NVIDIA’s $108 Billion Footnote
A blowout quarter, the guarantee almost nobody read, and the scorecard we froze the night before — graded, our misses first.
What NVIDIA told the world on Wednesday
Start with the part everyone saw. NVIDIA earned $2.22 a share against the $2.09 Wall Street expected, on $96.2 billion of revenue — more than double a year ago — and then guided next quarter to $108 billion, the largest revenue forecast in its history. Management said next year’s growth will run around 70% not because demand tops out there, but because supply does — in the CEO’s words, “the unconstrained is significant.” The stock rose 8.7% the next day. If you read nothing else about this quarter, that is the story the market traded on.
Now the part almost nobody read. Deep in the quarterly SEC filing — the 10-Q, published the same evening — NVIDIA disclosed that in August it signed guarantees capped at $105 billion to backstop a data-center campus in Pike County, Ohio, being built by SB Energy for a single tenant: an affiliate of OpenAI. A guarantee here means exactly what it sounds like: if the tenant defaults on defined portions of its lease and power payments across the site’s 20-year leases, NVIDIA pays. Add the pre-existing $3.5 billion program of similar backstops for AI cloud companies, and NVIDIA’s disclosed maximum guarantee exposure is now $108.5 billion — against $3.5 billion one quarter ago. The company also holds an option, at its sole discretion, to extend credit support to roughly 3.8 more gigawatts as the site grows. What NVIDIA gets in return: the campus exclusively hosts NVIDIA hardware.
NVIDIA’s maximum guarantee exposure, disclosed for the first time in the Q2 FY27 filing — $105bn backstopping an OpenAI data-center campus plus the $3.5bn AI-cloud program, with an option on ~3.8 gigawatts more. A quarter ago the figure was $3.5bn.
That one footnote reframes the quarter. NVIDIA is no longer only selling the AI buildout its customers are paying for — it is now, at meaningful scale, underwriting it.
The balance sheet is becoming the buildout
The guarantee is the headline item in a pattern that runs through the whole filing. In one quarter, NVIDIA’s disclosed commitments to finance, backstop, or pre-fund the AI buildout grew on every line:
Supply commitments more than doubled, to $279 billion — from $119 billion a quarter earlier — which the company’s written CFO commentary attributes “primarily related to the procurement of memory.” (Total future commitments across all categories: $366 billion.) $36 billion of “cloud service agreements” — six-year commitments to purchase cloud capacity from AI cloud partners, a contractual revenue floor for those partners; the same structure management described on the call as a minimum revenue guarantee “that gives lenders the confidence to underwrite the project,” with NVIDIA sharing in revenue above the floor. The partners “can unilaterally stop providing to us and sell to third-party customers at more advantageous rates,” meaning the floor binds NVIDIA, not them. Investments in other companies reached $99 billion, with $25 billion more committed — the privately-held slice alone is $51.2 billion, up $7.8 billion in the quarter, and on the earnings call management said AI labs funded partly off NVIDIA’s own balance sheet are expected to be “roughly 1/4 of our business next year.” Debt rose to $33.4 billion from $8.5 billion at fiscal year-end, after a $25 billion bond sale in June. And the filing’s exhibit list now includes a template legal document titled “Form of Residual Value Guaranty” — the instrument for guaranteeing what used GPUs resell for — filed as a form, with no dollar amount yet attached.
None of this is hidden; all of it is new. A company whose product already commands the whole industry’s capital spending is now also supplying floors under the loans, leases, and startups that buy the product. Two readings coexist, and both are true at once. The skeptical one: the demand NVIDIA reports is increasingly demand NVIDIA finances, and $108.5 billion of guarantees is risk that comes home exactly when customers weaken. The favorable one: the big guarantee wraps land, power, and shell — the layer of the AI buildout that holds value even if chips obsolesce — the exposure has a stated exit (it terminates if OpenAI achieves a satisfactory credit rating), and NVIDIA bought a moat with it: site exclusivity. Which reading wins is not decidable this quarter — but the size is now on the record, and it was the single most consequential disclosure of the night.
The margin story: the memory bill has arrived
One more thing the print settled. Memory prices — the DRAM and high-bandwidth memory that go into every AI system — have been surging for months. Wednesday put that bill on NVIDIA’s own income statement: management guided gross margin (the share of revenue left after production costs) from 75.0% today to 74% next quarter, bottoming at 71–72% the quarter after, recovering to 72–73% next fiscal year only as NVIDIA’s own announced price increases take effect. The CFO called it “extreme pricing conditions in memory,” with increases “headed even higher into next year.” Three to four points of margin, surrendered to a component. For anyone tracking where the profit in the AI buildout migrates, this is the clearest evidence yet that memory suppliers are taking a bigger cut of every AI dollar — and that NVIDIA can pass some, but not all, of it through.
So what is NVIDIA worth after this?
Our published fair value for NVIDIA was $163, set in July when the stock traded near $203 — built on three scenarios: a bear case worth $104 a share (the AI capex cycle turns), a base case worth $163 (revenue roughly triples by fiscal 2031), and a bull case worth $210 (the most favorable operating path we could defend). We re-ran that model on Wednesday’s actual numbers, changing only what the print itself changed: this year’s revenue re-anchored to the reported quarter plus the guide; next year at +50% (a haircut to management’s ~70% — for execution on a supply forecast, not doubt about demand — and taking the 70% at face value changes the answer by less than a dollar); margins trimmed for the memory path; the balance sheet updated; the risks re-weighted in both directions — the guarantee stack and tripled commitments push weight toward the bear case, while the record guide, a broadening customer base, and China’s removal from the outlook push weight toward the bull case.
The result: the scenarios barely move ($105 / $163 / $210), and the weighted fair value lands at $162, from $163. The biggest guide in the company’s history moved our estimate of the business’s value by less than a dollar. That is not a paradox — it is what happens when the same phone call that raises the revenue path also prices in the margin squeeze, and when the questions that actually separate $162 from $228 — what discount rate the dominant AI franchise deserves, and what its terminal decade looks like — are questions no single quarter can answer. At Thursday’s $227.98 close, the market pays 40.7% more than our estimate. Concretely: $228 sits above even our bull scenario. Either the market is applying a cheaper price of risk than we can defend — our own model reaches only $199 even at a 10% cost of equity — or it is re-rating a terminal future beyond what any earnings print can evidence. Reasonable people sit on both sides of that line; no print settles it.
The scorecard we published the night before — graded, misses first
Two nights before this note, we published a pre-event note — What $213 Has to Believe — that froze, in public, exactly what we would score after the print: nine questions with pre-written rules for what silence would mean, the market reads we relied on, and a hard bound on how far our fair value could move on any outcome ($149–$175). This section grades that scorecard, our misses first — because a scorecard you only publish when it flatters you is not a scorecard.
What we got wrong. We asked the right question about the wrong layer: our pre-registered concern was a guarantee on GPU resale values; the $105 billion that actually arrived wraps the real estate and power under the GPUs. The resale-value instrument exists as a filed template — with no number yet.
A frozen market read, resolved against us — and a claim we are retiring. The pre-note reported what the options market was paying for movement (±5.9%) against the median actual move of the last eight prints (3.2%), and the direction base rate (the stock had fallen after 6 of its 8 prior reports). The reaction delivered +8.74%, up — outside what the market paid, against the base rate. We record that as data, not as a forecast miss: our own framework holds that no single trading window is forecastable, and one draw can neither refute nor validate a pricing read. The deeper conclusion is about the verdict itself. Across our options ledger, the “event volatility is rich/cheap” call has been right 11 times in 19 — a coin flip — so we are retiring it as a house claim. Going forward, pre-event notes will present the priced move and the historical distribution as information for the reader, and grade the outcome afterward as another data point, with no skill claimed in between.
What the framework got right. The pre-note argued the filing, not the earnings number, would carry the news — it did. Its one market read that leaned the right way was the options market’s own probability distribution, which tilted toward a melt-up. And the fair-value bound held: $162 sits mid-band.
Rail colors:answered by the eventnot testable at this event (pre-defined)missed
1Does the residual/support guarantee get a real number in the filing?
Yes — beyond the question: $108.5bn of total maximum exposure ($105bn OpenAI/SB Energy site credit support + the $3.5bn AI-cloud program), plus a sole-discretion option on ~3.8GW more. The GPU resale-value instrument exists as a filed template (Exhibit 10.1) with no dollar amount yet.
2Does NVIDIA’s book of private-company stakes cross $55bn?
No. $51.2bn on the balance-sheet caption (up $7.8bn from $43.4bn); $47.9bn on the footnote’s own rollforward (up from $42.3bn). Rising fast; threshold not crossed, and not the flat-or-down print that would have suggested risk moving off the balance sheet.
3Does customer concentration broaden or tighten?
Broadened. One direct customer ≥10% of revenue (at 16%) versus three customers totaling 54% last quarter; first-half top-three 44%. Caveat: one customer was reclassified between segments with prior periods recast, so the quarter-on-quarter comparison is imperfect — the direction survives it.
4New writedowns against the supply commitments?
None material ($985m of ordinary-course provisions) — while the commitments more than doubled to $279bn, “primarily related to the procurement of memory” per the written CFO Commentary. The exposure grew: exactly the outcome our pre-written rule flagged as risk-increasing, not risk-cleared.
5What happens to the China business?
Written to zero in the outlook, by the company. H200 shipments to China were <1% of data-center revenue (PRC-restricted, tariffed 25%, a $0.4bn charge taken), and the written outlook is “not assuming any Data Center compute revenue from China.”
6Is the customers’ capital spending plateauing?
Not testable at this event — that reads from the customers’ own filings in late October. Frozen going in at +87% year-over-year growth; management commentary (top-5 capex ~$800bn 2026 → $1.3tn 2027) recorded, not scored.
7Is NVIDIA losing accelerator share to custom chips?
Not directly testable; the proxy read favorably — NVIDIA’s data-center growth (+117%) exceeded its customers’ capital-spending growth (+68% to +110%). No share-loss signal.
8Does a rated GPU-loan bond exist yet?
No. The $500bn financing platforms remain “subject to definitive agreements,” and this filing shows the support running through NVIDIA’s own guarantees, commitments, and equity instead of a rated bond market.
9What does NVIDIA’s credit protection cost versus peers?
Not observable this window — the market for insuring NVIDIA’s debt is thin; per the pre-written rule this is recorded as unmeasured, never inferred from the stock.
Three market graders resolve on their own calendar and will be added to this page when they do: the day-two trading decomposition, the options-ledger entry, and the final calibration of the pre-print probability distribution (Friday’s expiry; at $227.98 it reads at its 85th percentile).
What would change our view
Symmetric and checkable, both directions. Toward the price: OpenAI achieving the credit rating that terminates the guarantees — the structure’s own exit executing; a rated GPU-backed bond printing with an arm’s-length residual assumption and NVIDIA absent from the loss waterfall; the margin trough arriving above 72%; hyperscaler capex growth repeating above +50% at the late-October reports. Away from the price: the ~3.8GW option exercised or guarantee exposure passing ~$150bn; the private-stakes book crossing $55bn (next tested at the November report); any drawn guarantee or credit event at a supported counterparty; next year’s margins missing the guided 72–73%; concentration re-tightening above 50% on a comparable basis.
Recommendation
Fair value $162 (from $163), range $105–$210, against a $227.98 close — a 40.7% gap. We are constructive on the business — which just guided to the largest quarter in its history — and cautious on the equity at a price above our most favorable scenario, now with a $108.5 billion guarantee stack on the balance sheet that did not exist at our July valuation. Nothing here is a trade recommendation. The scorecard above, misses included, is logged for public tracking; the next test of this name arrives with the November report.
Sources & Method
Print and filing facts. NVIDIA Q2 FY27 press release, Form 10-Q (accession 0001045810-26-000075, filed 2026-08-26), and the written CFO Commentary (archived at research/nvda-q2-fy27/data/cfo_commentary_q2fy27.pdf), pulled 8/26–8/27 via archived scripts (research/nvda-q2-fy27/data/score_10q_pull.py; full filing text and extraction windows archived alongside). Guarantee, commitment, concentration, China, and balance-sheet figures quote the filing; the memory-procurement attribution quotes the CFO Commentary.
Call figures. Earnings-call figures not in the written documents (the Q4 margin trough, fiscal-2028 margin and ~70% growth, “roughly 1/4 of our business”) are attributed to the call and corroborated across three independent transcriptions (archived with provenance at research/nvda-q2-fy27/data/tells/transcript.md). EPS consensus and print verification: Robinhood earnings API.
Prices. Official SIP closes via Robinhood ($209.66 on 8/26; $227.98 on 8/27); the reaction move is close-to-close, with the day-two decomposition to be inserted with its own dateline.
Valuation. data/model_rerun.py — the July valuation’s archived engine with event-traced input changes documented line-by-line; results at data/model_results.json, written before the prose. The fair value of NVIDIA — Priced for the Clock Not Breaking moves $163 → $162 via that archived rerun and its Signal Ledger row (asof 8/27).
The pair. The frozen scorecard: research/nvda-q2-fy27/data/rubric.json (2026-08-25) and this note’s data/scorecard.json. The options-ledger record cited (vol verdict 11 of 19) is the house earnings-options ledger. Prior TON618 notes: the pre-event note (8/25) and the July valuation linked above; background on the financing structure and the buildout’s economics: The Fifth Seat (8/14), The AI Capex Payback Clock (7/19), The Losing Bets That Built the World (8/12) — cited for provenance; none is required reading for this note.
Disclosures
Information only. TON618 Capital. This report is for information purposes only. Nothing here is an offer to sell or a solicitation of an offer to buy any security, fund interest, or digital asset, and nothing here is personalized investment advice or a recommendation regarding any instrument.
Publisher’s exclusion. All research is published solely as general, impersonal information of regular circulation. It is not tailored to the objectives or circumstances of any individual and is not issued in connection with compensation from any client. The Fund has no clients and distributes all research free of charge. On that basis it publishes in reliance on the publisher’s exclusion from the definition of “investment adviser” under the Investment Advisers Act of 1940 (§202(a)(11)(D); cf. Lowe v. SEC, 472 U.S. 181 (1985)).
Registration & conflicts. TON618 Capital is not registered as an investment adviser or broker-dealer in any capacity. The Fund is a Bitcoin fund and may hold or transact in the securities or digital assets it discusses. This report is a post-event note on NVIDIA (NASDAQ: NVDA) and references, among others, Microsoft (MSFT), Alphabet (GOOGL), Meta Platforms (META), Amazon (AMZN), Micron (MU), Broadcom (AVGO), Oracle (ORCL), CoreWeave (CRWV), Intel (INTC) and Cisco (CSCO); the Fund holds no position, long or short, in any of them, and has no economic interest in the price of any security named here. The Fund receives no compensation from any party in connection with its research.
Use of AI. Artificial intelligence is used in the creation of this research. All methodology and data integrity are reviewed and approved before publication by TON618 Capital’s Chief Investment Officer, Keyth Beck; errors may nonetheless occur, and readers should verify independently.
CFA. This report was prepared to align with CFA Institute analytical standards (methodology only). CFA® and Chartered Financial Analyst® are registered trademarks owned by CFA Institute. That reference describes the analytical framework applied; it does not imply the report was prepared, reviewed, or authored by a CFA charterholder, and the report is not issued, reviewed, endorsed, certified, or approved by — nor affiliated with — CFA Institute.
Risk & feedback. Past performance is not indicative of future results. Digital assets and equities are volatile and may result in total loss of capital. Corrections and feedback are welcome — please direct them to CIO Keyth Beck at keyth@ton618capital.com.