Google expands its TPU lineup with chips for AI agents and giant contexts
Google adds TPU 8tx and 8i Max to its AI chip family. One targets confidential inference, the other very large contexts, with a clear goal: better serving AI agents and cloud customers.
On November 9, 2026, Google unveiled two new eighth-generation TPU systems, TPU 8tx and TPU 8i Max, complementing the TPU 8t and 8i announced earlier this year. The first targets confidential inference for regulated sectors, the second very large context sizes and compute density in data centers. The challenge goes beyond specs: it’s about adapting infrastructure to AI agents, which consume more memory, require more security, and tolerate less network latency.
Deux systèmes pour deux goulots d’étranglement
TPU 8i Max tackles the “memory wall,” the ceiling that appears when models must handle too much data at once. Google states this system doubles compute, SRAM, and HBM bandwidth compared to TPU 8i, while fitting in the same rack footprint. Concrete translation: more power per square meter and per watt, a criterion that has become central in data centers.
On the other hand, TPU 8tx targets use cases where confidentiality cannot be just a marketing argument. Google presents it as a system designed for secure and confidential inference, with regulated companies and online administrations in mind. The hardware integrates protection directly at the hardware layer to keep sensitive data and model weights isolated during computation.
Le pari des contextes massifs, ou quand les agents lisent trop de pages
The rise of AI agents changes the game. These systems no longer respond to isolated queries; they chain actions, consult tools, cross-reference documents, and keep long conversations in memory. Google says it wants to address this evolution with TPU 8i Max for “massive context windows,” context windows that allow a model to work on much larger volumes of text, code, audio, or video.
In its communication, the company also links this movement to Gemini, its flagship model, which supports a window of 1 million tokens. A token corresponds to a unit of text processed by the model; 1 million tokens represent about 750,000 words, or nearly 1,500 pages. At this stage, the promise is mainly industrial: keeping latency under control while models ingest more information.
Google pousse aussi l’ouverture logicielle
The new systems do not come alone. Google highlights a set of tools designed to reduce adoption friction, with support for PyTorch, open-source TorchTPU, compatibility with Hugging Face, and support for inference engines like vLLM and SGLang. The goal is clear: so developers don’t have to rewrite their code to use TPUs.
The company also announces operational tools, including Agent Compute Engine, which automatically suspends inactive agent sessions, and TPU Sync, a model state management layer aimed at reducing context transfers and first response time. On GKE, Cloud Pathways and dynamic slicing aim to simplify managing large TPU clusters. A way of saying Google no longer just sells silicon but a complete stack, from code to rack. All without forgetting that agents love to take a break at the worst moment.
Ce que change cette montée en gamme pour le marché cloud
With four TPU systems launched in a single year, Google seeks to lock in a finer position in the AI accelerator market. The 8t and 8i laid the foundation for training and reasoning; 8tx and 8i Max add two layers highly demanded by customers: compliance and density. This segmentation responds to a simple market reality: not all models have the same needs, and not all customers can freely move their data from one cluster to another.
Google’s announcements also reveal an underlying tension in AI infrastructure: models become more capable but demand more memory, bandwidth, and security guarantees. In this context, the differentiator is no longer just raw speed. It lies in the ability to properly execute very different workloads without breaking the bill or compliance.
Key points
- TPU 8tx and 8i Max complete the generation launched in April 2026.
- 8i Max doubles compute, SRAM, and HBM compared to 8i.
- 8tx targets confidential inference for regulated sectors.
- Google mentions four TPU systems launched in one year.
- Gemini supports 1 million tokens.
- Google Cloud highlights PyTorch, GKE, and Cloud Pathways.
By the numbers
- 4 systems — TPUs announced in a single year, according to Google, November 2026.
- 1 million tokens — context window mentioned for Gemini, according to Google.
- 2x — compute, SRAM, and HBM of TPU 8i Max versus TPU 8i, according to Google.
- 2x — latency reduction announced for TPU 8tx versus TPU 8t, according to Google.
- 80% — price-performance improvement for TPU 8i compared to the previous generation, according to Google in April 2026.