Blackwell

in English: Nvidia Blackwell

advancedwidespread

Blackwell is Nvidia's chip generation unveiled in 2024, on which a significant portion of today's major AI systems train and run in data centers.

In plain terms

Blackwell is not a program or an AI model, but hardware: a generation of Nvidia graphics processing units (GPUs). These chips perform the massive calculations required for chatbots, image generators, and AI agents.

Think of it as a new generation of engines in the automotive industry. The passenger does not see the engine, but it determines how fast the car is, how much fuel it consumes, and the cost of a trip. Similarly, the chip determines how quickly an AI service responds and how much its operation costs.

Compared to its predecessor, the Hopper generation, Blackwell performs more calculations with the same amount of power. The data center version is often sold by Nvidia not as individual chips but as entire liquid-cooled server racks. The 2025 GeForce RTX 50 graphics cards are also built on the same chip design (architecture), but in the news, the name almost always refers to the data center version.

How it works

A single datacenter Blackwell GPU contains more than 200 billion tiny switches (transistors) and performs thousands of simple calculations simultaneously. AI models require exactly this kind of parallel work.
The chips are assembled with large amounts of very fast memory so that huge models can fit and data does not get stuck along the way.
Several dozen such chips go into one server rack, 72 in the best-known configuration. A fast internal network connects them, so they work almost like a single giant computer.
The cabinets generate so much heat that they are cooled by liquid instead of air. In the largest data centers, thousands of such cabinets can stand, making power supply and cooling a separate engineering task.
Cloud providers and AI companies use this capacity to train and serve models. You only see the result: the answer on your screen.

An everyday example

When you ask a chatbot to summarize a long contract, your request runs to a distant data center. The answer is likely calculated in seconds by Blackwell chips. In economic news, you encounter the name when Nvidia's quarterly results are explained by demand for Blackwell.

Who is affected

bank customer, investorNvidia's stock price and the entire AI investment wave are largely driven by how many such chips major technology companies order. → Money

entrepreneur, executiveThe price and availability of AI services depend partly on how much and what kind of expensive computing capacity is available to providers.

developerThe type of cloud-rented machines determines what size model you can run or train, how fast, and at what cost. → Code

everyoneSuch chips underlie the speed of everyday AI tools and the power consumption of data centers.

A common misconception

MythBlackwell is an artificial intelligence, meaning a smart program like ChatGPT.

In realityBlackwell is only the hardware on which such programs run; by itself it knows nothing, as intelligence is provided by the models and agents running on it.

Limits and risks

Extremely expensive: the price of a full server rack is measured in millions of dollars, so it is mainly purchased by large cloud providers and well-funded AI companies.
Its high power and cooling demand places a significant burden on data centers and the energy grid.
Due to US export restrictions, chips cannot be freely shipped to every country, making them a geopolitical issue as well.
It becomes obsolete quickly: Nvidia releases a new generation roughly every year, so a top-tier machine today will be mid-range in a few years.

Similar, but not the same

Large language model – The language model is the software that generates text. Blackwell is the hardware on which this software runs.
Nvidia Hopper (H100) – Hopper is the previous chip generation introduced in 2022, on which the 2023–2024 AI boom was built. Blackwell is its stronger and more efficient successor.
Processzor (CPU) – The CPU is general-purpose and performs few tasks quickly one after another. The Blackwell GPU performs countless simple calculations simultaneously, which is what AI needs.

Related concepts

Broader

Grafikus processzor (GPU), MI-chip, computing capacity, Data center

Narrower

Blackwell Ultra, GB200 NVL72, GeForce RTX 50

On the edge

Nvidia, Hopper (Nvidia H100), Vera Rubin (Nvidia Rubin), CUDA, chip export restriction

A brief history

Nvidia introduced the Blackwell platform as the successor to the Hopper generation in March 2024 at the GTC conference. It is named after American mathematician and statistician David Blackwell.

Where things stand

At the end of 2026, Blackwell and its enhanced version, Blackwell Ultra, are among the main workhorses of large AI data centers. Meanwhile, Nvidia has already unveiled its successor, the Vera Rubin platform, with deliveries promised for the second half of 2026.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Nvidia orders cloud clients to sell at least 51% of their services to non-Chinese firmsNov 3, 2026 · Agent Age
Foxconn AI factory runs on Nvidia cluster; Nvidia and Foxconn target eight-week AI factory deploymentOct 26, 2026 · Ipar

Also known as

Blackwell · Nvidia Blackwell · Blackwell-chip · Blackwell GPU · Blackwell platform · B200 · GB200 · Blackwell Ultra

Sources

Wikipedia: Blackwell (microarchitecture)

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