Cerebras files, and the customer list is a customer

Cerebras Systems filed an S-1 with the SEC on 17 April, targeting a Nasdaq listing under the ticker CBRS.¹ The headline numbers are respectable: $510M in 2025 revenue, up 76% year on year, and $87.9M of net income, a rare thing, a profitable AI-infrastructure IPO candidate. The filing also discloses a multi-year compute agreement with OpenAI valued at "in excess of $20.0 billion" over its term, and an equity component that Axios, citing people familiar with the matter, pegs at roughly 10% of the company on a fully diluted basis.² Axios also reports the bankers are circling a $35B target valuation.

I spent yesterday afternoon reading the S-1. The structural story is mostly on pages 38 through 54, in the sections nobody quotes.

What the filing actually says

Start with the revenue. The $510M figure is real, in the sense that it is audited and GAAP. It is also, per the "Customer Concentration" disclosure on page 41, approximately 71% attributable to a single customer referred to throughout as "Customer A", which the Risk Factors section later clarifies is OpenAI.³ The next two customers, G42 and a US national-lab consortium, account for a further 19% combined. That leaves roughly 10% of 2025 revenue, about $51M, distributed across the rest of the book.

The $510M figure is real, in the sense that it is audited and GAAP.

This is not unusual for a company selling frontier-scale training and inference capacity. It is, however, worth saying out loud: Cerebras in 2025 is, for most commercial purposes, an OpenAI vendor with a public-sector sideline.

The $20B compute agreement is structured as a take-or-pay arrangement with tranche-based capacity commitments and, this is the part I want to read carefully, an "Equity Consideration" provision under which OpenAI receives warrants exercisable into Cerebras common stock upon satisfaction of specified delivery and performance milestones.⁴ The filing does not disclose the strike price. It does disclose that the warrants, if fully exercised, would represent approximately 9.8% of post-IPO diluted shares outstanding. Axios's "10% stake" is the filing's 9.8%, rounded up.

The warrants vest against Cerebras delivering capacity, not against OpenAI paying for it. Which is a slightly strange arrangement and I think it produces slightly strange outcomes.

What this is a case of

The inference-economics frame predicts that as training cost flattens and inference cost becomes the binding constraint, hyperscalers and frontier labs will move to lock in silicon supply through structured commitments that look like customer contracts but function as vertical integration.⁵ The CoreWeave–Microsoft arrangement is the canonical example; the Lambda–Microsoft deal is a variant; the Anthropic–Amazon Trainium commitment is a third.

Cerebras–OpenAI fits the pattern and extends it. The take-or-pay locks in revenue. The warrants align Cerebras's capital structure with OpenAI's capacity needs. The combined effect is that OpenAI gets silicon optionality outside the Nvidia stack at a cost, equity dilution at Cerebras, that OpenAI does not pay in cash and that Cerebras's public-market shareholders will.

The frame holds cleanly here. What it predicts, that the "customer" in a frontier-compute S-1 is functionally a counterparty, a shareholder, and a pricing-setter simultaneously, is exactly what the filing discloses. OpenAI sets the volume (via the tranche commitments), receives the upside (via the warrants), and anchors the valuation narrative (via the $20B headline). The bankers' $35B target is, approximately, $20B of contracted OpenAI revenue times a 1.75x enterprise-value-to-backlog multiple, which is the maths you'd do if you thought this was a compute-infrastructure company. It is not obvious that is the right maths.

Where the frame gets interesting

The performativity frame, spend at scale being materially influential regardless of delivery, cuts in a direction I wasn't expecting. A $20B multi-year commitment from OpenAI to Cerebras reads, on one view, as OpenAI diversifying away from Nvidia. On another view, it reads as OpenAI underwriting a public-market comparable that will help value OpenAI's own next round. If Cerebras trades at $35B on $510M of revenue, a 68x multiple, then the read-through to OpenAI's own revenue multiple is, let's say, supportive.

This is the part where I notice that OpenAI's warrants vest on Cerebras delivering capacity rather than on OpenAI paying for it. If OpenAI's cash commitments soften, because model economics shift, because a cheaper architecture emerges, because OpenAI's own revenue disappoints, OpenAI still gets the equity upside from Cerebras hitting its build-out milestones, which Cerebras will hit because the capital to hit them is what the IPO is for.

I don't want to overread this. Take-or-pay means OpenAI pays whether it uses the capacity or not. But "pays" in these structures typically means "owes", the timing and form of payment is negotiable, and the S-1's Related Party disclosures hint at some flexibility.⁶ The warrants are the cleaner instrument. The warrants vest on delivery.

The safety-positioning wrinkle

One other thing worth flagging. The S-1's "Business" section makes a specific claim I hadn't seen Cerebras make before: that its wafer-scale architecture offers "deterministic inference behaviour and auditable memory-access patterns" suited to "regulated and safety-sensitive deployments."⁷ The national-lab revenue, 8% of 2025, is invoked as the proof point.

This is the ai-safety-as-market-position frame in a slightly unusual application. Usually we see it at the model layer: Anthropic's RSP, OpenAI's preparedness framework, defence-market pitches about controllable systems. Cerebras is making the argument at the silicon layer: that the hardware itself has safety-relevant properties that distinguish it from GPU-based inference. Whether the technical claim holds is above my pay grade. The market-positioning claim is clear, and it is aimed at the one customer segment that cannot buy from a company that is 71% an OpenAI vendor: the US defence and intelligence community, which will want an independent supply line.

The frame predicts Cerebras will use IPO proceeds to build out a government-vertical sales motion that deliberately reduces OpenAI concentration. The S-1's "Use of Proceeds" section allocates $400M to "expansion of federal and regulated-industry go-to-market capacity," which is consistent.⁸

What to watch

Three things.

First, the pricing on the OpenAI warrants when the final prospectus lands. A nominal strike price and a low vesting bar would confirm the read that this is structured financing dressed as a commercial contract. A strike at IPO price with milestone gates that require genuine delivery is a different animal.

Second, the customer-concentration number in the first post-IPO 10-Q. If OpenAI drops from 71% toward 50% because the federal pipeline materialises, the inference-economics frame is doing work; if it stays at 71% or rises, Cerebras is a single-customer company with a listing.

Third, whether any other frontier lab, Anthropic, Google, a Chinese lab operating through an intermediary, appears as a named customer in subsequent filings. The pattern I'd expect, if this IPO prices well, is that the OpenAI template gets replicated: warrants-for-capacity arrangements become the standard way frontier labs lock in non-Nvidia silicon. Cerebras is the test case. The S-1 is the template.

It is, as filings go, a reasonably honest document. It tells you who the customer is, who the shareholder will be, and that those are the same entity. The rest is reading comprehension.

Footnotes

¶