Google's tensor processing units are now running inside Meta's data centers, according to a report from Tom's Hardware published on October 22, 2026 — the first time Google has sold its custom AI chips to an external customer as deployed hardware rather than through its own cloud. The move marks a historic shift in the AI hardware market, where Nvidia's GPUs have long been the default choice for training and serving large models.
For years, the assumption in the industry has been that Nvidia's CUDA software ecosystem made its GPUs irreplaceable, even as alternatives emerged. That assumption is now being tested by the two companies best positioned to break it. The deal, coming on top of Meta's enormous AI infrastructure spending this quarter, shows how seriously hyperscalers are pursuing compute independence — a story you can follow in more depth with the latest AI developments covered daily on our homepage at aibuzzwire.news.
From TorchTPU to Steel and Silicon
The groundwork for this moment was laid publicly just two days earlier. On October 20, Google and Meta jointly open-sourced TorchTPU, a native PyTorch backend for Google's TPUs, announced on the official PyTorch blog by the TorchTPU team. The project began at Google in mid-2025 in collaboration with Meta's PyTorch team, and it does something no previous effort managed cleanly: it lets existing PyTorch workloads run on TPUs with a trivial code change, essentially swapping `.to("cuda")` for `.to("tpu")`.
That matters because Google's earlier attempt, PyTorch/XLA, launched in 2020, required intrusive source-code modifications, suffered from long compile times, and lacked support for PyTorch's native eager execution mode. Many teams that experimented with TPUs ended up returning to Nvidia GPUs simply because the developer experience was too painful. TorchTPU was explicitly designed to fix that, with no JAX dependencies in the runtime and torch.compile integration built on Torch Dynamo rather than the older Lazy Tensor approach.
What Meta Gets Out of the Deal
The technical results Google and Meta have publicized are notable. According to details reported by Yahoo Tech, the TorchTPU stack has been validated against 5,000 Hugging Face models, and Google's DeferAndFuse execution mode delivers performance gains of 50 percent or more over strict eager execution. Vendor case studies show Meta's Llama 3 405B model reaching 57 percent model FLOPs utilization on clusters of 128 to 1,024 Ironwood TPUs — figures that, while self-reported, indicate the chips can run frontier-scale training workloads competently.
The economics are the bigger story. TPUs have long been rumored to cost significantly less than top-end Nvidia GPUs, and a company with Meta's compute appetite — covering both its AI research ambitions and its recommendation systems — has every incentive to diversify away from a single supplier. Bloomberg reported as far back as November 2025 that Meta and Google were discussing deploying TPUs in Meta's data centers, and the collaboration on TorchTPU was reported by Reuters in December 2025. What has changed now is that the deal has moved from discussion to deployment, with Tom's Hardware reporting that the chips are physically installed and running in Meta facilities.
A Crack in Nvidia's Armor
Nvidia still commands the overwhelming majority of the AI accelerator market, and its data center revenue remains the benchmark every competitor chases. But the dynamics are shifting. As TechCrunch put it in its coverage of the TorchTPU launch, Google has joined Amazon in attacking Nvidia's stronghold on the software layer that keeps developers tied to its hardware. Meta, meanwhile, gains negotiating leverage: every TPU rack it installs is a rack of GPUs it did not have to buy at Nvidia's pricing.
The pressure is coming from multiple directions at once. Amazon has its Trainium chips, Google has TPUs, and a wave of startups are building inference-specific accelerators. What makes the Meta deployment different is scale and visibility. Meta is one of Nvidia's largest customers, and if its engineering teams conclude that TPUs can carry a meaningful share of both training and inference workloads, other large buyers will pay close attention.
The Caveats
None of this means Nvidia's position collapses overnight. The PyTorch blog itself cautions that while models may run out of the box on TPUs, achieving peak performance still requires hardware-aware optimization, because model shapes, sharding strategies, and kernel implementations are typically tuned for GPU architectures. The TPU instruction set differs from Nvidia's, and the ecosystem of tools, libraries, and experienced engineers remains heavily GPU-centric.
There is also the question of supply. Google has historically reserved TPUs for its internal workloads and its cloud customers, and manufacturing capacity for custom accelerators is not unlimited. Whether Google can or wants to supply TPUs at a scale meaningful to Meta's total compute footprint — which dwarfs most companies' — remains an open question.
What to Watch Next
The immediate signals to monitor are whether other hyperscalers follow Meta's lead, how Nvidia responds on pricing, and whether TorchTPU's open-source momentum translates into a genuine migration of PyTorch workloads onto TPU infrastructure. The AI hardware market has been waiting years for a credible alternative to Nvidia at the top end. With TPUs now running in a second company's data centers, that wait may finally be over.
Stay Ahead of AI
The AI hardware race is accelerating fast, and the winners and losers of the next decade are being decided now. Stay ahead of AI with independent coverage of the chips, models, and deals reshaping the industry.
Read more AI news →
