Trainium

in English: AWS Trainium

advancedspreading

Trainium is Amazon's own custom-designed chip tailored for artificial intelligence, which it rents out via its cloud service (AWS) for training and running AI models.

In plain terms

Training large AI models requires massive computation, performed by specialized chips. The market is dominated by Nvidia's graphics processors, which are expensive and often hard to obtain. Amazon therefore designed its own chip, called Trainium.

It is like a large bakery chain that, instead of buying an oven off the shelf, has one built to suit its own breads. The resulting oven can do less than a general-purpose one, but it performs that one task more cheaply and with less electricity.

You cannot buy Trainium in a store, and companies cannot purchase it either. It operates exclusively in Amazon data centers, and its computing time can be rented via AWS.

Its significance goes beyond the chip itself: this way Amazon is less dependent on a single supplier, and customers receive an alternative. For example, Anthropic performs a significant part of training and running its Claude models on such chips, but it also uses Nvidia and Google chips.

How it works

The chip is designed by engineers at Annapurna Labs, owned by Amazon, while manufacturing is done by an external chip factory.
Trainium is a purpose-built chip: its circuits are specialised for the matrix operations that neural network training and inference consist of, and it is poorly suited to almost anything else.
Amazon connects tens of thousands, even hundreds of thousands, of chips via a high-speed network in massive data centers, allowing them to work together as a single giant computer.
Developers rent such machines on AWS and run their models written in standard frameworks (for example, PyTorch) on them using the AWS Neuron software package.
The customer pays for the used computing time, while Amazon handles cooling, power supply, and maintenance.

An everyday example

When you ask something of Claude, or a coding agent works on your program for hours, it is quite possible that the answer is generated in an Amazon data center on Trainium chips. You see nothing of this, at most you feel that the service has become faster or cheaper.

Who is affected

developerIf you train or run a model on AWS, Trainium may be cheaper than GPU, but you must adapt your code to Amazon's own software environment. → Code

entrepreneur, executiveThe price of AI services largely depends on chip costs, so when selecting a cloud provider, it is worth knowing what hardware lies behind it.

bank customer, investorDeveloping proprietary chips is a key battleground in the competition between tech giants like Amazon and Nvidia, affecting stock prices as well. → Money

everyoneThe price and availability of AI tools depend partly on how much cheap computing capacity is being built behind the scenes.

A common misconception

MythTrainium replaces Nvidia chips, and Amazon no longer relies on them.

In realityAmazon continues to buy and rent out large quantities of Nvidia chips, while Trainium offers an alternative alongside them. Most AI developers today still build on Nvidia hardware and its software.

Limits and risks

It is available only in the AWS cloud, so anyone building their system on it becomes heavily tied to a single provider.
Its software base is younger and narrower than Nvidia's CUDA ecosystem, so porting an existing model may require extra work.
Performance and cost data provided by the manufacturer are also for marketing purposes; actual savings depend on the specific task.
Massive data centers built on Trainium raise significant environmental and energy supply concerns due to their electricity and water demand.

Similar, but not the same

GPU (grafikus processzor) – A GPU is a more general-purpose chip that companies can purchase and install in their own machines, whereas Trainium is designed specifically for AI computation and can only be rented from Amazon.
AWS Inferentia – Inferentia is Amazon's other AI chip family, designed for running already trained models. Trainium was originally intended for training, but is now widely used for inference as well.
Google TPU – The TPU is Google's own chip for similar purposes, available primarily in Google Cloud, not in Amazon's.

Related concepts

Broader

AI chip, AI accelerator, Cloud service, computing capacity

Narrower

Trainium2, Trainium3

On the edge

GPU (grafikus processzor), Google TPU, AWS Inferentia, Project Rainier, Data center, Model training

A brief history

Amazon announced Trainium at the re:Invent conference at the end of 2020, designed by Annapurna Labs, the Israeli company acquired in 2015. The first generation became available for rental in 2022. The second was unveiled at the end of 2023 and became widely available from the end of 2024. The third was announced at the end of 2024 and became available at the end of 2025. Around the same time, at the end of 2025, Amazon also stated that it is already working on the fourth generation.

Where things stand

Several generations of Trainium operate at scale in Amazon data centers, with Anthropic being the largest user. Nvidia still holds the majority of the market, while Trainium is one of the most important challengers.

Status: 2026 November

Also known as

Trainium · AWS Trainium · Amazon Trainium · Trainium chip · Trainium2 · Trainium3 · Amazon AI-chip · Amazon MI-chip

Sources

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