Cerebras Systems' April 17 S-1 names OpenAI on three lines that almost never sit on the same page of a registration statement. OpenAI is a customer — a $20 billion-plus multi-year compute contract that the filing discloses as the company's largest. OpenAI is a lender — a $1,004,571,259 secured promissory note dated January 5, 2026, repayable in cash or in services. OpenAI is a shareholder — a warrant for 33,445,026 shares of non-voting Class N stock at $0.00001 a share, plus a separate 2,696,678-share warrant authorized after the year-end balance sheet, vesting toward roughly ten percent of the company as cumulative spend ramps to $30 billion. [1] [2]

The paper's Tuesday account of Cerebras making OpenAI validation and concentration the same fact read the S-1 as collapsing the marquee-customer story into the single-counterparty story. Wednesday's piece is what the same prospectus says, in regulator-blessed language, on adjacent paragraphs. OpenAI funds Cerebras's working capital. OpenAI buys Cerebras's chips. OpenAI's equity appreciates as the buying happens. The S-1 does not euphemize the structure. It names it.

Reuters' April 17 dispatch on the filing led with the IPO revival and noted, at the bottom, that "the ChatGPT creator will deploy 750 megawatts of Cerebras chips" under the headline contract. [3] CNBC reported, more bluntly, that "Cerebras issued OpenAI warrants to purchase up to 33.4 million shares of non-voting Class N stock" in December and "received a $1 billion loan from OpenAI, with a 6% annual interest rate, to build data center infrastructure." [4] Both pieces correctly described the warrant as performance-vesting. Neither separated the three roles into a single sentence the way the prospectus does.

The customer role is the largest disclosed contract in the company's history. Cerebras's order backlog is $24.6 billion, most of it tied to the OpenAI commitment. [3] The lender role is more unusual. The promissory note dated January 5 — disclosed as Exhibit to the S-1 — sets the principal at $1,004,571,259, the maturity at December 31, 2032, the interest at six percent annually, and the repayment terms at cash or "delivery of products or services." [2] OpenAI prepaid a year of compute and labeled it a loan. The shareholder role is the warrant. The strike price is one-millionth of a cent. The vesting schedule is tied to cumulative spend. At $30 billion of cumulative purchases, OpenAI's economic interest in Cerebras approaches ten percent. [1]

What makes the structure unusual is that it sits inside a prospectus heading the SEC reviews for plain-English readability. Cerebras's lawyers had to characterize the relationship for retail investors. They wrote it as a single paragraph. The paragraph names OpenAI three times. The risk-factors section, which a different desk at the SEC reviews, reproduces the same disclosure in the customer-concentration discussion: 86 percent of 2025 revenue came from two UAE-billed customers (62 percent MBZUAI, 24 percent G42); 86 percent of 2026 revenue is projected to come from OpenAI plus AWS. [5] The concentration moved. It did not diminish.

The mid-May roadshow has to price the package. The Information reported that Cerebras is targeting a $35 billion valuation and a $3 billion raise. [6] The lead underwriters — Morgan Stanley, Citigroup, Barclays, UBS — will face the question every prospectus with a single dominant counterparty has faced: what is the company worth without that customer, and how do you price the optionality of the warrant the customer holds? The paper's read on Tuesday was that the validation story (a marquee buyer) and the concentration story (a single counterparty) were the same fact. The S-1 itself, on Wednesday, says it more directly than analysts have: customer, lender, shareholder.

There is also the question of governance. Class N is non-voting. OpenAI's economic interest does not carry board representation or voting rights, which is the structural distinction Cerebras's S-1 leans on to argue this is not a controlled-company situation. [1] In practice, however, a counterparty whose contract represents the bulk of disclosed backlog and whose warrant becomes meaningful only as that contract is paid down does not need a board seat to exert influence. The paper's framing of the divergence: MSM treats the warrant as deal-engineering color; X treats the structure as OpenAI engineering a captive supplier. The S-1 itself, sitting between, says both at once.

For procurement officers reading the filing for risk-of-default exposure, the relevant disclosure is the working-capital deposit. OpenAI advanced about $1 billion against future deliveries; Cerebras booked it as a balance-sheet asset. [6] If Cerebras misses delivery, OpenAI's recourse is the note, the warrant clawback, or both. If OpenAI ramps faster than projected, the warrant accelerates. The IPO converts these mechanics from private contract to publicly disclosed cap-table choreography, which is why the filing matters beyond the chip-buyer audience.

Sell-side coverage on Tuesday emphasized the IPO valuation arithmetic; sell-side coverage on Wednesday will have to confront the disclosure architecture. Microsoft's FQ3 print, after the bell on the same day Cerebras filed, reopens its own counterparty-concentration window: the company disclosed in January that 45 percent of its $625 billion in remaining performance obligations was tied to OpenAI. [4] Two AI vendors — one a startup, one the most-capitalized public company in the world — now publicly tie multi-year revenue to one buyer. The prospectus will not be the last document this quarter to put OpenAI in three columns of the same risk table.

-- DAVID CHEN, Beijing