Cerebras takes the wafer to market

Cerebras Systems launched its IPO roadshow on 4 May, offering 28 million Class A shares at a $115–$125 range, which lands the company at roughly $26.6bn fully diluted at the top end. Morgan Stanley is left lead. The book is expected to price around 13 May, listing on Nasdaq under CBRS. The company is going out with a disclosed $20bn customer arrangement with OpenAI and named multi-year commitments from AWS and Meta.

This is the AI hardware IPO that didn't happen for two years, and the fact that it's happening now, and at this price, with these customers, is the structural story.

What is actually being sold. Cerebras makes wafer-scale chips: a single piece of silicon the size of a dinner plate, where everyone else cuts the wafer into hundreds of GPUs. The pitch has always been that for inference workloads with awkward memory-bandwidth profiles, one big chip beats many small ones networked together. The pitch has also always been that NVIDIA's CUDA moat makes this a hard sell into production. What changed, per the prospectus framing, is inference. Specifically, inference at latencies that make agent loops viable.

The customer concentration. The OpenAI line in the S-1 is the one to read carefully. A $20bn multi-year commitment from a single counterparty, against a company whose trailing revenue is a small fraction of that, is the entire equity story. It is also a textbook concentration risk that the risk factors will spend several pages on. I'd want to know, and the prospectus will tell us, in language drafted by people paid to be careful, whether this is a take-or-pay commitment, a capacity reservation with milestones, or a non-binding letter of intent dressed in firmer clothes. These are very different instruments and they justify very different valuations.

A $20bn multi-year commitment from a single counterparty, against a company whose trailing revenue is a small fraction of that, is the entire equity story.

The AWS and Meta lines matter for a different reason. AWS selling Cerebras capacity through Bedrock or a comparable channel means Cerebras has a distribution layer it doesn't have to build. Meta as a customer means someone with the in-house engineering depth to evaluate alternatives to H100s and B200s has chosen to deploy at least some inference on wafer-scale. Neither is OpenAI-sized in dollar terms, but both are signal in a way that a $20bn anchor from one customer is not.

Frame one: inference economics. This is the frame the deal is built for. The structural argument that inference cost, not training cost, is now the binding constraint on frontier AI economics has been the loudest theme in the market for eighteen months. Cerebras is the cleanest public-market expression of that thesis available. If you believe inference compute is going to be a larger market than training compute by 2027, and you believe NVIDIA's general-purpose architecture leaves margin on the table for inference-specialised silicon, Cerebras is the trade. The $26.6bn valuation prices in a meaningful slice of that thesis. It does not price in dominance.

The thing the frame predicts, and that the filing should let us check, is gross margin trajectory. Wafer-scale should, should, produce structurally better inference economics per token at scale, because you eliminate inter-chip networking overhead. If the S-1 shows gross margins expanding into the customer ramp, the frame fits. If margins are flat or compressing as volume grows, something in the unit economics isn't working the way the pitch says it is.

Frame two: AI performativity. The other frame this deal sits inside is less flattering. Cerebras is going public into a market where the scale of announced AI infrastructure spend has decoupled, somewhat, from the scale of demonstrated AI revenue. OpenAI's compute commitments across NVIDIA, Oracle, and now Cerebras run into the hundreds of billions over the coming years. OpenAI's annualised revenue, while large, does not obviously support that commitment profile on any conventional payback math. Which means one of three things is true: OpenAI's revenue is going to grow into the commitments, the commitments are softer than they look, or someone is going to be left holding capacity that doesn't clear.

Cerebras, by going public on the back of an OpenAI anchor, is asking public-market investors to take a view on which of those three is true. That is a meaningfully different proposition from the deal Cerebras would have done in 2024.¹

What the primary document will need to settle. I haven't yet seen the priced S-1, the roadshow version is what's circulating, but the disclosures that will determine whether this prices at the top of the range or in the middle are, in order:

The structure of the OpenAI commitment. Take-or-pay versus reservation versus LOI. The difference between these is roughly the difference between $26.6bn and $15bn of enterprise value.

The gross margin disclosure on shipped systems versus the gross margin on cloud-delivered inference. Cerebras has historically sold systems; the strategic shift is to sell tokens. These have very different margin profiles and the IPO pitch implicitly blends them.

Customer count outside the named three. AWS, Meta, and OpenAI are the headline. The depth of the second tier, research labs, sovereign deployments, enterprise, is what determines whether this is a three-customer business or a platform.

Capex commitments. To deliver $20bn of OpenAI capacity, Cerebras has to build it. The financing of that build, whether from IPO proceeds, customer prepayments, or debt, shapes the equity story materially.

What this is a case of. It is the first pure-play AI inference hardware IPO at scale. Groq has stayed private; SambaNova has stayed private; Tenstorrent has stayed private. Cerebras going first means Cerebras sets the comparable. If it prices well and trades well, the others follow within twelve months. If it prices at the bottom of the range or breaks issue, the window closes and the inference-silicon cohort stays private through another funding cycle.

It is also a case of the OpenAI-as-anchor-tenant model going public. Oracle's stock re-rated on the back of disclosed OpenAI commitments. CoreWeave's IPO was substantially an OpenAI-adjacency trade. Cerebras is the third instance of public-market capital being asked to underwrite OpenAI's compute roadmap by proxy. At some point, and I don't know where the point is, the public market will start pricing the correlation across these names rather than treating each as an independent bet on AI infrastructure.

What to watch. Pricing day, obviously. But more interestingly: the first 10-Q after listing, which will give the first look at how the OpenAI commitment converts to revenue recognition. And, separately, whether NVIDIA says anything about wafer-scale at its next earnings call. Jensen has been notably quiet on Cerebras for two years. Public-market competition tends to end that kind of silence.

Footnotes

¶