aigoogledevelopertutorialsproductivity

Prepare for Google TPU 8tx and 8i Max for Agentic AI

Guide to Google's new TPU 8tx confidential compute and TPU 8i Max high-density systems for agentic workloads on Google Cloud.

November 9, 2026 Yum News

Google expanded its eighth-generation TPU portfolio on November 9, 2026, with two new systems: TPU 8tx for secure confidential inference and TPU 8i Max for massive context windows. These address agentic AI needs where models handle long contexts, private data, and multi-step workflows.

TPU 8t and TPU 8i launched earlier in 2026 for training and low-latency reasoning. The new systems reuse that silicon in different packaging and networking setups. This marks the first time Google shipped four TPU systems in one year.

This guide explains what each system does, who needs it, and how developers can prepare to use them on Google Cloud.

What the new systems target

Agentic workloads differ from simple inference. Agents pause for tool calls, process million-token contexts, and often run on sensitive enterprise data. Standard accelerators hit memory walls or force data to leave encrypted environments.

TPU 8tx targets regulated industries and governments. It provides hardware-level confidential compute so model weights and private data stay protected even during active use. It pairs training-grade compute with the low-latency Boardfly network from TPU 8i.

TPU 8i Max combines two TPU 8i chips on one package. It doubles compute, SRAM, and HBM bandwidth within the same rack footprint. This raises density for workloads that process entire codebases or long video streams without stalling on memory bandwidth.

Close-up of a glowing AI chip on a circuit board

Key differences from earlier eighth-generation TPUs

TPU 8t focuses on large-scale pre-training with high goodput. TPU 8i prioritizes low-latency reasoning and sampling. The new systems specialize further.

TPU 8tx keeps data encrypted in use through hardware isolation and pairs with 5th-generation AMD EPYC CPU hosts. Organizations can fine-tune models, distill larger models, and run reinforcement learning on private data without moving it between clusters.

TPU 8i Max delivers double the compute and memory bandwidth of a single TPU 8i. It fits more silicon into existing rack space, which improves performance per square foot and per watt for high-context agent teams.

Together, the four systems cover the full agentic lifecycle: scale-up training on 8t, low-latency reasoning on 8i, confidential work on 8tx, and high-density long-context serving on 8i Max.

How to access and prepare

These systems are part of Google Cloud’s AI Hypercomputer. Availability follows the pattern of earlier eighth-generation TPUs, which became generally available later in 2026. Check the Google Cloud console and TPU documentation for current regions and quota.

Review your current TPU or GPU workloads in Google Cloud. Identify jobs that process sensitive data or long contexts.
Enable the relevant APIs in your Google Cloud project: Compute Engine and TPU API.
For confidential workloads, confirm your compliance requirements match TPU 8tx capabilities. Test with existing TPU 8t or 8i first if needed.
Update your software stack. Google open-sourced TorchTPU so PyTorch and Hugging Face models run natively on TPUs. vLLM and SGLang also have TPU backends.
Use the TPU agent suite (in preview) inside tools such as Google Antigravity or Cursor to onboard and diagnose bottlenecks with natural language.

For Android or client-side AI that still needs frontier models, see our earlier guide on adding Gemini 4 Argon to Android with Firebase AI Logic.

Secure data center hardware concept

Software stack improvements that ship with these systems

Google paired the hardware announcements with software that lowers the barrier for developers.

TorchTPU lets existing PyTorch code run without rewrites. The TPU agent suite automates onboarding and kernel tuning. TPU Sync, the open-source model-state layer, tiers KV caches across HBM, host DRAM, and storage so long-context agents do not stall. Cloud Pathways, now generally available on GKE, lets you orchestrate multi-slice TPU jobs from a single controller.

Agent Compute Engine suspends idle agent sessions in milliseconds and moves working memory off the chip. This cuts costs when agents wait on external tools or human input.

These tools work across the eighth-generation portfolio, including the new 8tx and 8i Max systems.

Practical steps for teams

Start by measuring your current bottlenecks. If memory bandwidth or context length is the limit, profile with TPU 8i and plan a migration path to 8i Max when quota opens. If data residency or encryption-in-use is required, document the private datasets and fine-tuning jobs that would benefit from TPU 8tx.

Test PyTorch models on existing TPUs using TorchTPU before requesting the new systems. Confirm that your inference engine (vLLM or SGLang) supports the TPU backend you need.

Monitor the Google Cloud blog and AI Hypercomputer documentation for exact availability dates, pricing, and supported regions. Pricing details for the new systems were not published in the initial announcement.

Tips for successful adoption

Keep model serving and training jobs on the system that matches their primary constraint. Do not force confidential workloads onto standard 8i or high-context jobs onto 8t if density or security is the goal.

Use GKE dynamic slicing, now generally available, to assemble or partition TPU topologies on demand. This reduces startup times and improves fault recovery.

Combine the new TPUs with Gemini models for agent orchestration. The same software stack that serves Gemini can now run on customer-managed TPU clusters.

Test small jobs first. The Boardfly network on 8tx and the dual-chip packaging on 8i Max change latency and bandwidth characteristics relative to earlier systems.

Conclusion

TPU 8tx and TPU 8i Max extend Google’s eighth-generation TPU line for the specific demands of agentic AI: confidential compute for sensitive data and higher density for long-context workloads. They reuse proven silicon in new system configurations so Google can respond faster to customer needs.

Developers can prepare today by validating workloads on existing TPUs, adopting TorchTPU and related open-source tools, and tracking Google Cloud availability. The combination of specialized hardware and an open software stack makes these systems practical for teams that already run Gemini or other models on Google Cloud.