Cerebras IPO and the Four Bottlenecks in Its Custom-Everything Architecture

Key Takeaways

• IPO oversubscribed ~20×, reflecting strong investor appetite
• WSE‑3 houses 900k cores, 44 GB SRAM, 21 PB/s bandwidth
• OpenAI service deal worth >$20 B for 750 MW compute capacity
• TSMC lock‑in limits wafer‑scale yield improvements to a single fab
• 2025 revenue 86% came from G42 and UAE customers

Pulse Analysis

Cerebras’ debut on the public markets arrives at a time when AI‑focused capital is flowing into every corner of the hardware stack. By pricing its shares amid a 20‑fold oversubscription, the company signals that investors see value in a differentiated approach that sidesteps the conventional GPU road‑map dominated by Nvidia. Yet the market is cautious: Cerebras targets a niche of ultra‑low‑latency inference, a segment where the performance premium must translate into tangible cost savings for customers to justify the higher price point.

The technical heart of Cerebras is the Wafer Scale Engine 3, a single 300 mm wafer that functions as one massive chip. With 900,000 processing cores and an unprecedented 21 petabytes per second of SRAM bandwidth, the WSE‑3 excels at the decode phase of transformer inference, where memory speed is the bottleneck. However, achieving usable yields on a wafer‑scale die required a decade of collaboration with TSMC, embedding sophisticated defect‑avoidance logic and custom interconnects. This deep fab partnership creates a lock‑in that makes it difficult for competitors to replicate the architecture without similar long‑term investments.

Cerebras’ business model now leans heavily on a cloud‑service agreement with OpenAI, valued at over $20 billion for 750 MW of compute capacity, plus a $1 billion working‑capital loan that funds data‑center build‑out. While the contract provides a sizable revenue runway, it also ties the company’s fortunes to OpenAI’s deployment schedule and the ability to operate large‑scale wafer‑scale clusters reliably. Coupled with the concentration of hardware sales in G42 and UAE institutions—accounting for 86 % of 2025 revenue—Cerebras faces a critical test: can its custom‑everything architecture scale operationally and financially beyond a handful of strategic partners? The answer will determine whether wafer‑scale chips become a mainstream AI inference solution or remain a specialized, high‑risk play.