Cerebras

Cerebras Systems is a Silicon Valley company, founded in 2016. It makes computer components specialized for computing artificial intelligence. Normal computer chips are about the size of a thumbnail. Cerebras instead builds a single chip roughly 21 by 21 centimeters — larger than a sheet of A5 paper. This component is called the Wafer Scale Engine, or WSE for short. It is the largest chip that is sold commercially.

The race against Nvidia

The market for AI computing power is dominated by a single company: Nvidia. Its graphics cards sit in almost every large data center running AI systems. This gives Nvidia enormous pricing power and has made the company one of the most valuable in the world. Cerebras is one of the few serious challengers.

For customers, dependence on a single supplier is a risk. During periods of tight supply, companies had to wait months for Nvidia hardware. Alternatives like Cerebras reduce this pressure, even if they only capture a small market share. That is why the company gets more attention in business news than its size would actually justify.

Then there is the financial story. Cerebras has raised billions in investor money and prepared for an IPO. However, a large share of its revenue comes from a single customer in the United Arab Emirates. Among analysts, this dependency is considered the company’s biggest question mark.

Why a single giant chip

Chips are cut from round slices of silicon called wafers. Usually, a wafer is divided into hundreds of small chips. Anyone who needs a lot of computing power then connects many of these chips via cables and circuit boards. This is exactly where the problem arises: sending data between two chips costs time and power.

Cerebras does not cut the wafer apart. The entire wafer remains a single component, with nearly a million compute cores on it. All data stays on the same piece of silicon and never has to take the detour over cables. The memory, too, sits directly next to the compute cores instead of in separate components. For AI models that constantly shuffle huge amounts of numbers back and forth, this is a noticeable advantage.

The price for this is manufacturing. Every wafer inevitably develops tiny defects. With small chips, the broken ones are simply discarded. With a chip that takes up the entire wafer, that’s not possible. That’s why Cerebras builds in reserve cores that replace defective spots. In addition, such a chip needs its own cooling and its own housing — you can’t just plug it into a normal server.

From language models to particle physics

Today, Cerebras comes up most often in connection with inference. This refers to the ongoing operation of an already trained AI model: you ask a question, the model answers. Cerebras advertises that it generates language model answers many times faster than standard hardware. Instead of words trickling out slowly, the text appears almost instantly.

Anyone who wants to try it out themselves doesn’t need to buy a chip. Cerebras rents out computing time over the internet, similar to other cloud providers. Some publicly available chat services run on it, and developers can test the speed directly in the browser.

Beyond that, the hardware is used in research institutions. National laboratories in the US use Cerebras systems for physics simulations, climate models, and the search for new drugs. Incidentally, a common misconception is that Cerebras is an AI model like ChatGPT. The company builds the machines that such models run on — not the models themselves.