Compute

Google’s New TPUs
Target Agent Security
and Memory Limits

TPU 8tx promises protection for data during processing; 8i Max targets larger contexts. Small service providers should treat both as roadmap signals, not customer-ready capabilities.

By Felix AroyanCompute & infrastructure·November 9, 2026

Google introduced TPU 8tx and TPU 8i Max on November 9, targeting data protection during processing and memory limits in AI agents. Neither has an announced price or availability date, according to TechTarget.

For small businesses selling AI-enabled services, those targets could address customer concerns about sensitive data and document-heavy workflows. But these aren’t customer-ready capabilities, and specifications don’t establish workflow performance or verified deployment protections.

Google’s announcement says 8tx targets confidential inference, while 8i Max targets massive context windows. The additions bring its eighth-generation accelerator portfolio to four systems. TPU 8t, for large-scale training, and TPU 8i, for real-time reasoning, arrived in April. Google claims they deliver up to 2.7 times better training performance and 80% better inference price-performance, respectively, versus its previous generation.

Protection during processing

Google says TPU 8tx combines the silicon used in TPU 8t with fifth-generation AMD EPYC CPU hosts. Its protections are designed to isolate sensitive data and model weights during active processing, not just storage or transmission. Network World describes it as Google Cloud’s first TPU with confidential computing capabilities.

Google says 8tx uses the low-latency Boardfly network for real-time reasoning, delivering “up to 2x lower latency” compared with TPU 8t, with a near-zero performance penalty for security protections.

Mark Lohmeyer, Google Cloud’s vice president and general manager for AI and computing infrastructure, told reporters the system encrypts data flowing from hosts to TPUs and protects active inference, according to Network World.

“Then that entire system is attested so that we know it is what it says it is,” Lohmeyer said.

Google says teams can fine-tune on private data, distill larger models into smaller ones and train agents through reinforcement learning on the inference system, without moving sensitive data between separate clusters.

More room for context

Google says TPU 8i Max packages two TPU 8i chips together, doubling compute, on-chip SRAM and high-bandwidth memory bandwidth while fitting twice the silicon into the existing rack footprint.

Network World reports 384 GB of high-bandwidth memory, 33% more than TPU 8i, alongside 768 MB of SRAM and 17.2 TB/s of memory bandwidth. Google cites agents working across an entire codebase or hours of video as workloads facing memory bottlenecks.

Google’s open-source TPU Sync software unifies chip memory, host memory, shared storage and networking. Google says it tiers KV caches, stored context information, off-chip in milliseconds, streams context between prefill and decode nodes, and synchronizes reinforcement-learning weights. It claims reductions exceeding 90% in KV-cache transfer latency and time to first token. Network World reports support for TPU v5, v6 and v7, not just v8.

Google also open-sourced TorchTPU in October. It says the software runs PyTorch code and Hugging Face models natively on TPUs, with TPU backends for the vLLM and SGLang inference engines.

For an owner assessing whether privacy concerns or document complexity are blocking sales, this is a roadmap to track, not a deployment promise. Larger memory isn’t proof of better customer-workflow results; confidential computing doesn’t by itself verify an entire service’s protections. Pricing, availability and cloud access remain the buying questions before providers can make new claims to customers.

Sources