On 9 November 2026, Google announced two further eighth-generation Tensor Processing Unit (TPU) systems: TPU 8tx, aimed at secure and confidential inference, and TPU 8i Max, aimed at very long context windows. In a post by Amin Vahdat, Google’s SVP and Chief Technologist for AI and Infrastructure, the company said it was the first time it had introduced four TPU systems in a single year; TPU 8t and TPU 8i were announced at Google Cloud Next in April. All performance figures below are Google’s own claims, not independent measurements.
What Google announced
TPU 8tx: confidential inference
- Google describes TPU 8tx as built for secure and confidential inference, and also suited to small-scale training. Its stated target is regulated enterprises and governments running frontier or proprietary models and autonomous agents alongside sensitive data.
- It uses the same silicon as TPU 8t, paired with 5th Gen AMD EPYC CPU hosts.
- Google says it adds hardware-level confidential compute that keeps sensitive data and model weights protected and isolated while in use.
- It uses the low-latency Boardfly network from TPU 8i. Google claims up to 2x lower latency than TPU 8t and a “near-zero performance penalty” from the confidential-computing protections.
- Google says teams can fine-tune models on private data, distill models and run reinforcement learning on one system without moving sensitive data between clusters.
TPU 8i Max: long-context inference
- TPU 8i Max combines two TPU 8i chips on a single package, targeting massive context windows and the highest compute density per rack.
- Relative to TPU 8i, Google claims double the compute, double the SRAM and double the HBM bandwidth, in the existing rack footprint.
Context from the same post
- Google says TPUs now serve nearly 100 customers, including frontier labs, trading firms, retailers and universities.
- It restates earlier claims for the April systems: up to 2.7x better training performance for TPU 8t and up to 80% better inference price-performance for TPU 8i, each versus the previous TPU generation.
- Software items listed include the open-source TorchTPU (native PyTorch and Hugging Face models on TPUs), TPU backends for vLLM and SGLang, the open-source TPU Sync memory layer, and generally available GKE dynamic slicing and Cloud Pathways on GKE.
What the announcement does not say
The post gives no pricing and no specific availability dates for either TPU 8tx or TPU 8i Max. Research offices should not assume access terms, regions, quotas or academic programme eligibility until Google publishes them. The benchmark and latency claims are vendor-reported and have not been independently reproduced in the source.
Why it matters for researchers and research offices
The points below are editorial interpretation, not claims from Google.
- Sensitive data. Hardware-isolated processing is relevant to work involving human-participant data, health records or controlled datasets, where data-protection and data-use agreements often restrict where data can go. Whether a given confidential-computing service satisfies a specific ethics approval, data-sharing agreement or institutional policy is a question for the institution’s information-security, privacy and research-compliance staff, and for Google’s documentation once available.
- Long-context workloads. Large context windows matter for analysing long documents, codebases or corpora in a single pass. TPU 8i Max is aimed at that serving problem, but the announcement does not describe research-specific use or access.
- Compute access and budgeting. University mention in the customer list is not a statement about academic pricing or grants. Offices that manage cloud credits and compute budgets should watch for pricing and regional availability before planning around these systems. See our related coverage of falling AI token costs and institutional budgets.
- Vendor concentration. Reproducibility and portability considerations apply when workflows depend on a single vendor’s accelerator; the open-source tooling Google lists (PyTorch and Hugging Face support, vLLM, SGLang) is relevant here.
Source
Amin Vahdat, “Our newest TPUs: 8tx and 8i Max, built for the agentic era”, Google blog, 9 November 2026. Technology.org also reported the announcement on 11 November 2026.
