Skip to content
C
Core Compute & Comms Hardware · AI Training Chips (Wafer-Scale)

Cerebras

📈 Unlisted🌍 United States

Develops the world's largest AI chip—the Wafer-Scale Engine (WSE)—where an entire wafer is a single chip.

⚔️ Core Moat

The wafer-scale chip (Wafer-Scale Engine) technology route is unique—the entire wafer is not diced but directly used as a single superchip, providing efficiency far exceeding that of traditional GPUs in certain AI training scenarios.

Company Overview

Founded in 2016 by Andrew Feldman and others, Cerebras is headquartered in California, USA. The company bet on a technological path completely different from NVIDIA's: instead of assembling small chips into large systems, it manufactures one single ultra-large chip at once.

Cerebras's WSE-3 (third-generation wafer-scale engine) is the largest chip ever manufactured—with an area of 46,225 square millimeters (larger than an iPad), 4 trillion transistors, and 900,000 integrated AI compute cores. In comparison, NVIDIA's H100 has a chip area of 814 square millimeters.

Wafer-Scale Approach vs. GPU Cluster Approach

Technical ApproachRepresentativeAdvantagesDisadvantages
Traditional GPU clusterNVIDIAMature ecosystem, high flexibilityHigh interconnect communication overhead
Wafer-scale single chipCerebrasNo interconnect communication required, high training efficiencySignificant manufacturing yield challenges

Cerebras has demonstrated better cost-performance than GPU clusters in certain specific large-model training tasks (such as sparse models and scientific computing). Cerebras's customers include G42, a UAE-based AI company, and multiple national laboratories.