Just in
AIGPT-6.2 Astra and Sol Arrive as Astra API Prices Fall 20%CryptoBullish Earnings: Q3 Adjusted Revenue Up 10%, $300 Million BuybackTechNvidia Earnings: Q3 Revenue Up 103% to $115.8 BillionTechApple Outage: iCloud Sign In Issue Continues as App Store ClearsTechARTEX AI Agents Used in Key Theft and Deepfake Attacks, Amazon SaysTechNvidia National Security Board Committee Will Oversee Export Rules
Tech

Google TPU 8tx Targets Confidential AI as 8i Max Doubles Up

Google added two new eighth generation TPU systems: the TPU 8tx for secure, confidential inference and the TPU 8i Max, which packs two inference chips into one package.

USA Publishing Group4 min read
Share
Silicon wafer covered in a grid of iridescent chip dies, the kind of silicon behind Google's new TPU 8tx
Photo: Laura Ockel / Unsplash

Key takeaways

  • Google says the TPU 8tx keeps data and model weights protected and isolated even while the chip is actively processing them, and it pairs the chip with 5th generation AMD EPYC hosts.
  • The TPU 8i Max joins two TPU 8i chips on one package, which Google says doubles compute, SRAM and HBM bandwidth in the same rack footprint.
  • Google announced no pricing, availability dates or launch customers for either system, though Data Center Knowledge reports select customers can run the 8tx in their own data centers.
In this story 5 sections
  1. Two spinoffs of the chips Google showed in April
  2. TPU 8tx vs TPU 8i Max, feature by feature
  3. Pricing and launch dates Google left out
  4. Why confidential computing could be Google's opening
  5. Which one fits which buyer

Google now sells four versions of its eighth generation TPU, and the two it added on November 9 are aimed at very different buyers. The TPU 8tx is built for organizations that cannot let their data or model weights sit exposed while a chip works on them. The TPU 8i Max is built for the opposite pressure: models that need huge context windows and as much compute as can fit in a rack.

Google laid out both systems in a Google Cloud blog post by Amin Vahdat, its chief technologist for AI and infrastructure. The company says this is the first year it has introduced four TPU systems. Here is how the two new ones compare, and who each one is really for.

Two spinoffs of the chips Google showed in April

Both are built on chips Google revealed in April 2026 at its Next conference. The TPU 8t handles training, and Google says it delivers up to 2.7 times better training performance than the previous generation. The TPU 8i handles inference, with what Google describes as 80% better price performance.

The new pair does not start from scratch. The 8tx uses the same silicon as the 8t, and the 8i Max is two 8i chips joined on one package. What changes is the packaging, the host hardware and the job each system is tuned for.

TPU 8tx vs TPU 8i Max, feature by feature

TPU 8tx: confidential inference first

Google calls the 8tx "optimized for secure and confidential inference," and says it also suits small scale training. The central claim is that security is built into the hardware layer, so data and model weights stay protected and isolated even while the chip is actively processing them.

That matters because encryption usually covers data at rest and in transit. The weak point has been the moment data is decrypted for computation. Data Center Knowledge describes the 8tx as adding hardware level encryption to close that gap. Google did not name the specific technology behind it in its post.

Other details Google and the trade press shared:

  • Hosts run on 5th generation AMD EPYC processors, according to Google.
  • The chip uses Google's Boardfly network, which Google says cuts latency by up to half compared with the TPU 8t.
  • According to Data Center Knowledge, Google will let select customers deploy the 8tx inside their own data centers.

TPU 8i Max: density and memory

The 8i Max goes after a different limit. Long context windows eat memory and memory bandwidth fast. By putting two TPU 8i chips on one package, Google says it doubles compute, on chip SRAM and HBM bandwidth compared with a single 8i, while fitting twice the silicon into the same rack footprint.

Mark Lohmeyer, who runs AI and computing infrastructure at Google Cloud, told Data Center Knowledge the design works by "adding more lanes to the highway." The same report puts the 8i Max at 384 GB of HBM, a figure Google's own post does not list.

Pricing and launch dates Google left out

Neither chip came with a price, a general availability date or a list of regions in the materials published Monday. Google also did not name launch customers for either system. It did say its TPUs now serve nearly 100 customers, including frontier AI labs, trading firms, retailers and universities.

That gap is worth keeping in mind. The performance claims are Google's own, and buyers will want independent benchmarks once the hardware is in customers' hands.

Why confidential computing could be Google's opening

Nvidia still dominates AI chips, though it is busy this week with a board investigation involving Jensen Huang. Google is not trying to beat Nvidia on every front. It is picking specific needs where a custom chip can stand out.

Security is one of them. Banks, hospitals and government agencies have been slow to move sensitive AI work to the cloud. Dave McCarthy, a group vice president at IDC, told Data Center Knowledge that "if Google can come out with a strong confidential compute story, it kind of gives them a little bit of an edge."

The wider backdrop is the enormous spending on AI infrastructure, from Google's own Wyoming data center plan to Oracle's $20 billion fundraising. Investors have grown nervous about that spending, as the recent AI stock selloff showed, so chips that bring in new paying customers matter more than raw speed records.

Which one fits which buyer

  • Regulated enterprises and governments that want AI on sensitive records: the TPU 8tx is the one to watch, especially if the on premises option becomes widely available.
  • AI labs and companies running long context models: the TPU 8i Max is the better match, since memory and rack density are its whole point.
  • Teams training large models: neither new chip replaces the TPU 8t, which remains Google's training workhorse.
  • Everyone else renting inference: the standard TPU 8i, with its price performance focus, is still the sensible default.

Google has not said when either system ships, so the next real test is whether named customers show up. For more on the companies building the AI economy, see our tech section.