Blackwell

Blackwell is the name of a chip generation from the American company NVIDIA. It concerns special chips for artificial intelligence, that is, for programs that learn patterns from huge amounts of data. Such chips don’t compute particularly cleverly, but perform extremely many simple calculations simultaneously – and that is exactly what AI systems need. NVIDIA names every new chip generation after a person from science; in this case it was the mathematician David Blackwell. The generation was unveiled in March 2024 and delivery began at the end of 2024. When business news talks about supply bottlenecks for AI chips, it’s usually about Blackwell.

Why a chip generation moves stock prices

NVIDIA today sells more computing power for AI than any other provider. Its market share in chips for training large models has been above 80 percent for years. Blackwell was the product with which the company aimed to defend this position. That is why every figure about it is closely watched by analysts and investors.

The customers are few, but very large companies: Microsoft, Amazon, Google, Meta and a handful of specialized data center operators. They don’t order individual chips, but entire server racks for double-digit million-dollar sums. A single Blackwell system of the GB200 NVL72 series fills a whole rack and costs around three million dollars. Anyone wanting to build an AI model on the level of the best-known chatbots today needs thousands of them.

For this reason, demand for Blackwell is considered an early indicator for the entire AI market. If orders decline, that suggests the big corporations are scaling back their investments. Conversely, every report of sold-out capacity in 2024 and 2025 was read as evidence of continued growth. The chip has thus become an economic metric, not just a technical product.

Two chips in one housing

The most striking feature of Blackwell is its construction. Until now, such a chip consisted of a single piece of silicon. Blackwell consists of two pieces sitting closely side by side, coupled with a very fast connection. Externally, they behave like a single chip. The reason is a physical limit: with today’s machines, a piece of silicon cannot be exposed larger than about 800 square millimeters.

The second major lever is the precision of the numbers. AI models compute with billions of numerical values, and these values don’t need to be especially precise. Blackwell can compute with as little as four bits per number, meaning sixteen distinguishable levels. That sounds crude, but it is sufficient for many tasks. The fewer digits a chip has to process, the more calculations it can perform per second and the less power it consumes.

Added to this is the connection between the chips. A large AI model doesn’t fit into a single chip but is distributed across many. These must constantly exchange intermediate results. With Blackwell, NVIDIA couples up to 72 chips so tightly that, for the software, they behave like a single giant computing unit. A fitting comparison is a large kitchen: the fastest cooks are of little use if the ingredients travel too slowly from station to station.

Blackwell in the news and in everyday life

You won’t be buying Blackwell directly. The chips sit in data centers and are rented out by the hour. Anyone using an AI chatbot or generating an image is thus using them indirectly. Companies, too, mostly just rent computing time rather than buying their own hardware.

In the news, the name mainly appears in three contexts. First, in NVIDIA’s quarterly results, where Blackwell’s revenue and ability to deliver move the stock price. Second, in export restrictions: the US does not allow the sale of the most powerful variants to China, which is why weakened special models exist. Third, in power consumption, since a server rack full of Blackwell chips draws more power than a small residential building.

A common misconception is to confuse Blackwell with a graphics card for video games. The technology shares common roots, but the data center variants have no display outputs and are designed solely for AI computations. Another misconception is to regard a chip generation as a long-term standard. NVIDIA introduces a new architecture roughly every year; Blackwell is already being followed by the next one, named Rubin.