DataCentreNews India - Specialist news for cloud & data centre decision-makers
India
Google expands TPU line-up with two new AI systems

Google expands TPU line-up with two new AI systems

Tue, 17th Nov 2026 (Today)
Joseph Gabriel Lagonsin
JOSEPH GABRIEL LAGONSIN News Editor

Google has expanded its eighth-generation Tensor Processing Unit line-up with TPU 8i Max and TPU 8tx, bringing its TPU releases this year to four systems.

The new systems target different AI workloads in Google Cloud. TPU 8tx is aimed at secure and confidential inference, while TPU 8i Max is designed for workloads that need large context windows and greater compute density within the same rack footprint.

Google said the launch reflects broader demand from customers using AI for model training, inference and autonomous agents. It added that TPUs are now used by nearly 100 customers, including frontier laboratories, trading firms, retailers and universities.

New systems

TPU 8tx is built on the same silicon as TPU 8t and is paired with fifth-generation AMD EPYC CPU hosts. Google said it is intended for regulated enterprises and government users that want to run models and agents alongside sensitive data.

According to Google, TPU 8tx protects data and model weights while they are in use through hardware-level confidential computing. It also uses the same Boardfly network introduced for lower-latency reasoning workloads, delivering up to half the latency of TPU 8t with a near-zero performance penalty, the company said.

The set-up is intended to let customers run inference and smaller-scale training on the same system. Organisations can fine-tune models with private data, distil larger models into smaller ones and train agents through reinforcement learning without moving sensitive data across separate clusters, Google said.

TPU 8i Max is designed to address what Google described as the memory bottleneck in AI systems. It combines two low-latency TPU 8i chips into a single package, doubling compute, SRAM and high-bandwidth memory bandwidth compared with TPU 8i, according to the company.

Google said the result is a system for workloads in which multiple agents work across large code bases or extended video inputs. It added that the design fits twice the silicon into the existing rack footprint, allowing data centres to deploy more compute per square foot and per watt.

Broader portfolio

The launch expands Google's eighth-generation TPU family to four products: TPU 8t for large-scale training, TPU 8i for lower-latency reasoning, TPU 8tx for confidential computing and TPU 8i Max for high-context workloads.

Google said this system-level approach allows it to respond more quickly to changing customer requirements than waiting for a full new chip generation. Rather than redesigning silicon each time, it modifies the surrounding compute, networking, storage and packaging to create more specialised systems.

"With computing moving from human to agent scale, businesses are asking for accelerators that can reason across AI models' massive context windows," Google said.

"In addition, some highly regulated customers are asking for specialized encryption capabilities for unique workloads."

Software push

Alongside the hardware additions, Google outlined a broader effort to make TPUs easier for developers and cloud customers to use. It said PyTorch code and Hugging Face models now run natively on TPUs through the open-source TorchTPU project, while TPU back ends are available for open-source inference engines including vLLM and SGLang.

Google also said a TPU agent suite in preview can plug into coding tools to automate onboarding, diagnose bottlenecks and tune kernels. For production use, its open-source XLA Serving Engine allows customers to train models in Python and serve them in a C++ runtime across TPUs, GPUs and CPUs.

Other software updates include Agent Compute Engine, which suspends and resumes idle agent sessions to reduce infrastructure costs, and TPU Sync, which manages model state across high-bandwidth memory, host memory, storage and network resources. Google said TPU Sync can reduce key-value transfer latency and time to first token by more than 90%.

It also pointed to the general availability of Cloud Pathways on Google Kubernetes Engine, extending distributed runtime tools used internally for Gemini to Google Cloud customers. In a separate update, Google said GKE dynamic slicing now lets users reconfigure TPU topologies on demand, cutting job start-up times by up to five times and improving fault recovery by up to 4.5 times.

Competitive backdrop

The expanded TPU range comes as cloud groups and chip suppliers compete to meet demand for AI computing infrastructure, particularly for workloads that need lower latency, larger memory footprints and stronger data controls. Google said the broader range reflects increasingly specialised demand as customers move from traditional model training to inference-heavy and agent-based applications.

Google also said it has been using AI tools from Google DeepMind in its chip design process, including AlphaChip reinforcement learning systems and Gemini-based design tools. According to the company, those tools have helped cut parts of chip layout work from weeks to hours.

"Shipping four TPU systems and a unified, open software stack in a single year only happens when silicon architects, networking engineers, compiler teams and Google DeepMind researchers build side by side," Google said.