Google revealed two [1] new eighth-generation [2] Tensor Processing Units, the TPU 8t and TPU 8i, during its Cloud Next 2026 event.

The launch marks a strategic effort to reduce reliance on external hardware providers and increase the speed of generative AI deployment. By developing specialized silicon for both training and inference, Google seeks to optimize how AI agents operate at scale.

The announcement took place during the annual conference held April 23-24, 2026 [3]. The company split the new hardware into two distinct models to address different stages of the AI lifecycle. The TPU 8t is designed specifically for training large-scale models, while the TPU 8i focuses on inference, the process where a trained model applies its knowledge to new data.

Industry analysts said that these chips are intended to keep Google competitive as competition with Nvidia intensifies. Beyond raw performance, the new architecture aims to improve the environmental footprint of AI workloads, which are notoriously energy-intensive.

This hardware shift comes as Google pushes its Gemini enterprise agent platform. The integration of eighth-generation [2] silicon allows the company to make generative AI systems more efficient rather than simply increasing their size. This approach focuses on the quality of processing, and the reduction of latency for end users.

The TPU 8t and 8i represent the latest iteration in Google's long-term strategy to vertically integrate its software and hardware stacks. By controlling the chip design, Google can tailor the hardware to the specific mathematical requirements of its latest transformer-based models.

“Google revealed two new eighth-generation Tensor Processing Units, the TPU 8t and TPU 8i.”

The introduction of bifurcated chips for training and inference suggests that Google is moving away from general-purpose AI acceleration toward highly specialized workloads. This strategy allows for greater energy efficiency and lower operational costs, which is critical as the industry faces scrutiny over the massive power demands of data centers.