A TPU, a tensor processing unit, is Google’s custom chip for large matrix math.
By Tony Yu / tkal · November 5, 2026 · One-minute visual guide
Perfect waffles, no soup~
Large matrix math ♡
↓
Good fitmodels that train for weeks or months
vs
Not suitedhigh-precision arithmetic, frequent branching
A TPU is shaped around one kind of math.
A simplified picture: the waffle iron is Mochi’s analogy, and the chip details come from Google Cloud.
Keep this bitA TPU is a tensor processing unit: Google’s custom chip, designed for the large matrix operations often found in machine learning.
Why care?
Only if you train huge models. Google Cloud points TPUs at models that train for weeks or months, and you reach them through Compute Engine, Google Kubernetes Engine and Gemini Enterprise Agent Platform. If you only chat with a bot, you never pick this hardware.
The limitation
One job means other jobs go elsewhere. Google Cloud lists workloads that require high-precision arithmetic as not suited, and linear algebra programs that require frequent branching too. Like a waffle iron: lovely with batter, no use for soup.
Good fit
Not suited