While GPUs (like Nvidia H100s) are the industry standard, Google designs its own AI accelerators called TPUs (Tensor Processing Units).
TPUs are explicitly designed for the matrix multiplication operations required by neural networks. They offer massive cost-performance benefits, particularly for training large foundational models.
The latest 8th-generation TPUs (announced April 2026) are split into two specialized variants:
These are interconnected via the new Virgo Network fabric, designed for high-performance AI cluster scaling with Managed Lustre storage delivering up to 10 TB/s throughput.