The neocloud credit cycle: what broke on October 14, 2026

On Wednesday, October 14, 2026, CoreWeave’s stock fell 15.46% to $65.78, within 9% of its 52-week low of $60.55. Nebius fell 10.08% to $186.01, IREN fell 12.2%, Oracle fell 8.83%, and the Philadelphia Semiconductor Index fell 4.30%, its sixth down day in a row. Less than eight hours after the U.S. close, in Seoul, SK hynix and Samsung Electronics each fell more than 10%, and the KOSPI triggered a 20-minute market-wide circuit breaker at 12:38 p.m., its tenth of 2026 and the first since July 29.

The headlines blamed three things: a hot inflation print, a Gartner forecast that most neoclouds will not survive as independent companies, and a chart of CoreWeave’s bond yields that Elon Musk answered with one word. All three mattered, but none of them explains why the move was this large, or why it started in a bond before it reached the stock.

My claim in this article is that the AI trade now has a credit layer under it that did not exist two years ago, and that this credit layer set the price on October 14. Demand for GPUs is as strong as it has ever been. Rental prices are rising, CoreWeave’s CEO said in August that its near-term capacity “remains effectively sold out,” and Nebius raised its on-demand GPU prices by about 20% on average on October 1. The change came in the price of the money that turns that demand into data centers.

I wrote a broader piece on October 11 about how the hyperscalers moved from paying for AI out of cash flow to financing it like infrastructure. This one goes one level down, to the companies that do the building with the least equity and the most debt.

Thirty-six hours, in order

Time (ET unless noted)	Event
Tue Oct 13, close	CoreWeave’s 9.625% notes due 2032 close at a 14.5% yield after touching 15% intraday, per a chart Chamath Palihapitiya posts on X. Musk replies “Yikes.” CoreWeave stock closes at $77.81.
Wed Oct 14, 8:30 a.m.	BLS reports September CPI: +0.6% m/m and 3.7% y/y headline; core +0.3% m/m and 2.6% y/y. Gasoline is up 9.1% in the month and 34.2% over the year.
Wed Oct 14, morning	Gartner publishes a forecast that “over $3.5 trillion” of AI infrastructure spending could be exposed to impairment by the end of 2030, and that over 80% of neocloud providers will cease independent operations by 2030.
Wed Oct 14, session	CoreWeave’s 2032 notes trade below 80 cents on the dollar, per Trace data cited by Bloomberg. Futures-implied odds of a Fed hike on October 28 reach about 50% intraday and close near 37%, up from 22% the day before, by one count (single source).
Wed Oct 14, close	CRWV −15.46%, IREN −12.2%, NBIS −10.08%, ORCL −8.83%, SOX −4.30%, NVDA −3.16%, Nasdaq −1.88%, S&P 500 −1.11%. On the Nasdaq, 521 stocks hit new lows and 27 hit new highs.
Wed Oct 14, evening	Reuters reports a memo in which CoreWeave CEO Michael Intrator writes that markets can “misunderstand and act irrationally” and that the backlog is “large enough to more than support its debt obligations.” CRWV trades at $61.87 after hours, down another 5.9%.
Thu Oct 15, 9:07 a.m. KST	KOSPI opens 5.59% lower and triggers a program-trading sidecar.
Thu Oct 15, 12:38 p.m. KST	KOSPI circuit breaker halts all trading for 20 minutes. KOSDAQ’s halted at 12:23 p.m.

In that sequence, the bond moved first, on Tuesday, before the CPI print and before Gartner. The size of each fall also followed leverage, and the companies that carry the most debt per dollar of equity fell hardest. Nvidia, with net cash and a 75% gross margin, fell 3%; Oracle, rated one notch above junk, fell 9%; and CoreWeave, with about $35 billion of debt against a $36 billion market capitalisation at the close, fell 15%. A pure demand scare would hit Nvidia first, because Nvidia sells the chips, but this selloff hit hardest where the equity sat furthest down the capital structure.

The rates shock underneath

The first cause has nothing to do with AI. The Federal Reserve, now chaired by Kevin Warsh, raised its target range by 25 basis points to 3.75%–4.00% on September 16, by a 12–0 vote; the effective federal funds rate moved from 3.63% to 3.88% on September 17, and SOFR fixed at 3.89% on October 13. The September CPI then showed headline inflation re-accelerating for a second month, driven by energy: the BLS energy index is up 20.4% over twelve months. Brent futures settled at about $106 on October 14. Middle East oil flows through the Strait of Hormuz remain constrained, and the EIA’s October outlook has Brent averaging $105 in the fourth quarter. On the Treasury’s par curve, the 2-year closed October 14 at 4.82%, the 10-year at 5.19% and the 30-year at 5.52%.

For most of the AI trade, rates matter through valuation: a stock whose cash flows arrive in 2030 is worth less when the discount rate rises. For a neocloud, rates matter twice, because they lower the present value of the equity, and they raise the cost of the debt that funds every new megawatt. A company that borrows at SOFR plus 550 basis points to buy GPUs pays about 9.4% today. Another hike would raise that floating-rate cost on the next reset, while the revenue on a contract signed last quarter stays the same. The credit problem was already there, and the hot CPI print removed the hope, priced in over the summer, that falling base rates would ease it.

Rates also act through a less obvious channel. The marginal buyers of high-yield AI debt are credit funds, insurers’ private-credit sleeves and multi-strategy hedge funds running capital-structure trades, and all of them finance themselves at short rates. When the front end of the curve reprices higher, their carry on a 9.6% bond shrinks, and the bond has to fall until the carry works again.

What a 9.625% bond at 80 cents says

CoreWeave sold $1.25 billion of 9.625% senior unsecured notes due July 15, 2032 at par on June 11, 2026, alongside €2 billion of 8.5% notes with the same maturity. Four months later the dollar notes traded below 80. The arithmetic is simple, and it gives the cleanest single number in this story.

With semiannual coupons and settlement on October 14, 2026, the bond has 5.75 years to run, and the table gives the yield at each price.

Clean price	Yield to maturity	Spread over today’s ~5.0% Treasury
100 (June issue)	9.62%	~4.6 pts
90	12.08%	~7.1 pts
85	13.44%	~8.4 pts
80	14.91%	~9.9 pts
78	15.53%	~10.5 pts

The 14.5% close on October 13 corresponds to a price of about 81.4, and “below 80” on October 14 means a yield above 14.9%. The modified duration at that price is about 4.0, so each further point of yield costs roughly 3.2 points of price.

A spread of about 10 percentage points can be turned into a rough default intensity with the standard reduced-form approximation, spread ≈ λ(1 − R), where λ is the annual hazard rate and R is recovery. With recovery of 30%–40% for senior unsecured debt, λ comes out at 14%–17% a year, and the implied cumulative probability of default over five years is 50%–57%. This arithmetic is behind the headlines that put the default odds above 50%, with two caveats. First, these are risk-neutral probabilities, so they include a premium for bearing the risk and for illiquidity, and they are usually well above realised default rates. Second, the unsecured notes rank behind the secured GPU facilities, which have first claim on the GPUs and the customer contracts in the subsidiaries that own them, and that pushes expected unsecured recovery down. The number is the price at which the marginal holder will take the risk, and the price at which CoreWeave would have to refinance in the unsecured market this week. It does not forecast a default.

The rest of CoreWeave’s capital structure was priced very differently. On September 18, less than four weeks earlier, CoreWeave sold $3.7 billion of 2.875% convertible notes due 2033, with a conversion price of about $97.85, 22.5% above the $79.88 reference price. A six-and-a-half-year claim on the same company paid 2.875% in cash because buyers were also paid in equity volatility, which convertible arbitrage funds can strip out and sell. With the stock at $65.78, the conversion option on those notes is now 33% out of the money, and the notes behave more like the straight debt they sit beside. Two markets priced the same company in different ways. The equity-volatility market paid CoreWeave for its uncertainty, and the credit market charged it for that uncertainty. On October 14 the credit market’s view won.

The capital stack of a neocloud

CoreWeave is the best example to study, because it is the largest, the most indebted, and the most thoroughly disclosed. At June 30, 2026, its 10-Q showed $35.1 billion of debt carried on the balance sheet (principal $35.6 billion) and $16.3 billion of operating lease liabilities; Stock Titan’s summary of the filing puts leases signed but not yet commenced at about $35.5 billion, starting between 2026 and 2029. Cash was $5.5 billion, plus $1.4 billion restricted. Revenue backlog was $103.7 billion.

The debt comes in layers, and each layer has its own price, as the table shows.

Layer	Example	Price	Who holds the risk
Investment-grade GPU loan	DDTL 4.0, $8.5 billion, March 2026, due 2032; rated A3 by Moody’s	SOFR + 225 bp	Lenders to a vehicle that is non-recourse to CoreWeave, backed by a contract with an unnamed investment-grade customer
Syndicated delayed-draw term loan	DDTL 5.0, $3.1 billion, May 2026, due 2031; Ba2 / BB+	SOFR + 450 bp	Banks and loan funds; full parent guarantee; backs contracts with two non-investment-grade customers, per The Energy Mag
Syndicated delayed-draw term loan	DDTL 5.5, $2.6 billion, August 2026, due 2031; Ba2 / BB+	SOFR + 550 bp, issued at 96–97 after marketing at S+425–450 and 99	Banks and loan funds; guaranteed by CoreWeave; backed by contracts averaging about three years
Senior unsecured notes	9.25% 2030, 9.0% and 9.75% 2031, 9.625% 2032, €8.5% 2032	9.6% at issue in June; ~15% on Oct 14	High-yield funds
Convertible notes	2.875% due 2033, $3.7 billion, September 2026; plus $4.0 billion of 2031 and 2032 converts with a $119.60 conversion price	Low cash coupons plus an equity option	Convertible arbitrage funds, hedged by shorting the stock
Equity	At-the-market program for up to 35 million shares; Nvidia and Jane Street placements	Dilution at market	Shareholders
Leases	$16.3 billion recognised; about $35.5 billion signed but not yet started	Embedded in rent	Data center landlords and their lenders

The spreads on CoreWeave’s secured loans show a trend. Its first delayed-draw facility, in 2023, priced about 750 basis points above DDTL 4.0, by the company’s own March 2026 presentation. Spreads then fell for three years as lenders learned the asset class, reached S+225 in March 2026, and have widened by 325 basis points in the five months since, as each new facility was backed by weaker customers on shorter contracts. DDTL 5.5 is backed by contracts that average about three years, financed with a loan of about five.

The same split shows up across the industry. Lambda’s $926 million term loan, backed by an unnamed investment-grade offtaker, priced at SOFR + 300 and was rated Baa2 in August. IREN financed its Microsoft contract with $3.65 billion of loans rated A by Fitch at a blended 6.0%. Nebius’s first secured facility, $775 million in July, priced at SOFR + 250 against an investment-grade customer’s contract.

The spread between the top and the bottom of that table defines the business model. A GPU loan at SOFR + 225 is priced on the customer, because the lenders look through the special-purpose vehicle to a take-or-pay contract with an investment-grade counterparty. A GPU loan priced at SOFR + 550 is priced on the GPUs and on CoreWeave’s guarantee, with a customer list that, in the August deal, Bloomberg reported as Anthropic, Jane Street and Hudson River Trading. An unsecured bond at 15% is priced on what is left after the secured lenders are paid.

Brannin McBee, the co-founder who runs financing at CoreWeave, told the Financial Times that investment-grade ratings on more chip-backed debt would open the insurance market, whose capital he called “unlimited,” adding: “that’s all the money in the world that sits in the insurance tranche.” CoreWeave’s strategy is to move as much of the stack as possible into the top row, where the customer’s rating sets the price. That works for capacity contracted to Microsoft, Meta or Google. It does not work for capacity contracted to unrated AI labs on short tenors, and the numbers below show that this slice is growing.

The maturity mismatch the bond market is pricing

CreditSights analyst Jordan Chalfin put the market’s concern in one question to the FT: “If there’s an overbuild, why would [customers] renew their contract with CoreWeave?” His concern is tenor, and the problem becomes visible when every clock in a neocloud’s business goes on one axis.

• Data center leases run 10 to 20 years. CoreWeave has about $35.5 billion of them signed and not yet started.
• Debt mostly matures between 2030 and 2033. The secured DDTLs amortise from contract cash flows; the unsecured notes and converts are bullets.
• GPU depreciation is straight-line over six years at CoreWeave. Nebius uses five, up from four before 2026.
• Nvidia’s product cycle is now annual: Vera Rubin was ramping into production as of August 2026, with Rubin Ultra targeted for 2027 and Feynman for 2028 on the roadmap shown at GTC in March.
• Customer contracts changed. The backlog was built on multi-year take-or-pay contracts, and about 41% of the $103.7 billion is expected to be recognised within 24 months. But the higher prices CoreWeave announced on September 17, about $40 million of annualised revenue per megawatt, came on what the company called “short-dated customer contracts of approximately three to six months.”

The equity market and the credit market read this trade-off in opposite ways. To an equity investor, short tenor at a high price is optionality, because the contracts reprice upward every quarter while prices keep rising. To a creditor, a short tenor gives almost none of the protection of a contract. A lender to a GPU vehicle wants the revenue locked for the life of the loan, because the collateral, a GPU, loses value on a known schedule and becomes obsolete on an unknown one. The September pricing release was therefore good news for shareholders and a warning to creditors.

Unit economics per megawatt

To see what is at stake, I price one megawatt of IT load. The inputs below are rough, and each has its source.

• Capex. A GB300 NVL72 rack costs just under $5.0 million on purchase orders and about $5.7 million deployed, and draws roughly 130–140 kW. Seven racks, about 500 GPUs, fill a megawatt, so about $40 million per MW of IT, before storage, core networking, and the building (which a neocloud usually leases rather than owns).
• Revenue. CoreWeave’s short-dated Q3 contracts earn about $40 million per MW per year. IREN reports “>$20m revenue per MW (IT)” on its recent three-year contracts but does not say whether that is per year; its Microsoft contract works out to about $9.7 million per MW per year. Nebius’s Q2 landmark deals carry $20–25 million of annual contract value per MW.
• Margin. CoreWeave’s adjusted EBITDA margin was 59% in Q2, after power and lease costs.

The revenue figures pass a sanity check. A GB300 rack holds 72 GPUs, so seven racks hold about 500. At $40 million per MW, each GPU earns about $9 an hour at full utilisation, close to the $9.50 on-demand list price Nebius now charges for a B300. At $20 million, it earns about $4.50. For older hardware, at roughly 1.3 kW per H100 including its share of the server and network, a megawatt holds about 770 H100s; at the Silicon Data index level of about $2.85 an hour, it grosses about $19 million a year, and at the SemiAnalysis one-year contract low of $1.70 in October 2025, about $11.5 million.

Revenue per MW per year	EBITDA at 59% ($M)	Capex payback on EBITDA	Pre-tax profit: 6-yr life, 9.6% debt ($M)	6-yr life, 15% debt	4-yr life, 15% debt
$40M (CoreWeave short-dated, Q3 2026)	23.6	1.7 yrs	+14.0	+12.4	+9.1
$30M	17.7	2.3 yrs	+8.1	+6.5	+3.2
$20M (IREN, Nebius multi-year)	11.8	3.4 yrs	+2.2	+0.6	-2.7
$11.5M (an H100 MW at the Oct 2025 rental low)	6.8	5.9 yrs	-2.8	-4.4	-7.7

Illustration, per MW of IT load. Inputs: $40 million of GPU capex per MW, 75% debt-financed ($30 million), straight-line depreciation, 59% EBITDA margin (CoreWeave Q2 2026 adjusted), interest on the full debt balance with no amortisation. The building is leased and its rent is inside the EBITDA margin. Taxes are ignored.

The table supports three conclusions. First, at today’s prices the economics are very good, and a $40 million-per-MW contract pays back the GPUs in under two years of EBITDA. Nebius quotes a payback of 1 year 10 months on Q2 deals at $20–25 million per MW, which implies lower capex or higher margins than I assume here, and prepayments help. Second, the equity is extremely sensitive to the renewal price, because the fixed charges (depreciation and interest) do not move with it. When revenue halves, pre-tax profit per MW falls by more than 80%, to about zero at a 15% cost of debt, and below zero with a four-year life. Third, the useful-life assumption matters a great deal, because a move from six years to four adds about $3.3 million per MW per year of depreciation, which is more than the entire profit in the $20 million case.

The table also explains the gap that critics point to in CoreWeave’s numbers. CoreWeave’s realised revenue in Q2, $2.575 billion, annualises to about $10.3 billion, or about $6.9 million per MW on 1.5 GW of active power. That is far below $40 million, but the two numbers measure different things. Active power counts capacity that is energised but not yet delivered to a customer, the older fleet earns legacy prices, and a facility megawatt is not an IT megawatt. The $40 million figure describes the marginal contract and the $6.9 million figure describes the average, and the debt is serviced by the average.

Where CoreWeave’s bond sits in the credit market

The high-yield indices put the 15% yield in context. On October 13, the ICE BofA US High Yield index had an option-adjusted spread of 316 basis points, up from 281 at the end of 2025. Single-B bonds traded at 317, and CCC bonds at 1,267. CoreWeave’s 2032 notes, at roughly 990 basis points over Treasuries, trade closer to the CCC index than to the single-B index, although the company’s unsecured notes are rated in single-B territory by S&P, according to an industry tracker. A bond that trades closer to the next rating category down than to its own shows that the market expects a downgrade.

The rest of the high-yield market is calm. Investment-grade spreads were 83 basis points. Meridian Arc, a data center vehicle with a lease backstopped by Google, priced $5.7 billion of secured bonds at 6.25% in April. CleanSpark sold about $2.3 billion of five-year notes backed by a Meta-guaranteed lease at an 8.25% yield on a $10 billion order book as recently as September 19. The market is still open to AI infrastructure, but it prices each deal by the credit of whoever pays the rent, and it has stopped giving the neoclouds credit for their own balance sheets.

The credit risk sits with the labs

Every neocloud’s revenue rests on a short list of customers. In the first half of 2026, CoreWeave’s largest customer was 40% of its revenue and the second was 23%. Nscale’s IPO filing showed 52% of first-half revenue from one customer. Lambda’s backlog rose from $15 billion in June to $50 billion in September, mostly on a $35 billion commitment from Anthropic. Concentration on customers rated A or better, as Microsoft, Meta and Google are, makes the financing cheap, and concentration on AI labs makes it expensive.

For this reason, one story about one lab moved the whole group a week before October 14. On October 8, Axios reported that OpenAI’s annualised revenue was $50 billion, about $20 billion below the figure widely reported in September, and explained that the gap came from accounting conventions for sales through cloud partners. CoreWeave fell 7.8% that day and IREN fell 7.7%. OpenAI raised $122 billion at an $852 billion valuation in April, and its compute plans have moved from Sam Altman’s earlier projection of roughly $1.4 trillion to about $600 billion of compute spending through 2030, as CNBC reported in February. On either number, the payer for much of the neocloud backlog is a company whose free cash flow is negative and whose funding depends on the next round.

Anthropic’s position is similar but newer, and in the last quarter it became the anchor tenant of the second tier: Lambda ($35 billion), Nscale ($45 billion for a West Virginia campus, per Nscale’s filing), TeraWulf (a 20-year lease worth about $19 billion, expected to be supported by investment-grade credit), and, by press reports, Fluidstack. Anthropic’s revenue has grown faster than any software business I know of, and its commitments have grown faster still.

Rothschild & Co Redburn’s Alexander Haissl, who put Sell ratings on CoreWeave and Nebius on September 21, made the bearish argument most clearly. As TipRanks summarised it, the highest-paying customers are increasingly venture-funded labs buying compute for training, and training demand lasts only as long as outside capital does, while inference must be paid for by end users. On that view, the price of a GPU-hour is now partly a function of the price of AI venture capital, and the neocloud bond is a levered claim on the next funding round of two or three labs.

The counter-argument is also strong. The labs that pay these prices are growing revenue at a rate no software category has matched, and they are buying compute because their models are capacity-constrained in production as well as in training. CoreWeave says it signed more than $25 billion of new commitments early in the third quarter, and no customer has yet missed a payment. Creditors still cannot underwrite the labs the way they underwrite Microsoft, so the cheap tier of the capital stack is closed to the labs’ capacity.

Nvidia as the lender of last resort

The main supplier also sits inside the capital structure, because Nvidia owns equity in most of the neoclouds that buy its chips. It bought $2 billion of CoreWeave stock at $87.20 in January 2026, and it disclosed a 9.3% stake in Nebius in July, after a $2 billion investment in March. It guarantees $860 million of Nscale’s lease obligations in Texas, is putting $1 billion into Nscale’s pre-IPO convertible notes, due by mid-November, and has agreed to buy residual unsold CoreWeave capacity worth $6.3 billion through April 2032. In July it launched a program under which it rents back unused partner capacity at a set price in return for a share of the partner’s cloud revenue.

This support is rational for Nvidia. It earns a 75% gross margin, its data center revenue was $89.0 billion in the quarter to July, and a few billion dollars of support keeps its customers’ customers solvent and its order book diversified away from the hyperscalers, who are building their own chips. For a creditor, the support is a soft floor under the collateral, because a backstop on capacity does not guarantee the debt. For an equity investor in Nvidia, it means that Nvidia has financed part of the demand it reports. Vendors have done this in every infrastructure cycle, and it is the line item to watch in Nvidia’s filings.

The cohort, side by side

The companies in the neocloud group have very different capital structures, and the differences explain most of the dispersion in their stocks.

Company	Oct 14 close (move)	Key customers	Representative financing	Cost
CoreWeave (CRWV)	$65.78 (−15.5%)	Two customers 63% of H1 revenue	9.625% notes 2032; DDTLs S+225 to S+550	~15% unsecured, Oct 14
Nebius (NBIS)	$186.01 (−10.1%)	Microsoft ($17.4–19.4B), Meta (up to ~$27B)	$5.75B converts at 0.50% and 4.50%, Aug 2026; $775M secured at S+250	Mostly equity-linked
IREN	$30.56 (−12.2%)	Microsoft ($9.7B), Nvidia ($3.4B)	$3.65B IG GPU loans, blended 6.0%; $2.8B non-IG at ~9%	6%–9% secured
Cipher Digital (CIFR)	$9.94 (−12.3%)	Fluidstack (Google backstop), AWS	7.125% secured notes 2030; $850M 0% converts due 2032, priced Oct 14	Lease-backed
TeraWulf (WULF)	$11.54 (−12.6%)	Fluidstack (Google backstop), Anthropic	$3.2B 7.75% secured notes 2030	Lease-backed
Applied Digital (APLD)	$20.33 (−9.4%)	CoreWeave, IG hyperscalers (tracker)	6.75% and 7.0% secured project notes, 2031	Lease-backed
Galaxy (GLXY)	$18.01 (−8.7%)	CoreWeave	$3.5B 9.875% secured notes 2031	Lease-backed, CoreWeave credit
Hut 8 (HUT)	$74.87 (−10.5%)	Fluidstack/Anthropic (Google backstop), an IG tenant	$3.25B and $4.25B IG project notes	Investment grade

Sources: Nebius and its August convertible; IREN’s IG financing; Cipher’s convertible; TeraWulf’s notes; Applied Digital; Galaxy; Hut 8; closing prices from StockAnalysis.

The table holds three business models with different risks.

The GPU renters (CoreWeave, Nebius, Lambda, Nscale) own the chips and sell compute by the hour or by the contract. They carry the full risk of GPU obsolescence and repricing. Nebius has financed itself mostly with convertibles. Its August deal raised $5.75 billion with the purchasers’ options at coupons of 0.50% and 4.50% with conversion prices near $313 and $325, against a stock that now trades at $186. Its customer prepayments are also large. It expects more than $9 billion of them in 2026, and about 70% of its landmark Q2 deals included prepayments covering 50–60% of the related capex, according to its Q2 shareholder letter. A prepayment is customer-provided financing at a zero coupon, and it is the cheapest money in the industry. Nebius’s converts are now deep out of the money, so its next raise will price more like debt than equity.

The landlords (Applied Digital, Cipher, TeraWulf, Hut 8, Galaxy, Core Scientific) lease powered shells to a tenant on 10-to-20-year terms and do not own the GPUs. Their debt is project finance, secured on one campus, serviced from a lockbox of rent, and priced on the tenant. When the tenant is Fluidstack with a Google backstop, or an investment-grade hyperscaler, the bonds price at 6%–8% and some are rated investment grade. When the tenant is CoreWeave, the bonds price at 9%–10%, and Galaxy’s Helios II notes carry 9.875% for a phase leased to CoreWeave. The landlords’ credit is the neoclouds’ credit at one remove, and it is the transmission channel that matters most if CoreWeave’s cost of capital stays high.

The hybrids (IREN) do both. IREN shows the cleanest example of tiered pricing in the group. Its Microsoft-backed GPU financing is rated A by Fitch at a blended 6.0%, while its non-investment-grade facility for its Mackenzie, British Columbia site, led by Blue Owl, reportedly carries about 9%. The same company borrows at two prices, 300 basis points apart, depending on who the customer is. IREN also had a site-level problem in the week before October 14: on October 8, Yahoo Finance reported that SemiAnalysis had compiled user reports of outages at its Prince George and Mackenzie sites in British Columbia. From its October 6 close to October 14, IREN fell about 26%.

The private companies are under the same pressure. Firmus, an Australian neocloud with Meta and OpenAI contracts, cancelled a planned IPO of about $5 billion on the Australian Securities Exchange, announced October 9, after cutting its price from A$11 to A$9 a share. Nscale filed for a New York listing in September at a valuation target of up to $35 billion and is still waiting. Lambda moved its IPO to 2027 and is raising up to $4 billion privately. These are the cases CoreWeave’s CEO meant when he wrote that difficulties and delays at some rivals’ IPOs and debt financings “are bleeding into the narrative around CoreWeave,”.

Oracle, the investment-grade neocloud

On October 14 Oracle traded like a neocloud. It closed at $130.55, down 8.83%, on volume about 61% above its three-month average. S&P cut Oracle to BBB−, one notch above high yield, on July 9, citing a projected free-cash-flow deficit of about $42 billion in fiscal 2027 and about $260 billion of new lease commitments. Its five-year CDS reached a record 227 basis points in late September. On September 24, CNBC reported that Oracle had sent a force-majeure notice to the developer of its Project Jupiter campus in New Mexico, a Blue Owl unit, and the FT reported that about $18 billion of debt tied to the project trades at stressed levels.

Oracle matters to the neocloud story for two reasons. First, it is the largest buyer of AI capacity whose own credit is close to the line, and about half of its $638 billion of remaining performance obligations is OpenAI, per S&P. Second, a downgrade to high yield would force many investment-grade funds to sell its bonds, which would reprice the whole category of AI-linked credit at once. Oracle connects the neocloud credit problem to the index-level credit market.

Memory, Korea, and the same trade in another market

The Korean selloff on October 15 looks at first like a separate event, because memory is the one part of the AI supply chain that is printing record profits right now. Micron reported fiscal fourth-quarter revenue of $54.2 billion and a GAAP gross margin of 86.8% on September 30, and guided to $61.5 billion for the current quarter. Samsung’s preliminary third-quarter operating profit was 107.4 trillion won, the first time a Korean company has topped 100 trillion won in a quarter. TrendForce expects conventional DRAM contract prices to rise another 10%–15% in the fourth quarter, and estimates that memory will account for 47% of cloud providers’ capex in 2026.

SK hynix and Samsung still each fell about 10% intraday, and three mechanisms connect them to the neocloud bond.

First, memory is now one of the largest costs in an AI server, and the neoclouds and the labs are among the marginal buyers whose demand sets the price. TrendForce puts HBM and server memory at about half of DRAM bit supply in 2026. A memory maker earning an 87% gross margin is earning it from customers whose cost of capital just rose by several hundred basis points. The market prices memory stocks on the expected peak of the cycle, and a credit event among the buyers is a signal about where the peak is.

Second, the Korean index is concentrated. Samsung and SK hynix together were 55.67% of the KOSPI’s market capitalisation on June 22, up from 35.22% at the start of the year, so a single long position in AI memory has become most of a national index. The KOSPI closed 2025 at 4,214.17 and set a record close of 9,114.55 on June 22. It closed at 6,941.39 on October 2 and at 6,008.81 on October 14. On October 15 it traded as low as 5,506.23 intraday, down 8.36%, and its circuit breaker at 12:38 p.m. was the tenth of 2026. The Korean close was not available when I finished this article; the latest level I found, from a single source, was 5,589.19, down 6.98%, after the halt.

Third, positioning and leverage tie Korea to the neoclouds, because the two markets share holders. Foreign investors had net sold about 199 trillion won of Korean stocks this year through October 12, most of it in SK hynix and Samsung. Margin loans stood at 34.25 trillion won on October 13, forced liquidations were the highest since July 31, and the VKOSPI was at 53, per Aju Press. A similar sequence happened in July. Situational Awareness LP, the AI-focused fund run by Leopold Aschenbrenner, held CoreWeave, Broadcom, Intel and other AI names according to its filings, and was reported to hold SK hynix and Nebius. After a July drawdown it sold the bulk of its stock portfolio to Citadel on July 30, as Reuters reported. In the same week the KOSPI fell 10.84% on July 28 and hit its circuit breaker on July 29, the day SK hynix reported a profit miss. The same levered, crowded factor sits under the neoclouds, the U.S. memory names and the Korean index. When holders of that factor de-lever, they sell in all three markets at once, and the correlation between stocks with very different fundamentals goes to one.

Korea’s own triggers added to the move. Samsung’s buyback ended on October 6 and SK hynix’s 40-trillion-won buyback was due to expire in mid-October, per the Korea Herald, and reports about the Chinese memory maker CXMT added to fears about the length of the cycle. SoftBank, whose OpenAI exposure is about $65 billion, fell 9.3% in Tokyo the same day.

Gartner, Burry, and the depreciation question

The Gartner forecast gave the day its headline, and it needs a careful reading. The forecast’s core claim is that AI infrastructure owners would need $8.5 trillion of cumulative revenue through 2030 to cover capex, financing and a 12% return on invested capital, against at most $3.6 trillion of revenue attributable to AI data centers. It estimates about $1 trillion of AI infrastructure capex in 2026 and $6 trillion more from 2027 to 2030, and it does not expect the 12% hurdle to be met before the end of 2032. On neoclouds specifically, it says “cost pressure and a lack of a differentiating business model will force over 80% of neocloud providers to cease independent operations.” “Cease independent operations” includes being acquired, which is how most capital-intensive cycles consolidate, so the forecast does not predict defaults, and Gartner names no company.

The depreciation argument is older. Michael Burry argued in November 2025 that the economic life of AI accelerators is two to three years, against the five to six years most operators use, and that hyperscalers will understate depreciation by $176 billion over 2026–2028. Operators use different lives. CoreWeave depreciates over six years, Nebius moved from four to five years in 2026, Amazon shortened part of its server fleet from six to five years in 2025 “due to the increased pace of technology development,” and Microsoft lengthened the lives of its data center buildings from 15 to 25 years from July 2026.

The market evidence on GPU life runs the other way this year. H100s, a 2022 design, rent for more in late 2026 than they did a year ago: the SemiAnalysis one-year contract index rose from $1.70 in October 2025 to $2.35 in March 2026, and Silicon Data’s index showed about $2.85 in mid-October. SemiAnalysis reported that some H100 contracts have been extended to 2028. That is strong evidence that a four-year-old GPU still earns money in a supply shortage. It is weak evidence about what it earns once the shortage ends, because scarcity sets rental prices in a shortage and the cost of the next-best chip has little effect on them. For a lender, the relevant test is the price of an H100-hour when Rubin capacity is plentiful, and nobody has observed that price yet. The new CME futures on the Silicon Data H100 and B200 rental indices, announced for an October 5 launch, may eventually give a forward curve for exactly that.

What I think the market is saying

Taken together, the message of October 14 is more specific than “the AI bubble is bursting.”

1. Demand for compute did not change. Rental prices are rising, capacity is sold out, Micron and Samsung are printing records, and TSMC reported a 69% rise in third-quarter profit on October 15. Every operating datapoint released in the week was strong.
2. The price of money changed. The Fed hiked in September, inflation re-accelerated on an oil shock, and the market started to price another hike. A business that borrows at floating rates to buy assets with a short economic life is the most rate-sensitive business in the AI supply chain.
3. The credit market is separating customers from operators. Debt backed by Microsoft, Meta, Google or an investment-grade lease still prices at 6%–8% and sometimes earns an A rating. Debt backed by AI labs, or by a neocloud’s own balance sheet, now prices at 9%–15%. The gap between those two tiers is the most important price in the AI trade, and it widened sharply this month.
4. Equity holders are the most levered exposure to that gap. CoreWeave’s equity is worth about as much as its debt. A rise in its marginal cost of capital from 9.6% to 15% does not affect the contracts it has already financed, but it affects every megawatt it has not yet financed, and the guidance assumes many of those.
5. The contagion runs through shared holders. The neoclouds, U.S. memory stocks, the bitcoin-miner landlords and the KOSPI share investors who are levered and long the same factor, so stocks with record profits fell as hard as stocks with record losses.

The bullish reading is that this is a financing problem with a known solution. CoreWeave’s DDTL 4.0 showed that GPU debt backed by an investment-grade contract can earn an A3 rating and price at SOFR + 225; IREN’s Microsoft financing showed the same at 6%. The more capacity is contracted to highly rated buyers, the more of the stack moves into the cheap tier, and McBee’s “insurance tranche” becomes real. The bearish reading is that the capacity being added now is contracted to the buyers who cannot get that rating, on tenors too short to finance, and that the cheap tier is full.

The calendar that will decide it

Date	Event	What to look for
Oct 27–28	FOMC meeting	Whether the Fed hikes again; the minutes of September said most participants saw another hike by year-end
Late October	Alphabet, Microsoft, Meta, Amazon Q3 results	2027 capex guidance; any change in third-party capacity purchases
Oct 29	Samsung final Q3 results	HBM4 volumes and pricing for Vera Rubin
Nov 11	CoreWeave Q3 results	Interest expense against the $860–940 million guide; how much of the $25 billion of new commitments is on long tenors; any DDTL refinancing
Mid-November	Nvidia’s $1 billion Nscale note funding; Nscale IPO timing	Whether the IPO window reopens for the second tier
November	Nebius and IREN quarterly results	Prepayments received; pricing on new contracts; the next financing
Ongoing	Trace prices on CoreWeave’s 2032 notes; Oracle five-year CDS; the CCC index	Whether 15% is the peak or a level

Caveats

Several figures come from a single secondary source: the post-CPI hike odds, the latest Korean intraday level, IREN’s ~9% facility rate, the claim that Situational Awareness held SK hynix and Nebius, and the press reports of Anthropic’s compute deals. Bond prices for most of the cohort’s notes are not public without Trace or Bloomberg access, so I report only the CoreWeave levels that Bloomberg and the FT reported. The per-megawatt model is an illustration with stated inputs and does not value any company. The risk-neutral default probabilities read a price and do not forecast a default. The Korean market had not closed when the facts on this page were last checked.