Definition
A TPU is a custom Google chip built to run the heavy math behind AI faster and more cheaply than ordinary processors.
At a glance
- A custom Google chip purpose-built for AI, not a general-purpose brain like your laptop’s CPU.[1]
- It does one thing fast and efficiently: the large matrix (tensor) math behind machine learning.
- You rent TPUs through Google Cloud rather than buy them — AI computing as a service.
How it works
A CPU is a generalist; a TPU is a specialist that does only AI math, but does it very fast and on far less electricity.[2] Google’s early TPUs delivered many times the performance-per-watt of standard chips.[4] They run inside Google’s data centers, powering both AI training and everyday use.
TPU vs GPU
GPUs (mostly NVIDIA) are the flexible all-rounder: available on every cloud with the widest software support.[3] TPUs can be cheaper and faster for the right workload, but only run on Google Cloud — flexibility versus savings.
Bottom line
TPUs can be cheaper and faster for big, repetitive AI work, as long as you’re willing to build on Google Cloud — a commercial choice, not a technical one.