B300

The B300 is a computer chip made by the American company Nvidia. It is not intended for ordinary computers, but for enormous data centers, meaning halls full of networked computers. There it performs one very specific task: it carries out the immense number of computing steps that programs like ChatGPT require. Such programs consist at their core of billions of numbers that are calculated together with every request. An ordinary processor does this one after another; the B300 performs tens of thousands of these calculations simultaneously. The name belongs to Nvidia’s Blackwell chip family, named after the mathematician David Blackwell. The B300 is the revised version of the one-year-older B200 and is often also called Blackwell Ultra.

Why a single chip moves stock prices

Anyone who wants to train a large AI model today can hardly avoid Nvidia. The company holds a market share of well over eighty percent in these specialized chips. That’s why the B300 is not an ordinary component, but a bottleneck for an entire industry. When Nvidia delivers, Google, Microsoft, Meta and OpenAI can expand their data centers. When it doesn’t, their plans get pushed back by months.

Then there’s the price. A single chip of this class costs a five-figure dollar amount depending on the configuration. Large operators don’t buy hundreds of them, but hundreds of thousands. That’s why chip names like B300 regularly appear in business news, even though hardly anyone ever holds such a device in their hands. Order volumes are considered a leading indicator on the stock market for how much money is currently flowing into AI.

A common misconception: the B300 does not make AI smarter. It makes it faster and cheaper to operate. A model’s capabilities depend on its training and data, not on the chip. The chip only determines how long that training takes and what it costs.

What’s different on the inside compared to the predecessor

The most important difference from the B200 is memory. The B300 has around 288 gigabytes of so-called HBM memory, meaning particularly fast working memory located directly next to the compute core. The B200 came with about 192 gigabytes. More memory means: a larger AI model fits completely on a single chip instead of having to be spread across several. This saves time, because the chips need to communicate with each other less.

A second trick is the precision of the numbers. AI models don’t calculate with exact decimal numbers, but with deliberately coarser ones. The B300 supports especially economical number formats with only four or eight bits per value. Roughly speaking: you round more aggressively and gain speed as a result. For AI, this level of precision is usually sufficient, because small rounding errors barely change the overall result.

The B300 is rarely sold individually. What’s common are ready-made server racks like the GB300 NVL72 system, in which 72 of these chips are connected together. Such racks consume over a hundred kilowatts of power, as much as several dozen households. That’s why they are cooled with water instead of fans.

The B300 in the news and in everyday life

You never encounter the chip directly. But indirectly, constantly: every chatbot answer, every generated image and every automatic translation is computed on chips like this. All you notice is the wait of a few seconds.

In the news, the B300 mainly appears in three contexts. First, in Nvidia’s quarterly results, where demand for the Blackwell lineup determines the revenue forecast. Second, in export restrictions, since the US only allows the sale of such chips to China in a limited way. Third, in announcements of new data centers, where the number of chips ordered serves as a measure of the project’s size.

Nevertheless, there is competition. AMD builds similar chips with its MI series, and Google uses its own components called TPUs. So far, this has changed little about Nvidia’s lead, mainly because of the CUDA software that developers have grown accustomed to over the years. The successor to the B300 has already been announced and is called Rubin.