Google is committed to both TPUs and GPUs, leveraging both for training and serving AI models, while maintaining flexibility in their infrastructure.

Google has the largest TPU cluster with a very advanced infrastructure for their Tensor Processing Units (TPUs), which are tightly connected across multiple data centers in regions like Nebraska and Iowa, allowing for efficient training of models like GPT-5, but their data centers are spread across multiple locations, complicating multi-datacenter training compared to tightly connected clusters.

Google primarily uses its own TPUs for AI models but also supports Nvidia chips, balancing performance and computational needs.

The TPU program was initiated to meet the demand for high compute power for inference and training in AI applications.

Our seventh generation TPU, Ironwood, delivers 10x the performance of the previous generation and will be available to Google Cloud customers later this year.

Google Cloud will offer a virtual machine with a 180 teraflop TPU version 2 device, running TensorFlow programs quickly.

The software stack at Google is highly optimized for their internal use cases, but it has not been released publicly, limiting external developers' access to the same efficiencies.

Google's infrastructure advantage allows it to deliver AI models at a cost-effective price point, enhancing performance and product quality.

Google has open-sourced the Gemma model, which is designed to leverage the massive matrix-multiply unit of TPUs, resulting in a vocabulary size that is larger than necessary for its model size.