Google Launches Two TPU Models, Accelerates Enterprise AI Expansion

Alphabet's Google Cloud has made a flurry of announcements at its annual developers' conference, launching a new generation of self-developed chips alongside enterprise-grade AI agents' tools, simultaneously challenging Nvidia, OpenAI, and Anthropic on both hardware and software fronts.

On Wednesday, in Las Vegas at Google Cloud Next 2026, Google Cloud unveiled two new products of the eighth-generation tensor processing units (TPU): the TPU 8T, specifically designed for AI model training, and the TPU 8i, optimized for the inference stage, both expected to be available later this year. This is the first time Google has separated training and inference tasks onto independent chips, marking a significant shift in its AI hardware strategy.

At the same time, Google also introduced a series of AI agent development tools, including the Gemini Enterprise Agent Platform, directly targeting the corporate automation market.

The release of the new chips comes as the demand for AI inference expands rapidly. Mark Lohmeyer, vice president of Google Cloud’s computing and AI infrastructure division, said: "The key is how to achieve the lowest response latency at the lowest cost per transaction. Transaction volume is surging, and the cost per transaction must decrease significantly to achieve scale." The two new chips will be officially launched later this year.

The Separation of Training and Inference Leads to a Significant Leap in Chip Performance

Google's move to split the eighth-generation TPU into two independent products is a direct response to the increasingly differentiated trends in AI workloads.

Amin Vahdat, Senior Vice President and Chief Technology Officer for AI and infrastructure, wrote in a blog post: "With the rise of AI agents, we believe the industry will benefit from chips optimized for the specific needs of training and inference."

The TPU 8t is optimized specifically for AI model training, claiming to be able to "compress the development cycle of cutting-edge models from months to weeks".

In terms of performance, the performance per watt of TPU 8t has increased by 124% compared to the previous generation, while TPU 8i has increased by 117%. Compared to the seventh-generation Ironwood TPU released in November last year, the performance of TPU 8t is increased by 2.8 times under the same price, while the performance of TPU 8i is increased by 80%.

Training chip TPU 8t can combine up to 9600 chips into a system. Google stated that when deploying such a large-scale system, electricity has become a core constraint for data centers, making a higher energy efficiency ratio essential.

TPU 8i is primarily aimed at inference scenarios, suitable for running AI models and handling tasks of AI agents. Its architectural design focuses on large-capacity die stacking memory. Each chip integrates 384MB of static random access memory (SRAM), three times that of the previous generation Ironwood.

Both chips are scheduled to be officially supplied externally later in 2026.

Alphabet’s CEO, Sundar Pichai, mentioned in his blog post that the architecture is designed to "deliver massive throughput and low latency needed for running millions of agents simultaneously in a cost-effective manner." The increase in on-chip storage means the chip doesn't need to frequently fetch data from external sources, which is crucial for AI tasks requiring multi-step inference.

The Comprehensive Deployment of AI Agent Platform Directly Targets OpenAI and Anthropic

On the software side, Google has released a suite of complete enterprise AI agent tools to directly confront the corporate market presence of OpenAI and Anthropic.

According to Bloomberg, several startup founders said that Silicon Valley engineers typically switch between Anthropic’s Claude Code and OpenAI’s Codex for AI