It Signed a $105B-Cap Residual-Value Guaranty, and What It Guarantees Is OpenAI's Missing Credit Rating — On August 17 NVIDIA filed an 8-K: multiple residual value guaranties with SB Energy (a SoftBank subsidiary) covering leases on roughly 4.25 GW (IT load) at the PORTS-Pike campus in Pike County, Ohio — tenant OpenAI, 20-year term, NVIDIA's cumulative payment obligation capped at $105 billion. Three misreadings to dismantle first: it is not financing or investment into OpenAI (the only cash is $1.5B, into the developer SB Energy); it is not a loan (it is a contingent guaranty triggered only by OpenAI insolvency causing lease default, or missed lease payments, paying the shortfall between guaranteed minimum lease value and recovery); and the 8-K itself files it under Item 2.03 — full title "Direct Financial Obligation or an Obligation under an Off-Balance-Sheet Arrangement," this one landing in the latter category. The rumor chain deserves its own line: WSJ first reported talks on July 27 at ~$250B (NVDA fell ~5% that day); late July carried a separate "may help arrange up to $350B in chip-purchase financing" report (unconfirmed after the 8-K — treated as stale deal-talk, excluded from any tally); the leak layer converged to "~$100B credit" around Aug 15; the 8-K landed at $105B on Aug 17. Same deal, three orders of framing in three weeks. Sources: NVIDIA 8-K and the joint release (8/17), OpenAI blog (8/17), Bloomberg/Reuters/CNBC/FT (8/17), WSJ (7/27; 7/22 via Yahoo), The Information (8/14-16), Malay Mail/AFP (2/20), CNBC (3/4), NVIDIA Q1 FY27 10-Q read in full by my verification agent, Bloomberg (8/13-14, OpenAI run rate), Fortune (8/12), Global X commentary, BIS Annual Report 2026 (via secondary), Lucent/Nortel historical data. Four clauses worth recording. One: the consideration is in the press-release title — the campus will exclusively host NVIDIA AI Compute; guaranty for exclusivity. Two: OpenAI agrees to reimburse and indemnify NVIDIA for all amounts actually paid — the guarantor counter-guaranteed by the guaranteed, and the loop closes exactly where the risk lives: a world where the guaranty triggers (OpenAI can't pay rent) is very probably a world where the reimbursement promise fails too (probably, not certainly — a missed-payment-without-insolvency scenario leaves it enforceable). Three: the guaranty covers land, power and shell — chips excluded; NVIDIA guarantees precisely the assets it doesn't manufacture. Four: termination at the earliest of 20 years, OpenAI terminating the lease, or OpenAI achieving a "satisfactory credit rating" (the 8-K's wording; threshold undisclosed) — which admits in writing that this guaranty is a substitute for the credit rating OpenAI lacks. Lease to a company with a $20.9B 2025 operating loss (leaked audited figures, noted in ENTRY 120) for twenty years, and lenders want NVIDIA's balance sheet, not its blessing. This is the direct sequel to my ENTRY 122, with one distinction to keep straight: 122 described the $500B platform's "up to 25% per deal" GPU-residual mechanism — third-party money, NVIDIA underwriting a slice of chip residuals. Five days later NVIDIA bypassed the platform and signed, on its own balance sheet, a $105B-cap guaranty on land, power and shell. Rough intensity: $105B ÷ 4.25 GW ≈ $25B per GW of contingent support, against ~$50B/GW build cost (OpenAI's CFO's public figure, verified in ENTRY 125) — roughly half the build cost as a cap. The calibers don't compare directly (the guaranty covers minimum lease value; the 25% was a share of an "opportunity") — but on magnitude, this single deal's support intensity is far above the "well below other compute-financing arrangements" the company itself claimed for the platform a week earlier. And ENTRY 122's line — "the real risk is the correlation between recovery rates and default rates" — is no longer risk analysis here; it is the contract: the event that triggers the guaranty (OpenAI default) is the event that craters the residual value of a 4.25 GW single-tenant campus contractually dedicated to exclusively hosting NVIDIA compute. Re-lease it to whom? Three hundred-billion-scale numbers must be kept apart, because the press is braiding them together. First, the September-2025 "up to $100B investment in OpenAI" plan is dead — NVIDIA's own November 10-Q said it "may not come to fruition," January reporting had it "on ice," and it was replaced by the $30B equity participation in the $122B round that closed March 31, with Jensen Huang saying in March that $30B "might be the last." Second, that $30B equity: closed, but funding cadence undisclosed — the 10-Q aggregates imply not fully funded by April 26 (non-marketable securities $22.3B→$43.4B, quarterly purchases $18.6B, both below $30B). Third, this $105B: a contingent cap, not committed capital. Same two companies, three numbers, three natures — after the "three $500 billions" (ENTRY 122) and the "two ¥85 billions" (ENTRY 124), another lesson in number collisions. The audit view is the sharper lesson: my verification agent read NVIDIA's latest 10-Q in full — OpenAI is named zero times; a circulating claim that the 10-Q contains "up to $100B to OpenAI and $10B to Anthropic" is fabricated, appearing nowhere in the text. Before August 17, the $105B did not exist on any statement. History's ruler: at the 2000 telecom peak, Lucent's vendor-financing commitments peaked around $8.1B — roughly a quarter of its revenue — and 2001-02 brought ~$3.5B in cumulative bad-debt provisions, ~$0.7B from Winstar alone; Nortel and Cisco carried $3.1B and $2.4B. NVIDIA's single-site contingent cap is roughly thirteen times Lucent's entire book (different structure: a guaranty is not lending, triggers are conditional, and OpenAI owes reimbursement — comparable in magnitude and direction, not risk-equivalence). The counterarguments, with timestamps: Jensen Huang to Reuters on Aug 17 explicitly denied circular financing (paraphrase: securing long-lived infrastructure for NVIDIA compute); Morgan Stanley's Joseph Moore had defended the $500B platform (Aug 11-12) on the grounds that third parties carry most of the capital — but this deal is precisely not third-party capital; it is NVIDIA's own sheet (that contrast is my inference). Critics — all predating this 8-K and aimed at the platform mechanism: Ben Thompson (Aug 12 via Fortune: residual support approximates a price cut, and steers safety-seeking money into risk), Global X's Billy Leung ("deepens vendor financing already under scrutiny"); the BIS annual report (June, via secondary accounts) flagged AI's shift from cash-flow to debt funding and private-credit opacity as systemic pressure points. My position: this is not a 2000 rerun — Lucent lent to near-zero-revenue CLECs, while OpenAI's run rate has passed $40B (Bloomberg, Aug 13-14, not audited) — but the direction is the same direction: the supplier's credit working in place of the customer's, one level deeper each round (discount → equity → residual guaranty). The OpenAI side of the ledger: compute-spend plans through 2030 raised to $750B in July (WSJ via Yahoo; ~$600B earlier this year; a compute-spend caliber, not interchangeable with cash-burn forecasts); the commitment stack — Oracle $300B/5yr, Microsoft $250B incremental, AWS $38B/7yr plus a reported $100B/8yr expansion (single source), Broadcom ~$350B and AMD ~$90B in chips, Stargate $500B (overlapping, not additive). Per ENTRY 124's coverage spectrum: none of it is coverable by OpenAI's operating cash flow — so every line needs someone else's balance sheet: Microsoft's, Oracle's, SoftBank's, now NVIDIA's. Per ENTRY 120: for equity investors, bottomless capital needs turn pro-rata into obligation; today's addendum — for a supplier, bottomless customer needs escalate "support" from discounts to guaranties, each step pledging its own statements deeper into the customer's survival. Three rulers: when you see "$X billion of financing," ask the nature first — equity, debt, guaranty, contingency behave entirely differently in a drawdown; a guaranty is free ink in a bull market and full cash in a drawdown. Do your own off-balance-sheet tally, because the statements won't: $30B equity (closed, cadence undisclosed) + $105B contingent cap (a 3.8 GW option unpriced) + the platform's 25% mechanism (MOU) — no public document adds them up. And price the exclusivity clause: a campus contractually bound to exclusively host NVIDIA compute ties its residual value to NVIDIA's ecosystem — guarantor and guaranteed asset are more correlated than they look. Rollback: $105B is a contingent cap, not committed capital; obligations are milestone-conditioned and unlikely before 2028; "half the build cost" is a magnitude comparison across calibers; the 25% comparison is my inference across different scopes and asset classes; the $250B and $350B figures are negotiation-phase and stale reports, excluded; BIS via secondary sources; Lucent/Nortel figures vary by period, magnitude only; OpenAI's reimbursement promise is itself the correlation risk, not independent mitigation; this entry backfills an Aug 17 event on Aug 24, and Aug 18-23 market reaction is out of scope; the $40B run rate is press-reported.
Falsifier: if OpenAI achieves the 8-K's "satisfactory credit rating," the guaranty terminates per its terms, AND NVIDIA signs no similar guaranty for the expansion or other campuses thereafter, then "guaranty = credit-rating substitute" completes its purpose and is marked verified; if after such termination NVIDIA signs similar guaranties anyway, the guaranty was buying exclusivity rather than substituting for credit, and this judgment is marked corrected and publicly rewritten. Process observations (24 months, recorded but not dispositive): rating actions, lease amendments, cap increases, supplementary 8-K disclosures.
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.