$108.5bn of Guarantees, $366bn of Commitments (28 August 2026): Nvidia’s Filing Numbered the $105bn OpenAI Lease — and the $36bn Program Beside It Went on Pause

Nvidia’s second-quarter fiscal 2027 filing, published with results after the close on 26 August 2026, put company numbers where inference had been. Maximum gross guarantee exposure: $108.5bn, of which $105bn is the Ohio residual value guaranty behind an OpenAI lease and $3.5bn covers land, power and shell obligations for other AI cloud partners. Total future commitments: $366bn, with the supply and capacity line alone rising from $119bn to $279bn in a single quarter, “primarily related to the procurement of memory.” And a $36bn line called “AI cloud agreements” that did not exist in the previous filing. A day later the Wall Street Journal reported that Nvidia had paused some deals inside that newest program, fewer than two months after launching it, and the shares fell 3.3% to $217.55 on Friday 28 August after rising 8.7% the session before. Nvidia’s response was that the July business model “is still in place and continues to evolve due to high demand.” The story is no longer one guarantee to one tenant. It is a disclosed ledger of four different ways a chip vendor can stand behind its own demand — and the newest of the four is being redrawn while the ink is still wet.

Before the tenant’s numbers, the argument about this filing was definitional, and the definition still matters more than the headline. A debt guarantee and a residual value guaranty are not the same instrument, and the difference is not pedantry. A debt guarantee makes the guarantor liable for a loan when a borrower stops paying, and the amount owed is knowable in advance: it is the outstanding balance. A residual value guaranty makes the guarantor liable for the difference between a contracted floor value and whatever the asset actually fetches. The amount owed is therefore unknown until the day it is calculated, and it depends on a market price rather than a loan schedule. Reading the actual 8-K rather than the headline changes which question you have to ask.

What actually happened: the guarantor filed its own numbers

Until 26 August the $105bn was a single 8-K disclosure and everything around it was inference. The CFO commentary filed with second-quarter fiscal 2027 results replaced the inference with a table, and the table is the story. The revenue lines from the same release are covered separately in the quarter itself and the index day that followed; what follows here is the balance-sheet half of it. Nvidia now discloses four distinct categories of forward obligation, and only one of them is the thing most coverage has been arguing about.

Start with the guarantees, because they are the smallest number and the most discussed. The company reports land, power and shell guarantees for certain AI cloud partners’ data centre lease obligations with a maximum gross exposure of $3.5bn, and separately the August 2026 guarantees providing credit support for roughly 4.25 gigawatts at SB Energy’s PORTS-Pike campus in Ohio, capped at $105.0bn. Total maximum gross exposure: $108.5bn. The filing repeats the two features that make this instrument different from a loan guarantee — the obligations become effective in phases as conditions are met, including data centres becoming ready for service, with the first expected in fiscal 2029; and the exposure declines as OpenAI fulfils lease payments. It also quantifies the other side of the trade for the first time: each generation of infrastructure deployed at the site could represent roughly 1.5 million GPUs, or approximately $150bn to $200bn of Nvidia revenue, with an option to provide credit support for a further 3.8 gigawatts as the site scales.

Then the commitments, which are an order of magnitude larger and drew a fraction of the attention.

The supply and capacity line is the one that moved: $119bn a quarter earlier, $279bn now, an increase the company attributes primarily to the procurement of memory. That is worth holding next to the memory complex’ own price action earlier in the month, described further down this page. A purchase commitment made by the largest buyer in the market is, read from the other end, the supplier’s order book — and $160bn of incremental committed memory demand appearing in a single quarter is a fact about pricing rather than about sentiment. Ninety-two billion dollars of the $279bn falls due in the remainder of fiscal 2027 alone.

The $36bn line, and the clause that made it contentious

The genuinely new disclosure is a category that did not appear in the prior quarter’s filing at all. Nvidia describes it in its own words: it has partnered with leading AI clouds to enable broader access to its infrastructure for AI startups, model builders, enterprises, research organisations and sovereign customers, and under those agreements it earns revenue on the upfront sale of the infrastructure and, if certain criteria are met, participates in revenue share generated by the AI clouds from their third-party customers. Alongside it sit data centre leases of approximately fifteen years, expected to commence between fiscal 2028 and fiscal 2029, which the company says it expects to reassign to third parties.

One day after that table was published, the Wall Street Journal reported that Nvidia had paused deals inside the programme, which it calls the AI Compute Partnership and announced in July 2026. As set out in that reporting, the economics work like this: the partner and Nvidia agree a base hourly rate that covers the partner’s costs, and Nvidia takes 50% of any revenue above that rate, on agreements running about six years. Sharon AI and Firmus Technologies were named among the participants. An Nvidia spokesperson said the new business model introduced in July “is still in place and continues to evolve due to high demand.” Both statements can hold at once: a programme can remain a stated business model while specific deals inside it stop being signed.

The part that matters mechanically is not the revenue split. It is the term reported alongside it — that participating providers could lease the GPUs only to customers Nvidia approved, with a preference for spreading capacity across several smaller AI companies rather than letting one large customer take most of it. The same reporting describes internal concern about antitrust exposure; treat that element as a news report rather than an established fact, because the structural point does not depend on it. A supplier that sells a chip has completed a transaction. A supplier that takes half the upside above a cost floor and also has a say in who the buyer may serve has not simply sold a chip; it has taken a position in the buyer’s market. Competition law in most jurisdictions treats those two things very differently, which is a sufficient reason for a programme six weeks old to be rewritten rather than merely repriced.

The tenant’s quarter: where the question started

The structure filed on 17 August has three parties, and until 19 August the market had numbers for only two of them. Nvidia’s side is disclosed in an 8-K: a cap of $105bn, roughly 4.25 gigawatts of IT load, a trigger that cannot fire before ready-for-service conditions expected around 2028. SB Energy’s side is a construction and lease-management job. The third party is the tenant, and OpenAI does not file with the SEC, so its capacity to pay 20 years of rent had been a matter of inference rather than arithmetic.

The Journal’s report closed part of that gap. The figures, as carried by Yahoo Finance on 19 August, are these.

Three things about that table are load-bearing, and a fourth is not.

The first is that these are press-reported figures for private companies, not audited filings. They deserve to be treated as approximately right rather than exactly right, and the operating loss in particular is a construction that depends heavily on how research compute is expensed rather than capitalised. The second is that the sequential growth rate, not the level, is what disappointed: 18% quarter on quarter is a pace most public companies would take, and the report itself notes it lagged several fast-growing listed technology businesses. It was measured against expectations set by a company that had been compounding faster than that. The third is that the loss grew faster than the revenue — up about $3bn sequentially against roughly $1bn of extra revenue — which is the signature of a business whose cost line is being set by capacity decisions rather than by demand.

The fourth thing, the one that is not load-bearing, is the league table. Anthropic passing OpenAI on quarterly revenue is a genuine fact and a poor analytical tool here, because the guarantee is written over a specific lease with a specific tenant. Nothing in the Ohio structure improves or deteriorates because a competitor grew faster. What matters for that footnote is the tenant’s own ability to pay rent, and what matters for the $3tn aggregate is whether the sector’s revenue is compounding fast enough to absorb the commitments — a question about the sum, in which both companies sit on the same side of the ledger.

Why the counterparty rose and the suppliers fell

The 19 August session is the cleanest illustration of the asymmetry this piece has argued from the start, because it ran opposite to intuition. The listed company most directly exposed to OpenAI’s revenue is Oracle, which carries a disclosed multi-year cloud agreement widely reported at around $300bn. On the day the tenant’s growth disappointed, Oracle rose more than 1% while the technology sector fell about 0.8%. Semiconductor names, meanwhile, extended a slide that had begun with a roughly 5% drop the previous session. The index moves were small in both directions — the Nasdaq Composite fell 0.10% and the S&P 500 rose 0.29% — which is itself part of the point.

The explanation is the one that separated Nvidia’s 2.34% from the memory complex’s 9% a day earlier, and it concerns the form of the claim rather than the size of the exposure. A signed cloud contract is a receivable with a schedule attached: it has a counterparty, a term and a legal remedy. Its holder can be paid late, renegotiated with, or in the worst case impaired, but the claim itself is documented. A memory maker’s order book, by contrast, is an expectation derived from somebody else’s stated capital-spending plans. It has no term and no remedy, and it is revised the moment the plan is revised. When news arrives that the ultimate source of demand is compounding more slowly than assumed, the documented claim reprices modestly and the undocumented expectation reprices violently — regardless of which one is nominally larger.

There is a second reason the day was not uglier, and it came from the bond market rather than the technology tape. On 19 August the Treasury doubled the per-operation cap on its long-end buybacks from $2bn to at least $4bn for operations running 9 September to 4 November, and the 30-year yield fell close to 10 basis points to around 5.19%, with the dollar sliding to a three-month low and gold reaching its highest level since early June. Every commitment in the $3tn ledger is a promise to pay for capacity that generates nothing until roughly 2028, so its present value moves inversely with the long end. The tenant’s growth rate and the discount rate are two independent inputs to the same calculation, and on 19 August they moved in opposite directions. That rate mechanism is traced in full in the term-premium storyline, and it is the same channel that puts the US dollar on the other side of this trade.

How the aggregate landed, and why the suppliers paid for it

The 8-K disclosed on 17 August was, in market terms, a non-event for the company that filed it — Nvidia closed that session at $225.01, fractionally lower, and the AI hardware complex around it actually rose hard. SanDisk gained 8.88% that day, Western Digital 5.35%, Marvell 5.54% and Micron 4.13%.

Then the Wall Street Journal published an analysis of the footnotes in the most recent securities filings of nine large technology companies — Meta, Alphabet, Amazon, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and AMD — and put their combined off-balance-sheet commitments, mostly AI-related, at roughly $3tn. The comparison that gave the number its force was not the absolute size but the ratio: about five times the roughly $600bn of capital expenditure those firms reported across their most recent twelve-month periods, and roughly triple what they owe under outstanding leases and long-term borrowings taken together. Two buckets account for most of it — on the Journal's split, roughly $1.2tn of leases that have not yet commenced, and roughly $1.9tn of purchase commitments covering chips, construction and energy procurement.

None of that is hidden in the sense of being undisclosed. All of it comes from the companies' own filings. It simply sits in the notes rather than on the face of the balance sheet, because under current accounting a lease that has not commenced is not yet a lease liability, and a purchase obligation is not capital expenditure until something is bought. Nvidia's $105bn is a member of exactly this set, which is why the company classified it under Item 2.03 in the first place.

Overnight the selling moved to the companies that actually build the memory. On 19 August SK Hynix closed down 9.75%, Kioxia down 12.60%, SoftBank Group down 10.34% and Samsung Electronics down 7.82%. South Korea's Kospi fell 5.80% to 6,471.17 and Japan's Nikkei 225 fell 3.16% to 65,326.42. Taiwan Semiconductor, which makes logic rather than memory, fell 1.26%.

The second half of the explanation is where the prices started from. Through the 17 August close SanDisk was up roughly 653% year to date, Micron about 255% and Western Digital about 211%. Nvidia was up about 21%. Drawdowns of 7% to 9% arriving on top of those runs are a different phenomenon from the same percentage arriving on a flat position, and it matters for interpretation: a crowded position unwinding tells you about the crowd, not about the credit. No memory-specific operating news accompanied the move.

Scale discipline is required in both directions. The Nasdaq-100 finished 18 August about 3.8% below its record close of 30,660.60 set on 2 June, so this was a rotation within the AI complex rather than a broad market correction, and the S&P 500's 0.69% decline says the same. But the Philadelphia semiconductor complex and the Korean index are where the buildout is expressed most directly, and those are the places that moved.

What the filing actually says

Item 1.01 describes a multi-year partnership with SB Energy Corp to develop the PORTS Technology Campus, a large-scale AI data centre campus in Pike County, Ohio, which the document refers to throughout as the Portsmouth Site. Through the partnership and the credit support, the filing states, Nvidia has secured land, power and shell capacity at the site to host Nvidia AI compute infrastructure. An affiliate of OpenAI Group PBC is the tenant, and will use capacity supporting approximately 4.25 gigawatts of IT load to deploy Nvidia's full-stack DSX AI factory platform.

Then the operative paragraph. Nvidia entered into multiple residual value guaranties with SB Energy as lessor, relating to leases for approximately 4.25 gigawatts of IT load in aggregate. Each agreement generally becomes effective on commencement of the applicable lease. Nvidia may also provide credit support to secure approximately an additional 3.8 gigawatts, exercisable at its sole discretion. And the cap: Nvidia's aggregate payment obligation is cumulatively capped at $105bn for its initial commitment.

CNBC reported the wider commercial context the same day — SB Energy will build and manage the campus under a 20-year lease to OpenAI, SB Energy and SoftBank will build power sources supporting 10 gigawatts and invest at least $4.2bn in regional grid infrastructure, Nvidia will separately invest $1.5bn in SB Energy, and capacity is expected to come online in phases from 2028.

A residual value guaranty is not a debt guarantee

Here is the mechanism, in the filing's own construction. Two Trigger Events are defined: OpenAI's insolvency resulting in a default under a lease, or OpenAI's failure to make payments under a lease. On a Trigger Event, Nvidia pays an amount generally equal to any shortfall between the guaranteed minimum value of a lease and amounts recovered through a replacement lease or sale.

Read that formula carefully, because it contains two variables rather than one. The first is OpenAI's ability to pay, which is what the headlines focused on. The second is the recovery value of powered data-centre capacity in Ohio at the moment the first variable fails — and that second term is doing at least as much work. If OpenAI defaults in a world where AI compute is scarce and every hyperscaler wants 4.25 gigawatts of ready capacity, the re-letting recovers most of the guaranteed minimum and Nvidia's payment is small. If OpenAI defaults in a world where AI compute demand has disappointed, the recovery is poor and the shortfall is large.

What Nvidia can do when the trigger fires

The remedy set is where the structure gets genuinely interesting, and it is the part almost no coverage unpacked. On a Trigger Event, Nvidia may elect any of five paths.

The first option is the one that distinguishes Nvidia from a bank holding the same paper. A lender that ends up owning 4.25 gigawatts of data-centre shell has a problem to liquidate. Nvidia is the one entity that can plausibly fill that capacity with its own product and its own customers, and it has already told the market it regards the site as infrastructure secured for Nvidia compute. The optionality embedded in that remedy is worth materially more in Nvidia's hands than in anyone else's, which is a reasonable part of why the company was willing to write the guarantee at all.

The deferral option matters too, in a quieter way: it buys up to a year to see whether a distressed capacity market recovers before crystallising a loss, at the cost of carrying specified project agreement costs in the meantime.

The clause that reveals what this is for

Nvidia's obligations terminate on the earliest of four events: the 20th anniversary of lease commencement, OpenAI terminating the lease under its terms, OpenAI achieving a satisfactory credit rating, or other customary termination events.

That third condition is the tell. The guarantee exists because OpenAI does not currently carry a credit rating that would let a lessor and its creditors finance a buildout of this scale on the tenant's own covenant. Nvidia is lending its balance sheet's credit standing to a counterparty that has not yet built its own, and the contract says explicitly that the arrangement dissolves the moment that is no longer necessary. This is a credit bridge with a written expiry, not an open-ended commitment — and it reframes the whole structure as a financing cost that OpenAI is paying in the form of dependence rather than interest.

One more clause deserves flagging without editorialising. OpenAI has agreed to reimburse and indemnify Nvidia for any and all amounts actually paid to the lessor. That is a real recovery route in the failure-to-pay case. It is worth considerably less in the insolvency case, because an unsecured indemnity from an insolvent entity ranks alongside every other unsecured claim. Both Trigger Events sit in the same contract; the indemnity is not equally valuable against each.

The scale, against Nvidia's own numbers

The cap on the initial tranche alone is equivalent to roughly 54% of Nvidia's reported shareholders' equity at the most recent quarter end, and about 40% of total assets. Three caveats keep that from being alarmist and all three are load-bearing: it is contingent rather than drawn, it cannot begin to trigger until 2028 and then only as individual leases commence, and it is a cap on cumulative payments rather than an expected loss. But it is also only the initial commitment, with a further 3.8 gigawatts of credit support available at Nvidia's own discretion and no disclosed cap attached to that expansion.

Three repricings of one structure: July, 17 August, 18 August

The market has now seen this story three times in three weeks, and the differences between the reactions are the useful lesson.

On Monday 27 July 2026, after Bloomberg reported talks over a backstop of up to $250bn as part of arrangements potentially exceeding $750bn, Nvidia shares fell 4.99% to close at $196.51, removing roughly $250bn of market value and handing the world's-most-valuable-company title back to Apple. Nvidia's five-year credit default swap widened to a record 82 basis points on ICE Data Services pricing — the largest single-day intraday move since the contract began trading actively in November 2025, from around 40 basis points at the start of that month. Société Générale's head of US equity strategy, Manish Kabra, summarised the regime shift in a line that circulated widely: "For hyperscale computing companies, it's CDS, not EPS, that matters now."

Then the Wall Street Journal reported the guarantee would be cut to under $120bn, and the filed number came in at $105bn. By the time the 8-K landed on 17 August, the structure was smaller than the rumour, capped rather than open-ended, conditional on service dates three years out, and specified as a residual value obligation rather than a debt guarantee. An undefined backstop must be priced as an unbounded claim; a bounded and conditional one is priced closer to a written option. The information that arrived in August was mostly definitional, and definitional information is worth less than existential information.

The third reaction, on 18 August, is the one that completes the sequence, and it repriced neither the definition nor the existence of Nvidia's obligation. Nothing about the 8-K changed. What changed was the denominator: the same category of obligation, counted across nine companies, turned out to be roughly $3tn rather than one filing's $105bn. A capped and dated commitment by a single guarantor is a credit question with a bounded answer. Three trillion dollars of forward commitments spread across an industry is a funding question, and funding questions are settled in the bond market rather than in any one company's filings. That is why the third repricing hit hardest in the equities furthest from the guarantee and closest to the capital spending — and why it arrived in the same week that both the US and Japanese long ends sat at generational highs.

The channel to what you actually trade

Two channels, and they are not the same one.

The first is index concentration. A handful of names now drive most of the move in NAS100 and a large share of US500, so a repricing of the largest of them is an index event whether or not you hold the stock. The second, and the less obvious one, is credit. The Bank for International Settlements has examined exactly this family of arrangements, describing them as shadow borrowing — obligations that are economically akin to debt but largely reside outside corporate balance sheets — and noting that credit default swap spreads rose especially for hyperscalers with lower credit ratings. In its June 2026 Annual Economic Report the BIS warned that disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust, with potential knock-on effects on financial conditions.

The evidence that the credit channel is already live is not hypothetical. S&P downgraded Oracle to BBB-, its lowest investment-grade rung, with Oracle's five-year CDS around 215 basis points. Alphabet reported negative free cash flow for the first time since its IPO. A FactSet analysis published on 23 July put aggregate capital expenditure for five major hyperscalers above $690bn in fiscal 2026, more than 80% growth year on year, with free cash flow for most approaching zero or turning negative. When capital spending exceeds internally generated cash, the balance goes on somebody's balance sheet — or, as here, deliberately beside one.

That is also where this connects to the rates picture. Structures like this are ultimately financed in debt markets, so the level and slope of the curve is an input to whether the buildout pencils, which is the link between an Ohio data centre and the term premium at the long end of the Treasury curve. Higher long yields raise the discount rate on capacity that will not generate revenue until 2028, and the US dollar sits on the other side of the same rate expectations, which is why the interest-rate factor is one of the five the meter tracks. For the equipment side of the same buildout, the free cash flow question at Applied Materials is the same argument viewed from the supplier's end.

What would change the picture

One of the checkpoints on this list has now been partly answered, and the rest are still dated and checkable — none of them requiring a forecast.

The answered one is the demand side. The question was whether the tenant’s revenue was compounding fast enough to make 20 years of rent an ordinary obligation rather than a financed one, and the June quarter says: not yet, at $6.7bn of revenue against a $12.3bn operating loss, with growth of about 18% sequentially and a report that the pace picked up again in the third quarter after July model releases. That does not resolve the 2028 question, because a guaranty that cannot trigger for two years is not tested by two quarters of data. It does mean the next reported quarter is now a scheduled input into the same structure, and the specific number to watch is the gap between sequential revenue growth and sequential loss growth rather than either figure on its own.

Two of the checkpoints that stood on this list have now been answered, and answering them created a third. Nvidia’s second-quarter fiscal 2027 results landed after the close on 26 August 2026: revenue of $96.2bn against a guide of $91.0bn plus or minus 2%, data centre revenue of $89.0bn, and a fiscal third-quarter outlook of $108.0bn plus or minus 2% that assumes no data centre compute revenue from China. The commitments disclosure filed with those results is the more durable answer, because it turned an argument about one guarantee into a table with four categories and a schedule by fiscal year. What replaces the old checkpoints is narrower. The first is whether the $36bn of AI cloud agreements is still $36bn in the next filing: a programme reported paused six weeks after launch either resumes, shrinks, or reappears under a different heading, and the commitments table will say which without needing any commentary. The second is the discretionary 3.8-gigawatt expansion at the Ohio site, still a live decision and still the most direct statement the company can make about its own read on capacity demand. The third is whether supply and capacity keeps compounding: $119bn to $279bn in one quarter, with $92bn of it falling due inside the remainder of fiscal 2027, is a schedule that has to be funded out of operating cash flow that came in at $24.1bn for the quarter, down from $50.3bn the quarter before. And the two checkpoints that were never about Nvidia still stand — whether the roughly $1.2tn of not-yet-commenced leases across the nine companies in the Journal’s tally keeps growing, and whether the memory makers’ bookings language changes now that the largest buyer has committed $160bn more in a single quarter.

The honest summary is that a headline number was replaced by a contract, and the contract says something more specific and more conditional than the number implied. Whether $105bn of capped, staged, wrong-way-correlated residual value exposure is prudent is a judgement the filing does not make for you. What the filing does do is tell you precisely which question to ask, and the answer turns out to depend less on OpenAI's creditworthiness than on what a gigawatt of powered shell in Ohio is worth to somebody else in 2028. The week that followed added the other half of the lesson. Read alone, this contract is a bounded question about one company. Read as one entry in a roughly $3tn ledger of commitments that mostly live in footnotes, it is a question about how an entire buildout gets funded, and that question is answered at the long end of two government bond curves rather than in any filing. Which is why the sharpest price reaction landed on the companies that never signed the guarantee at all.

To learn how Pip Theory builds its fundamental currency-strength scores, see the methodology overview.

Educational macro context only — not investment advice.