Google splits eighth-generation TPUs for agents and confidential AI

By Daniel Okafor

What did Google change?

Google has disclosed two eighth-generation TPU variants, but omitted pricing and launch dates for both.

The TPU 8i Max targets agentic computing; the 8tx targets confidential inference.

Google launched the first eighth-generation pair in April: the 8t for large-scale training and the 8i for inference.

Lohmeyer said the eighth generation was the first TPU generation split by hardware specialty.

Google built TPUs for its own AI workloads and cloud customers, with each generation targeting the constraints of its time.

Mark Lohmeyer, Google Cloud’s vice president and general manager for AI and computing infrastructure, said deployment priorities now span compute, memory, latency, cost and security.

According to Google, its rack-level architecture optimizes CPUs, system DRAM, networking and shared storage around each workload.

Lohmeyer said: “You can't optimize that infrastructure just by focusing on the chip.”

What is the 8i Max built for?

Google designed the 8i Max for the massive context windows that agentic systems require.

Patrick Moorhead, founder, CEO and chief analyst of Moor Insights & Strategy, estimates that roughly two-thirds of AI compute now goes to inference.

That estimate lacks a disclosed workload mix and time series.

Google’s chip adds a second TPU to a multi-chip package, giving agents as much memory as possible.

Google specifies:

• The 8i Max provides 768 MB of on-chip static RAM, according to Google.
• It provides 17.2 TB/s of high-bandwidth memory bandwidth—17,200 GB per second—according to Google.
• Its package contains two TPUs, according to Google.

Google says the two-chip design lets groups of agents cooperate on software development and video creation.

The company also says large TPU deployments maximize compute per square foot of data-center space.

The announcement gives no density figure, workload baseline or customer result for that claim.

Google recently open-sourced TPU Sync, which orchestrates chip HBM, host DRAM, shared storage and the data-center network into one memory hierarchy across large-scale TPU deployments.

The announcement provides no deployment count or measured gain.

How does the 8tx protect sensitive workloads?

Google describes secure AI computing as an emerging customer workload.

Google says the 8tx focuses on secure and confidential inference for sensitive government and regulated-enterprise work.

Vahdat says teams can fine-tune models, distill large models into smaller ones, and train agents through reinforcement learning on one confidential system.

Sensitive data need not move between separate clusters, according to Vahdat’s Monday blog.

Moorhead said: “A bank or government agency can fine-tune, distill and run reinforcement learning on the same confidential system that serves the model, without moving sensitive data between clusters.”

Google says data and model weights remain encrypted and isolated at the hardware level during processing.

Lohmeyer claims the 8tx incurs a near-zero performance penalty and cuts latency by half against the 8t.

The confidential-computing claim remains a vendor claim because the disclosure gives no benchmark definition, comparison configuration, independent test or latency method for 8tx against 8t.

Who is using Google’s TPUs?

Google says more than 100 customers and partners now use its TPUs.

The group includes frontier labs, trading firms, retailers and universities, according to the company.

Anthropic plans to expand its use of next-generation TPUs beginning in 2027.

The California Institute of Technology has shifted some cancer research from its own HPC system to Google Cloud TPUs, Moorhead said after speaking with the team.

The disclosed customer count does not show accelerator numbers, workload volumes or service availability.

Google built the eighth generation with input from the Google DeepMind research lab and outside advice from cloud customers and channel-sales users.

Moorhead calls the shift a turning point. Agents, he says, return the CPU and host platform to the center.

The GPU is no longer AI computing’s only focal point.

The customer count signals broader adoption, but it mixes laboratories, regulated enterprises, retailers and universities without revealing deployment size, revenue, workload volume or service availability.

What can buyers verify?

Vahdat says the 8t offers up to 2.7 times better training performance than earlier TPU generations.

Vahdat says the 8i provides 80% better inference price-performance than earlier generations.

The claims compare different workloads, so their percentages do not measure the same improvement.

The disclosed figures give no benchmark definition, workload mix or prior-generation configuration.

Google has given no availability date or price for the 8i Max or 8tx.

Google still expects the earlier 8t and 8i to come online by year’s end.

That distinction leaves the two new products without a firm deployment timetable.

For enterprise buyers, the central caveat is straightforward: performance claims carry percentages, while the new accelerators still lack dates, prices and independently reproducible comparison data.