MI450

MI450 is the name of a computing chip made by the company AMD. Such chips are not found in home computers, but in large data centers – that is, in halls full of servers accessed over the internet. Their job is to run AI programs like chatbots or image generators. To do this, they must perform enormous numbers of simple calculations simultaneously, rather than a few complicated ones one after another. AMD plans to ship the MI450 starting in the year 2026. It is the company’s most important attempt to win customers away from competitor Nvidia.

AMD’s assault on Nvidia’s dominance

The market for AI chips is one of the most lucrative in the tech industry. Nvidia holds a share there that, depending on the estimate, lies between 80 and 90 percent. This dominance comes at a price: anyone who needs AI computing power must buy on Nvidia’s terms. Large customers are therefore actively searching for a second source.

This is exactly where the MI450 comes in. Unlike previous AMD chips, it’s not just about computing speed per chip. AMD promises a complete server-rack system that is directly comparable to Nvidia’s offering. This matters because modern AI models never run on a single chip, but on thousands linked together.

Expectations for the chip are therefore high and are reflected on the stock market. OpenAI, the company behind ChatGPT, announced an agreement in 2025 for six gigawatts of AMD computing power – with the MI450 as the first stage of expansion. Such announcements sometimes move AMD’s share price by double-digit percentages in a single day.

Memory, bandwidth, and entire server racks

With AI chips, it is often not computing performance but memory that is decisive. A large language model consists of billions of stored numerical values. These values must flow through the chip with every response. If the path between memory and processing unit is too narrow, the chip sits idle – experts call this a memory bottleneck.

The MI450 therefore uses HBM4, a new generation of stacked high-speed memory. Instead of sitting next to the chip, the memory modules lie in layers directly beside it on the same carrier substrate. AMD cites figures of around 432 gigabytes of memory per chip. That is more than Nvidia’s comparable generation provides for, and a deliberately positioned advantage.

Then there is the interconnection between chips. In the system called Helios, several dozen MI450s sit in one rack, connected via fast links. For the customer, the rack then behaves like a single, very large computer. The whole setup is liquid-cooled, because such a rack can draw over a hundred kilowatts of power – as much as many dozens of households.

Where the chip appears in the news

You will never buy an MI450 directly. You encounter it indirectly when using AI services in a browser: a chatbot’s answers are generated on such chips in a data center. As a user, you don’t notice which manufacturer is behind it.

In business news, however, the name appears frequently. It comes up in AMD’s quarterly reports, in reports about new data centers, and in analyses of the competition with Nvidia. This often concerns delivery times, since production of such chips depends on just a few factories, above all the Taiwanese contract manufacturer TSMC.

A common misconception is that the faster chip automatically wins the market. Software is equally decisive. Nvidia offers CUDA, a programming environment that has grown over years and to which many developers are accustomed. AMD’s counterpart is called ROCm and is considered less mature. Whether the MI450 succeeds therefore depends just as much on this software as on the hardware itself.