Skip to content
Theme
OnDevice Tools

SoftwareCloud & Services

Google TPU sales open up as Nvidia loses Jensen Huang

Google is selling its TPUs to third parties and bringing them to Oracle's cloud in 2027, a move that could win customers after Jensen Huang's resignation from Nvidia.

Illustration of a Google TPU chip beside an Nvidia GPU, with a cloud network in the background

Key takeaways

  • Jensen Huang has resigned as Nvidia CEO, with Colette Kress serving as interim CEO, according to 9to5Google.
  • Google is opening TPU sales to third parties and will bring TPU capacity to Oracle's cloud in 2027.
  • Google's eighth-generation TPUs include the 8t for training, 8i for inference, 8tx for confidential computing and 8i Max for large context windows.
  • Google open-sourced TorchTPU with Meta's PyTorch team to make it easier to run PyTorch code on TPUs.
In this article
  1. What happened
  2. Why Nvidia's change matters for Google
  3. The eighth-generation TPU lineup
  4. Software is the real battleground
  5. What it means for Google's cloud business

What happened

Jensen Huang has resigned as CEO of Nvidia, according to 9to5Google. The newsletter report describes the departure as humiliating and says Colette Kress is now interim CEO. The same report argues that Google could be one of the biggest long-term winners from the shake-up.

Google has spent years offering its Tensor Processing Units (TPUs), custom chips for AI training and inference, mainly inside its own Google Cloud. The company is now opening sales to third parties and has said it will bring TPU capacity to Oracle's cloud starting in 2027, according to the same report.

Why Nvidia's change matters for Google

Illustration: NVIDIA logo, Google logo, code, news and the labels Google Tpu, Jensen Huang, Oracle Cloud

Nvidia's lead in AI chips has rested partly on CUDA, its software platform that developers use to write code for its GPUs. That lock-in has made it hard for rivals to win customers, even when their hardware is competitive. With Huang gone, the report suggests customers may be more willing to consider alternatives.

Google is not claiming its TPUs beat Nvidia's top hardware on raw speed. The report says even the most powerful TPU, the 8i Max, does not match Nvidia's Vera Rubin for sheer performance. Instead, Google is competing on power efficiency in some configurations and on making it easier to move existing code onto its chips.

The eighth-generation TPU lineup

Google's eighth-generation TPUs come in four versions, each aimed at a different part of AI work. The TPU 8t is built for large-scale training. The TPU 8i targets low-latency reasoning, which is the inference stage where a model answers requests. A third chip, the TPU 8tx, is designed for confidential computing, which keeps data private while it is processed. The TPU 8i Max is made for very large context windows and packs the most compute per rack.

Google says the four systems together give customers a purpose-built accelerator for each stage of what it calls the agentic lifecycle. The company is not comparing the chips directly with Nvidia's offerings.

Software is the real battleground

In October, Google open-sourced TorchTPU, a project it developed with Meta's PyTorch team. The tool lets developers run familiar PyTorch workflows on TPUs with few code changes. It will not make every custom CUDA operation portable, and performance tuning will still be needed.

Still, the report argues that lowering the effort to try an alternative can change buying decisions. Customers do not have to abandon Nvidia completely for Google to gain TPU market share or for Nvidia to face more price pressure. Google is taking on Nvidia alongside Broadcom and MediaTek.

What it means for Google's cloud business

Google's cloud division has become more important as the company's search business faces pressure from AI. The report says specialised, more efficient TPUs could help offset a slowdown in Search as Google shifts toward Cloud. In the short term, Google will likely see a boost in sentiment from Nvidia's turmoil.

For readers, the practical effect is that AI companies may soon have a more credible second option for training and running models. That could affect cloud pricing and the range of hardware available. The Oracle cloud expansion is set for 2027, so the competitive impact will take time to appear.

Sources

  1. Jensen’s exit could be Google’s TPU moment9to5Google

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.