On a podcast, a venture investor put a $700 billion all-in capital figure against the $350 billion revenue target he cites — unaudited, unverified, and structurally more interesting than either number alone.
On the 20VC podcast, hosted by Harry Stebbings alongside Jason Lemkin, Rory O’Driscoll of Scale Venture Partners offered two figures that framed OpenAI’s capital position in terms drawn from his own analysis rather than from any company disclosure. O’Driscoll claimed the business“is going to consume, directly or indirectly, $700 billion worth of capex to get there — to get to $350 billion in revenue,” and that the $350 billion is his figure for a target, not a reported result. Every number in that sentence is his analytical estimate, made on a podcast, without reference to any OpenAI financial disclosure, audited statement, or verified contract.
The mechanism O’Driscoll described is commercially ordinary. He argued that “the only reason they’re able to do it — only burning 278 billion — is because other companies like Oracle, like NVIDIA, with kind of rev support and kind of backstop insurance, are able to say we’ll do the capex and lease it to you.” Leasing compute infrastructure from suppliers, vendor revenue support, and backstop arrangements are standard commercial structures across capital-intensive industries. Nothing in his description alleges any concealment, impropriety, or accounting irregularity by any party.
What makes the claim analytically durable is that O’Driscoll did not try to hide the measurement difficulty. His own summary — “if all that burn, all 700 billion of it, had to appear on the balance sheet… it would be even worse than that” — is a conditional, not a finding. It restates a structural property of lease accounting rather than a discovered discrepancy. The conclusion he reached was simpler and more portable than any single figure: “Intelligence is not cheap.”
The key insight: O’Driscoll’s most load-bearing phrase is “directly or indirectly” — a qualifier that concedes the measurement boundary is itself an estimate. When an analyst must write those words, the error bar sits on the definition of what is being measured, not only on the arithmetic. More data does not fix a definitional problem.
The gap is capex that is being paid for and does not appear where the spending happens.
The Structural Read
The interesting thing about this episode is not the scale of the numbers. It is the absence of a shared denominator for off-balance-sheet AI capex across the industry — and what that absence reveals about how the current infrastructure buildout is actually structured.
When a cost moves off a spender’s balance sheet under a lease arrangement, it does not move onto a shared ledger. It moves onto whichever counterparty holds the asset, and that counterparty reports it in its own category, on its own reporting calendar, under its own definition. Three independent observers described the same general mechanism this week and produced three figures that cannot be placed in the same table — because they describe different parties, different scopes, different periods, and different definitions. No total is computed here, and none honestly can be computed from those inputs. Adding them would produce a number that means nothing.
This is the denominator problem at its largest scale. A balance sheet measures what a company owns, not what a company uses. Those are genuinely different economic positions with different risk profiles, and it is correct that accounting conventions treat them differently. The difficulty is not that the convention is wrong. It is that a reader who wants to know how much capital an activity requires is asking a question the convention was never designed to answer.
In the Map of AI, the compute infrastructure layer is distinct from the model layer and the application layer. O’Driscoll’s framing is essentially a claim about how capital from the application layer (OpenAI’s business) is being financed through the infrastructure layer (supplier-held assets), with the model layer sitting in between. The measurement problem is inseparable from this vertical structure: each layer reports its own economics, and the full-stack capital requirement has no natural row in any single set of accounts.
The same structural shape has appeared at smaller scales in AI product reporting: a model whose published price per token holds flat while tokens consumed per task rise; a hardware product whose stated weight describes the visible device but not the compute unit beside it. In each case, a metric describes the object it is attached to, and moving a cost off that object improves the metric without removing the cost. O’Driscoll’s argument is that a balance sheet is the same kind of object: bounded by ownership, not by use. And it is worth being fair about this, because the obvious reading is the wrong one. That is not a loophole and it is not an evasion — it is what a balance sheet is for. Owning an asset and using one are genuinely different economic positions carrying different risks, and it is correct that they are accounted for differently. The difficulty is not that the convention is wrong. It is that a reader who wants to know how much capital the activity requires is asking a question the convention was never built to answer.
Three Implications
IMPLICATION 1 — THE DENOMINATOR IS UNDEFINED, NOT JUST UNKNOWN
The aggregate capital requirement for AI infrastructure is not merely unknown because no one has counted it. It is undefined — there is no single row it would go in across the multiple sets of accounts that together describe the activity. That is a harder problem than data quality. Disclosure reform or analyst effort can narrow an unknown; only definitional standardization can address an undefined metric, and that does not yet exist for off-balance-sheet AI capex at industry scale.
IMPLICATION 2 — SOFTWARE INTUITIONS DO NOT TRANSFER
The valuation and competitive frameworks the industry built on software businesses — high margin, low incremental cost, capital-light — were calibrated on a different capital structure. A business whose all-in capital requirement is anywhere in the territory O’Driscoll described is structurally different: different in how it funds itself, what it owes to counterparties, and how long it takes to discover whether the spending worked. That conclusion survives regardless of which specific figure proves closest to reality.
IMPLICATION 3 — ANALYST HONESTY HAS A COST IN QUOTABILITY
O’Driscoll’s use of “directly or indirectly” is the intellectually honest move — it flags that the two routes are not separately observable from outside the company. But that qualifier is also why a single confident headline number is more likely to circulate than his actual framing. The more precise the analyst, the less portable the claim. That asymmetry shapes how AI capital estimates propagate through media, and it is worth tracking as a systematic distortion, not a one-off imprecision.
Business Engineer Framework
The Map of AI — Where Capital Sits in the Stack
The Map of AI tracks 200+ companies across nine layers of the AI stack — from raw compute infrastructure through model development to application deployment. Understanding which layer holds a given cost, and how capital flows vertically across layers through leasing and backstop arrangements, is the prerequisite for reading any AI capital estimate honestly. O’Driscoll’s framing is a live case study in why the infrastructure layer’s economics cannot be read off any single company’s balance sheet.
Rory O’Driscoll’s $700 billion figure — an unaudited, unverified podcast estimate that is not an OpenAI disclosure and cannot be checked against any other published number — matters less as a data point than as a diagnostic: it surfaces a measurement architecture in which the full capital cost of building frontier AI has no natural home in any single set of accounts, is distributed across counterparties who each report it differently, and therefore cannot be honestly aggregated by any outside observer. That is the finding. The specific numbers may narrow or widen as more information becomes available. The structural gap in the denominator closes only if the industry agrees on what it is counting, and no such standard is visible in any of the three accounts described here.
Source and editorial note: All figures ($700B, $278B, $350B) are Rory O’Driscoll’s claims made on the 20VC podcast (September 2026). They are not OpenAI financial disclosures, are not audited, and are not verified here. The $350B figure is O’Driscoll’s estimate of a revenue target, not a reported result. These figures are not comparable to any other published estimate; no total is computed in this article, and none honestly can be. No wrongdoing, concealment, accounting irregularity, or regulatory non-compliance is alleged against any party. Leasing capex from suppliers, revenue support, and vendor backstop arrangements are ordinary commercial structures. Nothing in this article constitutes investment advice. No outcome is predicted.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
Single source. Every figure above is Rory O’Driscoll’s claim on the 20VC podcast — not an OpenAI disclosure, not audited, and not verified here. The $350 billion is his figure for a target, not a reported result. These figures are not comparable to any other published capex or backstop number, including the Nvidia backstop-universe figure covered separately on this site — different party, different scope, different source. No total is computed above, and one cannot honestly be computed from inputs of this kind.No wrongdoing is alleged against anyone. Leasing capex from a supplier, revenue support and vendor backstops are ordinary commercial structures, and accounting for what a company owns separately from what it uses is what a balance sheet is for — nothing above suggests concealment, impropriety, irregularity or non-compliance by OpenAI, Oracle, Nvidia or any other party. This is analysis of financing structure. It is not investment advice, it expresses no view on any company or security, and it predicts nothing. OpenAI’s, Oracle’s and Nvidia’s financials, any real contract, lease, term or counterparty, the contractual meaning of “rev support” and “backstop insurance”, any date for either figure and any company comment are not established and do not appear above.
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.
