Three of the biggest names in GPU cloud infrastructure have turned down Google’s offer to host its Tensor Processing Units. According to The Information, Nebius, Lambda, and CoreWeave each said no when Google came knocking, a striking rejection that complicates Google’s plan to push TPUs beyond its own data centers and into the broader AI compute market.
This is significant because Google has spent the past year trying to position TPUs as a serious alternative to Nvidia’s GPUs. Getting independent neoclouds to rack and rent them would have given Google instant distribution and credibility. Instead, the three companies most aggressively building out AI capacity all passed.
Who said no, and why it matters
Nebius, Lambda, and CoreWeave aren’t fringe players. They’re the neoclouds underwriting much of the current AI buildout, signing multi-billion-dollar deals with OpenAI, Microsoft, Meta, and others to host Nvidia GPUs at scale.
What stands out here is the unanimity. When three competitors independently reach the same conclusion, it tells you something about the economics:
- Customer demand is Nvidia-shaped. AI labs build on CUDA. Asking a neocloud to invest capex in TPUs means betting customers will rewrite their stacks. Few will.
- Margin math is brutal. Neoclouds make money on utilization. An idle TPU rack is worse than no rack at all.
- Strategic conflict. Google Cloud competes with these companies for the same workloads. Hosting Google silicon while Google poaches your customers is an awkward arrangement.
What Google was trying to do
Google’s TPU strategy has shifted noticeably this year. The chips were once an internal advantage, powering Gemini training and Google Search inference. Now Google wants TPUs in the hands of external customers, including through third-party clouds. Anthropic is the headline external user, with a reported commitment to use up to a million TPUs.
Getting Nebius, Lambda, or CoreWeave on board would have signaled that the rest of the market is ready to follow. The rejection signals the opposite: even the most growth-hungry compute providers don’t see TPU demand worth the bet, at least not yet.
The status quo this reinforces
Nvidia’s grip on AI compute just got reinforced. The neocloud business model, which barely existed three years ago, is now a roughly $50 billion category built almost entirely on Nvidia GPUs. Each new Nvidia generation gets pre-sold by these providers before it ships.
Google’s TPUs, by contrast, remain a captive ecosystem. Powerful for Google and its closest partners. Hard to access for everyone else. The neocloud rejection means that gap won’t close quickly through distribution.
What practitioners should expect
If you’re building AI products and considering compute strategy, a few things follow from this:
- TPU access stays narrow. If you want TPUs, you’ll get them through Google Cloud directly or not at all. No cheaper third-party rentals coming soon.
- Nvidia pricing power holds. Without credible alternative silicon at scale, GPU pricing pressure stays muted. Budget accordingly.
- Watch Anthropic’s results. The biggest external TPU bet in the market belongs to Anthropic. If Claude’s training and inference economics improve visibly, that could shift neocloud calculus in 2026.
- AMD becomes the more interesting wildcard. MI300 and MI350 deployments at neoclouds are growing. AMD has a software story that’s improving and a business model that doesn’t compete with its customers.
The bigger picture
Google has the best non-Nvidia AI chip in the world. The problem isn’t the silicon. It’s the surrounding ecosystem, the customer relationships, and the structural conflict of being both a chip vendor and a cloud competitor. The Information’s report makes clear that even cash-rich infrastructure companies racing to deploy capacity won’t take TPUs on faith.
Google’s next move will tell us whether this is a temporary setback or a durable ceiling on TPU adoption outside Mountain View. More details on the negotiations and the holdouts’ reasoning are at the original source.