Projections shared as part of a major computing deal — reported by the Financial Times — reveal more about the structure of AI infrastructure commitments than about any single number inside them.
What Happened
The Financial Times reported on 18 September 2026 — with Bloomberg and Techmeme carrying the story — that materials shared in July as part of a major computing deal show OpenAI projecting negative free cash flow of $278 billion between 2026 and the end of 2030. Those same materials, as reported, project revenue rising from $36 billion this year to $350 billion in 2030, and show spending of approximately $856 billion on computing power and infrastructure by the end of 2030, identified as the largest expense category. Every figure cited here is a projection drawn from those leaked confidential materials as reported by a news organisation. None is a reported financial result, none is audited, and none constitutes an official disclosure by the company — including the $36 billion, which is a projection for the current year rather than a reported figure.
OpenAI raised $122 billion in March of this year at an $852 billion valuation — that valuation is the price at which the March raise was done, as reported, and no share price or market capitalisation exists or is stated here, as OpenAI is a private company. The same reporting indicates the company is on track to exhaust that cash by 2028. That is a projection of one modelled path, not a prediction of events.
The provenance of these materials is the first and most important fact to hold. They were shared in July as part of a major computing deal. No counterparty or deal is named here, nothing here speculates about who received them, and nothing here claims what the deal was. But that origin — shared so that another party could decide whether to commit capacity over a period of years — is what determines what kind of document this is, and therefore what properties it has.
The key insight: A projection circulated as part of a major computing deal is an underwriting document, not a forecast in the ordinary sense. The question it is built to answer is not whether the business is attractive — it is whether a multi-year commitment will be honoured. Documents built for different purposes have different properties: different things get stress-tested, different things get stated conservatively, different things appear at all. Holding that distinction is the beginning of reading these numbers correctly.
The Structural Read
Four structural observations follow from these materials. They are stated as observations about document type, contract structure, spending constraints, and capital planning. None of them is a claim that any number is optimistic, pessimistic, inflated, conservative, misleading, or unrealistic, and none makes any claim about anyone’s intent in preparing them.
FDE Framework — Founder Layer
The Underwriting Document vs. the Forecast
A forecast exists to describe an expectation. An underwriting document exists so that a counterparty can decide whether a commitment will be honoured. These are different epistemic objects. The question an infrastructure supplier asks — will this company be able to pay for what it commits to take? — is answered by a cash-flow projection, not a revenue narrative. That is the question these materials were apparently prepared to address, and it shapes every number inside them in ways that are not visible from the numbers alone.
First: document type. The materials were shared as part of a computing deal, which means their audience was not a general investor weighing upside. Their audience was a party deciding whether to commit capacity — infrastructure, power, chips — over a multi-year horizon. The relevant question for that audience is credit quality, not growth narrative. A document prepared for that question will emphasise cash-flow visibility, commitment schedules, and the ability to honour obligations. That is a different emphasis than a pitch deck or an earnings presentation, and it does not make the numbers more or less reliable — it makes them differently structured.
Second: both sides of the contract are now visible. Infrastructure suppliers have increasingly presented long-term take-or-pay contract books as their principal asset, on the reasoning that a counterparty owes whether or not it takes delivery — converting the counterparty’s future obligation into the supplier’s balance-sheet strength. This week, for the first time, a document that appears to represent the other side of that class of arrangement has become visible through reporting. The structural observation is narrow: a take-or-pay book is only as good as the credit standing behind it, and a five-year cash-flow projection is precisely the document that describes credit standing. This is an observation about where two sets of disclosures meet. No claim is made that any contract will or will not be performed, no credit assessment is offered, no supplier or contract is named, and no claim is made that any supplier’s counterparty is this company.
Structural Observation
“A take-or-pay contract book is an asset on one balance sheet and a liability on another. The quality of the asset depends entirely on the credit of the liability-holder. A five-year free-cash-flow projection is the most direct description of that credit. Both documents now exist in the same week’s reporting.”
Third: the largest number is the least discussed. The projected spending of approximately $856 billion on computing power and infrastructure through the end of 2030 is described in the materials as the largest expense category. This figure is a cumulative total across a multi-year period. The projected 2030 revenue of $350 billion is a single-year figure. These two numbers cover different spans and are not comparable with one another; no comparison is drawn here, and combining them into a ratio, multiple, payback period, or any other derived figure would be arithmetic on incompatible quantities. What can be stated without any arithmetic at all is this: when a document explicitly names a spending line as the largest expense category, it states plainly which variable the plan is built around. The constraint being managed here is the cost of capacity. The revenue line in these materials rises steeply. It is not the part of the model under strain.
Fourth: one figure implies a deadline. The reported projection that the $122 billion raised in March is on track to be exhausted by 2028 is a statement about one modelled path. It is not a prediction that the company will run out of money, will need to raise capital, will fail, or will take any particular action; no future fundraising, listing, terms, or valuation is predicted here. The structural observation is narrower: a five-year plan whose internally modelled cash reaches 2028 is a plan that treats continued access to capital markets as an operational input — in precisely the same way it treats access to power and to chips as inputs. Each of those inputs has a supplier, a price, and an availability that can change independently. A plan of this shape depends on all three in parallel.
Three Implications
IMPLICATION 1 — THE PROVENANCE PROBLEM IS PERMANENT
Every set of numbers that surfaces from private-company AI operations will carry a provenance — the purpose for which it was prepared, the audience it was prepared for, and the question it was built to answer. Those properties determine what the numbers mean more than the numbers themselves. Investors, analysts, and infrastructure counterparties reading leaked materials without holding the document-type distinction will systematically misread what they are looking at. The materials reported this week were prepared to answer a credit question. That is the lens they require.
IMPLICATION 2 — CAPACITY COST IS THE BINDING CONSTRAINT, NOT DEMAND
A plan that names computing power and infrastructure as its largest expense category — at a projected scale of approximately $856 billion cumulatively through 2030, a figure that is not comparable to any single-year revenue number and is not combined with one here — is a plan organised around supply-side constraints. The question this document is actually answering is not “how fast will revenue grow?” The revenue projection rises steeply, and by the document’s own framing, that is not the hard part. The hard part is securing, financing, and honouring the capacity commitments that enable it. That inversion — demand is the easy assumption, supply is the binding variable — is the correct frame for reading the entire AI infrastructure build-out, not just this document.
IMPLICATION 3 — CAPITAL MARKETS ARE AN INFRASTRUCTURE INPUT
Treating a projected 2028 cash exhaustion as a modelling curiosity misses the structural point. A five-year operational plan that models capital access as an input alongside power and silicon is a plan whose risk profile is the intersection of three supply chains simultaneously: compute, energy, and finance. Each of those has its own pricing dynamics, availability cycles, and counterparty dependencies. The plan is robust only if all three remain accessible on terms the model assumes. That is not an argument that any of them will fail — it is a description of the plan’s shape, and it is the most useful single observation for anyone trying to understand what kind of bet the AI infrastructure build-out actually represents.
The Bottom Line
The numbers in these materials — each of them a projection from leaked confidential documents reported by a news organisation, none audited, none an official disclosure — will be debated for their scale and their trajectory. The more durable analytical move is to hold what kind of document this is before debating what it says: a set of projections prepared so that a capacity counterparty could decide whether a multi-year commitment would be honoured, now legible alongside the supplier-side disclosures that treat such commitments as their principal asset. The constraint this document describes is not revenue growth — the revenue line rises steeply and is, by the document’s own framing, the assumed part. The constraint is the cost, availability, and financeability of capacity itself. That is the sentence the $856 billion is writing, and it is a sentence about the entire shape of the AI infrastructure build-out, not only about one company’s five-year plan.
This article is business analysis of figures reported from leaked confidential materials. It is not investment advice, no view is expressed on any security, and no recommendation of any kind is made.
Sources: Financial Times — OpenAI
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity. Every figure above is a projection drawn from leaked confidential materials reported by the Financial Times and carried by other outlets. None is a reported result, none is audited, and none is an official disclosure — including the current-year revenue figure, which is a projection rather than a reported number. The materials were shared in July as part of a major computing deal. No counterparty or deal is named above, nothing above speculates about who received them, and nothing above claims what the deal was. The cumulative figures and the single-year figure cover different spans and are not comparable with one another; no ratio, multiple, payback period or other derived figure is computed above. Nothing above characterises these projections as optimistic, pessimistic, inflated, conservative, misleading or unrealistic, or makes any claim about anyone’s intent in preparing them. A projection of cash exhaustion describes one modelled path. Nothing above says the company will run out of money, will need to raise, will fail or will do anything, and nothing above predicts any future fundraising, listing, terms or valuation. Nothing above offers a credit assessment of any company, claims that any contract will or will not be performed, names any supplier or contract, or claims that any supplier’s counterparty is this company. OpenAI is a private company; the valuation cited is the one at which the March raise was done as reported, and no share price or market capitalisation exists or is stated. This is business analysis. It is not investment advice, no view is expressed on any security, and no recommendation is made.More Resources
