The Neocloud Fuse

How the AI Buildout Turns Re-Leasable Assets Into Single-Name Credits, Why the Market Charges 334 Basis Points for the Difference, and What Happens When the Unsecured Half Needs New Money

TON618 Capital Research · As of July 16, 2026 · Download PDF

A point-in-time read of a fast-moving credit episode, not a forecast with a date on it. Data: SEC filings (10-K, 10-Q, 8-K, S-1, Form D, FWP), FRED, PJM and its Independent Market Monitor, Lawrence Berkeley National Laboratory, rating-agency actions and methodologies, and primary reporting — every figure as-of dated. Fact ("Data shows…") is separated from opinion ("We believe…"). Where a figure exists only in press reporting and not in a filing, it is labelled REPORTED; where a deal is announced but not contractually binding, ANNOUNCED; where we have computed something from an unverified input, ESTIMATE. Independently fact-checked against primary sources before release; the Source & Verification Note documents that pass, the corrections it forced on this note's own draft, and what we could not verify at all.

§0 The Argument

A neocloud is a company whose business is renting out GPU compute — CoreWeave, Nebius, Lambda, Crusoe, and a lengthening tail — as distinct from a hyperscaler (Microsoft, Amazon, Google, Meta, Oracle) for whom cloud is one business among several. The model is simple to state: borrow money, buy NVIDIA GPUs, sign a long contract with an AI lab, use the contract to service the debt.

The central finding of this note is a spread that should not be possible if the consensus framing were right. On June 10, 2026, Applied Digital priced $1.59bn of senior secured notes at a 7.00% yield, backed by North Dakota capacity under a 15-year lease to CoreWeave. Eight days later, on June 18, CoreWeave itself priced $1.25bn of unsecured notes at 9.625%. Treasuries moved less than 10bp in between. Two claims on substantially the same economic activity, priced ~262bp apart in eight days — the difference being what secures them. And CoreWeave's own debt footnote makes the point without any estimate at all: its $8.5bn DDTL 4.0 facility, entered March 2026, prices fixed draws at "2.00% per annum plus a blended Treasury rate" — a contractual UST+200bp — while its unsecured notes carry a 10% effective rate.

Our house view. The market has already worked out that neocloud assets are money-good and neocloud companies are not. It lends against the first at investment-grade-like spreads and against the second at deep high-yield. The base case is therefore not that a repricing is coming — it is that the repricing has happened, did not stick, and remains idiosyncratic. US high-yield spreads gapped 63bp to 346bp on March 30, 2026 and then fully retraced to 271bp by July 15. Meanwhile Oracle's 5-year CDS sits at a REPORTED record ~198bp while the investment-grade index it belongs to trades at 79bp — the 7th percentile of the last twenty years. An IG issuer at ~2.5× its own index, at a record wide, inside a market at multi-decade tights. Both things are true at once. The question is not whether AI credit reprices; it is whether the idiosyncratic repricing generalizes.

The load-bearing judgment. The parent note held that private credit's risk was never inside the funds but in the plumbing around them. The neocloud version: the risk is not in the asset — it is in what the financing structure does to the asset. Investment-grade data-center ABS earns its rating from demonstrated re-leasability: KBRA's January 2026 rating of DataBank Series 2026-1 rests on 1,757 customers, largest at 3.9% of revenue, on a two-year weighted-average contract term. That is real estate with a deep tenant market. A neocloud deal is the structural inverse — one tenant, one long contract, a purpose-built plant — and it converts a re-leasable real-estate asset into a single-name credit exposure. The ABS market pays up for re-leasability. Neoclouds engineer it away, and the wedge is the bill.

Why the counterparty analysis still matters. The wedge tells you the market prices this. It does not tell you whether it prices it correctly — and that turns on who is on the other side of the contract. Here the disclosure record is startling: the ~$1.4 trillion of announced AI compute commitments is nameplate, not contracted. NVIDIA's $100B was a letter of intent that was never converted; only the first 1GW of AMD's 6GW binds; only ~1.3GW of Broadcom's 10GW; Stargate's $500B entity was never capitalized. Oracle carries a $638B backlog whose largest counterparty it has never named in an audited financial statement — the name appears only in a bond-marketing document (§4). The wedge exists because the contracts are ceilings and the counterparties are funded in a circle. §4 through §6 explain the wedge; §7 shows it on the tape.

The debates we carry honestly as overhangs: (1) if the market already prices the wedge, is there anything left to say — or is this note describing efficiency and calling it risk? (2) Does the circularity net out to nothing, vendor equity in customers being old and ordinary? (3) Is the wedge a credit judgment or merely a structural one — secured paper always beats unsecured, so is 334bp actually wide? (4) Does the March round-trip prove the market is discriminating, or that it has the attention span of a quarter? Each has a falsifier, stated where it lives (§11).

§1 The Story in Plain English

Read this section if you read nothing else; it makes the whole case without jargon.

Suppose you want to build a toll road. You borrow to build it and show the bank a contract: a trucking company has agreed to use it. The bank lends against that contract. Ordinary, and it works — provided the trucking company is obliged to pay, and able to pay.

Now notice something about the bank. It will happily lend you money secured on the road itself at a modest rate, because a road is a road: if you fail, someone else drives on it — though if you built it wide and heavy for one company's road-trains, the next user drives cars and pays car tolls. The road survives; the toll schedule does not. And the bank will charge you far more to lend against your company, because your company is, at bottom, the promise of one trucking firm. Same project. Two prices. The gap between them is the bank telling you what it actually thinks.

That gap is the subject of this note, and in the AI buildout it is roughly 334 basis points — measured not by us but by CoreWeave's own filings, which disclose both prices in the same quarter.

Why is the gap so wide? Two reasons, and both are documented.

First, the obligation is thinner than the headlines. Read the filings and a pattern emerges that press coverage misses. CoreWeave's contracts with OpenAI say the customer "committed to pay us up to approximately" a number. "Up to" is a ceiling; a contract permitting a customer to spend up to $11.9B is compatible with spending far less. AMD granted OpenAI a warrant over 160 million shares — but only the first 1 gigawatt of a 6-gigawatt program binds, and full vesting additionally requires AMD's stock to reach $600 (it trades at ~$501). NVIDIA's celebrated $100B was announced as a "letter of intent" in September 2025; its 10-K said in February 2026 there was "no assurance… that a transaction will be completed"; by May 2026, the word "OpenAI" had vanished from NVIDIA's filings entirely.

Second, the ability to pay rests on numbers that are not what they appear. The ~$1.4T of commitments sits against roughly $24B of annualized revenue — about 58×. And a "run-rate" is an annualized snapshot, not money received: reporting indicates OpenAI recognized $5.7B of actual revenue in Q1 2026 against a $25B run-rate characterization. Both are defensible; only one pays a bill.

And the trucking company is partly funded by the company that sold you the asphalt. NVIDIA has taken equity across its own customer base. OpenAI lent Cerebras $1B so Cerebras could build capacity for OpenAI, and a Cerebras warrant tranche vested when that loan funded. SoftBank borrowed against its OpenAI shares to fund OpenAI — and when lenders balked at valuing an unlisted collateral, SoftBank cut its own ask 40% to $6B; a month later, reporting said even that had stalled.

So what actually happens next? Not a default. The market has already voted, in the only way that counts: it charges 200bp for the road and 534bp for the company. The question this note asks is whether that vote spreads — from CoreWeave's capital structure to everyone else's.

§2 What a Neocloud Is, and How It Got Financed

Definitions: a hyperscaler is a large diversified cloud operator. A neocloud rents GPU compute as its primary business. RPO (remaining performance obligations) is contracted revenue not yet recognized. Take-or-pay means the customer pays whether or not it uses the capacity. A delayed-draw term loan (DDTL) lets a borrower draw in tranches. Non-recourse means the lender can seize the pledged assets but cannot pursue the parent. Nameplate is our term for an announced headline not shown to be contractually binding.

• The public arm. CoreWeave is the sector's only large public pure-play. RPO grew $15.1B (Dec-2024) → $50.0B (Sep-2025) → $60.7B (Dec-2025, +302%) → $98.8B (March 31, 2026). FY25 revenue $5.1B; Q1-26 revenue $2.078B (vs $982M Q1-25). 43 data centers, >850MW active, ~3.1GW contracted power. [CRWV 10-K, Q1-26 10-Q]
• The debt is large, fast-growing, and front-loaded. Total principal $25.149B at March 31, 2026, up from $21.615B at December 31, 2025 — +$3.5B in a single quarter. Current portion $7.547B; $6.066B of principal falls due in the remainder of 2026 and $5.652B in 2027. [Q1-26 10-Q]
• The private arm finances almost entirely outside public markets — and discloses almost nothing. Lambda: $500M GPU-backed SPV (Macquarie, April 2024), $275M syndicated senior secured (J.P. Morgan, August 2025), upsized to $1.0B (May 2026). Crusoe: $175M (Victory Park), ~$225M (Upper90), $750M (Brookfield), plus construction loans of $2.3B and $7.1B (J.P. Morgan-led) against Abilene. [Company releases; law-firm deal announcements; Form D]
• A structural fact that is itself a finding. Across all eight of those private facilities, not one discloses an interest rate or a maturity. A market financing gigawatts of correlated, single-theme infrastructure has developed without public pricing. Any spread or tenor assumption an outside analyst makes about private neocloud credit is unsupported by disclosure — including ours. We make none. This is precisely why §7 leans on CoreWeave: it is the only place the price is legible.
• Headline round size is not money received. Crusoe's Series E was announced as "$1.375 billion"; the Form D shows $1.027B actually sold as of October 21, 2025. Its Series D was widely reported as "$600M"; the filing shows $685.7M sold of $817.7M offered. Lambda has no Form D on EDGAR at all for either its reported $480M Series D or its reported $1.5B Series E — verified against the SEC submissions record, cause unexplained. [SEC Form D; data.sec.gov]

§3 The Collateral Converts: Power Early, Contract Late

The axis everyone argues about is the wrong one.

§2 ended with $25 billion of CoreWeave debt and eight private facilities with no disclosed pricing — and a question implied but not asked: what actually secures all of this? The public argument answers instantly — GPU depreciation — and fights over whether a 4-to-6-year useful life is honest when NVIDIA ships a new architecture roughly annually. Our companion note The Depreciation Question takes that up for the Mag7. It is not where neocloud credit risk concentrates, and we can now say why with evidence rather than assertion.

• Start with what the rating agencies actually secure. KBRA's Data Center ABS Global Rating Methodology (final, January 9, 2026) scopes collateral as "title to cash flowing real property interests, along with mortgages on the properties" plus "assignment of customer contracts, tenant leases." GPUs and IT equipment are not in scope at all. The framework estimates sustainable net cash flow on the properties and stresses it. In the rated market, the chips are not the collateral — and they never were. (Private GPU-backed facilities do pledge the chips — Lambda's Macquarie SPV is built on them — but no rated structure relies on them.)
• Early in a project's life, the collateral is power — and power is genuinely scarce. PJM's 2027/2028 Base Residual Auction cleared at the FERC-approved cap of $333.44/MW-day across every zone, the third consecutive capped auction, and the first in which the entire RTO including FRR areas fell short of its reliability requirement. PJM's Independent Market Monitor found that including forecast data-center load increased total auction revenues by $6.24bn, or 61.4%. Roughly 97% of the increase in forecast peak load for 2027/28 versus 2026/27 is data-center demand. [PJM BRA report, Dec 17 2025; IMM, Jan 5 2026]
• But the naive version of "power is the collateral" is wrong, and the data kills it. LBNL's Queued Up (2025 edition, YE2025 data) finds 549 GW already holds a draft or executed interconnection agreement and has not reached commercial operation. PJM separately discloses ~47,000 MW has completed its interconnection process and is free to move to construction. If interconnection rights were the scarce, value-carrying asset, that queue would clear. It does not. A lien on a queue position is a lien on something half a terawatt of competitors also hold.
• The scarcity has migrated downstream, to equipment. GE Vernova's Q1 2026 8-K — the only audited evidence in this file — discloses gas equipment backlog plus slot reservations rising 83 GW → 100 GW in a single quarter, guided to ≥110 GW by year-end 2026, against 4 GW shipped in the quarter. At ~16 GW/year of shipments against a 100 GW book, that is roughly a six-year queue. Turbines are scarce. Interconnection paper is not.
• The conversion is the moment the risk changes hands. A site with land, interconnection, and turbine slots is fungible: it re-leases. A finished GB200 hall — 72 GPUs per rack, ~130 kW observed at full load, direct liquid cooling mandatory — is a purpose-built plant. Physically it is not stranded (a hall built for 130 kW racks hosts 10 kW enterprise racks without difficulty; the hard retrofit is the other direction). Economically it is the problem: you re-lease at enterprise-colo rent against a capital stack sized for AI-training economics. The asset re-leases; the rent does not.
• So: the lender holds power early and the contract late, and neocloud lending is the business of financing exactly that conversion. That is what the §7 wedge prices. It is also why the ABS market's revealed preference is so pointed: DataBank Series 2026-1 earns its rating on 1,757 customers, largest 3.9%, two-year weighted-average term — demonstrated re-leasability, not a covenant from one name.

§4 The Ledger: Nameplate Versus Binding

Why the wedge exists. Every row sorted by what a filing — not a press release — will support.

• The scale check. ~$1.4T of commitments against ~$24B annualized revenue ≈ 58×. The buildout is far behind its own announcements: Stargate has >9GW planned and ~0.3GW operational — roughly 3% of plan, fifteen months in. [Epoch AI, April 17, 2026]
• Oracle's disclosure gap is the single most striking fact we found. Oracle's FY2026 10-K, filed June 22, 2026, contains zero occurrences of "OpenAI," "gigawatt," or "Stargate." So do all three FY2026 10-Qs, the Q4 press release, and the Q4 earnings call transcript. An EDGAR full-text search across every Oracle filing ever made returns exactly three hits for "OpenAI," two of them incidental. The one substantive hit is a Free Writing Prospectus dated February 1, 2026 — a bond-marketing document — listing "AMD, Meta, NVIDIA, OpenAI, TikTok, xAI" among "our largest Oracle Cloud Infrastructure customers." Oracle names its largest counterparty when it is selling bonds, and not when it is filing audited financials.
• Oracle's concentration disclosure is true and useless. The FY26 10-K states: "No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024." Accurate — because the OpenAI contract is REPORTED to begin in 2027. It is in RPO; it is not yet in revenue. The regime permits a company to carry a $638B backlog roughly half of which S&P attributes to one counterparty, while truthfully reporting no 10% customer. The closest Oracle comes is a risk factor: "In certain OCI offerings, we are more concentrated among a number of large customers."
• The ceiling language, verbatim. CoreWeave's filed agreements read "committed to pay us up to approximately $11.9 billion" and "up to approximately $6.5 billion." The widely-cited $22.4B aggregate reconciles arithmetically ($11.9B + $4B + $6.5B) but has never appeared in an SEC filing, and the $4B middle component is not itemized in any filing at all.
• AMD's binding line, verbatim. "Concurrent with signing, Warrantholder agreed to a binding commitment to purchase… the initial one (1) gigawatt." One of six. As of March 28, 2026, zero warrant tranches had vested. AMD has never stated a dollar value; the ubiquitous "$100B" is analyst-derived — its CFO said only "tens of billions." The vesting schedule is redacted in the filed exhibit as [****]; anyone presenting a precise tranche table is reconstructing, not citing.
• NVIDIA's disappearing act. September 22, 2025: a "letter of intent" to invest "up to $100 billion… progressively as each gigawatt is deployed." November 2025 10-Q: "no assurance that we will enter into definitive agreements." February 2026 10-K: "no assurance… that a transaction will be completed." May 2026 10-Q: "OpenAI" appears zero times. NVIDIA has never filed an 8-K about the investment. Disclosed aggregate investment commitments rose $6.5B → $11.4B → $27B, none attributed to OpenAI. REPORTED: talks scaled to ~$30B; NVIDIA's CEO reportedly said $100B was "probably not in the cards" (March 4, 2026).
• A correction worth making loudly. Broadcom's "$10B mystery XPU customer," attributed across the financial press to OpenAI in late 2025, was Anthropic. Broadcom's Charlie Kawwas explicitly denied the OpenAI attribution on October 13, 2025 — the same day — and Hock Tan confirmed Anthropic on the December 11, 2025 call. "OpenAI" appears zero times in any Broadcom filing, ever. Similarly, Broadcom's +$119.6B RPO jump to $164.6B (May 3, 2026) is attributed in the 10-Q to "a long-term contract for custom AI accelerators" with no counterparty named — Meta is the likelier fit. Do not attribute it to OpenAI.

§5 The Circle: Vendor Financing as Demand

Not an allegation. A disclosure.

• The mechanic, concretely. A vendor takes equity in a customer; the customer signs a compute contract with an intermediary; the intermediary borrows against that contract; the intermediary buys the vendor's product; the vendor books revenue. Each leg is individually ordinary. Traversed end-to-end, the vendor has partially funded its own revenue, and the lender has advanced cash against a demand signal the vendor helped create.
• The cleanest documented instance is not NVIDIA — it is Cerebras. OpenAI is "contractually committed to purchase" 750MW. To build it, OpenAI funded a $1B secured promissory note to Cerebras (January 2026, 6%, maturing no later than December 31, 2032), and a Cerebras warrant tranche vested on the funding of that loan. The customer financed the supplier that serves the customer, and was compensated in equity for doing so. Disclosed in Cerebras's 10-Q, filed June 24, 2026. Not hidden — simply not what "$1.4 trillion of demand" sounds like.
• The leverage-on-leverage instance is SoftBank. Its funding of OpenAI is real and large — $7.5B + $22.5B = $30B completed December 26, 2025; a further $30B tranche in April 2026. But it borrowed against its OpenAI stake to do it. REPORTED (Bloomberg, May 8, 2026): SoftBank cut its own target for that margin loan by 40%, from $10B to $6B, on lender hesitancy — the reported sticking point being the difficulty of valuing an unlisted OpenAI as collateral. REPORTED (Bloomberg, June 10, 2026): even the reduced $6B then stalled. The mechanism matters: this is not lenders yanking a live facility, it is a borrower shrinking its ask and still not clearing. A slower signal but a more informative one — it prices the collateral, not the borrower.
• NVIDIA's stake is real; its disclosure is aggregate. Non-marketable securities show net additions of $17.9B in Q1 FY27 and a $42.3B balance; commitments rose to $27B. None attributed to OpenAI. The arithmetic is consistent with a partially-funded ~$30B position. It is not disclosure, and we state no funded figure as fact.
• OpenAI has now industrialized the loop. On May 19, 2026 it launched "Guaranteed Capacity" — selling 1/2/3-year compute guarantees to enterprises, converting its own compute obligations into customer-funded forward contracts. Whether that is prudent balance-sheet management or the securitization of a liability depends on enterprise demand behind it, which is unobservable.
• The steelman, which we take seriously. Vendor equity in customers is ancient and mostly benign — Intel, Cisco, and every semiconductor cycle in living memory. Lenders are not naïve: a committee lending against a CoreWeave contract knows what "up to" means and sizes the advance rate accordingly. And a ceiling is arguably the correct contractual form for capacity a buyer cannot perfectly forecast. The strongest form of this argument is that the §7 wedge IS the pricing — the market read the ceilings, charged 334bp, and the system is working. We engage that directly in §11, because it is the best objection to this note.

§6 The Counterparty Under the Ceiling

You are not underwriting a data center. You are underwriting an AI lab's forward cash flow.

• What is known is thinner than the discourse implies — but it has a visible expiry date. OpenAI and Anthropic are private; no audited public financial statement exists for either. OpenAI confidentially submitted a draft S-1 to the SEC on June 8, 2026 (OpenAI newsroom). A confidential submission is not public and its contents are unknowable until it flips — but the largest counterparty in this complex has started a process that ends in audited financials. We think this cuts against us more than for us: an IPO both removes the opacity we complain about and supplies a funding route independent of the circle.
• The best window is Microsoft's, and it is narrow. Microsoft holds "approximately 27 percent of OpenAI on an as-converted basis," equity-method, using hypothetical liquidation at book value, with "total funding commitments of $13 billion, of which $11.8 billion has been funded as of March 31, 2026." OpenAI-attributable P&L: −$4.1B (Q1 FY26), +$10.0B (Q2 FY26 — largely a non-cash dilution gain from the recapitalization, not cash earnings), −$19M (Q3 FY26). Microsoft now reports non-GAAP EPS excluding OpenAI results. Its loss-recognition lag is "up to three months" — even this window is stale by a quarter.
• Run-rate is not revenue. REPORTED: OpenAI recognized $5.7B of actual revenue in Q1 2026 against a $25B run-rate characterization. Funding figures conflict across sources ($110B/$730B vs $122B/$852B valuations); we do not adjudicate them.
• The relationship has been restructured twice, and each time the discipline loosened. October 2025: OpenAI "contracted to purchase" $250B of Azure, and Microsoft surrendered its right of first refusal to be OpenAI's compute provider. April 27, 2026: a second amendment — "OpenAI can now serve all its products to customers across any cloud provider"; the license became non-exclusive; Microsoft stopped paying revenue share to OpenAI; and an AGI declaration no longer alters Microsoft's rights, retiring the expert panel's leverage. This second amendment was never furnished in an 8-K — its only SEC trace is a single sentence of MD&A. The $250B figure appears in exactly one filing and is never repeated.
• Why this decides whether the wedge is priced right. Every take-or-pay contract in §4 is an asset whose value is the counterparty's ability to pay. If the counterparty's funding is circular (§5), its revenue a fraction of its run-rate, its largest partner has twice loosened its obligations, and its aggregate commitments run ~58× revenue — then the contract's enforceability is not the binding constraint. The counterparty's solvency is. A ceiling you cannot enforce against a debtor who cannot pay is not collateral at all. That is why the market charges 334bp for the unsecured claim and 200bp for the secured one.

§7 What the Credit Market Is Actually Saying

The section that carries the note. Every number here is either FRED or an SEC filing.

• The wedge, from CoreWeave's own debt footnote. Effective interest rates disclosed in the Q1-26 10-Q form a ladder that tells the whole story:

• The single cleanest fact in this note. The $8.5bn DDTL 4.0 facility — entered March 2026 by subsidiary CoreWeave Compute Acquisition Co. VIII, with MUFG as agent, non-recourse, secured by that subsidiary's equity and substantially all its assets — prices fixed-rate draws at "2.00% per annum plus a blended Treasury rate" and floating draws at SOFR + 2.25%. That is a contractual UST+200bp, disclosed, requiring no estimate. Against unsecured notes at a 10% effective rate, the same issuer pays roughly 334bp more to borrow against itself than against its assets. This is the neocloud credit market's verdict, and CoreWeave published it.
• The primary market said the same thing twice in eight days. Applied Digital priced $1.59bn of senior secured notes at a 7.00% yield on June 10, 2026, five times oversubscribed, backed by North Dakota capacity (Polaris Forge 1, +150MW) under a 15-year lease to CoreWeave; APLD has 400MW of CoreWeave leases in total. The prior tranche of the same project demanded 10% in November 2025 — 300bp of tightening in about seven months, reportedly on improved confidence in CoreWeave as tenant. Eight days later, on June 18, 2026, CoreWeave priced $1.25bn of unsecured notes at 9.625% (plus €2.0bn at 8.500%). Treasuries moved less than 10bp between the two dates, so the ~262bp coupon gap is not a rates artifact. [Bloomberg via Yahoo Finance; CRWV 8-K, June 18, 2026]
• CoreWeave's unsecured curve has not tightened at all. Coupons at issue: 9.25% (2030s, ~May 2025) → 9.00% (2031s) → 9.75% (2031s, April 2026) → 9.625% (2032s, June 2026). A year of "credit rebound" moved the unsecured coupon up, though tenor lengthened. Meanwhile its secured ladder fell 15% → 7%. The market repriced the collateral. It did not reprice the counterparty.
• Index spreads are tight — and that is not the same as calm. US high-yield OAS 271bp and investment-grade OAS 79bp, both July 15, 2026 (FRED, ICE BofA). On the only credit-spread series with genuine long history — Moody's Baa less 10-year Treasury, 160bp — that is the 7th percentile of the last twenty years and the 11th of the last ten. (A methodological note we think matters: FRED's ICE BofA OAS series now carry only a rolling ~3-year window, so no "tightest in twenty years" claim can be sourced from them. We use Baa10Y for long-run percentiles and say so. We do not claim "record tight" — the late 1980s were tighter.)
• But the repricing already happened once this year, and it did not stick. High-yield OAS gapped from 283bp (January 2) to 346bp on March 30, 2026 — roughly 63bp — then fully retraced to 263bp by mid-June and 271bp now. Investment grade did the same in miniature (79 → 94bp on March 16 → 73 → 79). Anyone claiming AI credit has not been tested in 2026 is not looking at the tape. It was tested in March and the market un-priced it.
• The juxtaposition that anchors this section. Oracle's 5-year CDS is REPORTED at a record ~198bp while the investment-grade index it belongs to sits at 79bp. An IG-rated issuer at roughly 2.5× its own index, at a record wide, inside a market at multi-decade tights. Corroboration: Oracle's 10-year bonds are REPORTED at ~6.52% (from ~5.34% at the start of 2026); against a 10-year Treasury of 4.55% that implies a ~197bp cash spread — essentially equal to the CDS print, which is meaningful because they are independent instruments. We flag honestly that sources conflict on the CDS level (125 / 139 / 174.9 / 198bp) and none carries a verified date; we quote ~198bp as REPORTED and date-unconfirmed.
• What this means. Spreads being tight and AI credit being repriced are both true simultaneously. The repricing is real, it is severe at the issuer level, and it has not propagated to the index. That is the disconnect our parent note built its argument around, in a sharper form: not placid public credit versus stressed private credit, but placid index versus stressed constituents.
• What we could not establish, and therefore do not assert: current secondary trading levels for any CoreWeave or Oracle bond (cbonds and FINRA TRACE are inaccessible without a subscription) — every CoreWeave pricing figure above is a primary-market coupon or yield at issue, not a secondary mark, and we do not imply otherwise; CoreWeave CDS levels; options skew for either name (a prior run's suggestion that ORCL's 25-delta skew is nearly flat is neither confirmed nor killed, and we do not use it); and 2026 year-to-date data-center ABS issuance. On 2025 volumes we resolve a conflict rather than report it: S&P's >$30bn global is coherent and RBC's $25bn is not, because US issuance alone was ~$26.2bn (~$15bn ABS + ~$11.2bn CMBS) — a global figure below the US figure is arithmetically impossible, so RBC's number is best read as ABS-only or a partial-year tally. (The US split itself is REPORTED and we could not reach its primary source; the resolution rests on it.)

§8 Two Falsifiable Tests

This note makes a claim. Two securities can break it. Each gets its own note; neither is prejudged here.

We are deliberately not valuing either name here. Stating a conclusion and then "testing" it in a companion we have already written would be theatre. What follows is the test design — what we would have to find to be wrong.

• Oracle is the purest test. It has the highest ratio of announced demand to demonstrated supply in the complex, and funds the gap with debt: FY2026 capex $55.66B, operating cash flow $31.98B, free cash flow −$23.69B. FY27 capex guidance is "~$70B net" — the CFO disclosing this is "net of customer prepayments and timing impacts expected at around $20B to $25B," i.e. ~$90–95B gross. Notably, "the prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion" — customers are prepaying for, or physically supplying, the GPUs. The falsifier: if that prepayment structure substantially de-risks Oracle's economics — closer to a toll operator on customer-owned assets than a principal — the thesis weakens at its most important node. The bear marker: S&P cut Oracle to BBB− on July 9, 2026 (outlook stable), naming OpenAI a "key credit risk," attributing ≈half of the $638B RPO to it, and forecasting FY27 free operating cash flow at ≈−$42B. ORCL closed at $124.21 on July 16, 2026.
• CoreWeave is the thesis in one security, and it already contradicts part of it. If our framing were simply "OpenAI concentration," CoreWeave would be the proof. It is not. Microsoft was ≈67% of FY25 revenue; OpenAI did not clear the 10% threshold through Q1 2026 — both the 10-K and Q1-26 10-Q say only that the company "expects OpenAI to be a significant customer in future periods." Despite a $22.4B headline. The falsifier: if CoreWeave's RPO ($98.8B at March 31, 2026, against $5.1B FY25 revenue) converts to cash at contractual rates with Microsoft as anchor, ceilings behave like floors and our central claim is wrong where it is most testable. The open threads: an unexplained Q1-26 reclassification moved deferred revenue $8.2B → $7.5B "primarily driven by reclassification of $1.3 billion to customer liabilities" (cause and counterparty undisclosed); capitalized interest grew 7.5× year-over-year ($97M in Q1-26 vs $13M in Q1-25, against $483M of contractual interest expense), which flatters reported earnings; and a securities class action (Masaitis v. CoreWeave, D.N.J., filed January 12, 2026) plus consolidated derivative litigation are pending with no accrual taken. CRWV closed at $72.91 on July 16, 2026.
• What we are not claiming. Neither company is alleged to have misstated anything. Every figure above is from its own filings. The argument is about what the disclosure regime permits to remain unsaid, not about misconduct.

§9 How It Most Likely Plays Out (Base Case)

• The repricing generalizes slowly, or not at all — and the wedge is the tell either way. The market has already sorted this complex into assets it will finance at UST+200bp and companies it will finance at UST+534bp. The base case is that this sorting spreads: more secured, non-recourse, SPV-level structures at tight spreads; more expensive and scarcer unsecured capital for the operating companies. That is not a crisis. It is a reallocation of who bears the conversion risk — from lenders to equity.
• Bifurcation, not contagion. Names with hyperscaler anchors (CoreWeave/Microsoft) and prepaid or customer-supplied hardware (Oracle's $75B) keep access. Names financing lab-facing capacity with unsecured debt and no anchor find that market closed first. The sector stops trading as a sector — which the June 10 / June 18 pair already demonstrates within a single name.
• The refinancing wall is the mechanism, not default. CoreWeave has $6.066bn of principal due in the remainder of 2026 and $5.652bn in 2027, against a $7.547bn current portion and total debt that grew $3.5bn in one quarter. Nothing needs to miss a payment. It needs to need new money. That is the fuse in this note's title: not a detonation, a refinancing.
• Watch the composition of the raise, not the size. CoreWeave's April 2026 $4.0bn 1.75% convertible against a 9.625% straight eight weeks later is the tell: at that spread the company is monetizing equity volatility rather than credit quality. That is a rational choice and a revealing one.
• Why this is slower than the equity tape implies. The private half of the market is drawn, illiquid, and not daily-marked — the same feature that turned the private-credit crash into a grind (parent note §4). Locked capital does not panic. It also does not inform.
• Why the March round-trip does not refute this. A 63bp gap-out that fully retraces in ten weeks is consistent with a market that discriminates within the theme rather than repricing the whole of it. The index never had to move, because the sorting happened underneath it.

§10 How Bad It Could Realistically Get (Bear Case, Bounded)

• The realistic worst case is a financing stop at the operating-company level, not a credit cascade. The assets are financeable — the 7.00% secured print, five times oversubscribed, proved as much. The companies are financed at deep high-yield with a near-term wall. A neocloud that cannot roll unsecured debt does not vaporize; it gets restructured, and its assets get taken by the secured lenders who always preferred them. Equity absorbs the conversion risk. That is the bear case, and it is survivable for the system.
• Three fast fuses would have to light together to exceed a sector workout: (1) an AI lab materially revises or walks a headline commitment, converting a ceiling into a demonstrated zero and repricing every comparable contract at once; (2) the vendor-financing chain fails to clear — SoftBank's OpenAI-share margin loan, cut 40% to $6B in May 2026 and reportedly still stalled in June, is the live precedent; (3) the stress transmits to the insurers holding the paper — the second-round channel the ECB warned about, and which our parent note documented at ~$1T of US life-insurer private-credit exposure.
• Why we bound it short of systemic. The exposure is large but concentrated in identifiable places; hyperscaler equity buffers are enormous; and Oracle at BBB− is still investment grade, with real, diversified, cash-generative businesses beneath the AI bet. This is not a banking-solvency structure.
• The honest tail we cannot size. If Oracle's OpenAI contract — roughly half of a $638B RPO, per S&P — were materially impaired, the mark would be unlike anything in the private-credit episode we described eight days ago, because it would hit a single investment-grade issuer's forward revenue base rather than a diversified loan book. No credible source offers an aggregate loss estimate for this complex. We decline to invent one.

§11 What Would Break This Thesis

Stated plainly, because a thesis worth publishing is one worth trying to break.

• The strongest objection: the wedge IS the market working. Secured paper always beats unsecured — that is what "secured" means. If 334bp is simply the ordinary structural premium rather than a judgment about AI, this note has dressed up a truism. Our answer: the level and the direction both argue otherwise. CoreWeave's secured ladder fell 15% → 7% while its unsecured coupon went 9.25% → 9.625% over the same period; an ordinary structural premium does not widen while the asset side tightens 800bp. And Applied Digital's identical project repriced 300bp tighter in seven months on tenant confidence, while the tenant's own paper did not move. That is a judgment about the counterparty, not about seniority. But we concede the objection is not fully answerable without secondary marks we could not obtain (§7), and a reader who thinks 334bp is unremarkable for a B-rated issuer is making a fair point we cannot conclusively rebut.
• Power re-leases, and the conversion never bites. The single best test we have found says the shell re-leases fast: Abilene re-tenanted from OpenAI to Microsoft in 21 days after the ~600MW expansion was reportedly scrapped (Bloomberg, March 6, 2026), with Crusoe announcing a 900MW campus for Microsoft on March 27, 2026, bringing the site to ~2.1GW — the exact figure OpenAI originally targeted. The tenant changed; the megawatts did not. Meta, the reported alternative, took 1.6GW of Crusoe capacity elsewhere in June. But this test is weaker than it looks: it re-leased land and power rights at development stage, before the specialized capex was sunk, and the terms were never disclosed — so "re-leased at or near original terms" is unverifiable and we do not claim it. No finished, liquid-cooled, GPU-specific hall has ever been publicly re-tenanted after a tenant walked. That experiment has not been run, and it is the one that would settle §3.
• OpenAI lists, and the opacity argument dissolves. The confidential S-1 of June 8, 2026 could convert the largest counterparty in the complex into an audited, markable, publicly-funded entity — answering §6's unknowns and severing the SoftBank collateral problem at once. An IPO is the single most plausible path by which this note ages badly.
• The labs simply grow into it. At OpenAI's growth rate an unprecedented revenue base is not impossible. 58× is a statement about today, not a law of nature.
• Prepayment de-risks the principal. Oracle's $75B of prepaid or customer-supplied hardware may mean the most exposed-looking balance sheet in the complex is substantially insulated. That is the ORCL note's central question and we have not answered it here.
• What would confirm us instead: a second rating action naming counterparty concentration; a lab formally revising the $1.4T; CoreWeave's next unsecured print coming wider still while its secured prints tighter; the wedge appearing at a second issuer; CoreWeave's RPO converting materially below contractual rates; or a second instance of vendor-chain financing failing to clear the way SoftBank's margin loan has.

§12 What to Watch (The Dashboard)

• The wedge at CoreWeave's next print. Secured versus unsecured, same issuer. If the gap widens, the market is hardening its view of the operating company. If it closes, we are wrong. This is the single cleanest instrument on the board and it requires no vendor data.
• Whether the wedge appears at a second issuer. One company's capital structure is an anecdote. Two is a market.
• Rating actions naming counterparty concentration. S&P/Oracle, July 9, 2026 was the first. A second makes this a sector view rather than an issuer view.
• Oracle's CDS against the IG index. ~198bp versus 79bp today (the former REPORTED and date-unconfirmed). Convergence refutes; divergence confirms.
• Whether the $1.4T is ever formally revised. Unrevised since November 2025, despite NVIDIA going $100B → ~$30B, Abilene losing ~600MW, and memory deals remaining LOIs. Silence is its own signal.
• CoreWeave's OpenAI revenue share. It has never cleared 10%. The quarter it does is the quarter ceilings started behaving like floors. The cleanest falsifier of §4.
• OpenAI's S-1. Confidentially submitted June 8, 2026. The flip to public would replace most of §6's inference with audited fact, in either direction. The single most consequential scheduled event for this thesis.
• NVIDIA's next 10-Q. "OpenAI" has appeared zero times since the FY26 10-K. Reappearance with a funded number would materially change §5.
• Stargate's operational gigawatts. ~0.3GW of >9GW planned. The conversion rate is the thesis, measured directly.
• Oracle's FY27 free cash flow against S&P's ≈−$42B forecast. The most quantified bear prediction on the tape.
• The unexplained CRWV $1.3B reclassification, and whether capitalized interest keeps compounding faster than the debt.

§13 House View

The call. The neocloud credit market has already decided that these assets are financeable and these companies are not — and it has published the price of that distinction in CoreWeave's own filings: UST+200bp secured and non-recourse, roughly UST+534bp unsecured, in the same quarter. The base case is not a coming repricing but the generalization of a repricing that has already occurred at the issuer level and has not reached the index: high-yield at 271bp and investment grade at 79bp (7th percentile of twenty years) alongside Oracle's REPORTED record ~198bp CDS. Expect bifurcation over roughly four to eight quarters — secured, SPV-level, non-recourse structures at tight spreads; scarce and expensive unsecured capital at the operating companies; conversion risk migrating from lenders to equity. The realistic worst case is a financing stop at the operating-company level that restructures mid-tier neoclouds and forces a mark at an investment-grade issuer, short of a systemic credit event, unless a lab walks a headline commitment, the vendor-financing chain fails to clear, and the insurer channel transmits — together. Confidence: moderate. The wedge is disclosed and unambiguous; the counterparty analysis rests on filings that are unusually clear about what binds. What we lack is secondary market data, any test of a finished GPU hall re-leasing, and a second issuer showing the same wedge.

Where the risk actually lands. Follow the wedge one layer down and it points somewhere specific. If the lenders are right — the contracts are ceilings, the counterparties are shaky — the loss that implies does not vanish because the secured creditors sidestepped it. It moves to whoever sits below them, and below the unsecured bondholder sits the equity. A secured lender who takes the plant and an unsecured lender who takes a haircut have, between them, handed the residual conversion risk to the shareholder. **The credit market has priced that risk at ~334bp. The question this raises — and pointedly does not answer — is whether the equity market has priced it at all, or whether it is still capitalizing the nameplate backlog as though the ceilings were floors. That is not a question a thematic note can settle, and we do not try: it is a single-name question, and we take it up directly in the companion valuation notes on Oracle and CoreWeave** (§8), which are built to test exactly this and are deliberately not prejudged here. We flag the direction only — that the interesting mispricing, if there is one, is more likely to be found in the equity than in the credit that has already moved.

What would change our view (observable, with thresholds): CoreWeave's secured/unsecured wedge narrowing at the next print (→ the market is warming to the operating company; we are wrong at the cleanest test); OpenAI's S-1 flipping public and showing a fundable business (→ the opacity and circularity arguments weaken at once; the fastest way we are wrong); CoreWeave disclosing OpenAI above 10% of revenue at contractual conversion (→ ceilings behave as floors); a formal downward revision of the $1.4T (→ confirms us, and dates it); Oracle's FY27 FCF landing materially better than S&P's ≈−$42B (→ prepayment is doing more work than we credit); a finished, GPU-specific hall re-leasing at or near original rent (→ §3's conversion is not the risk we claim).

The steelman against us. The bulls may be right that we have described an efficient market and called it a fuse. Every fact in §7 is a market correctly distinguishing secured from unsecured claims and charging accordingly; Applied Digital's 300bp of tightening is a market rewarding a real improvement in tenant quality; the March round-trip is a market that tested the theme and passed it. On that reading, nothing here is a warning — it is a photograph of price discovery working, and the S&P downgrade is one committee's opinion, not a regime. We hold our view because a market that will lend against the plant at 200bp and against the owner at 534bp is telling you it does not believe the owner's contracts, and because those contracts are ceilings written by counterparties funded in a circle (§4–§6). But the honest position is that the wedge is evidence of pricing, and whether that pricing is right is a judgment we cannot settle from public data. §11 is where this breaks, and we would rather mark it than bury it.

What the analysis implies — a lens, not a recommendation. This note is written to explain a financing structure, not to advise anyone to act on it, and it makes no recommendation to buy, sell, or avoid any security. Three understandings fall out of the work. First, the capital structure is the opinion: where an issuer's secured and unsecured claims price differently by hundreds of basis points at the same moment, the gap is the most honest credit research available and it costs nothing to read. Second, an RPO is a disclosure, not a receivable — it tells you what was signed, not who signed it, whether they must pay, or whether they can; the same regime that produced Oracle's $638B backlog also permitted a truthful statement that no customer exceeded 10% of revenue. Third, the index is not the market: high yield at 271bp and Oracle at a reported record CDS are the same tape, and aggregate calm is compatible with severe idiosyncratic stress. The aim is comprehension of how this buildout is financed, so a reader can judge it for themselves.

§14 Disclosures

Information only. This material is published by TON618 Capital 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; material conflicts are disclosed where they exist. As of the date of this note the Fund holds no position in Oracle, CoreWeave, Applied Digital, NVIDIA, AMD, Broadcom, Microsoft, Amazon, Cerebras, GE Vernova, or any other single security named herein; it maintains a strategic Bitcoin allocation, and to the extent this note bears on global liquidity conditions that inform that allocation, the reader should weigh that as a material interest. The Fund receives no compensation from any party in connection with this research.

Use of AI. Artificial intelligence is used in the creation of this research. A material conflict is disclosed here: this note analyzes the compute contracts, vendor financing, and counterparty credit of AI laboratories, and the AI tooling used to produce it is made by one of the companies discussed. 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.

Source & Verification Note

This note is the sequel to The Private Credit Reckoning (July 8, 2026) and takes apart the accelerant that note identified in its own §8 (Accelerant I: the AI & data-center financing loop). It synthesizes four completed primary-source research sweeps: OpenAI's compute web (Oracle's FY2026 10-K and three FY2026 10-Qs, AMD's 8-K and filed warrant exhibit, and NVIDIA, Microsoft, Broadcom, CoreWeave, Amazon and Cerebras filings pulled directly from EDGAR); the private neocloud financing complex (Lambda and Crusoe, from raw Form D primary_doc.xml parsed on sec.gov); the power-constraint question (PJM auction reports and its Independent Market Monitor, PJM Planning Committee materials, LBNL's Queued Up 2025 edition, GE Vernova's Q1-2026 8-K via EDGAR, KBRA's methodology and a live rated deal); and credit pricing (FRED, and CoreWeave's Q1-26 10-Q debt footnote and June 2026 8-K).

High-confidence, filing-verified. CoreWeave's DDTL 4.0 pricing ("2.00% per annum plus a blended Treasury rate"; SOFR+2.25% floating; $8.5bn; non-recourse; CoreWeave Compute Acquisition Co. VIII; MUFG agent), the full effective-rate ladder (15/11/9/9/7% secured vs 10% unsecured), total principal $25.149bn at March 31 2026 (from $21.615bn), the $7.547bn current portion, $6.066bn due in the remainder of 2026, Q1-26 interest expense $483M with $97M capitalized (vs $13M), the June 18 2026 unsecured pricing ($1.25bn at 9.625%, €2.0bn at 8.500%), the April 2026 $4.0bn 1.75% convertible and the $1.0bn Jane Street equity at $109.00, the pending Masaitis securities action and consolidated derivative litigation with no accrual — all DISCLOSED. Oracle's RPO series ($138B → $455.3B → $523.3B → $552.6B → $638B) and verbatim recognition schedule; the absence of "OpenAI," "gigawatt," and "Stargate" from the FY26 10-K and all three FY26 10-Qs, and the three-hit EDGAR full-text result whose only substantive instance is the February 1 2026 FWP; FY26 capex $55.66B, operating cash flow $31.98B, FCF −$23.69B; the $75B prepaid/customer-supplied figure. AMD's warrant terms (160M shares, $0.01 strike, $600 final tranche, binding only for the initial 1GW, zero tranches vested as of March 28 2026, Exhibits E and F redacted). NVIDIA's LOI language across three successive filings and its disappearance from the Q1 FY27 10-Q. Microsoft's 27% as-converted stake, HLBV method, $13B/$11.8B funding, the three quarterly P&L figures, the "up to three months" lag, both restructurings. Broadcom's 1.3GW-of-10GW contractual commitment and the zero-hit EDGAR result. CoreWeave's RPO series, 67% Microsoft concentration, and verbatim "up to" ceiling language. Amazon's $38B + $100B/8yr disclosure. Cerebras's "contractually committed" language and the $1B OpenAI promissory note. The Crusoe/Lambda Form D discrepancies. GE Vernova's gas backlog 83→100 GW in one quarter against 4 GW shipped. FRED levels: HY OAS 271bp and IG OAS 79bp (July 15 2026), the March 30 peak of 346bp, and Baa10Y at 160bp.

Deliberately hedged, not asserted as fact: the ~$300B Oracle–OpenAI contract terms (REPORTED, WSJ ~September 10 2025; never confirmed by Oracle in any filing); the S&P downgrade details of July 9 2026 — the rating, "key credit risk" language, ≈half-of-RPO attribution and ≈−$42B FY27 forecast are consistent across multiple carriers, but S&P's own release 403s automated retrieval and we have not read it directly; Oracle's CDS level, where four sources conflict (125 / 139 / 174.9 / ~198bp) and none carries a verified date — we quote ~198bp because it independently coheres with a ~197bp implied cash spread (10-year bonds REPORTED at ~6.52% against a 4.55% Treasury), but the yield input is itself REPORTED; the Applied Digital ComputeCo 3 spread of ~260–275bp and CoreWeave's ~534bp, both ESTIMATES — the 7.00% is a yield of unknown tenor and the 9.625% assumes par issue (the April 9.75% add-on reportedly priced at 102), though the ~262bp raw coupon gap requires no assumption at all since Treasuries moved <10bp between June 10 and June 18; the Abilene ~600MW cancellation and Microsoft re-tenanting (REPORTED; Crusoe's own release names Microsoft and 900MW while later reporting says 700MW — unresolved; the same-parcel identity is inferred, not disclosed, and Oracle disputes the cancellation framing); OpenAI's $5.7B Q1-2026 recognized revenue and ~$24B annualized figure (REPORTED; valuations conflict at $110B/$730B vs $122B/$852B and we do not adjudicate); NVIDIA's reported scaling to ~$30B and the "probably not in the cards" quote (REPORTED, primary text not read). The Abilene tenant identity is not named by us: deal counsel (Kirkland & Ellis, for the Blue Owl funds) says only "Fortune 100 hyperscale tenant."

What we could not verify, and therefore did not assert. (1) Secondary marks: cbonds and FINRA TRACE are inaccessible without subscription — every CoreWeave and Oracle pricing figure here is a primary-market coupon or yield at issue, not a secondary trading level, and §7 says so. (2) Options skew: not obtained; the prior suggestion that ORCL's 25-delta skew is "nearly flat" is neither confirmed nor killed and is not used — a near-flat skew on a name at a record-wide CDS would be surprising, and surprising claims need verification before print. (3) The finished-hall test: no GPU-specific, liquid-cooled hall has ever been publicly re-tenanted after a tenant walked; Abilene tested land and power rights at development stage, which is a materially easier test, and its terms were never disclosed. (4) KBRA's re-leasing and residual assumptions are paywalled; we confirmed the methodology's scope, date, final status and factor headings only, and we attribute rating-driver claims to KBRA specifically — Moody's, S&P, Fitch and Morningstar DBRS positions were not retrieved and we do not generalize to "the rating agencies." (5) Moody's and Fitch action on Oracle in 2026: none found, on one search pass — absence of evidence, not evidence of absence. (6) 2026 YTD data-center ABS issuance. (7) ERCOT and MISO queues were not researched, and Abilene sits in ERCOT — so the PJM capacity evidence is not drawn from the market the decisive test case occupies, which is a real analytical seam. (8) GPU residual values: deliberately out of scope — see The Depreciation Question; this note argues the credit risk sits in the conversion rather than the chip and rests on no residual-value view. (9) Whether NVIDIA has funded any OpenAI amount: the $17.9B net additions and $11.4B→$27B commitment progression are consistent with a partially-funded ~$30B, but that is arithmetic inference, not disclosure, and we state no funded figure. (10) OpenAI's announcement pages are Cloudflare-blocked to automated retrieval; all OpenAI-issued language here is sourced via counterparties' filings or secondary reporting.

A methodological constraint readers should know about. FRED's ICE BofA option-adjusted-spread series now carry only a rolling ~3-year window (BAMLH0A0HYM2 returns 795 observations beginning July 17, 2023, regardless of the requested start date — verified against the API metadata, not a truncation cap). No "tightest in twenty years" claim can be sourced from those series today. We therefore use Moody's Baa less 10-year Treasury (10,133 daily observations back to January 2, 1986) for long-run percentiles, and we state that basis explicitly rather than implying the HY and IG figures carry that history. We also decline "record tight" — the late 1980s were tighter.

Corrections made during verification, worth propagating: Broadcom's "$10B mystery XPU customer" was Anthropic, not OpenAI — denied by Broadcom on October 13 2025 and confirmed as Anthropic on December 11 2025; the stale attribution still circulates widely. Broadcom's +$119.6B RPO jump to $164.6B (May 3 2026) names no counterparty and is likelier Meta. Oracle's FY27 capex figures of ~$70B and ~$90–95B are not in conflict — the former is net of customer prepayments, the latter gross. "Backlog" and "RPO" are not interchangeable for Broadcom (~5× apart). Headline round sizes systematically overstate Form D proceeds: Crusoe's Series E was $1.027B sold against a $1.375B headline. On data-center securitization volumes we resolve rather than report the S&P/RBC conflict: >$30bn global for 2025 is right and $25bn is not, because US issuance alone was ~$26.2bn — though the US split is REPORTED and load-bearing for that resolution. And PJM's "11.7 GW" and "9.3 GW" RRI figures are not in conflict: they are 11,793 MWE (nameplate summer values) and 9,361 MW UCAP (unforced capacity) from the same PJM slide, an implied ~79% derate — any pitch quoting nameplate megawatts is quoting the larger number.

Corrections made to this note's own draft, disclosed because the note criticizes others for the same class of error. The adversarial pass and the subsequent research forced five. (1) A draft of the credit-market section stated that AI-adjacent bond yields moved "from ~7% to ~10%." This was backwards — Applied Digital's ComputeCo notes moved from ≈10% to ≈7%, i.e. tightened, and the corrected fact became load-bearing evidence for the note's central wedge rather than against it. (2) The draft omitted OpenAI's confidential S-1 submission of June 8, 2026, a material public fact five weeks older than the note's own dateline; it is now in §6, §11 and §12 and identified as the most plausible route by which this note ages badly. (3) The draft asserted that SoftBank's "lenders cut that margin loan 40%"; the reporting says SoftBank cut its own target on lender hesitancy, and the reduced $6B then stalled. (4) A facility count of "seven" was corrected to eight. (5) An earlier draft's central claim — "the collateral is a contract, and the contract is a ceiling" — was the wrong axis and has been replaced. The power research established that the collateral is power early and the contract late, and that neocloud lending finances the conversion between them; the credit-pricing research then found that conversion already priced at ~334bp inside a single issuer's filings. The note was rewritten around the evidence rather than defending the original framing. That earlier draft also conceded the power question as unrebuttable; it is now answered, in a way that neither side of the original debate had it right.

Market levels are July 16, 2026 official closes, pulled from the Schwab market-data API (not from search results): ORCL $124.21 · CRWV $72.91 · NVDA $207.40 · AMD $500.94 · AVGO $374.45 · MSFT $401.10. Credit levels are FRED as of July 15, 2026. A verification pass surfaced conflicting third-party equity prices for AMD (~$483.59) and AVGO (~$394.28); the authoritative feed confirmed the values above.

Companion notes: The Private Credit Reckoning (July 8, 2026) supplies the framework this note inherits; The Depreciation Question covers useful-life restatement. Single-name treatments of Oracle and CoreWeave — the two falsifiable tests designed in §8 — follow separately and are deliberately not prejudged here.

Version 1.0 · analyst: TON618 Thematic Research.