GB300

GB300 is the designation for a computing module from the company Nvidia. It is not a device one sets up at home, but a component for enormous data centers. Such modules take on the computational work behind programs like ChatGPT. The name is made up of two parts: a main processor called Grace, which controls the processes, and graphics chips called Blackwell, which do the actual heavy computing. Both sit tightly connected on a shared board. The series was unveiled in 2024 and 2025 as the successor to the GB200 and H100 models.

Why a chip became a stock market topic

Anyone who wants to train a large AI model today needs thousands of such computing modules. Nvidia has been the dominant provider of this type of hardware for years. As a result, a substantial part of the AI boom hinges on a single product line. When Nvidia announces a new generation like GB300, stock prices react across the entire industry.

A single module costs a five-figure sum depending on configuration. A fully equipped server rack runs into the millions. Companies like Microsoft, Amazon, Google, and Meta buy tens of thousands of units of these. These very expenditures show up in quarterly earnings as so-called capital expenditures, meaning money spent on long-lived acquisitions.

Power consumption is also important. A rack of GB300 modules draws more than a hundred kilowatts, as much as several dozen households. That is why the availability of electricity and cooling water now also plays a role in deciding where new data centers are built.

Grace and Blackwell working together

A normal computer has a main processor and, next to it, connected via a slow slot, a graphics card. In the GB300, the main processor and graphics chips sit directly next to each other and share a very fast data pathway. The advantage: data does not have to be constantly copied back and forth. With AI models containing hundreds of billions of numerical values, this exact shuffling back and forth is often the real bottleneck.

On top of that comes a special kind of working memory that is stacked directly on the chip. It delivers data many times faster than the memory in a laptop. The computing units themselves work deliberately imprecisely: they calculate with heavily shortened numbers. For AI this is sufficient, and it doubles the speed compared to more precise computation.

On its own, a single module remains too small for the largest models. That is why 72 graphics chips are connected via an in-house network called NVLink into a single rack. To the software, this rack then looks almost like one single, very large computer. This design is named GB300 NVL72.

Where the name shows up in the news

One practically never encounters a GB300 directly. It is used indirectly whenever one asks an AI chatbot a question, has an image generated, or requests a translation. The request travels over the internet into a data center, gets computed there on such hardware, and comes back as a response.

In business news, the name usually appears in three contexts. First, in Nvidia’s quarterly earnings, when it comes to demand and delivery times. Second, in announcements of large data centers, for instance when a company reports buying a hundred thousand chips. Third, in export controls, because the United States restricts the sale of the most powerful AI chips to China.

A common misconception is to confuse the GB300 with a gaming graphics card. Both come from Nvidia, but their purposes differ: a gaming card is meant to display images on a monitor, whereas a GB300 has no display connector at all. It is a pure computing machine for data centers.