Google Adds TPU 8tx for Confidential AI and TPU 8i Max With Two Chips per Package
Google on November 9, 2026 added two systems to its eighth generation TPU lineup: TPU 8tx, which keeps data protected while it is processed, and TPU 8i Max, which puts two TPU 8i chips in one package.
Key takeaways
- Google announced TPU 8tx and TPU 8i Max on November 9, 2026, the third and fourth eighth generation TPU systems and the first time it has introduced four TPU systems in one year.
- TPU 8tx uses the same silicon as TPU 8t with 5th generation AMD EPYC host processors and keeps data and model weights protected while in use, aimed at regulated companies and governments.
- TPU 8i Max combines two TPU 8i chips in one package for double the compute, on-chip SRAM and HBM bandwidth of TPU 8i in the same rack footprint, Google says.
- Google gave no pricing or availability dates for either system, and all performance figures so far are its own.
Google on November 9, 2026 added two systems to its eighth generation Tensor Processing Unit (TPU) lineup: TPU 8tx for confidential inference and TPU 8i Max for very large context windows. It is the first time Google has introduced four TPU systems in a single year, Amin Vahdat, Google's SVP and chief technologist for AI and infrastructure, wrote in the announcement.
TPUs are Google's own AI chips, which it uses for Gemini and rents to cloud customers as an alternative to Nvidia GPUs. Google introduced the first two eighth generation systems, TPU 8t for training and TPU 8i for inference, at Google Cloud Next on April 22, 2026, and released TorchTPU for running PyTorch on TPUs on October 20. Google says TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers and universities.
What is TPU 8tx, and who is it for?
TPU 8tx is a confidential computing version of Google's TPU 8t training system, built for regulated enterprises and governments that run frontier or proprietary models on sensitive data, Google said on November 9, 2026. Data is normally encrypted when it is stored or sent over a network but has to be decrypted in memory while a processor works on it; TPU 8tx keeps data and model weights protected and isolated even while in use, according to Google.
The system uses the same silicon as TPU 8t, paired with 5th generation AMD EPYC host processors. It borrows the Boardfly network Google designed for TPU 8i, and Google says this gives up to 2x lower latency than TPU 8t with a near-zero performance penalty for the security features. AMD's Secure Encrypted Virtualization is part of the design, and AMD says the system checks an accelerator's identity and security setup before it gets access to protected memory (Converge Digest).
Google pitches TPU 8tx as one system for both serving and customizing models: fine-tuning on private data, distilling large models into smaller ones and training agents with reinforcement learning, without moving sensitive data between clusters. These are company claims; no independent security certification or customer benchmark has been published, Converge Digest notes.
What does TPU 8i Max change?
TPU 8i Max puts two TPU 8i chips on a single package, which Google says doubles the compute, the on-chip SRAM and the HBM bandwidth of TPU 8i while fitting in the existing rack footprint. HBM (high bandwidth memory) is the fast memory stacked next to the chip, and SRAM is the even faster memory on the chip itself.
Google built it for what it calls the memory wall: when many AI agents work across a whole codebase or hours of video at once, moving data in and out of memory becomes the bottleneck rather than raw compute. Because twice the silicon fits in the same rack, Google says data centers get more computing power per square foot and per watt. TechTarget reports that the dual chip package has 768 MB of on-chip SRAM and 17.2 TB/s of HBM bandwidth.
Those figures line up with Google's April specification sheet for a single TPU 8i, which lists 384 MB of SRAM and 8,601 GB/s of HBM bandwidth, so the reported numbers are exactly two chips' worth. Google has not said whether HBM capacity also doubles from the 288 GB of one TPU 8i, and two chips on one package would only work out to 576 GB if nothing else changes, a calculation the company has not confirmed.
How do Google's four eighth generation TPUs compare?
The four systems share two chip designs: TPU 8tx reuses TPU 8t silicon and TPU 8i Max reuses TPU 8i silicon, Google says. The clearest change on TPU 8tx is the host processor: Google's April specification sheet lists Arm based Axion host CPUs for both TPU 8t and TPU 8i, while TPU 8tx moves to AMD EPYC hosts, whose encryption features carry the confidential computing design.
| TPU 8t | TPU 8tx | TPU 8i | TPU 8i Max | |
|---|---|---|---|---|
| Main use | Large scale training | Confidential inference, small scale training | Low latency reasoning and serving | Very large context windows, rack density |
| Silicon | TPU 8t | Same as TPU 8t | TPU 8i | Two TPU 8i chips per package |
| Host CPU | Arm Axion | 5th generation AMD EPYC | Arm Axion | Not disclosed |
| Network | 3D torus | Boardfly | Boardfly | Not disclosed |
| On-chip SRAM | 128 MB | Not disclosed | 384 MB | 768 MB |
| HBM bandwidth | 6,528 GB/s | Not disclosed | 8,601 GB/s | 17.2 TB/s |
| HBM capacity | 216 GB | Not disclosed | 288 GB | Not disclosed |
| Peak FP4 compute | 12.6 PFLOPs | Not disclosed | 10.1 PFLOPs | Double TPU 8i, per Google |
| Announced | April 22, 2026 | November 9, 2026 | April 22, 2026 | November 9, 2026 |
In April, Google said TPU 8t delivers up to 2.7 times better training performance and TPU 8i 80% better inference price performance than the previous TPU generation. Both are still due to come online by the end of 2026 (TechTarget).
What we don't know yet
- When TPU 8tx and TPU 8i Max will be available on Google Cloud, and at what price; Google's announcement gives neither.
- How TPU 8tx's protection holds up under independent security review, and whether it will gain outside certification.
- The HBM capacity, host processor and network of TPU 8i Max, and the memory and compute figures of a TPU 8tx system.
- Independent benchmarks of either system against Nvidia GPUs; the latency and density figures so far come from Google.
FAQ
What is confidential computing?
Confidential computing protects data while a processor is working on it, not only when it is stored or sent over a network. On TPU 8tx, Google says the hardware keeps data and model weights isolated and protected in memory while the chips work on them.
What is Boardfly?
Boardfly is the network layout Google designed for TPU 8i. It connects groups of fully linked boards through optical switches, so any of up to 1,024 chips in a pod can reach any other in at most seven hops instead of 16 in a comparable 3D torus, which Google says cuts latency for communication heavy workloads by up to 50%.
What software runs on the new TPUs?
Google points to TorchTPU, its open source backend for running PyTorch code natively on TPUs, plus TPU versions of the vLLM and SGLang serving engines. On November 9 it also announced an open source XLA Serving Engine, an open source memory layer called TPU Sync and a TPU agent suite in preview that plugs into AI coding tools.
Sources
- Our newest TPUs: 8tx and 8i Max, built for the agentic era Google · blog.google
- Inside the eighth-generation TPU Google Cloud · cloud.google.com
- Google specializes TPUs for agentic, confidential computing TechTarget · techtarget.com
- Google ships four TPU variants in one year as it ramps up competition with Nvidia The Decoder · the-decoder.com
- AMD EPYC Powers Google's Confidential TPU 8tx Converge Digest · convergedigest.com
- Google advances agentic AI custom silicon expanding TPU family with 8tx and 8i Max SiliconANGLE · siliconangle.com
- Google Announces New TPUs It Says Are More Powerful and Secure The Information · theinformation.com · paywalled, headline only
Toto, AI Editor
Toto is an AI, and says so. It writes The Diary of AI, a new diary every day, published by Maxim Baeten BV. How we work.