Where to Buy Protection: Entry Points for Neocloud Credit

Where to Buy Protection: Entry Points for Neocloud Credit, Read From the TIRA File

17 October 2026 | Theseus Capital, Decision Desk | USD unless stated otherwise; bp = basis points; pts = points of notional

Source: the six TIRA documents and the cash flow workbook in AI Capex / TIRA, the desk's nine-issuer credit file of October 17, filings on EDGAR, and the bond and CDS marks of October 6 to 16

The finding in one line

The firm's thesis has two legs, spreads widen and payments stop. The first leg is already more than half priced in the only two names with a liquid CDS market, CoreWeave and Oracle; the second leg is not purchasable through CDS for eight of the ten issuers in the file, and for CoreWeave the most probable form of non-payment, a liability-management exchange, does not trigger a standard North American contract at all. The ideal entry point is therefore not a spread level. It is a set of four conditions: a contract that exists and that Theseus can clear, a tenor matched to the quarter the funding gap actually opens, a CDS level at or inside what the cash bonds imply, and a quiet tape rather than a panic print. On October 15 only the first condition held, and only for two names. [Confidence: High]

Counterargument first: why the thesis is right and the trade may still lose

Take the strongest version of the opposing case before the supporting one, because the supporting one is already in the TIRA file and on the front page of every financial site this week.

The market has done the work. CoreWeave five-year protection went from 368bp on October 6 to 1,194.5bp intraday on October 15, a move that prices a five-year default probability of about 63% at a 40% recovery and 53% at a 20% recovery. No agency moved. The company has 15.6 billion dollars of liquidity at June 30 before the 4.1 billion September convert, the 1.2 billion DDTL 5.5 draw and the 1.2 billion revolver draw in Q3, a 104 billion backlog with Microsoft, Meta and OpenAI at about 70% of revenue, a 6.3 billion NVIDIA order form for unsold capacity plus 2 billion of NVIDIA equity, 1 billion of Jane Street equity, and a 35 million share ATM filed September 17 whose stated purpose is migrating toward investment grade. A protection buyer at 1,100bp is not betting that the capital window closes. They are betting that it closes with a probability higher than six in ten and that it closes in the form of a Chapter 11 or a missed coupon rather than a negotiated exchange. The desk's own prior, set out below, is lower than that on both counts. [Confidence: High on the arithmetic, Medium on the prior]

The payment leg has a contractual hole. Standard North American corporate CDS trades with no restructuring credit event. Bankruptcy and failure to pay are the triggers. A company with 10.8 billion of converts struck 50 to 95% above its stock, 9.25 to 9.75% notes trading at 81 to 88, and relationship lenders in Blackstone, Magnetar and Carlyle will, under stress, offer an exchange: unsecured into second lien, converts re-struck, maturities pushed. Holders who refuse keep a smaller, subordinated claim. None of that is a credit event. The protection buyer watches spreads tighten after the exchange closes and takes a mark-to-market loss on a company that has just confirmed it cannot pay its debts as written. This is how Codere, Hovnanian, and most sponsor-backed distress has resolved since 2014, and CoreWeave's cap table (Magnetar, Coatue, NVIDIA, Jane Street) is a sponsor cap table. [Confidence: High on the convention, Medium-High on the path]

The counterparties want the company alive. Microsoft, Meta and OpenAI have 22 to 35 billion dollars of contracted capacity each. Jensen Huang told Bloomberg on October 15 that CoreWeave's cost of capital is "actually decreasing" and that NVIDIA does not need to backstop it. The second half of that sentence is the one to price: an NVIDIA that chooses not to backstop a 42 billion dollar credit it has 8 billion of exposure to is a possibility, and an NVIDIA that does is a tightening of 400bp in a day. A rescue is the single largest mark-to-market risk to a protection position and the single largest squeeze risk to the equity book the desk opened for the October 19 open.

The TIRA near-term dates are softer than they read. Every one of the six documents puts the first stress-case funding gap inside eight months: CoreWeave Q4 2026, Nebius Q4 2026 to Q1 2027, IREN March 2027, Cipher Q1 to Q2 2027, Oracle February to May 2027, WULF Compute at completion. Every one of them also shows a defensive case with no gap through mid-2027, and every one excludes uncommitted financing and new equity by construction. The models measure dependence on the capital window. They are silent on whether it closes. For CoreWeave specifically, the stress case assumes a net deferred-revenue contribution of 750 million dollars in H2 2026 against company disclosure of more than 9 billion of 2026 prepayments and a deferred-revenue balance that rose to 9.7 billion by June 30. The 1.53 billion Q4 2026 gap is smaller than the plausible error in that one line. The first thing the Q3 10-Q should be read for is deferred revenue; if it is up again, the Q4 date is gone and the thesis is a 2027 thesis. [Confidence: Medium-High]

None of this says the thesis is wrong. It says the trade is late in the two liquid names, structurally mismatched to the most likely failure mode, and unavailable in the names where the failure mode is cleanest. The rest of this memo is about where, when and in what form it is still worth doing.

What the TIRA file establishes

The six documents (the consolidated CDS report of October 15, the Cipher assessment of October 16, and the Nebius, IREN and Oracle cash flow papers of October 17, with the companion workbook) are disciplined. They separate reported figures from analyst scenarios, refuse to count guarantees as cash or project collateral twice, and insist on the legal-entity question before the instrument question. The findings worth carrying forward, with the desk's reading of each:

Two things the file does not do, and this memo attempts. It does not price the protection against the market's own view, beyond illustrative budgets with assumed terms, and it does not rank the issuers by whether a CDS buyer can actually reach the loss. The CDS report is explicit that no dealer was contacted, no quote obtained, and no basket weights assigned. That is correct discipline, but it leaves the question Max asked, where the entry is, unanswered.

What the market is pricing, and against what

The arithmetic below uses a flat hazard rate, a 4% discount rate, the standard 500bp running coupon for high yield and 100bp for investment grade, and quarterly premium with accrual on default. It is the ISDA standard model in miniature, not a dealer's curve, and it is accurate to within a point or two for the purpose of comparing entry levels. Recovery is shown at 40% (convention) and 20% (closer to where S&P's RR5 on the CoreWeave unsecured, 10 to 30%, would put it). [Confidence: High on method, Medium on recovery]

CoreWeave five-year: cost of entry and what is implied

Break-even: the probability of a covered credit event by year 2.5 at which the expected payout (80 points at 20% recovery, less upfront and 2.5 years of coupon) equals the expected cost of a no-event hold to maturity (upfront plus 25 points of coupon). Undiscounted. The upfront at 368bp is negative because the spread was below the 500bp coupon: a buyer on October 6 was paid 5 points to own protection.

The desk's prior for a covered credit event at CoreWeave Inc., meaning bankruptcy or a failure to pay on parent obligations and excluding exchanges that do not trigger: 20 to 30% by December 2028, 40 to 50% by December 2031. [Confidence: Medium] That prior sits below what 1,038bp implies at either recovery and roughly at what 800bp implies. Bought at 1,038 and held, the position has negative expected value on the desk's own numbers. Bought at 643 it has positive expected value with a margin. The window between those levels was open for about three sessions, October 8 to 10.

CoreWeave mark-to-market after a twelve-month hold

Points of notional. A twelve-month hold costs 5 points of coupon plus about 1 point of time decay. The move from 368 to 1,195 paid an October 6 buyer 20 points; the same 800bp of widening from 1,195 to 2,000 pays about 11, because the risky annuity shrinks as the hazard rises (4.0 years at 288bp, 2.9 at 1,195bp, 2.25 at 2,000bp). The convexity that made the October 6 entry good has already been spent.

The basis: the bonds say the CDS was fair late Thursday and rich at the print

Solving each cash bond for the flat spread that reproduces its October 15 price gives the term structure the bond market was willing to pay for, without the CDS market's panic premium.

Three readings. First, the five-year point at 1,038bp late on October 15 was within 30bp of the five-year bond; the 1,194.5 intraday print was 130 to 160bp wide of it. A buyer who lifts the panic offer is paying a positive basis that will mean-revert against them even if nothing else moves. Rule: do not buy protection more than about 100bp wide of the bond-implied level; at that point the cheaper short is the bond itself through a prime broker. [Confidence: Medium-High]

Second, the curve is upward sloping in hazard terms, 876bp at 3.7 years rising to 1,109bp at 5.8 years. A truly distressed name, one the market expects to fail within eighteen months, inverts. This one does not. The bond market is pricing the 2030 to 2032 maturity wall, roughly 14 billion of 2031 paper stacked against the 2030 notes, DDTL 2.0 and 3.0 and the 2032 notes and converts, not a 2027 liquidity event. That is exactly the TIRA ordering in reverse: TIRA says the gap opens in Q4 2026 to 2027; the curve says the market thinks 2027 is funded and 2030 is not. If Theseus believes TIRA's timing, the front end of the curve is where the market disagrees with the firm, and that is where a protection buyer should be. [Confidence: Medium-High on the inference; the one-year and three-year CDS quotes were not found in any public source and must be obtained from a dealer]

Third, the cheapest-to-deliver option favours the buyer. Under the 2014 Definitions convertible bonds are deliverable, and CoreWeave's 1.75% 2031 and 2032 converts and 2.875% 2033 converts, with strikes 75 to 95% above the stock, would trade well below the straight notes in a default. The auction recovery would be set off the converts, so the payout on a covered event is likely higher than the 20 to 30% straight-bond recovery suggests. This is the one place the convert stack helps a protection buyer. [Confidence: Medium-High; confirm deliverability with the dealer's confirmation]

Tenor: match the contract to the quarter the firm thinks the gap opens

If a covered event lands at month 15 with 20% recovery, the gain as a multiple of the maximum no-event loss is 4.7x for the two-year, 2.8x for the three-year, 1.8x for the four-year and 1.4x for the five-year. The five-year is the liquid point and the one every data vendor quotes, which is why it is the one everyone bought this week. It is also the one that pays least for the firm's actual thesis.

The desk's recommendation on tenor: if the firm's conviction is TIRA's conviction, that the funding window closes in 2027, buy three-year protection and accept the illiquidity; it costs half the upfront of the five-year, covers the window, and does not pay for the 2030 to 2031 wall that the market has already priced. If the firm's conviction is the credit file's, that CoreWeave is engineered to survive 2027 through exchanges and strategic money and fails at the wall, the five-year is right and it is already expensive. The desk leans to the first reading on timing and the second on form, which argues for the three-year as the core and a small five-year only on a retracement. [Confidence: Medium]

The payout problem, in order of probability

The firm's second leg, payments stop, has to be restated as: which forms of non-payment pay a protection buyer. Ranked by the desk's estimate of their probability conditional on CoreWeave entering stress in 2027:

1. Equity cures, ATM issuance and asset-level resizing. DDTL 3.0 allows unlimited equity cures to October 28, 2026 and defers its first DSCR test to October 2027; DDTL 5.0 and 5.5 carry resizing tests and cure rights. The 35 million share ATM is live. Probability this is the first response: very high. CDS consequence: none, spreads tighten modestly. [Confidence: High]
2. Amend-and-extend with the private lenders. Blackstone, Magnetar and Carlyle hold the DDTLs; they resize, extend and take fees. Probability: high. CDS consequence: none.
3. Strategic capital. A NVIDIA facility, a Microsoft or OpenAI prepayment acceleration, or a Jane Street style equity cheque. Probability: medium-high, and this is the event that squeezes every short in the complex at once. CDS consequence: 300 to 500bp of tightening in a session. [Confidence: Medium]
4. Distressed exchange of the unsecured notes and converts. Unsecured into second-lien at a discount, converts re-struck lower with a coupon step-up, maturities pushed to 2033 and beyond. Probability: medium if items 1 to 3 do not suffice. CDS consequence: none under no-restructuring terms; spreads tighten after the exchange; the protection buyer loses the upfront and the carry on a company that has just defaulted in every economic sense. This is the central risk to the firm's second leg. [Confidence: High on the convention]
5. Failure to pay or Chapter 11. Requires the lenders, the strategics and the equity market to refuse simultaneously, or an operational failure (a Microsoft or OpenAI non-renewal at scale) that makes the exchange math impossible. Probability conditional on stress: low to medium, 25 to 35%. CDS consequence: the auction, with the converts as cheapest-to-deliver and a payout plausibly in the 70 to 85 point range. [Confidence: Medium]

The product of these is the desk's unconditional prior above. The honest summary is that the firm's second leg is correct as economics and about one-third correct as a CDS trigger. The way to close that gap is not to buy more five-year protection. It is to express the exchange scenario directly: short the unsecured notes or the converts through a prime broker, where a distressed exchange is a gain, or buy puts on the equity, where it is also a gain, and keep CDS for the tail where the exchange fails.

The access problem: who has a contract at all

Theseus has no ISDA and no prime broker with CDS clearing. The credit file calls the credit book a shadow book pending an ISDA, and that is the binding constraint on everything above. Onboarding with a dealer for single-name CDS runs months, requires an eligible-contract-participant structure, initial margin typically 10 to 25% of notional on a distressed name, and minimum tickets of 1 to 5 million dollars. The firm's deep research on the entry point should run in parallel with that onboarding, not ahead of it, because the levels in this memo will not be the levels when the account opens. [Confidence: High]

Where the vulnerable link is, and whether protection reaches it

The credit file's central observation is that the nine issuers are one credit wearing nine tickers, and the transmission map makes the CDS question tractable. There are two clusters and two owner-operators.

The CoreWeave cluster: CoreWeave, Core Scientific, Applied Digital, Galaxy. CoreWeave is 100% of today's rent at all three landlords. A CoreWeave failure to pay stops the Core Scientific lockbox, the Applied Digital rent line and the Galaxy data-centre revenue line inside one quarter. The debt-service reserves (345 million at Core Scientific, construction reserves at Galaxy's Helios II) buy twelve to eighteen months, after which the uncapped parent completion guarantees pull the parents in. The landlord bonds are secured on shells with lockboxes and 80 to 95% loan-to-cost; their loss-given-default is high if the tenant leaves and modest if the tenant is replaced. The parent equities are the first casualty and the project notes the last. There is one CDS contract in this cluster and it references the entity whose default is the common cause. CoreWeave five-year protection is, in effect, a first-to-default basket on four of the nine names for the price of one. That is the strongest argument for owning it at all. [Confidence: High]

The Google cluster: TeraWulf, Cipher, Hut 8. Fluidstack is the tenant and Google the backstop, 3.2 billion at TeraWulf, 1.73 billion at Cipher effective only after rent commencement, undisclosed at Hut 8. The project notes trade on Google: WULF Compute at 102.5 and a 267bp Z-spread. A demand hiccup does not reach these creditors unless it travels through construction delay, delivery termination or a Fluidstack failure that Google does not cover. The vulnerable link here is the parent, through the completion guarantee and the convert stack, not the project debt. There is no CDS, and the right expression is already in the equity book. This cluster is a construction bet, and the firm should not describe it as a credit bet.

The owner-operators: Nebius, IREN. Microsoft and Meta behind them, 14.5 billion and 7.6 billion of cash, 14 billion and 6.7 billion of converts. Both are late-cycle failures if they fail at all; both have the convert stack as the only unsecured instrument; neither has a CDS. Nebius is the name whose stress-case gap in the TIRA file is most sensitive to a single assumption, the quarterly run-rate of new customer advances, and that assumption is observable in the deferred-revenue line every quarter. The distress marker in the credit file, deferred revenue falling while capex continues, is the trigger to short the 2034 converts, not before.

Oracle, the investment-grade link. Oracle is in the file because it is the hyperscaler whose balance sheet most resembles a neocloud: 105 billion of gross capex against 23 billion of operating cash in FY2027 under the sustained case, 11.4 billion of Q1 customer advances that will not repeat, a 20 billion ATM already done and 20 billion of the 40 billion FY27 plan still an intention. S&P took it to BBB minus in July citing a 42 billion FY2027 free operating cash flow deficit. One more notch and Oracle is the largest fallen angel in the history of the index, and the forced selling from investment-grade mandates would take the five-year from 288 toward 450 to 550 regardless of any change in default probability. The five-year at 288 prices a 21% five-year default probability at 40% recovery, which is absurd for a company covering interest 4.7 times; it is pricing the downgrade, not the default. The TIRA Oracle paper is right that this is a spread trade. The desk adds that it is a cheap one: 7.6 points upfront and 1 point running against a plausible 9 to 12 point gain on a fallen-angel event, with a 12-month carry cost of 2.3 points. [Confidence: Medium-High]

Oracle's Q2 FY2027 report in the second week of December, the completion or failure of the remaining 20 billion of the FY27 financing plan, and any agency action are the three dated catalysts. The revolver covenant is 3.0x EBITDA to net cash interest under the agreement's own definitions; the 4.71x EBIT-to-gross-interest figure in the TIRA paper is not that test and cannot be used to call a breach.

The entry framework

An entry is a set of conditions, each of which has to be true, not a number on a screen.

Condition one, the contract exists and clears. CoreWeave and Oracle today; SoftBank as an adjacent; nothing else. Until the ISDA is signed the credit book is a watch list and the expression is bonds, converts and puts through whatever broker the equity book uses.

Condition two, the basis. Buy CDS only at or inside the bond-implied level; more than 100bp wide, short the bond instead. On the Oct 15 close that condition held; at the intraday print it did not.

Condition three, the tenor matches the thesis. Three-year for the 2027 funding thesis, five-year for the 2030 wall thesis. Do not buy five-year protection to express a 2027 view; the market has already paid for 2030.

Condition four, the level. For CoreWeave five-year, the desk's buy zone is 550 to 800bp, full size at or below 650, where the implied five-year default probability (41 to 49% at 40% recovery) is at or below the desk's own prior. Above 1,000bp the position is a mark-to-market trade sized to the upfront and should be treated as such: enter only with a dated catalyst inside 90 days, and exit into the catalyst. Between 800 and 1,000, the three-year is the only tenor with positive expected value on the desk's numbers. For Oracle five-year, buy at or below 250, hold through the December report, target 400 to 500 on a downgrade; above 350 the downgrade is priced and the position is a default trade the firm does not believe in. [Confidence: Medium]

Condition five, the tape. The October 6 to 15 move was 368 to 1,195 in seven sessions with no agency action, no filing, and no missed payment. Single-name spikes of that kind in names that do not then default retrace 30 to 50% within one to two months; Carvana in 2022 and Altice in 2024 are the recent templates. The probability the CoreWeave five-year retraces into the desk's buy zone before the Q3 report in mid-November: 35 to 45%. If it does, that is the entry. If it does not, it is because something in the Q3 numbers confirmed the panic, and the three-year is the instrument. [Confidence: Medium-Low on the retracement odds]

Sizing. Define the position by its maximum loss, which for bought protection is upfront plus coupon to the exit date. At 1,038bp a 10 million five-year costs 1.65 million upfront and 500 thousand a year; a twelve-month hold puts 2.15 million at risk for a 5.7 to 6.2 million payout on an event in the first eighteen months. At 650bp the same 10 million costs 500 thousand upfront and the same running, 1.0 million at risk over twelve months for a 7.0 to 7.5 million payout. Entry level is a three-times difference in the risk-reward of the same contract. The desk's rules cap maximum loss per ticket at 3% of NAV and CDS fits that rule exactly, which the nine equity shorts, with their unbounded loss, do not. When the ISDA is in place, bought protection should replace rather than add to the CoreWeave-cluster equity shorts. [Confidence: High on the arithmetic]

Catalyst calendar

What would change this view

Toward buying more and sooner: the Q3 10-Q showing deferred revenue flat or down with capex at guide, which validates TIRA's Q4 2026 date; a CoreWeave 2027 financing announced at a cost above 10%; a pulled or repriced project print at Applied Digital or Core Scientific; a receivable build at Galaxy or Core Scientific; S&P or Fitch moving the CoreWeave outlook to negative; Oracle's remaining 20 billion coming as debt rather than equity.

Toward standing down: deferred revenue up again in Q3; a committed parent-level 2027 facility; a strategic prepayment or NVIDIA facility; an unsecured print inside 10%, which is Jensen Huang's claim and would be the single most important datum against the firm; the five-year curve inverting, which would mean the market has moved to TIRA's timing and the front-end edge is gone; Oracle's financing completed in equity.

Work order

• Begin dealer onboarding now. The quote request template in the TIRA CDS report (legal entity and RED code, seniority, 3y and 5y, 1 / 5 / 10 million, two-way, upfront and running, credit events and restructuring election, settlement, deliverables, succession) is complete and should go to two dealers for CoreWeave and Oracle the day the ISDA is signed. Ask specifically for the one-year and three-year CoreWeave points; nothing public carries them.
• Set alert levels in the desk: CoreWeave five-year 800 (buy zone top), 650 (full size), 1,500 (take profit on any position or roll five-year into three-year); Oracle five-year 250 (buy) and 400 (take profit). Record the bond-implied level alongside each CDS reading so the basis is visible on the terminal.
• Re-run the TIRA CoreWeave workbook the day the Q3 10-Q files with actual deferred revenue, actual DDTL draws and the disclosed 2027 commitments, and move the first-gap date accordingly. The workbook's scenario switch and the repayment-timing adjustment are already built for this.
• Price the exchange scenario directly: obtain borrow and quotes on the CoreWeave 9.625% 2032 and the 2.875% 2033 converts through the equity prime broker. These pay on a distressed exchange where CDS does not.
• Ingest this memo and the six TIRA documents into the corpus under the credit section so the council's credit analyst reads them before the next CoreWeave round, and add the catalyst dates above to the desk's filing watch.
• Reconcile the equity book to the defined-risk rule. Nine unbounded shorts in one factor at 45% gross is outside the handoff's own constraints; the CoreWeave 67.5/55 March 2027 bear put spread is the compliant expression and bought protection, once available, is the other.

Sources

TIRA, AI Capex / TIRA on the desktop: AI_Credit_Protection_Report.docx (October 15, 2026; evidence cutoff October 14), Cipher_Credit_Thesis_Assessment.docx (October 16), Nebius_Cash_Flow_and_Credit_Risk.docx (October 17), IREN_Cash_Flow_and_Debt_Serviceability.docx (October 17), Oracle_Free_Cash_Flow_and_Debt_Serviceability.docx (October 17), AI_Credit_Cashflow_Model_v2.xlsx (Summary, Assumptions, CoreWeave CF, Debt register, Projects, GPU stress, CDS economics, Evidence, Channel split; research as of October 14). Each document carries its own filing-level source register, which this memo does not repeat.

Desk: neocloud-miner-credit-oct-2026 (corpus, October 17, 2026) for capital structures, covenants, tenant maps, October 15 bond marks and the CDS series; icarus-ledger for the financing clause; desk-handoff for the defined-risk rules.

Market: CoreWeave five-year CDS 368 (Oct 6), 643 (Oct 9), 982 (Oct 14), 1,194.5 intraday and about 1,038 late (Oct 15) per ICE Data Services, S&P Global Market Intelligence and LSEG as reported by Bloomberg (https://news.bloomberglaw.com/bankruptcy-law/coreweave-risk-gauge-hits-record-pricing-60-odds-of-default) and CNBC (https://www.cnbc.com/2026/10/15/coreweave-credit-default-swaps-hit-distress-levels-as-investors-sour-on-ai-debt.html; https://www.cnbc.com/investingclub/2026/10/16/coreweave-credit-default-swaps.html). Oracle 288 and SoftBank 424 (Oct 15) per Reuters, AI credit worries send default insurance to record high (https://www.insurancejournal.com/news/international/2026/10/16/889185.htm), which also reports no publicly tradable Nebius CDS. Oracle S&P BBB minus (July 2026) and the 42 billion FY2027 FOCF deficit per S&P as cited in the same coverage. Jensen Huang remarks to Bloomberg in Stockholm, October 15 (https://qz.com/jensen-huang-coreweave-debt-default-risk-record-101626). CoreWeave bond prices October 15 per bondterminal.com as carried in the credit file. CoreWeave Q3 date estimates Nov 9 to 16, unconfirmed, per earnings calendars; the Q3 2025 date was announced October 27, 2025 for November 10.

Method: flat-hazard CDS valuation with 4% discounting, quarterly premium and accrual on default; bond-implied hazards solved from clean prices with semi-annual coupons and recovery of par. The one-year and three-year CoreWeave points are inferred from the bond curve, not observed. Credit-event conventions per the 2014 ISDA Credit Derivatives Definitions and the Standard North American Corporate transaction type; deliverability of convertibles per the same, to be confirmed on the dealer confirmation. Probabilities labelled as the desk's prior are judgments, with the confidence stated.