OpenAI advanced Cerebras $1 billion at 6% interest. If Cerebras repays in compute capacity rather than cash, the loan is forgiven. That is not a loan in any ordinary sense, it is a prepaid forward, and oil and gas producers have used the same instrument for forty years.
The Cerebras S-1 is the cleanest public evidence yet that compute is being financed like a commodity, in contractual structure rather than in metaphor. The timing is not coincidence. Global inference spend overtook training spend in early 2026, call it the Inference Flip, and markets attract commodity-style financing once they become large, scarce, and physically deliverable. Compute now clears that bar. The S-1 shows what the financing looks like in practice, and this post walks through the instruments.
A forward strip and a prepaid forward
The OpenAI deal pairs a capacity contract with a working capital loan. Both are familiar shapes in commodity finance.
The capacity contract commits Cerebras to deliver 750 megawatts of inference capacity in tranches between 2026 and 2028, each running three to four years and extendable to five, anchoring revenue into the early 2030s. Cerebras carries $24.6 billion of remaining performance obligation tied to the deal on its balance sheet. That is a strip of long-dated, physical-settled forwards on compute capacity, the same shape any commodity trader would recognise: the buyer locks in delivery, the seller locks in volume, and the price is fixed at contract.
The loan is the more interesting half. OpenAI advanced $1 billion at 6%, forgiven on physical delivery of compute, accelerated to immediate cash repayment if the underlying capacity contract is terminated. Settled physically by default, in cash only as a fallback. In oil and gas this instrument has a name: the Volumetric Production Payment, or VPP. A financier paid an upstream producer cash upfront, and the producer owed a fixed volume of barrels over time. If the wells produced, the obligation cleared in kind; if they did not, the producer owed cash back with interest. Anadarko and Chesapeake built large parts of their balance sheets this way in the 2000s, and mining used a close cousin called streaming, with Wheaton Precious Metals building a $30 billion company prepaying miners for fixed streams of silver and gold production.
The Cerebras loan is the same template. The producer is Cerebras, the barrels are megawatts of inference capacity, and the financier is OpenAI. The novelty is not that someone signed forward contracts on compute, since large customers have been signing multi-year commitments for years. The novelty is the size, the public disclosure, and the fact that this single counterparty relationship anchored a $56 billion IPO that traded to nearly $95 billion fully diluted on Day 1.
This framing is not metaphor. The 2024 G42 agreement filed with the SEC called the upfront cash advance a “Prepaid Sum,” held in trust by Cerebras and offset against deliveries. The 2025 income statement then carries a $363 million non-cash gain from the extinguishment of a forward contract liability tied to that arrangement, the line item that swings GAAP net income from a $76 million non-GAAP loss into $238 million of headline profit. The auditors classify these prepaid capacity commitments as forward contracts because that is what they are.
Why compute fits the template
Two conditions make a market amenable to this kind of financing.
First, scarcity. The inference capacity shortage runs at least through 2028 and likely longer. Vera Rubin starts shipping in low volume in July, with broader availability not until 2027. Capacity coming online before then is already pre-sold. Supply does not catch demand on any realistic ramp scenario.
Second, verifiable physical delivery. The seller is on the hook for chips and power, not for tokens or useful work, and hardware allocation, power allocation, and uptime against spec are all measurable and enforceable. That is enough for the financing to work.
The Cerebras-OpenAI contract is structured exactly along these lines. Data centre rent, power, leasehold improvements, and security are pass-through costs reimbursed by OpenAI, and Cerebras recognises them as revenue on a gross basis. The seller carries delivery risk on chips and power; the buyer carries the operating-cost risk on the facility. That is a tolling structure in operational form.
The underlying does not need to be a fully fungible commodity. The first oil and gas VPPs specified production from particular wells with particular quality grades, and mining streams specified output from specific mines. Standardisation came later in both cases, once exchanges did the work of homogenising the underlying. Compute is at the early-bilateral stage of that arc.
Concentration, workload, renewal
Two risks in the Cerebras case deserve careful naming, because each has been mischaracterised in the standard coverage.
The first is concentration. Cerebras 2025 revenue was 86% from two related Abu Dhabi entities, MBZUAI and G42, with MBZUAI alone holding nearly 78% of accounts receivable at year-end. The 2024 IPO collapsed when CFIUS, the US national security review for foreign investment, examined this exact relationship, and the 2026 S-1 replaces that concentration with OpenAI on three legs at once: customer, lender, and warrant-holder of up to 12% fully diluted. The IPO does not diversify counterparty risk, it substitutes it. A commodity producer with one buyer, one financier, and one equity-aligned counterparty is exposed on every leg of the capital stack to the same name, a lesson energy producers learned the hard way through the 2000s. The deeper logic is a swap of surpluses. OpenAI has equity capacity but compute scarcity; Cerebras has compute capacity but capital scarcity. The contract trades each side’s surplus for the other’s shortage. That is why this template works for less-capitalised producers and not just well-capitalised ones.
The second risk is workload, and contrary to the standard framing, it sits with OpenAI rather than Cerebras. The Cerebras edge is on low-latency inference, and the economics work best on models that fit entirely in on-wafer memory. The hardware runs larger models through external memory and multi-chip scaling, but the cost structure compresses with scale: 70B already requires four CS-3s, and 400B-class deployments need many more. The same megawatts deliver fewer effective users at frontier scale. If frontier inference shifts toward scales that erode those economics, the contract still pays out: Cerebras delivers chips and power, OpenAI gets the capacity it bought, but OpenAI is left with capacity optimised for a workload mix that may be the wrong one by 2028. This is the buyer-side problem in a tolling structure, where oil and gas refineries get paid for processing capacity regardless of whether the feedstock turned out to be the right grade. The Cerebras-OpenAI contract operates the same way.
Cerebras does carry an adjacent risk: renewal. This contract executes regardless, but the next contract, and every customer after, depends on the wafer remaining competitive for the workloads of the late 2020s. Equity value sits in that question, not in this contract settling as written. That is the model-size-ceiling debate framed correctly: workload-fit risk for OpenAI, renewal risk for Cerebras, not a delivery risk on the prepaid forward.
From bilateral to exchange
The valuation rests on a portfolio of forward contracts and prepaid deliveries with one counterparty. That is the structure of a commodity producer with a long-term offtake agreement, not the structure of a merchant chip company selling boxes. Other analysts have made the broader observation that Cerebras should be priced as infrastructure rather than as a chip-specialist. This piece pushes the point further: the specific instruments the S-1 reveals are commodity-finance instruments with forty-year precedents, not generic infrastructure offtake. This is not a complaint about Cerebras, it is recognition that compute has reached the stage where commodity finance becomes the natural funding structure. Capacity-redemption loans, capacity forward strips, and capacity-tied warrants are the instruments, and they will repeat. The next inference provider that signs a multi-billion capacity contract will reach for the same toolkit, and the next hyperscaler funding a buildout against a committed offtake will too. The template has already started flexing. AMD signed a warrant-for-capacity deal with OpenAI in October 2025 without a working capital loan, because AMD did not need the cash. Cerebras needed both. The loan is the anchoring mechanism for less-capitalised producers; the warrant is the alignment mechanism. Both appear in different combinations depending on the producer’s capital position.
The Cerebras S-1 is the proof of concept. The next milestone will be when these structures clear in standardised, exchange-traded form. CME has not listed a compute future, ICE has not opened a capacity benchmark, and bilateral and lender-buyer structures remain the state of play. But bilateral prepaid forwards always come before standardised futures. They did in oil, in gas, and in metals. They are doing it now in compute.
very interesting! compute just became fungible enough to trade like barrels