Google announced two new server configurations on November 9 for its eighth-generation Tensor Processing Units, the custom chips Google builds to run artificial intelligence. Both configurations reuse chip designs the company introduced in April. One adds hardware-level isolation for private data, and the other puts two inference chips in one package that doubles compute, on-chip memory, and memory bandwidth within the same rack footprint, according to Google.
Reusing April chips for privacy and larger workspaces
Google built each new system around a chip it introduced in April. TPU 8tx uses the same silicon as the TPU 8t training chip, and TPU 8i Max combines two TPU 8i inference chips. The new configurations change the host processors, networking, and packaging around that silicon.
Google says the TPU 8tx uses hardware-level isolation to keep customer data and model weights protected while they are in active use. Standard security scrambles data while it sits on a hard drive or travels across a network, then unlocks it so the processor can read it. The 8tx pairs the TPU 8t silicon with fifth-generation AMD EPYC processors and the Boardfly network from the TPU 8i, and Google claims up to two times lower latency than the TPU 8t with a near-zero performance penalty. Teams can fine-tune, distill, and run reinforcement learning on one system, so sensitive data does not have to move between clusters. Google positions it for confidential inference and small-scale training, which makes banks, hospitals, and government agencies the obvious candidates.
The TPU 8i Max places two TPU 8i chips in a single package. Google says it delivers twice the compute, on-chip SRAM, and HBM bandwidth of the TPU 8i while fitting twice the silicon into the existing rack footprint, which Google says means more compute per square foot and per watt. Google targets massive context windows, such as many agents working across a large codebase or hours of video, where a processor can stall waiting for data to arrive. Google has not yet published absolute compute, memory capacity, or power figures for the 8i Max.
Four machines for different stages of AI work
These two new setups join the original TPU 8t model-training system and TPU 8i prompt-answering system that Google introduced in April. Google says the 8t delivers up to 2.7 times better training performance than the prior generation, and the 8i delivers 80 percent better inference price performance (Google, 2026). The four systems give buyers dedicated hardware options for building large models, generating fast replies, protecting regulated data, and running complex multi-step AI assistants.
Google also pointed to recent software updates that make these chips easier for outside teams to use. TorchTPU, which Google open-sourced in October, lets companies run PyTorch code and Hugging Face models on Google's chips with no code rewrites, according to Google. Google also added out-of-the-box TPU backends for the open-source serving engines vLLM and SGLang, released an open-source agent suite in preview that works with Claude Code, Cursor, and Google Antigravity to find bottlenecks and tune custom kernels, and made Cloud Pathways and GKE Dynamic Slicing generally available to help teams coordinate large TPU clusters.
These software updates give current Google Cloud customers easier ways to run their AI models today, even though Bloomberg reported that Google expects the two new systems in 2027.
Practical value compared with other cloud providers
These systems make sense if your workloads already run on Google Cloud, you have a clear requirement for hardware-locked privacy or long-context AI assistants, and your timeline can wait until 2027. Google's announcement includes no pricing, regional availability, or availability date for either system, and Google has not published benchmark results for either one.
Nvidia GPUs are already offered by the major cloud providers and run on mature software tools that most engineering teams know well. Competing cloud providers also sell their own custom chips for customers committed to their platforms. Google's pitch is specialization and rack density for two specific bottlenecks. The tradeoff is a 2027 arrival and a toolchain your team may know less well than Nvidia's.
How long will it take your engineering team to move its software onto Google's chip tools, and what delivery date, region, and price will you require in writing before either system enters your 2027 plan?
Google has published no head-to-head comparisons against rival chips for either system, so plan to test the privacy controls and memory bandwidth on your own workloads once Google opens access.
Sources: Google. "Introducing TPU 8i Max and TPU 8tx." Google Blog, 9 Nov. 2026, blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/.
