It Builds the World’s Largest Chip.
86% of Its Revenue Comes From the UAE.

In December 2025, OpenAI signed a deal with Cerebras: a multi-year agreement worth more than $20 billion, under which 750 megawatts of AI compute would run on Cerebras systems.

And yet 85% of this chip company’s 2024 revenue of $290 million came from a single customer — G42 of the United Arab Emirates. In 2025, on revenue of $510 million, two UAE-linked entities accounted for 86%. One customer, or two. For the maker of the world’s largest AI chip, that is an extraordinarily concentrated customer list.

Even so, the company has the world’s largest chip. Measured against NVIDIA’s newest GPU, the B200, it is 58 times the size, with 2,625 times the memory bandwidth, on a single processor — the largest commercial chip in the history of silicon.

What the S-1 presents is the two faces of Cerebras at once: a company taking on the NVIDIA ecosystem with raw size and bandwidth, and a company with an extraordinarily lopsided customer base.

Ten questions. We lay out only what the S-1 says.

Question 1

What is this company, really?

In one line: a company building the world’s largest chip to take NVIDIA head-on.

An ordinary AI chip (NVIDIA’s B200, for example) is a small processor, hundreds or thousands of which are wired together to train AI. Because the connections between them are physical wires, moving data around takes time.

What Cerebras builds is a single chip measuring 46,225 mm² — 58 times the size of the B200. The industry calls it the “Wafer-Scale Engine” (WSE). The idea is to use an entire silicon wafer (the base disk semiconductors are cut from) as one chip.

Two inventions make this possible.

· Cross-reticle interconnect: normally a wafer is diced into small pieces; Cerebras does not dice it, and during fabrication makes the whole thing communicate as one chip.· Fault-tolerant architecture: turning a whole wafer into one chip inevitably leaves defects, so the design detects them, isolates them, and routes around them.

The latest generation, the WSE-3, has 900,000 cores, 44 GB of on-chip memory, and 21 PB/s of on-chip memory bandwidth. Against NVIDIA’s B200, that is 250 times the on-chip memory and 2,625 times the memory bandwidth (the S-1’s own comparison — note this is on-chip SRAM, not the B200’s total HBM capacity).

Why does sheer size help? Because there are no wires to cross. An AI computation moves hundreds of billions of numbers at high speed. Wire the chips together and that is where it jams. Fit it all on one chip and it does not.

It solves “wires make it slow” through physical size.

Question 2

What does it compete on?

Cerebras competes in three distinctive ways.

Compete on size

NVIDIA’s AI chips are the de facto standard for training and inference worldwide. A software environment called CUDA has accumulated over many years, and developers use CUDA to run their AI. The whole ecosystem assumes thousands of GPUs wired together.

Cerebras refuses to fight this battle head-on — it sidesteps it. If one chip is enough, you need neither GPU-to-GPU communication nor CUDA. Its software environment is a proprietary compiler called CSoft, which compiles directly to the WSE from PyTorch (the standard AI development library).

The S-1 states its edge plainly:

“

many customers have achieved more than 10 times faster training time-to-solution compared to leading GPU systems of the same generation.

CEREBRAS S-1 (APR 2026) — BUSINESS, OUR COMPETITIVE STRENGTHS

The founders’ letter goes further, claiming more than 1,000x faster on some workloads.