What is Cerebras? The giant AI chip, explained

In plain words

Cerebras is an American company that makes hardware for AI, not a chatbot or a model. It sells computers built around its own chip, and it rents them out as a cloud service.

Chips are normally made many at a time on a round disc of silicon called a wafer, which is then cut into lots of separate chips. A big AI model then needs many of those chips, wired together, constantly passing data between them.

Cerebras does the opposite: it keeps the wafer in one piece. Its chip is a square about 21.5 cm on each side, with around 900,000 small computing cores and fast memory spread across the whole surface. Data stays on one chip instead of travelling between many, which helps AI models write their answers very quickly.

Everyday example

One big hall instead of many small rooms

Imagine a wedding team split across fifty small rooms in different buildings, phoning each other to coordinate every dish. Now put everyone in one huge hall where they can just turn and talk. Same people, far less waiting. Many GPUs wired together are the scattered rooms; a Cerebras wafer is the single hall.

Where the example breaks: One hall is not unlimited: very large models still need several wafers linked together, and a single flaw on such a big chip has to be worked around, not thrown away.

How it actually works

Making a chip the size of a wafer means solving problems normal chips avoid.

Living with defects

Every wafer has a few faulty spots. Normal makers throw away the chips that land on them. Cerebras adds spare cores and spare connections, and the chip routes around faulty areas.

Memory beside every core

The WSE-3 chip has 44 GB of fast SRAM memory spread among its cores. Model weights sit right next to the cores that use them, instead of in separate high-bandwidth memory chips, so less time is spent moving data.

Systems, then clusters

Each wafer is mounted in its own system with power and cooling, such as the CS-3, which fills part of a data-centre rack. Several systems are linked to hold models too big for one wafer.

Sold as speed

Most customers do not buy the hardware. They use Cerebras's cloud for inference, paying for answers delivered at high tokens per second, which suits coding tools, agents and live chat.

Founded

2015, by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe Fricker

Based in

Sunnyvale, California

The WSE-3 chip

Introduced March 2024: 4 trillion transistors, 900,000 cores, 46,225 square millimetres, made by TSMC

Inference cloud

Launched August 2024

Groq, Nvidia and Google's TPU chase the same goal, faster AI, by different routes. See AI accelerators.

Questions people ask

Why is the Cerebras chip so big?

To keep everything on one piece of silicon. Passing data between separate chips is slow and uses energy. By using the whole wafer as a single chip, Cerebras keeps a model's numbers and the cores that use them close together, which speeds up answers.

Is Cerebras better than Nvidia?

They are built for different strengths. Nvidia GPUs are general-purpose AI chips used for both training and running models, with widely used software. Cerebras is built mainly for very fast answers from a model, especially for each individual user. Which suits better depends on the job and the cost.

Who uses Cerebras?

Mostly AI companies and research centres. OpenAI signed a multi-year deal in January 2026 for 750 megawatts of Cerebras systems to give its customers fast answers. Amazon Web Services and the Abu Dhabi AI firm G42 and the university MBZUAI are also major customers.

Is Cerebras a public company?

Yes, since 14 May 2026, when its shares began trading on the Nasdaq stock exchange in New York after it raised about 5.5 billion dollars. Before that it was privately owned by its founders, staff and investors.