Google Cloud Unveils 8th-Gen TPUs and Gemini Enterprise Agent Platform (Comprehensive)

Two Types of 8th-Generation TPUs and Enterprise AI Agent Platform Unveiled

"Supporting AI Agent Adoption for Every Business"

Google Cloud has unveiled its 8th generation Tensor Processing Units (TPUs), along with an enterprise AI agent platform built on this new hardware, aiming to boost the efficiency of AI agent (assistant) adoption by businesses.

The appearance of TPU 8th generation. TPU 8t (left) and TPU 8i (right). Google Cloud.

On April 22 (local time) in Las Vegas, United States, Google Cloud introduced two types of 8th generation TPUs at 'Google Cloud Next 2026'. These include the TPU 8t for training and the TPU 8i for inference. They support not only building high-performance AI models but also enable agent clustering—running multiple AIs as a team—and help solve complex inference tasks.

The 8th generation TPU delivers 80% better performance per dollar compared to its predecessor, the 7th generation 'Ironwood'. This allows enterprises to handle twice as much customer demand at the same cost.

The TPU 8t achieves three times the processing capability of the previous model by leveraging high computational throughput and shared high-bandwidth memory (HBM). With this chip, the development period for cutting-edge models can be reduced from several months to just a few weeks. The TPU 8i is equipped with 288GB of HBM and 384MB of high-speed SRAM, enabling the entire active working set of a model to be processed on-chip. This shortens the data transfer path between chips by more than half, allowing AI responses or outputs to be delivered instantly.

Both types of 8th generation TPUs, together with Axion (Google's in-house designed CPU) and NVIDIA's GPU, form the foundation of Google's AI hypercomputer.

Thomas Kurian, CEO of Google Cloud, explained, "We anticipated that as generative AI becomes more widespread, people would want systems optimized for training and inference. Expecting that power consumption would become a limiting factor in scaling AI infrastructure, we focused on maximizing energy efficiency from the design stage onward."

On the 22nd (local time) at 'Google Cloud Next 2026' held in the United States, Thomas Kurian, CEO of Google Cloud, gave a keynote speech. Google Cloud.

"A Full-Fledged Leap into the Era of AI Agents"

Google Cloud has also newly introduced its enterprise AI agent platform, 'Gemini Enterprise Agent Platform', built on the AI hypercomputer. This platform is an evolution of the previous AI development platform 'Vertex AI', integrating model selection and construction features, agent creation and integration, and new security functions. Agents deployed through the Gemini Enterprise Appare tightly integrated with IT operating systems, ensuring security is maintained even as services scale.

More than 200 models are available within the platform. It supports Google’s latest models—Gemini 3.1 Pro, Nano Banana 2, Lyria 3—as well as third-party models such as Anthropic’s Claude Opus 4.7.

Google Cloud aims to accelerate enterprise AI transformation by leveraging its AI-powered cloud technologies. 75% of Google Cloud customers worldwide use AI product suites to drive their businesses. Over the past 12 months, 330 Google Cloud clients processed more than 1 trillion tokens of data, while 35 companies processed 10 trillion tokens using Google models.

CEO Kurian emphasized, "We will support every company in adopting AI agents," adding, "We will provide a fully integrated, vertically optimized stack to maximize efficiency and enable large-scale operations."

This content was produced with the assistance of AI translation services.