Hopper

Hopper is a blueprint for computer chips. The company Nvidia developed it and brought it to market in 2022. Such chips are not found in phones or laptops, but in large data centers, meaning halls full of computers that are rented out over the internet. There they perform the computing work behind programs like ChatGPT. The best-known chip with this blueprint is called the H100. The name Hopper is a reference to Grace Hopper, an American computer scientist of the 20th century.

Why the H100 became scarce

When the boom around language-capable AI programs began in 2022 and 2023, all major providers suddenly needed the same thing: very many chips that compute very fast. Hopper was, at that time, the fastest product available on the market. Microsoft, Meta, Amazon, and Google ordered in quantities of tens of thousands. A single H100 cost around 30,000 dollars at times.

Demand exceeded what the factories could deliver. Those who hadn’t ordered early waited months. For start-ups, access to these chips became more important than money or good ideas. Some investors bought computing capacity and offered it to the companies they invested in.

For Nvidia, Hopper was the reason for an unprecedented rise. Revenue multiplied within two years, and the company’s stock market value rose above three trillion dollars. That’s why the term also appears in business news, not just in technical articles.

What sets Hopper apart from a normal processor

An ordinary processor in a laptop has few, very flexible computing cores. It processes tasks one after another and can handle almost anything. A Hopper chip, by contrast, has thousands of simple computing units that work simultaneously. They can only handle a few types of calculations, but do so massively in parallel.

This fits exactly with what AI models need. Their computing work consists almost entirely of multiplications of large tables of numbers. This is precisely what Hopper has dedicated special components for, the so-called Tensor Cores. In addition, the chip can compute with especially coarsely stored numbers, which increases speed without noticeably worsening the result.

A single chip still isn’t enough for large models, though. That’s why hundreds of Hopper chips are connected together via fast cable links and work on the same task. Memory is often the bottleneck here: it limits how large a model can be. The successor, the H200, mainly received more memory for this reason, not more computing cores.

Hopper in news and successor models

You practically never encounter Hopper directly. Anyone typing an AI question into a chat window is nevertheless using these chips, just remotely in a data center. Cloud providers also rent out computing time on H100 chips by the hour, which is mainly used by research groups and companies.

The term appears in the news in two contexts. First, in Nvidia’s quarterly earnings, where demand for Hopper explains the revenue. Second, in politics: the United States has heavily restricted the export of the most powerful Hopper chips to China. In response, Nvidia developed weakened special variants for that market.

By now, Hopper is no longer the newest. Since 2024, there has been the successor architecture Blackwell, followed by further generations on a two-year cycle. A common misconception is that older chips become worthless as a result. In fact, Hopper cards keep running in data centers for years to come, because demand for computing power grows faster than supply.