CoreWeave's Real Risk Is Not Its Contracts. It Is the Next Loan

CoreWeave, Inc. (ticker CRWV) calls itself an AI cloud. In practice it rents out computing power. It buys NVIDIA GPU servers, installs them in data centres it mostly leases, and rents the capacity to AI labsCompanies whose main business is building large AI models, such as OpenAI and Anthropic. They need very large amounts of GPU time to train and run those models., hyperscalers (the largest cloud operators) and large companies. In the twelve months to 30 June 2026 it took $7.6B of revenue, up 115.3% on the year before. It reported an operating loss of -$230.7M and a net loss of -$1.9B. Over the same twelve months its spending on equipment, recorded as an outflow of -$20.6B, was nearly three times its revenue. Most of that money was borrowed.

This report is about how that borrowing works and whether it holds up. CoreWeave puts most new GPUs in special-purpose subsidiariesCompanies CoreWeave sets up to hold one set of GPUs, the contracts they serve and the loan that paid for them. Keeping each build in its own company lets lenders see exactly which assets and which customer payments stand behind their loan.. These subsidiaries borrow under delayed-draw term loans (loans drawn in stages as equipment is delivered), secured on the GPUs and on one or a few customer contracts. The loans amortise, meaning the principal is repaid in instalments, out of the customer’s payments. The lenders set debt-service-coverage tests of 1.15 to 1.35 times, which means the cash from the contract must be at least that multiple of the loan payments due. The parent company then borrows separately, through unsecured bonds (bonds not secured on any specific asset) now yielding close to 15% (the yearly return a buyer earns at today’s market price, and roughly what new borrowing of this kind would now cost). That money pays for the part of each build-out the lenders will not fund. It also covers leases, staff and the next contract’s deposit. Our reading is that the contract-backed loans in the subsidiaries are sound on their own terms, and that the risk sits in the parent’s unsecured borrowing. The full argument is in the Debate 1 chapter.

Independent forecasts. On 14 October 2026 Gartner forecast that spending on AI-optimised infrastructure-as-a-serviceRenting computing hardware over the internet, paid by the hour or by contract, instead of buying it. AI-optimised means the hardware is GPU clusters built for AI work., the market CoreWeave sells into, would grow from $66 billion in 2027 to $174 billion in 2030, and that this growth would be “too low for most AI compute providers to recover their AI investments before the current generation of AI semiconductors is replaced”, according to its press release¹³ [E44]. It predicted that over 80% of neocloud providers would cease independent operations by 2030. Neoclouds are the newer companies whose business is renting out GPU capacity, such as CoreWeave, Nebius, Lambda and Crusoe; for the hyperscalers, the same capacity is one part of a much larger cloud business. That forecast covers the group as a whole. CoreWeave is the largest of these providers and has contracted backlog. The forecast matters because it is the sharpest public statement of the bear case for the industry, and it came out on the day of the selloff.

Power. GPU capacity is limited more by electricity and buildings than by chips. Lawrence Berkeley National Laboratory’s Queued Up 2026 report¹⁴ found that for power projects completed in 2025, the median time from an interconnection request (asking to be connected to the grid) to commercial operation was over five years, and that only 13% of capacity that requested interconnection between 2000 and 2020 had been built by the end of 2025 [E46]. The report counts generators rather than data centres drawing power, but it describes the same grid queue. CoreWeave avoids most of that wait by leasing sites from developers who have already secured power: in its own illustration, a partner takes 12–18 months to deliver the building shellThe finished building, structure and roof, before the equipment that goes inside it., grid connection and cooling, and CoreWeave then needs one to two months to install and test the GPUs [E17]. That makes the developers’ delivery the step that sets the timetable, which the Debate 4 chapter tests.

The price of GPU time. Rental prices fell sharply after 2023 and have since recovered. Silicon Data’s H100 rental index, a daily standardised rate drawn from neocloud providers and marketplaces, stood at $2.53 per GPU-hourOne GPU rented for one hour, the unit in which GPU time is priced. on 17 October 2026, according to its index page¹⁵ [E45]. The H100 was introduced in 2022. CoreWeave told investors that it raised prices about 25% across its products in July 2026 and that pricing for prior-generation chips was “at or above where it was years ago” [E41]. The two sources agree that GPU time is scarce in 2026. Neither shows what a 2026 chip will rent for once the next two generations of chips are widely available.

Credit. The AI build-out is now financed in the bond market. Alphabet, Amazon, Meta and Oracle together issued more than $228 billion of bonds by 13 October 2026, more than twice their total for 2025, according to LSEG data cited by CNBC¹⁶ [E42]. A credit rating is an agency’s letter grade for how likely a borrower is to default, and the lower the grade, the more lenders charge. The upper part of the scale is called investment grade. CoreWeave’s corporate rating is B+A grade on S&P’s scale four steps below investment grade. Debt rated this low is often called high-yield or junk., below investment grade. CoreWeave competes for the same lenders with a weaker balance sheet. Its answer has been to borrow against specific customer contracts, so that the customer’s credit can support the loan’s rating. DDTL 4.0 was rated A3A grade on Moody’s scale, three steps above its lowest investment-grade rating. by Moody’s, according to the March 2026 release¹⁷ [E27], while DDTL 5.0 and 5.5, backed by a wider mix of customers, were rated Ba2A grade on Moody’s scale two steps below investment grade. [E29] [E23].

On 14 October 2026 CoreWeave’s shares fell 15.5% and the yield on its 9.75% notes due 2031 rose to about 14.9%, according to CNBC¹⁶ [E42]. The yield is the yearly return a buyer earns by paying today’s market price and holding the bond until it is repaid. A yield that high on a bond with a 9.75% coupon (the fixed interest it pays each year on its face value) means the bond was trading well below that face value. Bond investors were pricing in a real chance that they would not be paid in full. The chief executive, Michael Intrator, told staff in a memo quoted by CNBC that “we are confident in our backlog which more than supports all of our debt obligations.”

The first question in this debate is whether the cash CoreWeave collects from customer contracts covers the interest and principal on the debt raised to serve them over the contracts’ lives. The second is whether the company can keep financing new contracts on terms that still leave it a profit as credit tightens. Our view would be wrong in either of two cases. One is if, more than once, a contract financed with its own loans failed and the parent company had to make the payments on those loans from other cash. The other is if the parent company could no longer raise unsecured money (borrowing not backed by specific assets) at a cost its contracts can bear.

The first two links are well documented. The order forms with OpenAI and Meta filed in September 2025 commit the customers to pay, but “subject to any termination described below and satisfaction of delivery and availability of service requirements” [E32] [E34]. The master services agreements (MSAs) behind them set the general terms and were filed with some text redacted. They let either party terminate only for cause, meaning an uncured material breach (a serious breach of the contract that has not been put right) or the other side’s insolvency (OpenAI MSA, Section 10)¹⁸ [E33]. The Meta agreement¹⁹ adds that if Meta terminates for cause, CoreWeave refunds the pro rata (proportional) share of any prepayment [E35]. Neither filed agreement contains a right to cancel for convenience, that is, to cancel without cause. The order forms, which carry delivery dates and any credits for late delivery, are not public.

The credit agreements set how much can be lent against the contracts through covenants, which are ratio tests the borrower must keep passing. In the $2.6 billion DDTL 5.5 facility of August 2026, the borrowing subsidiary must keep a debt-service coverage ratio (the cash available for loan payments divided by the payments due) of at least 1.35 times and a “Projected Contract Value Ratio” of at least 2.40 to 1, according to the 8-K²⁰ [E22] and the credit agreement itself²¹ [E24]. The second ratio compares “Total Projected Contract Value” with the outstanding loan. The public copy redacts the definition of that value, so a reader cannot check what it counts. The $8.5 billion DDTL 4.0 facility of March 2026 requires coverage of at least 1.15 times [E25]. That lower bar fits its investment-grade customer (one rated a low risk of default) and its A3 rating. In DDTL 5.5, the debt service being covered leaves out the bullet repayment due at maturity, which is the single lump sum of principal owed when the loan ends. For periods not under contract, the projections may rest on the facility’s base-case assumptions. So even if the covenants hold, they do not show that the original contracts repay the whole loan.

Two things follow. First, the lenders have arranged to be paid before CoreWeave sees anything. In DDTL 5.5 a breach that would let the customer terminate, or a contract-value ratio below 2.40×, traps cash in the subsidiary from the next monthly payment date. A customer that stops paying is not one of the named triggers; it would show up through the coverage test or a termination. The parent’s income from a financed contract is the residual, and the residual stops first. Second, except in DDTL 4.0 and $882 million of vendor financing, the parent has guaranteed the loans [E07]. At 30 June 2026, $31.4 billion of the $35.1 billion carrying value of debt was recourse to CoreWeave, Inc. A contract that failed inside DDTL 2.0, 3.0, 5.0 or 5.5 would leave the lenders able to collect the shortfall from the parent. The stress can also reach the bonds. The indenture for the 9.625% notes due 2032²³ makes a payment default at maturity or an acceleration of debt the parent guarantees an event of default on the notes once it exceeds the greater of $200 million and 5% of twelve-month EBITDA, though the notes’ acceleration is annulled if the default is cured within 30 days (Sections 6.1(a)(4) and 6.2) [E60]. The 2033 convertible’s indenture²⁴ has the same threshold for the parent’s and its significant subsidiaries’ borrowed money, and allows 30 days after notice to cure before the trusteeA bank appointed to act for all the holders of a bond issue, for example in declaring a default and demanding repayment. or holders of 30% can accelerate (Sections 6.01(h) and 6.02) [E61]. Both exclude the financing subsidiaries from the subsidiaries they count, so a large enough accelerated recourse DDTL reaches the 9.625% notes through the parent’s guarantee.

The DDTL 5.5 facility shows where the structure is going. Its customer contracts average about three years while the loan runs about five, so the lenders are underwritingTaking on a risk and putting a price on it. Here the lenders are lending on the expectation that customers will renew contracts that end before the loan does. renewal, as CoreWeave’s release²⁵ says: lenders are “signalling confidence in the long-term value of NVIDIA GPUs … and a willingness to underwrite renewal risk” [E23]. The non-replacement clause decides who carries that risk. If a three-year contract is not renewed, the subsidiary must prepay, and if it cannot, the parent pays. Lenders are lending past the contract because the parent stands behind them.

The events ran over five weeks. On 17 September CoreWeave announced a $3.0 billion convertible bond and, the same day, a programme to sell up to 35 million new shares [E20]; the stock closed at $79.88 [E19]. The bond issue was enlarged to $4.2 billion with a 2.875% coupon, and the sale closed on 22 September [E18]. On 13 October the investor Chamath Palihapitiya posted that the 2032 notes had touched a 15% yield, and Elon Musk replied “yikes” [E42]. On 14 October Gartner published its forecast that most neoclouds would not survive as independent companies [E44]. The same day, US consumer prices were reported up 0.6% for September, according to Vantage Markets’ account²⁷ citing the Bureau of Labor Statistics [E43]. CoreWeave’s shares closed at $65.78, down 15.46%, on 107 million shares traded against a three-month average of 29.5 million. Oracle fell almost 9% and Nebius 10% the same day [E42]. No CoreWeave filing or customer announcement came out that day.

The shares recovered on Thursday 15 October, helped by NVIDIA’s chief executive Jensen Huang telling Bloomberg he was optimistic about CoreWeave, then fell almost 12% on Friday 16 October, ending the week down about a quarter and more than 65% below their peak shortly after the 2025 listing, according to CNN²⁸ [E56]. CNBC linked Friday’s fall to slower revenue growth reported by Anthropic, which is itself a CoreWeave customer. It also reported that brokers at Jefferies, JPMorgan and Cantor Fitzgerald called the fall overdone. Other analysts, and lawyers who work on these contracts, questioned whether take-or-pay terms are as iron-clad as assumed and whether backlog should be treated as revenue, according to CNBC²⁹ [E57]. That last question matters most here, and the filed agreements answer it only in part. They allow termination for cause, and the delivery terms that would decide a dispute are in order forms that are not public.

The selloff shows that raising money has become more expensive for the parent company, and the parent’s money is what the financing model needs most. The filings support three parts of the market’s concern: debt has risen faster than revenue, interest is rising faster than cash earnings, and growth depends on raising new unsecured capital every few months, now including new shares. The press’s “debt at 6.8 times sales” framing implies that existing loans are in danger, and the filings do not support that. Those loans are being repaid in instalments out of contracts with Meta, Microsoft, OpenAI and others. The company reported that it was in compliance with all its DDTL covenants at 30 June 2026 [E07], and the first of its unsecured debt falls due in June 2030. The 15% yield reflects the risk that the parent cannot keep funding growth and ends up holding guarantees on contracts that go bad. Nothing in the filings points to an imminent default. The bonds’ indentures (their governing contracts) do contain cross-acceleration clauses, which can make the bonds fall due at once if other debt is called in early. Because of them, a large contract going bad could put the bondholders at risk before 2030.

The strongest argument against our view goes like this. The protections written into the contracts are real, but they protect the lenders rather than CoreWeave. Every protective clause moves cash away from the parent company at the first sign of trouble. Meanwhile the parent company has guaranteed 90% of the debt, signed $35.5 billion of leases that have not yet started, and committed to spend $35–39 billion this year. Its business model requires it to keep raising unsecured debt, meaning borrowing that is not backed by specific assets. That debt now costs it about 15%. Its customer contracts were priced when it cost 9%. On this reading CoreWeave holds the riskiest position in the AI build-out: it is paid last, after the lenders, and it is heavily in debt itself. The contracts give most of the upside to customers and most of the protection to secured lenders. The parent company is left with whatever cash remains, plus the guarantees. If credit stays tight, CoreWeave has to slow its growth. A slower-growing CoreWeave would still owe $20.8 billion of bonds and convertibles (debt that can be exchanged for shares) at the parent level. It would then need its stabilised contracts, the ones running at their full rate, to produce far more leftover cash than they have shown so far.

Much of that is right, and it is why our confidence is medium. Two pieces of evidence keep it from being the whole story. First, the parent company has been able to choose where it borrows. In the weeks before unsecured yields peaked it sold convertible debt at 2.875%, and its single-customer, investment-grade facility (a loan backed by one customer with a strong credit rating) priced at SOFR (the benchmark US overnight lending rate) plus 2.25%, against SOFR plus 4.0–4.25% for the 2025 facilities [E16] [E25]. The multi-customer loans, DDTL 5.0 and 5.5 (delayed-draw term loans, which the borrower draws down in stages), cost more, at SOFR plus 4.50% and 5.50% [E28] [E22]. So the low rate depends on the customer being investment grade. Second, asset-level debt, which is secured on the assets of each build, is now the largest part of CoreWeave's new financing, and lenders price it on the customer's credit rather than CoreWeave's. If that market stays open, the parent company needs unsecured debt only for the equity slice of each build. That slice is the 15–25% customer prepayment plus whatever the DDTL’s cost cap leaves unfunded.

Once a contract has been delivered and is performing, meaning the customer is paying as agreed, it is built to cover the scheduled interest and principal on its own loan, as the loan covenants are written. It is unproven that the original contracts repay the whole loan, including any bullet payment (a lump sum due when the loan matures). For DDTL 5.5, repayment depends partly on renewal, replacement or other repayment funding. We hold this view with medium confidence because three things are not disclosed: the coverage ratios actually achieved, the cash distributed to the parent company, and the share of the backlog tied to fully delivered capacity.

The weakest link is the fifth step in the chain. That step is whether the cash left over after the loan payments reaches the parent in amounts that, together with new capital, fund the parent’s bonds, leases and the next build. It is unproven, and the October repricing made it more expensive. For that reason, the evidence status is limited.

The evidence for our view is primary: the company’s own credit agreements, filings and covenant disclosures. The evidence against it is mostly market pricing and industry forecasts. The proof that would settle the debate is subsidiary-level cash flows, and those are not public.

CoreWeave depreciates its technology equipment, the GPU servers and networking that make up most of its assets, over six years on a straight line, which spreads the cost evenly across those six years. It lengthened the life from five years to six on 1 January 2023, which cut 2023 expenses by $20 million, according to the accounting-policy note in the 2025 10-K¹ [E03]. Technology equipment had a gross cost (its cost before depreciation) of $33.8 billion at 30 June 2026, out of $52.6 billion of gross property and equipment [E09]. The debate is whether a GPU bought today keeps earning rental income for six years. The alternative is that a faster product cycle means the depreciation charge is too small and reported profit is overstated.

Depreciation is an accounting charge that involves no cash, so the life CoreWeave chooses does not change the cash it collects. It changes reported profit, and it changes who is exposed if a chip stops earning. The asset-level loans are repaid in instalments over the contract, which typically runs about five years. A five-year contract that the customer pays as agreed is therefore meant to repay its lenders before the sixth year. Earnings in the sixth year depend on the chance to re-rent the chip once the first contract ends. Shareholders carry that risk. In the DDTL 5.5 facility, which finances shorter contracts, the parent company’s guarantee also carries part of it, through the non-replacement clause.

One contract reduces the risk. Under a September 2025 order initially valued at $6.3 billion, NVIDIA must buy residual capacity, meaning capacity that CoreWeave’s own customers leave unsold, through 13 April 2032, subject to termination rights and delivery and availability-of-service requirements, according to the 8-K filed 15 September 2025³⁰ [E62]. That backstop is small next to $33.8 billion of technology equipment at cost. It also depends on a single counterparty, which supplies the chips and owns CoreWeave shares.

The evidence against the six-year life concerns how fast chips are replaced, and when the supporting evidence was gathered. NVIDIA now ships a new generation roughly every year. CoreWeave itself completed the first bring-upGetting the first systems of a new chip generation installed and running, before they go into service for customers. of the Vera Rubin NVL72NVIDIA’s newest system generation, sold as a full rack of GPUs linked to work as one machine. system in the second quarter of 2026 [E13]. Gartner’s October forecast rests on this point: providers will not recover their investment “before the current generation of AI semiconductors is replaced” [E44]. IRENA smaller rival of CoreWeave that rents out GPU capacity from data-centre sites it controls; Microsoft is one of its customers.’s 2026 annual report warns that “the economic useful life and residual value of our deployed equipment may be shorter or lower than we anticipate” as demand concentrates on the newest chips [E50]. The re-rental evidence also comes from 2025–2026, a period when every GPU found a buyer because power, not chips, was scarce. It shows that old chips earn money in a tight market. It does not show what they earn when power is plentiful and the newest chips are cheap per unit of work.

A shorter life would make a large difference to reported profit and a small one to cash. If the six-year life were wrong and technology equipment should be depreciated faster, twelve-month depreciation would rise and operating income would fall, as shown below.

The shock path in Part VI models a large AI-lab customer that stops paying from 2028. The lost revenue is a hypothetical $4.5 billion a year, about 14% of base-case revenue for that year. In that path 2028 operating income falls to -$1.94B, because depreciation and rent continue after the revenue stops. That is worse than the bear case for that year. The path does not test the loss of the largest customer, or whether CoreWeave could fund the debt and guarantee payments that would follow. The GPUs would be released to rent to someone else, and in today’s market they would find a buyer. For some of that capacity there is already a buyer under contract: NVIDIA is obliged to buy unsold capacity through April 2032, an obligation initially valued at $6.3 billion and subject to delivery and availability requirements [E62]. Whether the rest would find a buyer at a price high enough to keep up the loan payments depends on the market examined in Debate 2.

The alternative reading is that customer concentration is less dangerous than it looks, because the largest customers have the least reason to walk away. Each depends on CoreWeave for capacity it could not build in time, prepaid part of the contract, and can terminate only for cause. The evidence that would decide between the two readings is what a large customer does when its own capacity comes online. It could renew, extend, or use a clause like the one in Meta’s 10-Q, where a $14.72 billion contingent capacity purchase “may be reduced if the cloud service provider is able to sell such capacity to other customers” [E53]. Meta does not name the provider in that clause, and it should not be assumed to be CoreWeave.

Recent news about demand from these customers is mixed. CNBC reported on 16 October 2026 that Anthropic, which signed a multi-year agreement with CoreWeave in the first quarter [E14], had reported slowing revenue growth [E57]. That is a press report of another company’s figures, and we draw no conclusion from it. But news like this is what would come before a renewal problem.

Core Scientific, a former bitcoin miner that converted its sites into data centres for CoreWeave, leased CoreWeave about 590 MW of capacity across five sites under contracts first announced in 2024. At 30 June 2026 about 395 MW had started billing, according to Core Scientific’s 10-Q for the quarter³⁴ [E47]. So two years after the first contract, two-thirds of the capacity was live. Core Scientific also disclosed that CoreWeave accounted for 77% of its revenue, and that “certain CoreWeave license agreements were assigned to a special purpose vehicle financing structure while CoreWeave remained a primary obligor”. That means some of the license agreements were moved into a separate entity used for financing, and CoreWeave stayed directly responsible for the payments. It is the financing model described in Debate 1, seen from the landlord’s side.

The question is whether CoreWeave can turn its backlog into revenue on the schedule it has disclosed. CoreWeave reports the backlog as remaining performance obligations (RPO), meaning revenue under signed contracts that it has not yet recognised. The schedule calls for 41% of the $103.7 billion of RPO, about $42.5 billion, within 24 months of 30 June 2026 [E06]. That schedule is demanding. Revenue in the twelve months to that date was $7.6B. To meet the schedule, CoreWeave would have to average more than $21 billion of revenue a year for two years from the existing backlog alone. Management’s guided exit rate for 2026, the annual pace of revenue it expects to be running at by the end of the year, is $18.5–19.5 billion a year [E41].

Delays cost money even when contracts survive. Equipment bought for late sites is held as construction in progress, the balance-sheet line for assets not yet in use, which stood at $11.9 billion at 30 June 2026. The interest on it is capitalised, meaning it is added to the cost of the asset instead of being charged against earnings; that came to $79 million in the second quarter [E09]. Leases that have been signed but have not yet started total $35.5 billion and begin between 2026 and 2029. At one site, 393 MW was still undelivered at 30 June, due in phases in 2026 and 2028, and at other sites a further 355 MW was due between 2026 and 2028 [E08]. The DDTL 4.0 loan agreement goes further. Under it, principal starts being repaid on a “delayed data center site amortization date” for any site that misses its delivery requirements, according to the 10-Q debt note⁶ [E07]. A late site can therefore bring debt repayment forward, before its revenue starts.

The same constraint affects the rest of the industry. IREN delivered the first of four data-centre phases for its $9.7 billion Microsoft contract in August 2026, after delays caused by “industry-wide component shortages, principally relating to NVIDIA equipment”, and has grace periods for the remaining phases running to the beginning of the second quarter of 2027, according to its 2026 annual report³⁶ [E50]. Lawrence Berkeley National Laboratory puts the time from a grid connection request to operation at more than five years [E46], and every new site that does not already have power waits in that queue.

Workflow one: a lab or hyperscaler buys a dedicated multi-year cluster. This is most of CoreWeave’s backlog. The buyer needs a large cluster of a named chip generation, running by a set date, for five to seven years. The steps are: find a site with power, secure the chips, build and cool the hall, install and test the cluster, then run it with the buyer’s software. CoreWeave handles the first step by leasing sites from developers such as Core Scientific [E47]. It has been early with each new NVIDIA generation, including the first GB300An NVIDIA system generation, the one before Vera Rubin. deployment and the first Vera Rubin bring-up (getting the first systems of a new generation running) [E17]. NVIDIA has reinforced that supply advantage with its $2 billion share purchase and its commitment to help CoreWeave procure “land, power and shell” [E36]. CoreWeave charges per reserved cluster, with a 15–25% prepayment [E02]. Oracle sells the same thing at a larger scale. Its cloud infrastructure revenue more than doubled to $7.39 billion in the quarter to 31 August 2026, and its remaining performance obligations (revenue under signed contracts that it has not yet delivered) reached $664 billion [E51]. IREN sells the same product from sites it controls, and its $9.7 billion Microsoft contract shows Microsoft buying it from a smaller rival [E50]. The largest alternative of all is the buyer building for itself: Microsoft has $329.1 billion of data-centre leases yet to start [E54].

CoreWeave claims its edge in how much work it takes to get a cluster running and in how well the cluster performs. It cites SemiAnalysis’s ClusterMAX PlatinumA grading of GPU cloud providers published by SemiAnalysis, an independent research firm that covers chips and AI infrastructure. Platinum is its top grade. rating, its MLPerf results (an industry benchmark of AI system speed), and a total cost of ownershipEverything it costs a customer to use a system over time, including setup, running costs and the staff time to keep it working, on top of the rental price. “up to 47% lower than the average hyperscaler”, a figure management gave on the August call [E41] [E17]. These are the company’s own claims or ratings it cites. No customer has published a measured comparison. At the hardware level, customers can move their work easily, because they run their own models on standard NVIDIA systems, managed with Kubernetes and Slurm (widely used tools for scheduling work across many machines). CoreWeave has made moving easier on purpose, with zero-egress migration (moving data across without a transfer fee) [E15] and a private interconnect to Google Cloud [E13]. That makes it easy for customers to move workloads in, and just as easy to move them out when the contract comes up for renewal. What keeps a customer at CoreWeave is the contract.

Three conclusions follow, each limited to what is public. First, CoreWeave wins large expansions from existing customers, which shows that customers who have used it come back for more. Every documented decision is an expansion, though, so the record says nothing about how CoreWeave does when it bids against rivals for new work. Second, the strongest competitor is the buyer itself, and the buyer’s build-out is far larger than any neocloud’s. Third, the competitors face CoreWeave’s constraints: IREN’s delays, Nebius’s spending and Oracle’s bond issuance show the same bottlenecks in power, chips and capital [E50] [E48] [E42].

Nebius and IREN now finance GPUs the way CoreWeave does. Nebius’s first secured loan, $775 million at SOFR plus 2.50% due 2030, is backed by GPUs and an investment-grade customer’s contract [E65]. That rate is close to the SOFR plus 2.25% on CoreWeave’s DDTL 4.0. IREN has roughly $3.6 billion of investment-grade-rated GPU financing, and its parent company guarantees only a limited part of it [E50]. CoreWeave’s advantages are that it gets new chips early and that it has the largest track record of financing capacity and delivering it. Neither is a durable cost advantage, and a rival with time and capital could match both.

Pay is mostly equity (shares and share awards), and it is reported in two ways that give very different figures. The Summary Compensation Table values awards when they are granted. It puts the chief executive’s 2025 total at $3.45 million. The pay-versus-performance rules value unvested equity at what it was worth at year-end. Under those rules his “compensation actually paid” was $146.7 million. A $100 investment at the listing was worth $179.03 at year-end [E38]. Despite its name, compensation actually paid is not pay he received. It moved with the share price.

CoreWeave is more candid about its financing than most issuers. It files its credit agreements and customer master agreements, with redactions, and gives a detailed debt note. Two things count against it. First, the amount the company may borrow under its unsecured bonds is measured on “Run-Rate Adjusted EBITDA”. In the June 2026 offering materials, that measure took $3.6 billion of reported adjusted EBITDA and added $15.1 billion of projected profits from new contracts to reach $18.76 billion [E39]. The covenants therefore let the parent borrow against profits it has not yet earned. Second, the chief executive responded to the October selloff by calling the market “irrational” [E42]. He added no disclosure on the subsidiaries’ coverage, which was what bondholders were asking about. A securities class action filed in January 2026 and related derivative suits (claims brought by shareholders on the company’s behalf) are pending. The company calls them without merit [E12].

Insiders have been selling modestly under pre-arranged plans. Form 4 filings (the reports insiders file when they trade) show Intrator selling about 187,000 Class A shares at $85.86–$88.72 on 29 September 2026 under a Rule 10b5-1 trading planA plan, set up in advance, that sells an insider’s shares on a fixed schedule or formula. Because the timing is fixed beforehand, it gives the insider a defence against charges of trading on inside information. (Form 4)⁴⁰ [E58]. The chief financial officer sold about 22,600 shares at around $80 on 12 October, also under a 10b5-1 plan (Form 4)⁴¹ [E59]. Next to the founders’ holdings of tens of millions of shares, these sales are small. All they show is that the plans kept running through the selloff.

CoreWeave is a rental business that needs a great deal of capital. Its financing is better built than its critics allow, and it depends on outside money more than its management admits. When a contract performs, meaning the customer pays as agreed, its payments reduce the loan raised against it on schedule. The filings do not show that the original contract repays all of that loan. Growth needs a steady supply of new unsecured capital (money not tied to specific contracts or assets), and in October 2026 that capital started to cost about 15%. Whether this is a good business is unproven. Here a good business means one that earns more than its cost of capital, the return its lenders and shareholders demand, across a full cycle. CoreWeave has not yet earned an operating profit at scale, and it has no long record of customers renewing. Its main advantages are early access to new NVIDIA chips and the largest record of financed delivery. Well-funded rivals and its own customers can close both leads.

We disagree with the common view on two points, and add a caution on a third. First, the October selloff treated all of the debt as one burden. Most of the asset-level debt, the loans raised against specific assets, is matched to named contracts and is paid down in instalments from those contracts' payments. The filings do not show that those payments alone pay it off. The risk is at the parent company, which has to keep funding new builds. Second, and against the company's framing, calling the debt "asset-backed" does not make it ring-fenced, which would mean walled off from the parent so that a default stays with the assets behind the loan. About 90% of the debt is guaranteed by the parent, and the newest facility relies on that guarantee to lend beyond what its contracts cover. Third, the largest customer's share of revenue has fallen, but that fall may overstate how much CoreWeave's reliance on a few customers has eased. Meta and OpenAI have announced contract awards worth up to about $35 billion and $22 billion. Those are large next to the $104 billion backlog, the contracted revenue not yet earned, that the debt is repaid from. How much of each award is still left in the backlog is not disclosed.