Google Unveils TPU 8 Chips to Make AI Faster and Cheaper

Google today announced two new custom AI chips at its Cloud Next 2026 conference in Las Vegas — the TPU 8t for training and TPU 8i for inference — marking the first time the company has split these workloads into separate, purpose-built processors. The move is a direct challenge to Nvidia’s dominance in AI hardware.

Why Two Chips Instead of One

Every AI tool you use involves two distinct processes. Training is how companies build models — feeding them massive amounts of data so they learn patterns. Inference is what happens when you actually use the model, like asking Gemini a question or generating an image. These two jobs have very different hardware needs, and Google is now designing a dedicated chip for each.

The TPU 8t delivers 2.7 times the performance-per-dollar of Google’s previous-generation Ironwood chip for training. A single TPU 8t superpod scales to 9,600 chips with two petabytes of shared memory, and double the interchip bandwidth of the prior generation.

The TPU 8i targets inference with an 80% improvement in performance-per-dollar. It packs three times more on-chip memory than Ironwood, allowing it to hold larger context windows in fast-access storage — which directly translates to quicker responses from AI models.

Both chips deliver twice the performance-per-watt of their predecessors, addressing growing concerns about AI’s energy footprint.

A Nvidia Partnership Too

In a notable twist, Google also announced a partnership with Nvidia to bring Nvidia’s upcoming Vera Rubin chips to Google Cloud. The company is betting that customers want options — Google’s own silicon for cost efficiency, and Nvidia hardware for workloads already optimised for its ecosystem.

What This Means for You

You won’t plug a TPU into your laptop, but these chips power the AI tools you use every day. When inference gets 80% cheaper to run, that savings flows downstream — expect faster responses from Google’s AI products, more generous usage limits, and competitive pressure on other providers to match. If you use Gemini or any Google Cloud-powered AI service, the TPU 8i should make those experiences noticeably snappier once it rolls out later this year.

For more on how AI infrastructure is evolving, check out our guide to getting started with AI tools or explore the latest Claude and Anthropic updates.