Google TPU 8tx and 8i Max: Agentic Era Guide

Google expanded its eighth-generation TPU portfolio on 9 November 2026 with TPU 8tx and TPU 8i Max. These are the third and fourth systems in the generation, marking the first time the company has introduced four TPU systems in a single year.

TPU 8t and TPU 8i arrived in April for large-scale training and real-time reasoning. The new systems address two rising demands from agentic workloads: hardware-level data protection during processing, and enough memory bandwidth for million-token contexts.

Why four systems in one year

Google first designed TPUs more than a decade ago for internal needs such as voice recognition and translation. Eight generations later the workloads cover Gemini training and serving, agentic systems, Google Cloud customers, and direct customer deployments in their own data centers.

Nearly 100 customers now use TPUs, including frontier labs, trading firms, retailers, and universities. Feedback from external organizations has grown. With computing shifting from human to agent scale, customers ask for accelerators that handle massive context windows and, for regulated sectors, specialized encryption that protects data while it is in active use.

Google’s modular AI Hypercomputer approach lets engineers adjust packaging, high-speed memory, optical networking, and security around the same silicon base in months rather than the multi-year cycle required for a new chip.

TPU 8tx: secure and confidential inference

TPU 8tx is optimized for secure and confidential inference and works well for small-scale training. It targets regulated enterprises and governments that run frontier or proprietary models and autonomous agents next to sensitive data.

Normally data is encrypted at rest or in transit but must be unlocked in memory while a processor works on it. TPU 8tx builds protection into the hardware layer so sensitive data and model weights stay protected and isolated even during active processing.

It uses the same silicon as TPU 8t, pairs with 5th Generation AMD EPYC CPU hosts, and adopts the low-latency Boardfly network created for real-time reasoning. Google states this delivers up to 2x lower latency compared with TPU 8t, with a near-zero performance penalty.

Because the system combines training-grade compute with a fast inference network, teams can fine-tune models on private data, distill large models into smaller ones, and train agents through reinforcement learning without moving sensitive data between separate clusters.

TPU 8i Max: massive context and rack density

TPU 8i Max attacks the memory wall. When many AI agents work across an entire software codebase or hours of video at once, memory bandwidth becomes the bottleneck.

Google placed two low-latency TPU 8i chips onto a single package. The result doubles the compute, doubles the on-chip SRAM, and doubles the HBM bandwidth of a single TPU 8i. Processors stay supplied with data so large workloads do not stall.

Because the dual-chip package fits twice the silicon inside the existing rack footprint, data centers gain more computing power per square foot and per watt.

Together the four eighth-generation systems cover the agentic lifecycle: TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing, and TPU 8i Max for massive context windows and the highest compute density per rack.

Software that makes the hardware usable

Hardware alone is not enough. Google has opened TPU development to PyTorch users. TorchTPU is open-sourced so PyTorch code and Hugging Face models run natively on TPUs. Out-of-the-box backends exist for vLLM and SGLang.

A new open-source TPU agent suite (in preview) plugs into coding tools such as Claude Code, Cursor, and Google Antigravity. It automates onboarding, diagnoses bottlenecks, and autotunes kernels with plain language.

The open-source XLA Serving Engine lets teams train models in Python and serve them in a low-latency C++ runtime across TPUs, GPUs, and CPUs.

On industry InferenceX and AgentX benchmarks that test realistic agentic coding and long-context work, the native vLLM-TorchTPU stack has delivered strong price-performance results. Upcoming software optimizations on the same hardware are expected to improve those marks further.

Agent Compute Engine automatically suspends and resumes idle agent sessions, cutting costs when agents wait on tools or human feedback. TPU Sync, the open-sourced model-state layer, turns chip HBM, host DRAM, shared storage, and the network into one memory hierarchy. It tiers KV caches, streams context across prefill and decode nodes, and synchronizes RL weights, cutting KV transfer latency and Time to First Token by more than 90 percent.

Cloud Pathways is now generally available on GKE. It brings the DeepMind runtime behind Gemini to customers so they can orchestrate large TPU systems across slices from a single controller for pre-training, reinforcement learning, and multihost inference. GKE dynamic slicing, also generally available, uses optical circuit switching to assemble or partition topologies on the fly, reducing job startup times and accelerating fault recovery.

How the systems fit agent workflows

Start by matching the workload:

1. Use TPU 8t when the primary job is large-scale foundation-model training and maximum goodput (useful intelligence per watt).
2. Choose TPU 8i for low-latency real-time reasoning and serving.
3. Select TPU 8tx when the data or model weights must remain protected in memory during inference or small-scale training and fine-tuning.
4. Deploy TPU 8i Max when agents require very large context windows and you need maximum compute density inside the existing rack footprint.

Combine the hardware with the software stack above. TorchTPU removes the need to rewrite PyTorch code. TPU Sync and Agent Compute Engine reduce the cost of long-running, stateful agents that frequently pause. Cloud Pathways and GKE dynamic slicing simplify orchestration at scale.

For teams already exploring secure agent runtimes, the earlier Agent Compute Engine guide covers hardware-isolated microVMs that hold no long-lived credentials—an orthogonal but complementary layer to the confidential-compute features in TPU 8tx.

Practical tips

Treat the four systems as specialized tools rather than interchangeable chips. Profile your agent’s context length and sensitivity requirements before choosing a configuration.

Because the systems reuse existing silicon and the AI Hypercomputer packaging model, Google can ship new variants faster than a full new-chip cycle. Expect further configurations as agent workloads evolve.

Availability dates and pricing have not been published. Monitor the Google Cloud TPU product page and the original announcement for updates. Early access will likely follow the same phased pattern used for other recent AI infrastructure.

Conclusion

TPU 8tx and TPU 8i Max complete Google’s eighth-generation lineup for the agentic era. One protects data and model weights while they are in use; the other doubles compute and memory bandwidth inside the same rack footprint so long-context agents do not stall. Both sit on the modular system approach that has let Google ship four TPU systems in one year.

The accompanying open software—TorchTPU, the TPU agent suite, TPU Sync, Cloud Pathways, and GKE dynamic slicing—lowers the barrier for PyTorch developers and production teams. The hardware is only as useful as the stack that makes it reachable.

Sources

• Google Blog: Introducing TPU 8tx and TPU 8i Max (Amin Vahdat, 9 November 2026)
• Google Cloud Blog and product documentation on TorchTPU, Cloud Pathways, GKE dynamic slicing, and TPU Sync
• Official Google Cloud Tech YouTube broadcast linked from the announcement