A firm order, an option, and a vesting threshold that only triggers if both are fully exercised — the arithmetic is the story, and the structure is worth reading carefully.
What follows is analysis of reported deal terms. It is not investment advice, it expresses no view on any company or security, and it predicts nothing about exercise or delivery.
A warrant struck against a volume threshold is an ordinary commercial instrument, used precisely because it aligns a supplier’s equity with a customer’s purchasing. What follows describes how this one is calibrated and alleges nothing improper by either party, and imputes no motive to anyone.
On The MAD Podcast, Cerebras co-founder and CEO Andrew Feldman described the company’s agreement with OpenAI in the units that actually govern it: megawatts. “Our deal with them, with OpenAI, was because data center capacity is a limiting constraint, measured the way data centers are measured in megawatts,” Feldman said. “The deal is 750 megawatts, 250 megawatts in 26 on a multiyear lease, an additional 250 megawatts in 27 on a multiyear lease, an additional 28 on a multiyear lease.” Cerebras delivers this, in his framing, as a full cloud solution.
As reported in coverage of Cerebras’s S-1 and by CNBC — rather than read from the filing itself — the headline commitment has grown across successive reports: first disclosed in January 2026 as a deal worth over $10 billion covering up to 750 MW through 2028, and subsequently reported as a multi-year agreement worth over $20 billion. Both figures, and their sequence, matter; neither is “the” number, and the gap between them is itself informative. The reported structure also includes an additional option for 1.25 GW of capacity through 2030, on top of the 750 MW firm order.
As reported, Cerebras has granted OpenAI warrants covering up to 33.4 million non-voting shares at an exercise price of $0.00001 per share. The vesting condition ties to a volume threshold: 2 gigawatts of total capacity purchased by OpenAI. That is the number that makes the arithmetic worth doing explicitly.
The key insight: The firm order is 750 MW. The option is 1,250 MW. Those sum to exactly 2,000 MW — exactly the 2 GW vesting threshold. The warrant does not vest on the deal as signed. It vests only if OpenAI takes the firm order and exercises the entire option on top of it. A headline describing a potential ~10% customer stake is describing a conditional outcome that requires the agreement to become roughly three times larger in volume terms.
The highlighted bar is not a separate quantity. It is the other two added together, which is exactly what the vesting condition asks for.
The Structural Read
Start with what Feldman said about the unit of denomination, because it is the most structurally informative line in the exchange. When a contract is written in megawatts rather than in chips or dollars, the seller is selling access to a constraint rather than a product. Data center capacity is the scarce input; the contract attaches directly to that scarcity. That choice also has a precise mathematical consequence: because both the volume covenant and the warrant vesting threshold are denominated in the same unit, the reconciliation is exact. 750 plus 1,250 equals 2,000. The arithmetic works because the units match.
Now look at where the vesting threshold sits relative to the firm order. A warrant’s vesting condition tells you what the equity is actually paying for. Set the threshold at or below the firm commitment, and the equity rewards signing — it compensates for entering the agreement. Set the threshold at commitment plus the full option, and it rewards exercising — a different act, taken later, under conditions neither party can see at signing. Here, 750 MW firm plus 1,250 MW option equals exactly 2,000 MW — exactly the vesting threshold. The equity is not consideration for the agreement that exists today. It is consideration for an agreement that would have to become roughly three times larger in volume terms.
Andrew Feldman — The MAD Podcast
“Our deal with them, with OpenAI, was because data center capacity is a limiting constraint, measured the way data centers are measured in megawatts. The deal is 750 megawatts, 250 megawatts in 26 on a multiyear lease, an additional 250 megawatts in 27 on a multiyear lease, an additional 28 on a multiyear lease.”
The instrument itself deserves separate attention. An exercise price of $0.00001 per share is effectively zero. That means the consideration is not the cash payment — it is the purchasing. The customer pays in volume rather than in dollars. And because the shares are non-voting, the supplier is parting with economics while keeping control intact. Put those two design choices together and the structure describes a deal in which the customer is being compensated — in the one currency that does not change who decides anything — for becoming a committed, large-scale buyer.
A warrant structured this way is a common commercial instrument, and nothing here suggests impropriety by either party. Volume-linked warrants exist precisely because they align a supplier’s equity upside with a customer’s scale commitment. What is worth noting is the structural property: the “up to” and “potential” and “option” doing load-bearing work in every published account of the stake are not hedging language. They are the mechanism. The stake is genuinely contingent on an outcome that does not yet exist.
Map of AI — Enabler Layer
Where Cerebras sits in the AI stack
The Map of AI identifies nine layers from raw infrastructure through end applications. Cerebras operates at the compute and cloud-delivery layer — the Enabler tier that supplies scarce capacity to Founders (model builders like OpenAI) without competing on applications. Structuring the deal in megawatts and delivering it as a full cloud solution is a precise description of what an Enabler does: it sells access to a constraint so the Founder can focus on the model. The warrant structure then ties the Enabler’s equity upside directly to whether the Founder scales into that constraint at full volume.
Three Implications
IMPLICATION 1 — THE UNIT IS THE SIGNAL
When a major AI agreement is denominated in megawatts rather than chips or tokens, it marks a structural shift in what is being sold. The scarce input is capacity, not compute. As long as power remains the binding constraint for large-scale AI infrastructure, agreements written in megawatts will increasingly be the template — and the companies that control that unit will sit at the leverage point in the stack.
The reported commitment moved from over $10 billion to over $20 billion between published accounts. That movement is not a correction — it is a snapshot of a negotiation. Any single large number attached to a multi-year, multi-option infrastructure agreement is a ceiling on a conditional path, not a fixed quantity. The sequence of how a commitment is reported over time carries as much information as the figure itself.
IMPLICATION 3 — A NEW PATTERN IN AI SUPPLY AGREEMENTS
This is the third deal of this structural family covered here within a week: a cloud agreement in which a customer receives a warrant over its supplier, combined with a separation of economics from control through non-voting shares. The recurrence is itself the observation. Volume-linked, non-voting, effectively zero-exercise-price warrants are becoming a recognisable instrument in AI infrastructure contracting — one that aligns scale incentives without transferring governance.
Business Engineer Framework
The Map of AI — Founders, Distributors, Enablers
The Map of AI maps 200+ companies across nine layers of the AI stack. The Cerebras–OpenAI structure is a precise illustration of the Enabler–Founder relationship: an infrastructure supplier selling access to a scarce constraint (megawatts of compute capacity) to a model builder, with equity alignment calibrated to full volume uptake rather than to the act of signing. Understanding where any company sits in those nine layers tells you who holds the leverage — and what the contracts will look like.
Every deal term above is as reported, in CNBC’s coverage and in reporting of Cerebras’s S-1. This publication is reading reporting of that filing rather than the filing itself.The headline commitment grew between reports, from over $10 billion in January 2026 to over $20 billion subsequently; both figures appear above because neither is the final word on a multi-year agreement.The stake is conditional. The warrants are reported to cover up to 33.4 million non-voting shares at an exercise price of $0.00001, described as a potential holding of around 10 per cent, and they vest only when OpenAI has purchased 2 gigawatts of total capacity — which is the 750 MW firm order plus the entire 1.25 GW option. The words up to, potential and option are load-bearing and are kept throughout. Nothing above alleges impropriety by Cerebras, by OpenAI or by anyone, imputes any motive, or describes the arrangement as circular financing or a gimmick. A warrant struck against a volume threshold is an ordinary commercial instrument, and the passage above describes how this one is calibrated rather than passing judgement on it. No share price, market capitalisation, founder net worth or trading figure appears. How much capacity OpenAI has purchased to date, whether the option will be exercised, what happens if it lapses, Cerebras’s revenue, backlog, margins and costs, any OpenAI statement, the lease rates and any other customer’s terms are not established and do not appear. Nothing above predicts exercise, delivery, prices or anything about either company, and nothing above is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
