TPU (tensor processing unit)
Google's custom AI accelerator, designed for its own machine-learning workloads and offered to cloud customers through Google Cloud.
A TPU is an accelerator Google designed to run the matrix operations used in neural networks. Google uses TPUs internally and rents them to customers through Google Cloud. Successive generations have been offered in different configurations, and some are built with outside manufacturing and packaging partners.
Traders care because TPU volumes are a read-across for custom AI chip demand and for Google's capital spending and cloud revenue. News about external TPU customers can shift expectations for merchant GPU suppliers. Treat claims about volumes or customers as estimates until the company or a filing confirms them.
Worked example (illustrative numbers, not real data): a customer runs 2,000 accelerator-hours at an illustrative price of 2 dollars per hour, for a bill of 4,000 dollars. The same hours on a GPU at an illustrative 3 dollars per hour would cost 6,000 dollars.