OpenAI Isn't Short $1.4 Trillion — It's Long a Free Option. The Duration Mismatch Is on Oracle's, Broadcom's and CoreWeave's Books.

A $1.4 Trillion Liability That Exists in Exactly One Sentence

Every version of this story begins with the same number, and the number has exactly one parent. On 6 November 2025, Sam Altman wrote a public post containing the sentence: "We are looking at commitments of about $1.4 trillion over the next 8 years." [24] That is the entire provenance. The verb is "looking at." The qualifier is "about." The horizon is eight years — the same figure rendered elsewhere as Altman having "stated it would be spending $1.4 trillion on compute by 2033." [27] It has never been itemised in a filing, never audited, never reconciled line by line by anyone, anywhere. And no counterparty's SEC filing carries a matching obligation.

Three months later, OpenAI told its own investors something else. In February 2026 the company said it is targeting "roughly $600 billion in total compute spend by 2030" — a plan CNBC reports was "meant to more directly tie to its expected revenue growth." [22] DataCenterDynamics independently reports that OpenAI "has scaled back its compute spending plans to $600 billion over the next four years." [27] Management re-anchored the headline downward by more than half, in a quarter, and not one supplier has disclosed a termination charge, a penalty, or an impairment as a result.

That fact alone should reorganise how you read the rest of this. The received wisdom about OpenAI has hardened into a single sentence of its own: OpenAI has a balance-sheet duration mismatch — fixed, long-dated compute obligations funded by cancellable, consumption-based revenue. Four load-bearing words. We tested each one against the primary documents. Three of them fail.

"Balance-sheet" fails first, and fastest. Nothing in the public record puts these commitments on OpenAI's balance sheet. What the record does contain is a company that held more than $73 billion in cash and marketable securities at the end of the first quarter of 2026, up from $40 billion at the end of December — a step-up that came from the $122 billion round closed on 31 March 2026, not from operations. [31] [12] Its committed bank liquidity is a revolving credit facility of approximately $4.7 billion, undrawn at close. [12] There is no disclosed balance-sheet debt. A duration mismatch is a funding-structure pathology: long assets rolled on short funding, with rollover risk at every maturity. Permanent equity plus a $73 billion cash pile is the structural opposite of that. Whatever OpenAI's problem is, it is not that it has to refinance next Tuesday.

"Fixed" fails next, and it fails item by item, from the documents that carry legal liability. "Cancellable, consumption-based revenue" partly survives — and is moving against the thesis, because management is publicly terming the revenue book out. And "duration mismatch" turns out to be real, filed, dated and non-cancellable — on the suppliers' books, not OpenAI's, which is exactly where the Bank of England and the Bank for International Settlements say it is being warehoused and mispriced.

The inversion, stated plainly: OpenAI is not short $1.4 trillion of fixed obligations. It is long a very large, very cheap option on compute, which it has already begun exercising by not exercising. The fixed leg of this trade sits with Oracle, Broadcom, CoreWeave — and with the lenders who financed them.

Read the Deals as Documents: "Intends", "Term Sheet", "Up To", "Contracted"

There is a test for whether a commitment is a commitment, and helpfully, one of OpenAI's own suppliers wrote it down. Broadcom's 10-Q for the quarter ended 3 May 2026 defines a purchase commitment as an "unconditional purchase commitment to purchase goods or services… that are enforceable and legally binding on us and specify all significant terms, including fixed or minimum quantities to be purchased, price provisions, and the approximate timing of the transaction. Purchase commitments exclude agreements that are cancelable without penalty." [9] Enforceable. Legally binding. Fixed quantities. Price. Timing. Not cancellable without penalty.

Run the announcement set through that filter and most of the headline evaporates.

NVIDIA: a letter of intent, disowned by its own CEO, and pointing the wrong way

The September 2025 announcement is explicit about what it is: "NVIDIA and OpenAI today announced a letter of intent," under which "NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed" toward "at least 10 gigawatts of NVIDIA systems." [14] Three hedges in one clause — intends, up to, and conditioned on deployment — and the release closes by saying the parties "look forward to finalizing the details of this new phase of strategic partnership in the coming weeks." [14]

Five months later, in the document that carries Section 18 liability, NVIDIA's 10-K for the fiscal year ended 25 January 2026 says: "We are finalizing an investment and partnership agreement with OpenAI. There is no assurance that we will enter into an investment and partnership agreement with OpenAI or that a transaction will be completed." [8] The filing carries no OpenAI purchase obligation and no OpenAI-related liability. And in February 2026, Jensen Huang said it out loud: "It was never a commitment. They invited us to invest up to $100 billion and of course, we were, we were very happy and honored that they invited us, but we will invest one step at a time." [26] The Wall Street Journal reported he had privately emphasised the agreement "was nonbinding." [26]

And then the category error, which is the part almost nobody catches: the $100 billion is NVIDIA investing in OpenAI. It is a potential equity inflow. Every tabulation that adds it to OpenAI's outbound compute commitments is counting a possible source of funds as a use of funds. [14] [26] It should never have been in the column at all.

Broadcom: a term sheet, and no price exists

The joint release of 13 October 2025 is one sentence long where it matters: "The two companies have signed a term sheet to deploy racks incorporating the AI accelerators and Broadcom networking solutions," covering "10 gigawatts of custom AI accelerators," targeted to start in the second half of 2026 and complete by end of 2029. [16] No dollar value appears anywhere in the release. None. Any Broadcom dollar line in a $1.4 trillion tabulation — and in most of them it is the single largest — is a third-party extrapolation from gigawatts to money, performed by someone who is not a party to the contract.

Meanwhile, Broadcom's own commitments table is not empty. It shows $128,110 million of Broadcom's own unconditional, non-cancellable purchase commitments as of 3 May 2026 — $55,214 million falling in fiscal 2027, $72,870 million in fiscal 2028. [9] Broadcom's 10-Q does not name OpenAI anywhere and discloses no OpenAI backlog. [9] Sit with that for a second: the supplier has booked $128,110 million of hard, enforceable, legally binding obligations to its suppliers, and has booked nothing enforceable from OpenAI against them.

AMD: one gigawatt of six, and an option structure for the rest

AMD's is the cleanest binding/non-binding split on the record, because AMD's own 8-K draws it. The press release announces "a 6 gigawatt agreement." [7] The 8-K says: "Concurrent with signing, Warrantholder agreed to a binding commitment to purchase (directly or through its affiliates or Authorized Purchasers) the initial one (1) gigawatt of AMD Instinct MI450 Series GPU products." [6] One. Of six.

The other five gigawatts are not bought with a contract; they are bought with equity. AMD issued OpenAI a warrant "to purchase up to an aggregate of 160 million shares of common stock of the Company… at an exercise price of $0.01 per share," vesting in tranches only as purchases scale to six gigawatts, and "further subject to achievement of specified Company stock price targets that escalate to $600 per share for the final tranche." [6] OpenAI is being paid, in penny-strike equity, to buy. That is an incentive, not a liability. And the size check comes from AMD's own CFO, who described the entire six-gigawatt partnership as "expected to deliver tens of billions of dollars in revenue for AMD" — tens of billions, not hundreds. [7]

Oracle and Microsoft: the biggest lines, and neither is in an account

Oracle's $300 billion is the largest single line item in most tabulations, and it is a press report. WSJ reporting relayed by DataCenterDynamics: "OpenAI has reportedly agreed to buy compute power worth $300 billion from Oracle in a five-year-deal," and "the contract will start in 2027." [28] Neither Oracle nor OpenAI has ever confirmed that figure in a filing or a press release. And because it starts in 2027, essentially none of it is a current-period cash obligation.

Microsoft's $250 billion has the strongest verb in the entire complex, and both principals use it in identical words: "OpenAI has contracted to purchase an incremental $250B of Azure services, and Microsoft will no longer have a right of first refusal to be OpenAI's compute provider." [17] [18] "Contracted" appears nowhere else across the deal set. Two principals choosing that word in matching language is strong evidence a contract exists. But it is unverifiable in the accounts: Microsoft's 10-Q for the quarter ended 31 March 2026 books no such purchase obligation as a receivable, an RPO item, or a contract asset. What it does disclose is an equity-method stake of approximately 27 percent, and "total funding commitments of $13 billion, of which $11.8 billion has been funded as of March 31, 2026." [5] No term, no schedule, no minimum draw, no cancellation terms have ever been disclosed for the $250 billion. It therefore cannot be placed on a duration axis at all — which is fatal for a report whose entire subject is duration.

CoreWeave: the only filed dollar figures — and even they say "up to"

One counterparty, and one only, has ever put a dollar figure on an OpenAI commitment inside an SEC filing. CoreWeave's 10-K: "in March 2025, we entered into a master services agreement with OpenAI, a private company, pursuant to which OpenAI has committed to pay us up to approximately $11.9 billion through October 2030." [4] And a second, September 2025 order form under which "OpenAI has committed to pay us up to approximately $6.5 billion through May 31, 2031." [3] [4]

The operative words are up to. And CoreWeave's own filings tell you exactly what "up to" means, because the parallel Meta disclosure spells it out: Meta "initially committed to pay us up to approximately $21 billion (inclusive of access to new computing capacity through December 20, 2032 and the exercise of an existing option to access additional computing capacity through April 10, 2032)." [3] Headline contract value includes optional expansions. It is a ceiling, not a floor.

Add the two OpenAI figures — our arithmetic, not a disclosure — and you get $18.4 billion. That is the total of every OpenAI compute commitment ever quantified by name in any SEC filing by any counterparty on earth. Against the $1.4 trillion headline, it is roughly 1.3%. Both components are ceilings, and Forbes reports that CoreWeave contracts worth $22.4 billion "contain termination clauses allowing either party to exit 'for cause' (e.g. delays), giving OpenAI exit mechanisms." [33]

Counterparty	Announced	The operative words	Binding on OpenAI?	In a counterparty's SEC filing?
NVIDIA	Up to $100bn / at least 10 GW	"a letter of intent"; "intends to invest up to"; "as each gigawatt is deployed"	No — and it is an INFLOW to OpenAI, not an outflow. Huang: "It was never a commitment."	No obligation. FY2026 10-K: "no assurance that we will enter into an investment and partnership agreement"
Broadcom	10 GW — no dollar figure exists	"The two companies have signed a term sheet"	No purchase contract disclosed; any dollar line is a third-party extrapolation from gigawatts	OpenAI not named; no OpenAI backlog. Broadcom's own purchase commitments: $128,110m
AMD	6 GW agreement	8-K: "a binding commitment to purchase… the initial one (1) gigawatt"; the rest rides on a $0.01-strike warrant vesting to a $600 share price	1 of 6 GW. The other five are incentivised, not obliged	Yes — the 8-K, but no dollar figure. CFO sizes the whole 6 GW at "tens of billions"
Oracle	$300bn over five years, starting 2027	"reportedly agreed" — WSJ reporting; never confirmed by either party	Unknown; starts 2027, so essentially no current-period cash obligation	Never named. RPO rose $138bn → $638bn, attributed only to "certain significant cloud contracts"
Microsoft	$250bn of incremental Azure services	"OpenAI has contracted to purchase" — the strongest verb in the set, used identically by both principals	Probably a contract exists — but no term, schedule, minimum draw or cancellation terms disclosed anywhere	Not booked as receivable, RPO or contract asset. 10-Q discloses only ~27% equity stake and "$13 billion" of funding commitments
AWS	$38bn over seven years; capacity targeted before end-2026	"Representing a $38B commitment" / "this new $38 billion agreement"	Called a commitment and an agreement — not an LOI or term sheet	No Amazon filing in our source set corroborates the figure
CoreWeave	$11.9bn (to Oct 2030) + $6.5bn (to May 2031)	"OpenAI has committed to pay us UP TO approximately…" — a ceiling, as CoreWeave's parallel Meta disclosure makes explicit	Hardest evidence on the record — and capped, with Forbes reporting "for cause" exits across $22.4bn of contracts	Yes — the only counterparty that names OpenAI with a dollar figure, inside its insolvency/bankruptcy credit-risk factor
Stargate	$500bn over four years	"a new company which intends to invest"; SoftBank has "financial responsibility", OpenAI "operational responsibility"	No — and already trimmed: Texas, UK and Norway sites scaled back, no penalty reported	Not an OpenAI balance-sheet obligation in any filing

Build the Fixed Leg Yourself: The Only Honest Way to Size This

We are not going to hand you a number and ask you to trust it. We are going to hand you the switch.

The size of OpenAI's fixed leg is not a fact you discover; it is a function of the evidentiary standard you are willing to accept. Set the bar at quantified by name in an SEC filing and the answer is $18.4 billion — CoreWeave, twice, both "up to." Set it at a principal calls it contracted or a commitment and you get Oracle's press-reported $300 billion, Microsoft's $250 billion, the AWS $38 billion "commitment," and CoreWeave's $22.4 billion aggregate. [28] [17] [15] [33] That is our sum, and it is our arithmetic and not anybody's disclosure: roughly $610 billion.

Look at where that lands. It lands within a rounding error of the "roughly $600 billion in total compute spend by 2030" that OpenAI itself now guides investors to. [22] Two completely independent routes — bottom-up from the strongest available deal language, and top-down from what management privately tells its own shareholders — converge on the same figure. That convergence is the single most persuasive number in this report, and it is not the headline.

To get from there to $1.4 trillion you must do four specific things, and it is worth naming them because they are the whole argument. You must count NVIDIA's $100 billion inbound investment as an OpenAI outflow. [14] You must invent a dollar price for Broadcom that Broadcom never stated and that exists in no source. [16] You must bind five gigawatts that AMD's 8-K explicitly did not bind. [6] And you must move SoftBank's balance sheet onto OpenAI's, because the Stargate release says the project "intends to invest $500 billion over the next four years," in a separate company, where "SoftBank [has] financial responsibility and OpenAI [has] operational responsibility." [13] Do all four and you reach the headline. Do none of them and you reach management's own plan.

Quantified by name in an SEC filingDefensible fixed leg (our base case)Rebuilding the $1.4tn headlineReset

Oracle — press-reported, starts 2027estimate300$bn

Microsoft Azure — "contracted", unbookedestimate250$bn

AWS — "a $38B commitment"estimate38$bn

CoreWeave — filed by name ($18.4bn) to Forbes' aggregate ($22.4bn)fact22.4$bn

NVIDIA — letter of intent (and an INFLOW to OpenAI)assumption0$bn

Broadcom — term sheet, no dollar figure exists anywhereassumption0$bn

AMD — the five gigawatts the 8-K did NOT bindassumption0$bn

Stargate — "intends to invest", SoftBank has financial responsibilityassumption0$bn

Defensible fixed leg610.4$bn

Total once you add the soft items610.4$bn

Distance still to travel to the $1.4tn headline789.6$bn

Fixed leg ÷ management's own $600bn plan to 20301.02×

Years of current revenue (~$22.8bn annualised) to cover the fixed leg26.8×

Our arithmetic, not a disclosure. The base case sums only the line items a principal calls "contracted"/"a commitment" or a filing quantifies: ~$610bn — within a rounding error of the "roughly $600 billion in total compute spend by 2030" OpenAI now guides investors to. To reach $1.4tn you must count NVIDIA's inbound investment as an OpenAI outflow, invent a price for Broadcom that exists in no source, bind five gigawatts AMD's 8-K explicitly did not bind, and move SoftBank's balance sheet onto OpenAI's. The Broadcom and AMD scenario values in "Rebuilding the $1.4tn headline" are our own inventions, shown precisely to demonstrate that inventing them is what the headline requires.

Now run coverage against it. Q1 2026 revenue was $5.7 billion. [30] [31] Annualise it — our arithmetic — and you get roughly $22.8 billion, which is consistent with OpenAI's own statement on 31 March 2026: "We are now generating $2B in revenue per month." [12] Against a defensible fixed leg of roughly $610 billion, that is about twenty-seven years of current revenue. OpenAI's answer is growth: it projects more than $280 billion of 2030 revenue, roughly evenly split consumer and enterprise. [22] Against $13.1 billion of recognised 2025 revenue, that is about a 21x increase in five years — again, our arithmetic on OpenAI's own two figures. [22] Forbes, as an outside estimate, puts the revenue needed to sustain all of the announced commitments at approximately $577 billion by 2029. [33]

These are large numbers and we are not going to pretend otherwise. But notice which question they answer. They answer can OpenAI grow into this. They do not answer is OpenAI legally on the hook for this, and the second question is the one everyone has been assuming the answer to.

The One Leg That Survives — and It Is Moving Against the Thesis

Give the received wisdom its strongest form, because in one place it is right.

OpenAI's revenue book is about as short-dated as a revenue book gets. Its own Services Agreement, effective 1 January 2026, provides that "Notice of non-renewal or scope reduction must be given at least thirty days before the start of the next Renewal Term," and that "If Customer reduces its license count, quantity, or minimum commitment, OpenAI may adjust or remove discounts offered to Customer based on its prior purchase." [19] Read that twice. An enterprise customer can not only walk on thirty days' notice at renewal — it can shrink its minimum commitment, and the only remedy stated anywhere in the published terms is that OpenAI claws back a discount. That is not a take-or-pay contract. That is a subscription.

Consumer is worse, structurally. OpenAI reports "more than 900 million weekly active users, and over 50 million subscribers." [12] Our arithmetic on OpenAI's own two numbers: that is a paid conversion of roughly 5.6%. Something like 94% of the user base contributes no subscription revenue at all, and the 50 million who do are on month-to-month terms. Meanwhile the API is metered by usage — the most cancellable revenue there is, since a customer cancels simply by making fewer calls.

So: short-dated, thin, and consumption-based. The hypothesis is right about the shape of the revenue leg.

It is wrong about the direction of travel, and it is wrong in a way that matters, because the company is publicly dismantling the premise. Enterprise already "makes up more than 40% of our revenue, and is on track to reach parity with consumer by the end of 2026." [12] And in May 2026 OpenAI launched Guaranteed Capacity, which "includes certainty of access to compute based on spend levels, and customers can draw down from this commitment across the portfolio of OpenAI products" — sold as annual spending commitments spanning one to three years, with discounts scaling by commitment duration. [34] That is a company deliberately buying duration on the revenue side by paying for it with margin.

The launch is itself evidence. Management looked at the same asymmetry everyone else is writing about and shipped a product against it. A thesis whose central premise the subject is actively engineering away is not a thesis to print — it is a snapshot with an expiry date on it.

Two honest caveats, both of which cut our way and against us in turn. First, a one-to-three-year ceiling on the revenue side still does not reach cost commitments running to October 2030 and May 2031. [4] [3] The gap narrows; it does not close. Second, and this is the finding rather than the caveat: OpenAI has never published churn, gross or net revenue retention, cohort retention, or its own backlog, and there is no audited financial statement in the public record. Every retention percentage circulating about OpenAI traces to sources with no named methodology, and we have excluded all of them. The honest statement is that the renewal picture is undisclosed.

Which is the deeper asymmetry, and it is one of disclosure, not only of duration. The cost side of this trade is measured in filed, audited backlog with percentage-by-period runoff schedules — Oracle's $638 billion, CoreWeave's $98.8 billion. [1] [3] The revenue side is measured in blog-post run-rates: "We are now generating $2B in revenue per month." [12] In the same 6 November 2025 post that produced the headline, Altman said OpenAI expected "to end this year above $20 billion in annualized revenue run rate" and that the company "already had a million business customers." [24] No renewal, retention or net-revenue-retention figure accompanied it, then or since. When you compare an annualised exit run-rate to an audited multi-year backlog, you are not comparing two numbers. You are comparing two epistemic regimes.

OpenAI Is Long the Option — and the Scale Gap Is Still Real

Here is the structural claim, stated as sharply as the evidence allows.

An obligation that management can cut in half by fiat, in three months, with no counterparty disclosing a penalty, is not a fixed obligation. It is a demand call option that OpenAI is short — the right, not the duty, to take delivery of compute — and OpenAI has already demonstrated its willingness to let parts of it expire. It re-anchored the plan from the $1.4 trillion framing to "roughly $600 billion in total compute spend by 2030." [22] It scaled back Stargate projects in Texas, the UK and Norway. [29] Not one vendor has booked a termination charge or an impairment against any of it, anywhere in the public record.

And the counterparties have no incentive to make them. The named analyst Gil Luria of D.A. Davidson puts it as directly as it can be put: "They don't want OpenAI to go bankrupt, so their incentive is to renegotiate." [33] The named data-centre expert Daniel Golding adds the mechanism: "The 'big numbers' being announced are often larger than what's actually committed under contract, largely due to variables like share price, data center construction cost and GPU price." [33] Forbes reports that infrastructure partners "will most likely renegotiate the contracts and ensure they get at least some amount of business." [33]

Combine that with the funding structure. More than $73 billion in cash and marketable securities at the end of Q1 2026, up from $40 billion in December — permanent equity, not debt. [31] An approximately $4.7 billion revolver, undrawn. [12] No disclosed balance-sheet debt. There is no maturity wall here. There is nothing to roll. The word "balance-sheet" in the hypothesis is doing work the balance sheet does not support.

1. 6 Nov 2025

   The number is born, in one sentence[24]

   Sam Altman posts: "We are looking at commitments of about $1.4 trillion over the next 8 years." In the same post: "We expect to end this year above $20 billion in annualized revenue run rate." No itemisation, no filing, no audit — then or since.

2. 2 Feb 2026

   The largest line item is disowned by its counterparty[26][8]

   Jensen Huang: "It was never a commitment. They invited us to invest up to $100 billion… but we will invest one step at a time." NVIDIA's FY2026 10-K, filed weeks later, still says there is "no assurance" an agreement will be entered into at all.

3. 20 Feb 2026

   Management re-anchors to "roughly $600 billion"[22][27]

   OpenAI tells investors it is targeting roughly $600bn of total compute spend by 2030 — a plan "meant to more directly tie to its expected revenue growth". More than half the headline disappears in a quarter. No supplier discloses a termination charge, penalty or impairment.

4. 31 Mar 2026

   $122bn of permanent equity lands[12][31]

   The round closes at an $852bn post-money mark, anchored by Amazon, NVIDIA and SoftBank — the same names on the other side of the compute deals. Cash and marketable securities go from $40bn to more than $73bn. The revolver is ~$4.7bn and undrawn.

5. 3 May 2026

   Stargate sites trimmed; the CFO's question surfaces[29]

   WSJ reporting: OpenAI has scaled back Stargate projects in Texas, the UK and Norway, and CFO Sarah Friar warned leadership OpenAI "might not be able to pay for future computing contracts if revenue doesn't grow fast enough". Altman and Friar jointly: "This is ridiculous."

6. 20 May 2026

   OpenAI starts buying duration on the revenue side[34]

   Guaranteed Capacity launches: one-to-three-year spending commitments with duration-scaled discounts. The company is terming its revenue book out — the opposite of the direction the received wisdom assumes.

7. 8 Jun 2026

   The S-1 goes in, confidentially[11][10]

   Submitted in or around late May 2026, announced under Rule 135 on 8 June: "We recently submitted a confidential S-1… We have not decided on timing yet; it may be a while." The contractual-obligations table — the document that settles all of this — is not on EDGAR and will not be until at least 15 days before any road show.

Act two: the scale gap, which is real and enormous

None of the above makes OpenAI safe. It relocates the danger, and the relocated danger is severe.

What OpenAI has is not a duration mismatch. It is an operating-leverage and financing-gap problem of extraordinary size. Roughly $22.8 billion of annualised revenue — our arithmetic on the $5.7 billion first quarter — against a defensible fixed leg of roughly $610 billion. [30] Q1 2026 burn of $3.7 billion, more than half of revenue, on top of $8 billion burned across 2025. [30] [22] An operating loss of $9.3 billion in a single quarter, with stock-based compensation alone topping $2.3 billion. [30] Gross margin climbing, but only "from 33 to 39 percent." [30] A projected burn of $85 billion in 2028, per OpenAI's own projection as reported by TechCrunch. [25] And the physical check: announced capacity across NVIDIA, Broadcom, AMD and Oracle sums — our arithmetic — to about 30.5 GW, against an installed base that "ended 2025 with 1.9GW of compute capacity," having scaled from 200MW in 2023 and 600MW in 2024. [27]

The company knows. Per WSJ reporting relayed by DataCenterDynamics in May 2026, CFO Sarah Friar communicated concern to leadership that OpenAI "might not be able to pay for future computing contracts if revenue doesn't grow fast enough." [29] The same reporting says OpenAI missed internal targets for new users and revenue and "fell short of reaching one billion weekly active users for its ChatGPT service for year-end." [29] Altman and Friar rebutted jointly: "This is ridiculous. We are totally aligned on buying as much compute as we can and working hard on it together every day." [29] Note carefully what that rebuts. It denies internal misalignment. It does not deny the arithmetic.

And the funding loop is not clean. The round that set the $852 billion mark "was anchored by our strategic partners Amazon, NVIDIA, and SoftBank, with continued participation from our long-term partner, Microsoft" — the same names that appear on the other side of the compute deals. [12] CNBC reported Nvidia was "in discussions to invest up to $30 billion in OpenAI as part of the round that could value the company at a $730 billion pre-money valuation." [22] The BIS names the pattern in its own words: "the circular financing within the AI ecosystem." [21]

So the honest formulation is this. Can OpenAI grow into the plan? Genuinely unknown, and the required 21x by 2030 is the steepest sustained corporate growth path anyone has ever underwritten at this scale. But who eats it if it can't — that question has an answer, and the answer is the point of this report. Not OpenAI first.

Where the Duration Actually Sits: Oracle's Ladder, Broadcom's Own Book, CoreWeave's Lenders

A filed, dated, non-cancellable maturity ladder does exist in this story. Every element of the duration-mismatch thesis — the fixed leg, the long tenor, the legal enforceability, the financing against it — is present and documented. It is simply on the other side of the trade.

Oracle wrote our thesis into its own risk factors

Oracle's remaining performance obligations "were $638 billion and $138 billion as of May 31, 2026 and 2025, respectively," an increase Oracle attributes only to "certain significant cloud contracts that were entered into during the period." [1] Our arithmetic on Oracle's two disclosed figures: a $500 billion increase in a single year. OpenAI is never named.

The duration is disclosed, and it is long. Of the $638 billion, Oracle expects to recognise "approximately 12% as revenues over the next twelve months, 34% over the subsequent month 13 to month 36, 34% over the subsequent month 37 to month 60 and the remainder thereafter." [1] Our arithmetic on Oracle's own percentages: 88% of that backlog is more than a year out, and 54% is more than three years out. The shape held a quarter earlier — $552.6 billion of RPO at 28 February 2026 against $130.2 billion a year before, again with approximately 12% inside twelve months. [2] Almost none of this has been delivered or collected: Oracle states that "No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024." [1] Concentration in the backlog; none yet in the revenue.

And then Oracle, in its 10-K, writes the argument for us: "We enter into certain large, long-term customer cloud arrangements that require us to make significant infrastructure investments, including data center capacity… The economic returns on these investments are dependent on customer demand and the ability of our key customers to meet their contractual obligations." [1] It follows with a risk factor contemplating that "one or more of our key customers experiences insolvency, bankruptcy or other issues impacting their creditworthiness." [1] It even flags that "changes in interest rates could adversely affect the economics of certain long-term commitments or customer arrangements, particularly where pricing is fixed or committed." [1]

Suppliers do not write bankruptcy risk factors about customers whose obligations they regard as costlessly cancellable. Oracle has made a fixed, dated, capital-intensive investment against a long-dated backlog whose economics depend on a customer it will not name. That is the duration mismatch. It is Oracle's.

Broadcom is short the hard leg with nothing booked against it

Broadcom's $128,110 million of its own unconditional, non-cancellable purchase commitments — $55,214 million in fiscal 2027, $72,870 million in fiscal 2028 — meet Broadcom's own GAAP test in full: enforceable, legally binding, fixed quantities, price provisions, timing, and explicitly excluding anything "cancelable without penalty." [9] Against those obligations, Broadcom's 10-Q discloses no OpenAI backlog and does not name OpenAI at all. [9] Broadcom has, in other words, gone long inventory commitments against a term sheet.

CoreWeave finances against a customer it names in its bankruptcy risk factor

CoreWeave's unsatisfied RPO was $98.8 billion at 31 March 2026 — up from $60.7 billion at 31 December 2025 — of which 36% is expected to be recognised in the first 24 months and 39% between months 25 and 48. [3] [4] Its top two customers were approximately 65% of Q1 2026 revenue; Microsoft alone was approximately 67% of 2025 revenue. [3] [4] And in its counterparty-credit-risk factor, CoreWeave names OpenAI directly, as a "private company" of the type that "may have increased risk of insolvency, bankruptcy, or other issues impacting their creditworthiness," managing the exposure "through receipt of prepayments under our committed contracts, credit analysis and monitoring procedures… letters of credit, prepayments, and guarantees." [4]

The regulators have named the mechanism — and put it on the lenders

The Bank of England's July 2026 Financial Stability Report says it in one sentence: "The scale of this maturity mismatch risk across the debt financing of data centre facilities and AI chips, and how that risk is allocated across lenders, borrowers or tenants, will depend on the terms of the debt agreements." [20] Not on OpenAI's balance sheet. Across lenders, borrowers or tenants — an allocation question.

The scale of the exposure it describes is the reason this matters beyond one company. Bloomberg consensus for 2028 hyperscaler capex has been revised from "less than $600 billion" at the December 2025 FSR to "over $1 trillion" today. [20] The five AI hyperscalers were "3% of the stock of outstanding US IG debt at the end of 2025, but as of early May accounted for over 15% of year-to-date issuance." [20] AI issuers have accounted for "41% of non-refinance-related US high-yield (HY) issuance this year so far, despite only accounting for 1% of the JP Morgan HY bond index at the end of 2025." [20] The OECD estimated private credit's share financing AI investment rose from 9% in 2024 to 34% in 2025. [20] Barclays projects $240 billion of hyperscaler investment needs financed through IG issuance in 2026, against the $150–170 billion of senior debt the "big six" US banks typically issue in a year. [20] And JP Morgan estimates that "over $2 trillion of aggregate funding may be needed for the AI chips used to train and run AI models over the next five years." [20]

The BoE is explicit about the fragility: "there is also a potential source of fragility where long-term debt is issued to finance investment in AI assets which may have shorter lifecycles," and it notes the growing use of "off-balance sheet financing arrangements… securitised data centre and other asset-backed structures, special purpose vehicles," which "increases the complexity of identifying where risk ultimately sits." [20]

And the lenders funding all of it are not charging for any of it. BIS Bulletin 120 finds that private-credit lending to AI-related sectors has gone "from near zero to over $200 billion today" and projects it could reach roughly $300 billion to $600 billion by 2030 — yet loans to AI-related companies are "similar in terms of maturity (4.7 vs 4.8 years) and rate spreads (6.2 vs 6.1 pp)" versus everyone else. [21] Ten basis points. For a sector the BoE describes as a maturity-mismatch risk of unknown allocation. The BIS draws the only conclusion available: "either lenders may be underestimating the risks of AI investments (just as their exposures are growing significantly) or equity markets may be overestimating the future cash flows AI could generate." [21] It adds, on the off-balance-sheet structures: "leverage does not disappear by being out of sight." [21]

	OpenAI	Oracle	Broadcom	CoreWeave
Filed, quantified obligation on the public record[12][1][9][3]	None. No audited financial statements, no RPO, no backlog, no churn or retention ever published. Revenue is reported in blog-post run-rates: "We are now generating $2B in revenue per month."	$638bn of remaining performance obligations at 31 May 2026 — up from $138bn a year earlier, a $500bn increase in twelve months (our arithmetic on Oracle's two disclosed figures).	$128,110m of its own unconditional, non-cancellable purchase commitments at 3 May 2026 — meeting Broadcom's own GAAP test in full.	$98.8bn of unsatisfied RPO at 31 March 2026, up from $60.7bn at 31 December 2025.
Duration disclosed?[19][1][9][3]	No. Enterprise terms run on 30-day non-renewal notice; Guaranteed Capacity caps out at one to three years.	Yes: ~12% inside twelve months, 34% in months 13–36, 34% in months 37–60, remainder thereafter. Our arithmetic: 88% beyond a year, 54% beyond three years.	Yes: $55,214m falls in fiscal 2027 and $72,870m in fiscal 2028.	Yes: 36% in the first 24 months, 39% in months 25–48, the balance in months 49–84.
Cancellable without penalty?[22][29][1][9][33]	In practice, yes. Plan re-anchored from the $1.4tn framing to ~$600bn; Stargate sites in Texas, the UK and Norway trimmed. No vendor has booked a penalty, termination charge or impairment.	No. The infrastructure investment is already being made against the backlog.	No — by definition. Its commitments table "exclude[s] agreements that are cancelable without penalty".	OpenAI's side is an "up to" ceiling and, per Forbes, terminable "for cause" across $22.4bn of contracts. CoreWeave's own build-out is not.
Does the filing name OpenAI?[1][9][4]	—	Never. No RPO customer concentration disclosed; "No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024."	Never. No OpenAI backlog anywhere in the 10-Q — against $128bn of hard commitments to its own suppliers.	Yes — by name, inside the counterparty-credit-risk factor, as a "private company" that "may have increased risk of insolvency, bankruptcy, or other issues impacting their creditworthiness".
Who is warning about whom?[29][1][9][4]	Internally: the CFO reportedly warned OpenAI "might not be able to pay for future computing contracts if revenue doesn't grow fast enough".	"The economic returns on these investments are dependent on customer demand and the ability of our key customers to meet their contractual obligations" — plus a risk factor contemplating a key customer's "insolvency, bankruptcy or other issues".	Silence. It has gone long inventory commitments against a term sheet.	Manages the exposure "through receipt of prepayments under our committed contracts, credit analysis and monitoring procedures… letters of credit, prepayments, and guarantees".
Financing structure behind it[31][12][1][9][20][21]	More than $73bn of cash and marketable securities, funded by permanent equity; ~$4.7bn revolver, undrawn; no disclosed balance-sheet debt. Nothing to roll.	"Large, long-term customer cloud arrangements that require us to make significant infrastructure investments", with a risk factor on interest rates hitting "long-term commitments… where pricing is fixed or committed".	Its own supply chain — enforceable, legally binding, with fixed quantities, prices and timing.	Debt and prepayments against contracts with the customer it names in its bankruptcy risk factor. The BoE: the maturity mismatch's allocation "across lenders, borrowers or tenants… will depend on the terms of the debt agreements"; the BIS finds AI credit priced at 6.2pp versus 6.1pp for non-AI.

What Would Prove Us Wrong

The case against this report is not stupid, and we are not going to caricature it.

Two principals — OpenAI and Microsoft — chose the word "contracted" about $250 billion, in identical language, in public, where it could be sued over. [17] [18] Oracle's remaining performance obligations really did jump by half a trillion dollars in twelve months, and Oracle really did write an insolvency risk factor about an unnamed key customer immediately afterwards. [1] CoreWeave really does take "receipt of prepayments under our committed contracts" as its principal credit mitigant. [4] None of that is the behaviour of parties who believe nothing is binding. If these contracts are harder than their public wording suggests, our fixed leg is too small — and OpenAI's cash pile shrinks against the true one very fast, because $22.8 billion of annualised revenue, an $85 billion burn projected for 2028, and roughly 30.5 GW announced against 1.9 GW installed leave no margin for being wrong about which way the option points. [25] [27]

So here is the falsifier, and it is dated, specific, and on the way.

When OpenAI publicly files its registration statement, MD&A must itemise material cash requirements from known contractual obligations by type and by time period — under Section 11 liability, with underwriters' diligence behind it. That table settles this question and nothing else does. It is not available today: a confidentially submitted draft registration statement is not on EDGAR, and under SEC policy the issuer must publicly file the registration statement and all prior nonpublic drafts "at least 15 days prior to any road show." [10] OpenAI submitted its draft in or around late May 2026 and announced it under Rule 135 on 8 June 2026, saying only: "We recently submitted a confidential S-1. We expect it to leak so we're just announcing it. We have not decided on timing yet; it may be a while." [11] No source we accept gives a submission date, so we do not assert one.

If that table lands near $1.4 trillion, our central claim is dead. If it shows the NVIDIA letter of intent, the Broadcom term sheet, AMD's five unbound gigawatts or Stargate's stated intent converted into enforceable, non-cancellable purchase obligations meeting Broadcom's own GAAP test, our adjudications on all four collapse and the headline becomes a real liability figure. [9] If it lands near $600 billion, the headline was always the error.

Three other things would break us, and we name them without hedging. If OpenAI discloses a material multi-year revenue backlog — Guaranteed Capacity RPO with a real weighted-average life, committed contracts as a clear majority of the P&L — then the "cancellable revenue" leg was simply wrong about most of the business and the alleged mismatch was never a duration problem at all. [34] If any counterparty ever books an OpenAI-related contract-termination charge or impairment, that proves the obligations bite, and our reading of them as an option OpenAI is short is falsified. Conversely: another trim with no penalty confirms it. [29] And if anyone other than CoreWeave ever names OpenAI with a dollar figure in a filing — an Oracle RPO concentration, a Microsoft breakout of the $250 billion, a Broadcom or AMD backlog — then the counterparty-silence finding, which is among our strongest, falls, and the headline finally gains the audited corroboration it currently lacks entirely. [1] [5] [9]

Which leaves the trade. The IPO is where all of this gets repriced, and it keeps slipping for a reason that has nothing to do with compute contracts. OpenAI's last private mark is $852 billion post-money, set by the $122 billion round in late March 2026 — a private mark, not a price; by SEC rule a confidential draft contains no price at all. [12] [23] As of 8 June the company had been "gearing up to go public as soon as the fourth quarter of this year." [23] By 25 June it was "leaning toward delaying its public debut from later this year to next year," after SpaceX's stock fell to $153 having topped $225 the previous week — with Altman calling any cut to the trillion-dollar figure a "nonstarter." [32] Our arithmetic: that target sits roughly 17 percent above the last private mark. Meanwhile Anthropic filed a week earlier, closed at a $965 billion valuation, and has been marked at $1 trillion on Forge Global, while OpenAI was recorded at around $880 billion in April. [23] [25] And a tender offer at $852 billion is being arranged to "alleviate some near-term pressure for liquidity" that a delayed listing does nothing to relieve. [23]

The slippage is confirmed; the window is not. OpenAI itself says only that it has not decided. [11] But when that document lands, the first table anyone should turn to is the one nobody has been talking about — and the number in it will tell you whether the last two years of commentary were describing OpenAI's risk, or the risk of everybody who lent against it.