Google Adds TPU 8i Max and TPU 8tx to Its Eighth-Generation AI Chips: Giant Context and Confidential Computing

Google hasn’t designed a new chip, but it is rolling out two new machines from its existing silicon. This Monday, November 9, the company added TPU 8i Max and TPU 8tx to its eighth-generation AI accelerators in a blog post by Amin Vahdat, its chief technologist for AI infrastructure. The former targets massive context windows, while the latter focuses on sensitive data. Along with TPU 8t and TPU 8i, introduced in April, this brings the count to four TPU systems in a single year—a first, according to Google.

Two Configurations, Not Two New Chips

Google stresses the distinction: designing a chip takes years, whereas developing a system (how chips, memory, networking, and security are assembled) can be pulled off in a matter of months. Both newcomers therefore reuse existing silicon.

TPU 8i Max takes on the “memory wall.” When dozens of agents collaborate across an entire codebase or hours of video, memory bandwidth—rather than raw compute—becomes the bottleneck. Google addresses this by packing two TPU 8i chips into a single package, delivering twice the compute, SRAM, and HBM bandwidth of TPU 8i within the same rack footprint, resulting in more power per square meter and per watt. According to Network World, which interviewed Google Cloud’s Mark Lohmeyer, this configuration offers 384 GB of HBM, 17.2 TB/s of bandwidth, and 768 MB of SRAM; these figures do not appear in the official blog post.

TPU 8tx repurposes the chip from TPU 8t, designed for training, paired with 5th Gen AMD EPYC host processors. Its defining feature: confidential computing. Typically, data is encrypted at rest and in transit, but decrypted in memory during processing. TPU 8tx keeps data protected and isolated even during computation, including model weights. Google is pitching it to regulated industries and governments looking to run models and agents on their most sensitive data, and to perform fine-tuning, distillation, or reinforcement learning without moving that data. Leveraging TPU 8i’s low-latency network, Google claims up to 2x lower latency compared to TPU 8t.

Software: The Other Half of the Announcement

Alongside the hardware, Google rolled out a suite of software tools aimed at making its TPUs as frictionless to use as GPUs. It is open-sourcing TPU Sync, a layer that tiers model state (specifically the KV cache for long contexts) across chip memory, host memory, and storage, which Google says cuts time-to-first-token by over 90%. An open-source TPU agent suite, currently in preview, integrates with Claude Code, Cursor, and Google Antigravity to help migrate and optimize code. These are joined by the open-source XLA Serving Engine, the Agent Compute Engine software—which suspends idle agent sessions to free up chip resources—and the general availability of Cloud Pathways and dynamic slicing for TPU clusters on GKE, its managed Kubernetes service.

What’s Missing

Google provides no pricing, regions, or availability dates for either system. The cited performance improvements are internal measurements, with no published independent benchmarks. The company claims nearly 100 customers for its TPUs, including frontier AI labs, and promises to share more details during a livestream this Monday at 6:00 PM Paris time.