Cerebras WSE-3: Why Turn an Entire Wafer into a Single AI Chip?

language

eng

date

Aug 12, 2026

slug

cerebras-wse-3-wafer-scale-engine

author

status

Public

category

🤖 Computer Science

updatedAt

Aug 11, 2026 09:00 PM

In June 2026, South Korea's KOSPI index broke through 9,000. It has since fallen back into the 6,000s, but this was still a market few could have imagined just a year earlier, when the index was dismissed as the ‘Boxpi’ stuck around 2,000. The rally was led by semiconductor heavyweights such as Samsung Electronics and SK hynix, with DRAM at the center. Memory has become one of the most fundamental components of the AI era. Two major constraints we run into with LLMs every day are context length and speed. In tools such as Codex and Claude Code, context length is related to how much information a model can handle in a single session, while speed refers to how many tokens can be processed and generated per second. Of course, both are shaped not only by hardware memory but also by model architecture and software. At a time when speed matters this much, Cerebras has become one of the most compelling chips for inference—even if calling something at this scale a mere ‘chip’ feels a little strange.

Cerebras has signed a multi-year, 750 MW agreement with OpenAI and is collaborating with AWS on inference infrastructure, establishing itself as a major success in AI infrastructure investment. Why was Cerebras chosen from among so many NPU companies? Cerebras overturns conventional assumptions with the WSE (Wafer Scale Engine), connecting 84 reticle fields across a single wafer and using them as one enormous chip. The advantage of this structure is clear. According to Cerebras, its SRAM-based design delivers 1,000 to 2,000 times the effective memory bandwidth of an NVIDIA B200: 44 GB of SRAM and 21 PB/s of bandwidth. By dramatically reducing the memory-bandwidth bottleneck, it enables extremely fast inference. To get there, it accepts other trade-offs. SRAM cells are already approaching their scaling limits, so continuing to expand SRAM capacity on-chip is not easy. External I/O and system-level scaling remain difficult problems as well. And because the entire wafer is used as a chip, power delivery, cooling, and packaging all require custom designs.

Manufacturing takes the company down an equally difficult path, because wafer-scale systems do not easily benefit from conventional economies of scale. Defective dies appear even on ordinary wafers—and Cerebras uses the entire wafer as a chip? It turned what sounds like an absurd idea into reality. To avoid discarding a whole wafer because of a small defect, Cerebras disables defective cores among roughly 970,000 physical cores and routes around them with spare cores and redundant communication links. This leaves 900,000 active cores in the shipping product. The power-delivery and cooling modules must also be custom-built, while assembly and packaging require specialized processes. By any conventional measure, the conditions look almost impossible. Most ordinary ways of thinking would never have attempted it, but Cerebras did. Just how serious are they about fast tokens?

While countless NPU companies were betting on power efficiency and edge deployment in a fiercely competitive market, Cerebras broke through where competition was thinner. Anyone who has used Codex fast mode alongside regular Claude Code knows how much token speed matters. When Claude Code is slow, you give it a task, switch to something else, and then have to return later—raising the user's context-switching cost. At exactly the right moment, Cerebras packed in a massive amount of SRAM and secured a clear position in low-latency inference.

The strategy Cerebras demonstrates is worth considering in other fields as well. The objective is simple, and the company obsesses over it. The implementation is intuitive, yet everyone assumes it is impossible or irrational. Cerebras identifies every bottleneck and obstacle created by that implementation and solves them one by one. There is something awe-inspiring about turning an approach that seems impossible at first glance into reality.

What comes next for Cerebras? First, consider its competitors. NVIDIA is also moving aggressively into SRAM-based machines. In December 2025, NVIDIA did not acquire Groq; it entered into a non-exclusive inference technology licensing agreement, and Groq's founder and several key employees joined NVIDIA. Groq itself and GroqCloud continue to operate independently. Like Cerebras, Groq builds SRAM-based machines, but at the die level rather than the wafer level. NVIDIA's Groq 3 LPX scales by connecting multiple LPUs through C2C links and combining them with Vera Rubin GPUs. Cerebras would naturally welcome SRAM-cell scaling that allows more SRAM to fit on a wafer, but SRAM density scaling is likely to remain a bottleneck for some time. Another possible direction is expanding SRAM through wafer-to-wafer hybrid bonding. This is not an official Cerebras roadmap; it is my own speculation based on the current architecture. The technical and manufacturing challenges would be considerable, but it appears to be one possible path.

Cerebras feels like a pioneering ship setting sail in a different direction from everyone else in the Age of AI Exploration. It is navigating countless risks and adventures toward waters no one has crossed before. Will it ultimately reach a new continent?

Image source: Cerebras WSE-3 official page

Cerebras WSE-3: Why Turn an Entire Wafer into a Single AI Chip?

Related posts

My First PCB Prototype with AI: The Connector Was Harder Than the Design

ThunderScope: The Open-Source Oscilloscope That Moves the Display and Processing to a PC

What does the CPU think? A great introductory article on building a 16-bit CPU directly in C.