Google broadens its TPU lineup with 8tx and 8i Max for agentic AI

Google adds two systems to its TPU v8 lineup for confidential inference and giant contexts. The offering targets regulated enterprises and AI agent use cases, with a strong software focus.

Google has introduced two new TPU systems, the TPU 8tx and TPU 8i Max, bringing the total to four variants of its v8 generation announced this year. The goal is clear: to meet the rise of AI agents, million-token contexts, and the confidentiality needs of regulated enterprises. According to Amin Vahdat, chief technologist for AI and infrastructure, this is "the first time we have introduced four TPU systems in a single year." This pace speaks volumes about the pressure on AI infrastructure, where hardware can no longer arrive after software.

Deux systèmes, deux problèmes très concrets

The TPU 8tx, a new system in the lineup, targets confidential inference and small-scale training. Google aims it at banks, governments, and organizations handling sensitive data, with a simple promise: to keep data and model weights "protected and isolated even while in active use." The system uses the TPU 8t silicon but pairs it with 5th generation AMD EPYC hosts and the same low-latency network link as the TPU 8i. The announced result: up to twice less latency than the TPU 8t, with an almost negligible impact on performance.

Meanwhile, the TPU 8i Max tackles the notorious "memory wall," the bottleneck where memory bandwidth eventually slows down the heaviest workloads. Google chose to put two TPU 8i chips in a single package. The idea is to double compute, onboard SRAM, and HBM bandwidth to continuously feed agents that browse large code repositories or hours of video. All while keeping the same rack footprint. It’s the hardware version of the saying "more memory, fewer sighs."

Pourquoi Google insiste autant sur le logiciel

The novelty is not only in silicon. Google is also pushing a series of tools meant to lower adoption barriers against CUDA and habits built around Nvidia. Last month, the company open-sourced TorchTPU, so PyTorch code and Hugging Face models run natively on TPU, with backends for vLLM and SGLang. The new TPU agent suite, still in preview, connects to Claude Code, Cursor, and Google Antigravity to automate onboarding, diagnose bottlenecks, and adjust kernels using natural language.

Another highlighted component: TPU Sync, an open-source layer that distributes KV caches between chip memory, host memory, and storage. Google claims it reduces KV transfer latency and time to first token by over 90%. For agent workloads, this matters as much as raw compute gains, since interruptions and memory roundtrips quickly cost more than an extra watt.

Une gamme pensée pour l’IA agentique

Google now presents its v8 generation as a complete chain: TPU 8t for large-scale training, TPU 8i for low-latency reasoning, TPU 8tx for confidential compute, and TPU 8i Max for giant contexts. According to the company, TPUs represent a fast-growing business with "nearly 100 customers," from cutting-edge labs to universities, finance, and commerce players. Exact availability and pricing for the two new systems have not been disclosed [to be verified].

This strategy has a dual effect. First, it gives Google an extra argument against Nvidia on inference workloads, which now absorb a large share of AI compute. Second, it shows that the battle is no longer just about chip power but the entire system: network, memory, orchestration, and security. In this context, compute is no longer sold as bare silicon but as a complete architecture.

Key points

  • Google adds TPU 8tx and TPU 8i Max to its v8 lineup.
  • Amin Vahdat speaks of the "first time" with four systems in one year.
  • TPU 8tx targets regulated uses and confidential inference.
  • TPU 8i Max doubles compute, SRAM, and memory bandwidth.
  • TorchTPU makes PyTorch native on TPU without major rewrites.
  • TPU Sync promises over 90% latency reduction on KV caches.

By the numbers

  • 4 systems — Google’s v8 generation in one year, according to Google.
  • Nearly 100 customers — commercial base mentioned by Google, November 2026.
  • Up to 2x — latency reduction announced for TPU 8tx versus TPU 8t.
  • Over 90% — announced reduction of time to first token with TPU Sync.
  • 2 systems in 1 package — design principle of TPU 8i Max.

To read