Cerebras Files for $26.6B Nasdaq IPO — The Wafer-Scale AI Chip That's 57x Bigger Than Nvidia's H100

Cerebras Systems filed an S-1 on April 17 and is set to price its Nasdaq IPO on May 13 under the ticker CBRS. The company is selling 28 million shares at $115–$125, targeting a raise of up to $3.5 billion at a valuation of $26.6 billion — up from the $23 billion it commanded in a February venture round backed by AMD.

If it holds, this would make Cerebras one of the largest US tech IPOs of 2026, and the most significant pure-play AI chip listing since Nvidia’s GPU business became the defining infrastructure trade of the decade.

The Architecture Bet

Cerebras’s core argument is that GPUs are the wrong abstraction for large language model inference. The company’s WSE-3 (Wafer-Scale Engine 3) covers an entire 300mm silicon wafer — 46,225 square millimeters of die area. That’s 57 times larger than Nvidia’s H100. Fabricated on TSMC’s 5nm process, the chip integrates 44 GB of on-chip SRAM directly adjacent to 900,000 AI-optimized compute cores, eliminating the memory bandwidth bottleneck that limits GPU clusters running LLMs.

The result: for inference on large models, Cerebras claims its systems outperform equivalent GPU clusters at lower power draw. That claim has been validated at scale. In January, OpenAI signed a deal for Cerebras to supply up to 750 megawatts of AI computing power through 2028, with the total contract value exceeding $20 billion. That’s a single customer representing roughly 40 times Cerebras’s 2025 annual revenue.

The Numbers

The S-1 shows $510 million in revenue for 2025, up from an annualized ~$136 million run rate in early 2024. Non-GAAP net income was $237.8 million. For a hardware startup that spent years as a bet on a contrarian architectural thesis, turning profitable at this scale before going public is genuinely unusual.

The OpenAI dependency is real. A customer representing the bulk of revenue is a material risk, and the S-1 addresses it directly. But the pipeline behind it — other hyperscalers and large enterprises exploring alternatives to GPU clusters for inference workloads — is what the $26.6B valuation is pricing in.

Why the Timing Matters

Cerebras is going public at the moment when the AI infrastructure trade is at its most intense. Nvidia is reporting $37B+ per quarter in data center revenue. AMD is scaling its MI300 series. Intel is pushing Gaudi. Into that environment, Cerebras is arguing that wafer-scale isn’t just a faster GPU — it’s a fundamentally different architecture suited to the inference workloads that now dominate production AI deployment.

Whether that argument holds as Nvidia’s Blackwell and next-generation Rubin families scale, and as competing inference chips reach market, will determine whether $26.6B is conservative or ambitious.

Morgan Stanley, Citigroup, Barclays, and UBS are joint lead underwriters. Pricing is May 13 on Nasdaq.

What It Signals

Cerebras going public is a test of whether public markets will pay for differentiated AI silicon, or whether they treat every non-Nvidia chip story as speculative. Given that the company is profitable, has a $20B customer anchor, and is demonstrating real revenue growth, the risk profile is materially different from most hardware IPOs.

The more interesting question after pricing: who else is buying WSE-3 capacity, and at what scale?