Google debuts eighth-gen TPU chips for training and inference with big performance claims

When we talk about AI chips, the first name that often comes to mind is Nvidia. However, Google has also made a significant impact in the AI infrastructure space with the launch of its homegrown Tensor Processing Unit (TPU). The search engine giant has now introduced its eighth-generation TPUs at Cloud Next ‘26, designed for both training and inference.

As the company explains, TPU 8t and TPU 8i are the latest additions to the TPU lineup. Unlike most conventional AI chips that handle both training and inference, Google has adopted a dual-chip approach. TPU 8t is specifically designed for training AI models. Google says a single superpod can scale up to 9,600 chips, along with 2 petabytes of shared high-bandwidth memory (HBM) .

Meanwhile, TPU 8i is optimized for inference and AI task execution. According to Alphabet CEO Sundar Pichai, TPU 8i chips are able to “deliver the massive throughput and low latency needed to concurrently run millions of agents cost-effectively.” Google states that TPU 8i is designed for near-zero latency inference and can deliver up to 80% better performance per dollar compared to previous generations.

TPU 8t and TPU 8i are more powerful than their predecessor, Ironwood. As a result, supplying sufficient power to them in data centers could be challenging. To address this, Google says both chips feature integrated power management that can dynamically adjust power consumption based on real-time demand. Thanks to this capability, Google’s new TPU chips claim to deliver up to twice the performance per watt compared to the Ironwood chips.

Unlike Nvidia, which sells its AI chips to a wide range of companies and organizations, Google often reserves its TPUs for its Cloud customers seeking an alternative. At the same time, Google maintains close ties with Nvidia and has announced that it will be among the first to offer NVIDIA Vera Rubin NVL72 systems.