Cerebras Systems's April 17 S-1 refile for a proposed Nasdaq listing under the ticker CBRS is now two and a half weeks into its pre-roadshow window. The paper's Wednesday Day Two note reconstructed the OpenAI triangle as the single piece of disclosure that will price the mid-May listing. What the subsequent reporting has confirmed is that the triangle has not loosened since the refile — it has tightened. OpenAI's total compute commitment to Cerebras now runs to more than $20 billion over three years, up from a $10 billion January agreement now superseded. [1] The $1 billion working-capital loan at six percent annual interest remains. The 33,445,026-share warrant for Class N common stock at an exercise price of $0.00001 per share remains. [2] The customer who is the lender who is the potential shareholder is, on Day Three, still the same customer.

The economic structure in the S-1 is specific enough to reward reading. OpenAI has committed to 250 megawatts of compute capacity per year from 2026 through 2028 — a 750 MW base commitment — with an option to scale to 1.25 gigawatts through 2030. The warrants for up to 10 percent equity vest only if OpenAI's purchases reach 2 gigawatts. [1] The asymmetry is that warrant vesting is tied to the option tranche, not the base tranche. The more OpenAI buys, the more equity it earns at effectively zero cost. OpenAI also extended the $1 billion loan at 6 percent annual interest, repayable in cash or in goods and services. [1] Goods and services here means Cerebras chips. The loan can be repaid by Cerebras delivering the compute OpenAI has already agreed to pay to use. That is the circular revenue the S-1 discloses.

Underwriter selection is publicly fixed: Morgan Stanley, Citigroup, Barclays, and UBS are lead-book. [3] This is a senior-tier syndicate and a reminder that the 2024 attempt at listing — paused over G42 customer-concentration risk and CFIUS review — was not a case of Cerebras failing the investment-bank bar. It passed that bar with the same names. What it failed, on its first attempt, was the national-security review. The refile does not carry a CFIUS question because the concentration shifted from a UAE-linked buyer (G42, 87 percent of 2024 revenue) to a US-based buyer (OpenAI, projected to be a majority of 2026 revenue together with AWS). [1] The concentration did not shift. The jurisdiction of the concentration did.

The revenue trajectory tells its own story. Cerebras reported $510 million in 2025 revenue, up 76 percent year over year, and its first profit in eleven years. [1] The S-1 says OpenAI "represents a substantial portion of our projected revenues over the next several years." Backlog — the leading indicator that replaces trailing revenue in a registration statement — stands at $24.6 billion, most of it tied to OpenAI. [4] The OpenAI Master Relationship Agreement was signed on December 24, 2025. The $1 billion loan closed in January 2026. OpenAI revenue begins in 2026. When the filing describes OpenAI as a "substantial portion" of projected revenue, the phrase points at a forward projection that has not yet appeared in any audited period.

The governance structure is where the triangle's IPO implications sharpen. OpenAI's warrants are for Class N common stock — non-voting — and are tied to compute purchases rather than to cash-on-cash equity investment. The registration-rights agreement filed as an exhibit to the S-1 gives OpenAI the right to include its shares in future registered offerings and to demand a registration of its own. [2] This is the governance mechanism that converts operational relationship into capital-markets relationship. In practical terms, OpenAI has secured both price-locked compute and an equity stake whose exit mechanics are already contractualised before the listing.

The risk disclosure is the unusual part. Cerebras's S-1 does not bury the related-party concentration; it flags it explicitly as a risk factor. [1] The roadshow narrative has to manage that flag rather than disguise it. Underwriters priced the 2024 offering past the G42 CFIUS review by arguing that concentration was a diversification challenge on the path to listing, not a structural feature. The OpenAI triangle on the 2026 S-1 is a structural feature by disclosure. The mid-May pricing will test whether institutional buyers accept a customer-concentration architecture as the price of access to the fastest-growing AI chip alternative to Nvidia.

Read against the sector tape, the triangle is the clearest case of what OpenAI is doing across the compute stack. OpenAI has signed multi-year commitments with AMD (reported in October 2025), with Broadcom (reported in December 2025), with Oracle (reported earlier in 2026), and now, in scaled form, with Cerebras. The dollar magnitudes vary; the structural pattern — price-locked compute from diversified silicon suppliers, often with equity-linked terms — is consistent. The triangle is not a one-off. It is a category of deal OpenAI is executing across multiple counterparties. Cerebras's S-1 is the first one that has to be priced publicly.

The paper's position, held since Tuesday's initial feature, is that the triangle is not a scandal but a disclosed governance architecture that transfers valuation risk from the issuer to the buyer of the offering. If the roadshow prices at a range that accepts the triangle, the precedent is set for subsequent AI-chip issuers who have similar customer-lender-shareholder structures to disclose. If the roadshow prices at a discount, the market has said that one-counterparty financial architecture has a valuation cost even when disclosed in full. Neither outcome is yet determined. Both are governance precedents for the AI compute stack's next twelve months of listings.

For the paper, the next data point is the preliminary-prospectus price range when Cerebras files the S-1/A. The price range is the underwriters' first public estimate of what the market will bear on the triangle. It is scheduled for early May. Until then, the S-1 that was filed on April 17 remains the complete public record, the circular-revenue disclosure remains explicit, and the triangle remains the one thing the roadshow has to explain. Three days in, the explanation has not started.

-- THEO KAPLAN, San Francisco