Google Adds TPU 8tx for Confidential Inference and TPU 8i Max for Long-Context Agents

Google added two systems to its eighth-generation TPU lineup on Monday, November 9: TPU 8tx, built for confidential inference on sensitive data, and TPU 8i Max, which packs two inference chips into one package to handle very long context windows. The announcement targets two problems that enterprise AI leaders keep running into as agents move into production. Regulated companies want to run frontier and proprietary models next to sensitive data without exposing it, and agent workloads that read whole codebases or long document sets are hitting memory limits before they hit compute limits.

What are TPU 8tx and TPU 8i Max?

TPU 8tx is "optimized for secure and confidential inference" and is also suited to small-scale training, according to Google's announcement, written by Amin Vahdat, Google's chief technologist for AI and infrastructure. Google said it was built for regulated enterprises and governments that run frontier or proprietary models and agents alongside sensitive data, and that it lets teams fine-tune on private data, distill models and run agent reinforcement learning without moving data between clusters. Data Center Knowledge reported that it is built on the same silicon as TPU 8t, Google's training chip, with hardware-level encryption added.

TPU 8i Max is designed for what Google calls the memory wall. It combines two TPU 8i chips on a single package, doubling compute, high-bandwidth memory bandwidth and on-chip SRAM compared with TPU 8i, and fits twice the silicon into the same rack footprint, according to Google. Network World reported that it carries 384 GB of HBM, 33% more than TPU 8i, with 17.2 TB/s of memory bandwidth and 768 MB of SRAM.

How does confidential inference on TPU 8tx work?

Google said the design protects data while it is in active use by building security into the hardware, keeping sensitive data and model weights isolated. Mark Lohmeyer, Google Cloud's vice president and general manager for AI and computing infrastructure, told Network World that the system encrypts the flow of data from the host to the TPUs and secures the TPUs while they run inference, and that the entire system is attested. Network World described TPU 8tx as Google Cloud's first TPU with confidential computing capabilities.

The hosts are 5th Generation AMD EPYC processors, according to Google. TPU 8tx also adopts the low-latency network design used in TPU 8i, and Google claims up to two times lower latency than TPU 8t "with a near-zero performance penalty" for the security features. "You shouldn't have to trade off between security and performance," Lohmeyer told Network World. Google did not publish details of its encryption or attestation mechanisms, which security architects will want before approving regulated workloads.

Can enterprises run TPU 8tx in their own data centers?

Possibly. Data Center Knowledge reported, citing Lohmeyer, that Google Cloud would allow select customers to deploy TPU 8tx in their own data centers. That matters for banks, health systems and public agencies that cannot move certain data into a shared cloud, and it extends a pattern in which hyperscalers bring their own hardware into customer and sovereign facilities. Google did not say which customers qualify or when on-premises deployments would begin.

No general availability date, region list or pricing was given for either system in the coverage we reviewed. Google did say TPUs are now "a rapidly growing business with nearly 100 customers."

Why does long context need different hardware?

Agent workloads behave differently from chatbots. Lohmeyer described them to Network World as "bursty, long running, but often idle," sometimes waiting for days between steps. Google's announcement gives examples such as dozens of agents working across an entire codebase or hours of video, which require large key-value caches to hold context. More memory and bandwidth per package let those caches stay closer to the compute.

Google paired the hardware with software aimed at the same problem. Its announcement describes TPU Sync, which moves key-value caches off-chip and cuts cache transfer latency and time to first token by more than 90%, and Agent Compute Engine, which suspends and resumes idle agent sessions. GKE dynamic slicing is now generally available, with up to five times faster job startup and up to 4.5 times faster fault recovery, and Cloud Pathways is generally available on GKE, according to Google.

How are analysts reading the launch?

Dave McCarthy, a group vice president at IDC, told Data Center Knowledge that the pace of Google's 2026 chip launches shows its commitment and could help it stand apart from rival clouds. He called confidential computing an emerging requirement and said it could give Google "a little bit of an edge," noting that Nvidia is still developing its approach in this area. Constellation Research counted four new Google AI chips this year.

What should enterprise data and AI leaders take from it?

For teams in regulated industries, confidential inference is moving from a niche option to a standard question in cloud evaluations. Leaders should ask every provider how data and model weights are protected during inference, what attestation evidence they can see, and whether the hardware can run in their own facilities. For teams building long-running agents, the more useful metric is now memory capacity and cache handling per dollar rather than raw compute, and TPU 8i Max is aimed squarely at that. Because pricing and availability are not yet public, the practical step today is to put both systems on evaluation roadmaps for 2027 workloads and to test whether existing inference software runs on TPUs without major rework.

Frequently asked questions

What is Google TPU 8tx?

A TPU system announced November 9, 2026, optimized for confidential inference and small-scale training, aimed at regulated enterprises and governments.

What is TPU 8i Max?

Two TPU 8i chips in one package, with double the compute, memory bandwidth and SRAM of TPU 8i, designed for large context windows and agent workloads.

When will TPU 8tx and 8i Max be available?

Google did not give an availability date or pricing in its announcement.

Filed under: Data & AI Infrastructure, AI Governance & Trust, Agentic AI