Google Cloud has introduced its eighth generation of custom-built tensor processing units (TPUs), dividing them into two: the TPU 8t for model training and the TPU 8i for inference. These chips promise up to three times faster model training, an 80% better performance per dollar, and the ability to integrate a million TPUs in one cluster, potentially reducing energy use and costs.

However, Google is not planning to fully replace Nvidia’s GPUs on its cloud infrastructure. Instead, it will continue to offer Nvidia-based systems alongside its custom chips. This dual approach suggests that while Google aims to enhance its AI capabilities, it still recognizes the value of Nvidia's offerings.

The move reflects a broader trend among tech giants like Amazon and Microsoft who are developing their own AI chips but still rely on Nvidia for certain tasks. While some analysts predict this could eventually reduce Nvidia’s market share, for now, it appears that the GPU maker will continue to thrive in the cloud computing landscape.

Interestingly, Google is also working with Nvidia to improve software-based networking technology, Falcon, which aims to boost efficiency within its cloud infrastructure. This collaboration shows that even as tech titans develop their own solutions, they still recognize the value of partnering with industry leaders like Nvidia.