensor processing units directly to select customers for deployment in their own data centers. On paper, the move looked like a serious attempt to challenge Nvidia’s stranglehold on AI compute. In practice, however, the companies Google most needs to convince are not particularly interested.

Nebius, Lambda, and CoreWeave, three of the most prominent neocloud providers currently building out AI infrastructure at scale, have each indicated they have no immediate plans to adopt Google TPUs. Their reasoning, as shared with The Information, is straightforward and consistent across all three companies. Nvidia GPUs still account for the overwhelming majority of what their customers actually request, and shifting capital toward an alternative that serves a fraction of demand simply does not make financial sense right now.

Nebius chief revenue officer Marc Boroditsky put the situation plainly, noting that roughly 99 percent of demand his company sees comes in for Nvidia hardware. He added that TPU requests tend to arrive from people with prior Google experience rather than from the broader market, which limits how far that interest actually stretches.

CoreWeave VP of corporate development Nick Robbins framed his company’s position around risk-adjusted returns, pointing out that when 99 percent of the market leans toward one product, even a shift down to 90 percent still leaves the same answer standing. Lambda CFO Chuck Fisher kept it even shorter, saying the company simply “bleeds green,” a reference to Nvidia’s branding that needed no further elaboration.

The financial commitments already sitting on neocloud balance sheets reinforce this position further. CoreWeave alone spent nearly $15 billion in its full fiscal year of 2025, with the bulk of that going toward Nvidia hardware. Beyond current deployments, most neoclouds have additionally signed agreements to deploy Nvidia’s upcoming Vera Rubin chips, meaning their roadmaps already point in a direction that excludes TPUs.

Nvidia itself holds investor stakes in several of these companies, including CoreWeave, Lambda, and Nebius, which deepens the alignment considerably.

Google did find one taker, signing an agreement with Fluidstack to deploy TPUs in a New York data center, with Anthropic as the end user. However, that arrangement does not represent the broad neocloud adoption Google was reportedly seeking. For now, Nvidia’s grip on this segment of the market remains firmly intact.