‹ Cerebras Systems
Ch 3 · The Biggest Chip Ever Built
Chapter 3 · Competitive Position
The chip is 58x bigger than Nvidia's B200.
Cerebras's moat is a fundamentally different way to build a chip. Nvidia's moat is software (CUDA). Both bets are real.
↓ scroll to read
✦ The bottom line
Cerebras built the only commercial wafer-scale AI chip. 4 trillion transistors. 900,000 cores. 58x bigger than Nvidia B200. The architectural moat is real. The adoption moat is what they have to build next.
↓ the brief below
✦ Teach me
Architectural moat
Most chip companies (Nvidia, AMD, Intel) build small chips — typically a few hundred square millimeters. To scale up AI training, they connect hundreds or thousands of these chips with networking. Cerebras took the opposite bet: build one enormous chip, so big it covers most of a silicon wafer. No networking needed between cores — everything's already on-chip. That's wafer-scale integration, and Cerebras is the only company that's done it commercially. A moat built on architecture is durable — patents and engineering know-how are hard to copy. A moat built on adoption (like Nvidia's CUDA) is durable in a different way — once developers learn your tools, they don't want to switch.
Wall Street calls this
Wafer-scale integration
Architectural moats reward *technical performance.* Adoption moats reward *ecosystem.* Cerebras has the first. Nvidia has the second. Each kind protects in a different way.
Cerebras WSE-3 chip area vs Nvidia B200
58x
46,225 mm² of silicon vs Nvidia B200's ~800 mm² per chip (1,600 mm² combined for the dual-chip package). *2,625x* more memory bandwidth. *19x* more transistors. The *architectural* moat is measurable.
Source · 424B4 · Prospectus Summary — Our Solution · FY2025 · Filed May 14, 2026
Hardware specs can win in isolation. But AI software is built on top of years of investment in tools, libraries, and developer education. Nvidia spent 20 years building CUDA — the programming framework every AI researcher learned on. To win share from Nvidia, Cerebras has to convince developers to switch ecosystems. Their pitch is in the prospectus: their compiler turns PyTorch models directly into wafer-scale code, bypassing CUDA entirely.
From the prospectus · the software pitch
Cerebras Compiler compiles PyTorch models directly to the WSE, eliminating the need for CUDA or low-level programming. Our software platform makes wafer-scale computing simple to use. It spans the full AI life cycle — from model development through production deployment.
↳ Cerebras's bet: PyTorch is the actual lingua franca of AI, not CUDA. If developers can use PyTorch on Cerebras hardware without writing low-level code, the CUDA lock-in loses a step. This is the most strategically important sentence in the whole prospectus.
Source · 424B4 · Prospectus Summary — Software Platform · FY2025 · Filed May 14, 2026
How Cerebras got here
2016
Andrew Feldman and Sean Lie co-found Cerebras after their prior AMD-acquired startup SeaMicro. Both came from AMD's data center group.
2019
First wafer-scale chip (WSE-1) ships. The largest commercial chip ever built. Pundits called it impractical.
2021-23
WSE-2 then WSE-3 ship. Customers (mostly research labs, government, UAE) start buying real volumes.
Dec 2025
OpenAI selects Cerebras as its fast inference solution. Signs Master Reservation Agreement. Provides $1B working capital loan.
Mar 2026
AWS signs binding term sheet to become the first hyperscaler deploying Cerebras in its data centers.
May 15, 2026
IPO closes. Cerebras becomes a public company at a multi-billion-dollar valuation.
Strong
Real architectural moat. Real strategic-partner endorsements (OpenAI + AWS). The CUDA-bypass software claim is the most important bet to track.
You just finished
Chapter 3 · MOAT
The Biggest Chip Ever Built
you now read: architectural moat
Up next
Then
Chapter 5 · BEHIND THE NUMBERS
The Story Behind the Numbers
Chapter 6 · RISK
What Could Break the Brief