TPU

in English: tensor processing unit (TPU)

advancedwidespread

The TPU is Google's own chip, designed specifically for artificial intelligence calculations, so it performs these tasks faster and with less power than a general-purpose processor.

In plain terms

A TPU is a computer chip designed for a single task: performing the vast number of multiplications and additions required to train and run AI models. Google designs it and uses it mainly in its own data centers.

Imagine it like a pizzeria where the oven can only bake pizza but produces hundreds per hour. A general-purpose processor (CPU) is like a home kitchen where you can cook anything, but more slowly. A TPU is a specialized machine: good for fewer things, but highly efficient in that one area.

The best-known AI chip is the GPU, which was originally designed for video game graphics and later turned out to be excellent for AI as well. The TPU, by contrast, was designed for AI from the start, so for the same work it often uses less power.

You cannot buy a datacenter TPU in a store. You encounter it through Google services, such as Gemini, Search, or Translate, which run on such chips in the background, while companies can rent them in Google Cloud.

How it works

An AI model works internally with huge tables of numbers; these are called tensors, hence the chip's name.
TPU circuits are designed to multiply these tables simultaneously in parallel, not one by one in sequence.
Everything unnecessary for AI (such as graphical rendering) has been removed, so more useful computations are produced from the same amount of power.
A single chip is not enough: Google connects thousands of TPUs with a fast network, and together, as one giant machine, they train or run the model.
The developer does not program the chip directly but writes code in an AI framework (such as JAX, TensorFlow, or PyTorch), which the system translates into the TPU's language.

An everyday example

You photograph an Italian menu with your phone, and Google Translate shows it in Hungarian within seconds. The response likely comes from a data center where TPUs performed the necessary calculations. You only see that it was fast.

Who is affected

developerIf you train an AI model or run an agent in the cloud, the TPU can be an alternative to the GPU, with different pricing and software constraints. → Code

entrepreneur, executiveThe price and availability of AI services partly depend on which chips they run on and which cloud provider you are tied to.

bank customer, investorThe market for AI chips is a constant topic in stock market news, and the TPU is one of Nvidia's main GPU challengers. → Money

everyoneThese chips are partly behind the speed of the Google services you use every day and the power consumption of AI.

A common misconception

MythTPU is just another name for GPU, or a more powerful version of it.

In realityThe TPU is a specific chip type: a so-called application-specific integrated circuit (ASIC) designed exclusively for AI computations. It is not better than the GPU in all aspects, but is more efficient in a narrower scope of tasks, in exchange for less flexibility.

Limits and risks

An average company cannot buy a datacenter TPU and install it in its own server room: it is available essentially only through Google Cloud, which ties you to one provider. The only exceptions are a few very large customers with whom Google signs separate agreements.
Most AI software and specialists are set up for GPUs, specifically Nvidia devices, so switching to TPUs may require extra work.
A faster or more energy-efficient chip does not make a model smarter: the quality of answers depends on the model and the data.
Manufacturers' own speed and efficiency data are difficult to compare because everyone measures on tasks that favor them.

Similar, but not the same

GPU (grafikus processzor) – The GPU is more general-purpose, available for purchase by anyone, and suitable for many tasks, whereas the TPU is built solely for AI computation and is mainly available in Google Cloud.
NPU (neural processing unit) – The NPU is a small AI chip inside your phone or laptop, while the TPU is primarily a large-scale accelerator for data centers.
CPU (processzor) – The CPU is the general-purpose component of a computer that runs any program, but in AI's massive calculations, it is much slower than the purpose-built TPU.

Related concepts

Broader

AI accelerator, MI-csip, ASIC (application-specific integrated circuit), computing capacity

Narrower

Cloud TPU, Edge TPU

On the edge

GPU, NPU, data center, model training, inference

A brief history

Google began using the TPU in its own data centers in 2015 and unveiled it to the public in 2016. Since 2018, external companies have been able to rent it via Google Cloud.

Where things stand

The seventh generation (Ironwood) has been widely available for rental since the end of 2025, and in 2026 Google also presented the eighth generation. Google's own models run on TPUs, and other major AI developers, such as Anthropic, also use them in large quantities. GPUs still make up the larger part of the market, but the TPU is their best-known alternative.

Status: 2026 November

Also known as

TPU · tensor processing unit · Tensor Processing Unit · tensor processing unit · tenzorprocesszor · Google TPU · Cloud TPU · TPU-csip

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.