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.

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

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

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

We made these chips programmable via TensorFlow, allowing models to run on CPUs, GPUs, or TPUs with minimal modifications.

Google's TPU architecture is optimized for internal workloads, such as Search and YouTube, which means their TPU purchases are primarily for their own applications rather than for external sales.

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 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.

Azure is the first cloud to scale Nvidia's GB200 supercomputer, achieving the highest throughput for any cloud platform with 865,000 tokens per second.