H100

in English: Nvidia H100

advancedwidespread

The Nvidia H100 is a high-performance AI chip designed for data centers, on which a significant portion of large language models similar to ChatGPT are trained and run.

In plain terms

The H100 is Nvidia's accelerator card, specifically designed for training and running artificial intelligence. It is not a graphics card for your computer: it operates in server rooms, connected by the thousands.

Imagine it as the giant oven of an industrial kitchen. At home, you bake one tray of cake; here, thousands of trays go in at once, day and night. A standard processor is a skilled chef who knows how to do everything but can only do a few things at once. The H100 is an entire brigade repeating the same simple move in parallel at an incredible pace.

It became famous because when generative AI exploded in 2023, this was one of the world's most sought-after components. For a long time, the strength of companies was measured by how many H100s they could acquire, and the chip became a kind of unit of measurement for the AI race.

How it works

Training and using an AI model consists of many simple multiplications and additions that can be performed simultaneously in parallel.
The H100 executes these operations simultaneously across its thousands of small computing units, making it orders of magnitude faster than a traditional processor for such tasks.
The card has 80 gigabytes of very fast memory, which holds the model and the data currently being processed.
Typically, eight such chips are installed in a server, and the servers are connected by a high-speed network, so thousands of H100s work as a single giant computer.
A single chip can consume up to 700 watts, so large deployments require significant power supply and cooling.

An everyday example

When you ask a chatbot a question, your request travels to a remote data center, where an H100 or one of its successors most likely computes the answer. When the news says a company is building a "hundred-thousand-GPU" data center, for a long time that meant chips like these.

Who is affected

developerIf you train or run your own model in the cloud, you often rent H100 hours, and the price of this determines your costs. → Code

entrepreneur, executiveThe price and availability of AI services ultimately depend on how many such chips the provider can access.

bank customer, investorNvidia's stock surge and the AI investment frenzy were largely triggered by demand for the H100. → Money

lawyer, citizenSince 2022, the United States has required licenses for chip exports to China and a few other countries. The rules have changed several times since then, and the chip has become a key player in technological and geopolitical disputes. → Law

A common misconception

MythThe H100 is a very powerful graphics card that can be used for gaming or running AI at home.

In realityData-center hardware: it has no monitor output, is designed for server racks and industrial cooling, and cost tens of thousands of dollars per unit when new.

Limits and risks

No longer at the peak: Nvidia's newer generations (Blackwell, Rubin) are much faster and perform more computations with the same power.
It requires significant electricity and cooling, so the environmental impact and electricity bill of large deployments are substantial.
It is expensive, and in some countries it can only be shipped with an American export license, so not everyone can obtain it freely.
The number of chips alone says little: the result also depends on the software, the data, and the network.

Similar, but not the same

Processzor (CPU) – The processor is general-purpose: it solves many types of tasks but only a few at a time. The H100 mainly performs one type of simple calculation, but does so massively and in parallel.
gaming graphics card (GeForce RTX) – The gaming card is built for home computers and graphics rendering; the H100 is for servers and continuous AI computation.
Nvidia Blackwell – Blackwell is the new, more powerful chip generation after the H100, not a variant of the same product.

Related concepts

Broader

GPU, AI chip, Data center, computing capacity

Narrower

H800, DGX H100

On the edge

Nvidia, CUDA, H200, Nvidia Blackwell, Model training, export restriction

A brief history

Nvidia introduced the H100 in March 2022 as the first member of the Hopper chip family; the family is named after Grace Hopper, an American computing pioneer. During the generative AI boom in 2023, it became a scarce item, with waits of several months.

Where things stand

At the end of 2026, the H100 is no longer the latest generation, but it is still widely used in data centers. Its rental price has fallen significantly from the peak, and it is mainly used for running models and smaller training jobs.

Status: 2026 November

Also known as

H100 · Nvidia H100 · NVIDIA H100 Tensor Core GPU · H100 GPU · H100-as chip · H100 accelerator · Nvidia H100 AI-chip

Sources

NVIDIA: H100 Tensor Core GPU (official product page)

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