Is a wafer-scale chip the same thing as a GPU?

No. A GPU is one of many identical chips cut from a silicon wafer. A wafer-scale chip keeps the whole wafer intact as a single processor, which is why it’s tens of times larger than any individual GPU die.

How does Cerebras stop one manufacturing flaw from ruining the whole chip?

The chip is built with spare cores and spare routing. Any flawed core found during testing is permanently mapped out and traffic is rerouted around it — a “fail-in-place” design rather than discarding the whole chip.

Do I need to buy Cerebras hardware to use it?

No. Most developers access the chip through Cerebras Cloud, a hosted API that’s compatible with OpenAI’s client libraries. Only large enterprises typically license dedicated on-premises systems.

Is wafer-scale hardware used for training AI models too, or just inference?

Both. Cerebras originally built the Wafer-Scale Engine for training large models faster, and its recent partnerships — like the one with AMD — focus on real-time inference, where its low latency is the bigger advantage.

What Is a Wafer-Scale Chip — and How Is It Different From a GPU?

A wafer-scale chip is a processor built from an entire silicon wafer left in one piece, instead of being cut up into hundreds of individual chips the way a GPU or CPU normally is. The best-known example is made by Cerebras Systems, whose current chip, the WSE-3, is the largest single processor ever manufactured — and it’s built specifically to make AI models respond faster.

What it is

Most chips start life as one of dozens of identical rectangles patterned onto a round silicon wafer — a thin, polished disc of crystalline silicon, typically 300mm (about a foot) across. Once the circuitry is etched on, the wafer is sliced apart, or “diced,” into individual rectangles called dies. Each die becomes one chip: one GPU, one CPU. A wafer that size might yield dozens of separate GPU dies.

Cerebras skips the dicing step. Its WSE-3 keeps almost the entire wafer intact as a single chip, roughly 21.5cm across — about the size of a dinner plate, and, according to Cerebras, around 57 times the surface area of Nvidia’s flagship data-center GPU die. Manufactured for Cerebras by the semiconductor foundry TSMC on its 5-nanometer process, that single piece of silicon holds 4 trillion transistors and 900,000 identical processing cores, wired directly to 44GB of on-chip memory. Cerebras says that adds up to roughly 19 times more transistors than Nvidia’s B200 GPU.

How it works

Building a chip this size sounds like it should be impossible: the bigger a chip is, the more likely it is that at least one microscopic manufacturing flaw lands somewhere on it — normally enough to ruin that one die, so it gets discarded while its neighbors on the wafer ship fine. A flaw anywhere on a wafer-scale chip would seem to ruin the whole thing.

Cerebras’ answer is to design for defects rather than avoid them. A small fraction of the 900,000 cores are spares. During testing, any core or connection with a manufacturing flaw is mapped and permanently routed around, and traffic is rerouted through the working cores nearby — a “fail-in-place” design that keeps the chip running at close to full strength even though not every one of its 900,000 cores is perfect.

The real payoff is memory bandwidth. Running a large AI model means constantly pulling its parameters back and forth between memory and processing cores to generate each word, or “token,” of a response. On a regular AI hardware setup, those parameters usually live in separate memory chips wired to the processor over a relatively narrow connection, and large models often need several GPUs wired together to hold everything — adding more distance for data to travel. Because the WSE-3’s cores and memory sit on the same slab of silicon, Cerebras says it moves data internally at roughly 21 petabytes per second, letting it generate tokens far faster than typical GPU setups for the same model.

Why it matters

That speed shows up directly in how a model feels to use: Cerebras has demonstrated a 400-billion-parameter model streaming answers at roughly 2,500 tokens per second per user — output most GPU-based services can’t match. For a simple chatbot reply that difference is barely noticeable, but for AI agents that chain together dozens of model calls to plan and act, or for coding tools waiting on a long response, latency compounds fast, and a faster chip means a more usable product.

That’s also why wafer-scale and GPU hardware are increasingly paired rather than treated as rivals: a common setup splits an AI query into two stages, letting a rack of GPUs handle “prefill” — reading and processing a long prompt, a job that rewards raw throughput — while a wafer-scale chip handles “decode,” generating the reply token by token, a job that rewards low latency. Nvidia, the dominant maker of AI GPUs, doesn’t own the wafer-scale side of that pairing, which is part of why rival chipmakers keep teaming up with Cerebras instead.

Most developers never touch the physical hardware, tapping into it instead as one option in the broader race for AI compute. Cerebras sells access through a cloud API, documented at its developer portal, that lets anyone plug a model into an existing app. As of July 2026, Cerebras’ pricing page lists a free trial with $5 in starting credits, a self-serve Developer tier from $10, and a custom-priced Enterprise tier for production workloads — with exact per-model, per-token rates listed on the same page. The approach has drawn serious investor interest, too: Cerebras went public on the Nasdaq under the ticker CBRS in May 2026, one of the largest US tech IPOs of the year.

In the news

Cerebras’ hardware is showing up in fresh partnerships as chipmakers look for ways to speed up AI responses — see our report on AMD and Cerebras pairing chips for faster, disaggregated AI inference.

What it is

How it works

Why it matters

In the news

Frequently asked questions