Same Chips, Three Balance Sheets

§Same chips, three ways to pay for them

On 11 August 2026 CoreWeave told investors it had signed customer contracts worth about $104 billion¹, held 1.5 gigawatts of active power, and lost $626 million in the quarter. Two weeks later IREN closed a fiscal year with $707.0 million² of revenue, roughly 40 megawatts of AI capacity running, and a $702.6 million net loss driven by writing off the bitcoin miners it is tearing out. Nebius sat between them: $582.3 million³ of quarterly revenue, $8.0 billion of cash, and a $430.8 million half-year profit that came mostly from revaluing an equity stake. Three companies, one product, three completely different pictures.

All three buy Nvidia silicon at broadly the same prices and sell it on broadly the same contracts. The durable difference sits in the financing: CoreWeave pays 9.00% on $35.6 billion of debt, while IREN and Nebius pay under 2% on roughly $8 billion each.

The consensus treats these three as one trade. They are lumped together as "neocloudsA cloud company built for one workload. It buys the specialised Nvidia chips that AI models run on, puts them in a building with the power, cooling and networking they need, and rents the capacity out — usually on multi-year contracts to a small number of large customers. What it does not sell is the databases, developer tools and enterprise software that a general-purpose cloud wraps around the same hardware." (companies whose whole business is buying Nvidia chips, housing them, and renting the capacity out under contract), ranked by backlog, and discussed as though the question were which of them has the best contracts. That framing is half right and it hides the part that matters. Contract structure across the three is more similar than different — multi-year, prepaid or take-or-pay (the customer pays for the capacity it reserved whether or not it uses it), anchored on a handful of names that could build the capacity themselves. Where they genuinely differ is upstream of the contract, in what each company had to promise to raise the capital that bought the hardware. CoreWeave promised cash coupons. IREN and Nebius promised equity. Those are two different answers to the same question: who eats the loss if AI compute reprices. They produce two different companies in a downturn.

That difference decides three of the arguments that follow. It runs through the depreciation policies, which diverge enough to make reported margins non-comparable. It runs through the contract books, which cannot be compared at face value because one of the three records its largest arrangements as equipment leases rather than cloud services. And it runs through the moat question, where only one advantage in the group cannot simply be bought at today's prices.

1. 1The funding model, not the business model, is the real differentiator. CoreWeave's disclosed weighted-average interest rate — the blended rate across everything it has borrowed, each loan counted in proportion to its size — is 9.00% on $35.6bn of principal; IREN's computed weighted average is 1.64% and Nebius's 1.91%. On current balances that is roughly $3.6bn of annual cash interest against about $155m and $126m. High
2. 2The three contract books are not comparable as reported. CoreWeave's $103.7bn and Nebius's $37.5bn are ASC 606The US accounting standard that governs how a company recognises revenue from selling a service. It is the rule that produces the "remaining performance obligation" figure every neocloud quotes as its backlog. remaining performance obligations — signed revenue the company has not yet delivered and therefore not yet booked; IREN's headline arrangements are operating leasesA contract the accountants treat as renting out a specific asset rather than selling a service. The distinction matters here because a lease of named GPUs in a named hall is a different economic animal from a cloud subscription — and, crucially, its contracted value never appears in the RPO line that everyone compares. — $11.4bn of ASC 842The US accounting standard for leases. When a customer contract is judged to be predominantly the rental of specified equipment and space, it is accounted for under this standard instead of ASC 606, and its value sits outside the reported revenue backlog altogether. lease value plus only $5.1bn of RPO. Compare $104bn to $5.1bn and you have measured nothing. High
3. 3The investment-gradeThe top half of the credit-rating scale — a borrower, or a promise to pay, that the rating agencies judge unlikely to default. Below it sits "high yield," where lenders demand a much higher interest rate for the same money. The label can attach to a company, to a bond, or — as this brief shows repeatedly — to a single ring-fenced pool of contracts rather than to the operator that assembled it. anchor customer is paid for in price. A customer that large and that creditworthy anchors the whole build — and lenders will finance its promise cheaply. IREN's Microsoft contract implies about $9.7m of revenue per MW of IT loadThe power actually drawn by the computers themselves, measured in megawatts, excluding what the cooling and the rest of the building consume. It has become the industry's unit of sale: capacity is quoted in megawatts of IT load, and price in dollars of annual revenue per megawatt. A site's total facility power is always the larger number, which is why the two must never be compared directly. per year — the industry's standard unit, dollars of annual revenue for each megawatt of computing power sold; IREN says recent three-year deals clear above $20m and discussions run near $25m. The cheaper funding that Microsoft's credit unlocks is worth a fraction of the price gap. Moderate
4. 4Depreciation divergence makes reported margins un-rankable. Depreciation is the slice of a chip's purchase price charged against profit each year, and the size of the slice depends entirely on how many years the company assumes the chip will earn. CoreWeave depreciates technology equipment over six years; IREN and Nebius use five. Re-lifing CoreWeave's fleet to five years, which means re-running its accounts on the shorter clock its peers use, adds roughly $238m of quarterly depreciation and turns a $49m Q2 operating loss into about $287m. High
5. 5IREN's power is the only advantage here that cannot be bought at today's prices. Electricity was 3.5% of its FY26 AI Cloud revenue — $4.5m against $128.8m — because grid positionsLand that comes with a signed, scheduled right to draw a stated amount of power from the electricity network. The land is the cheap part. The right to draw power is the scarce part, because utilities grant it through a queue that runs years long and cannot be jumped by paying more. (land with a signed right to draw power from the network) bought for bitcoin mining now carry a product worth roughly thirteen times more per megawatt. Moderate
6. 6Nebius's funding advantage is a condition, not a moat. It raised $2.8bn on an at-the-market programmeA standing arrangement, usually shortened to ATM, that lets a listed company drip new shares into the open market at whatever the price happens to be that day, rather than raising a fixed sum in one announced offering. Cheap and flexible while the share price is high. It closes on its own when the price falls, because nobody wants to issue stock into weakness. (selling new shares into the open market a slice at a time) at a weighted-average $223.6 per share and $4.3bn of convertsConvertible notes: debt that pays a low interest rate because the lender also gets the right to swap it for shares if the stock rises far enough. The borrower buys a cheap coupon today and pays for it later in dilution. The low rates on IREN's and Nebius's borrowings are almost entirely this instrument. (bonds that pay a low coupon because the lender can swap them for shares) at 1.25–2.625%. Both prices are functions of its own equity price, which is exactly what a downcycle removes. Moderate
7. 7Ranking on durable structural quality: IREN first, Nebius a close second, CoreWeave third. CoreWeave would rank first on scale and contracted revenue and does not rank first here, because at 2.9× capex-to-revenue (nearly three dollars of building for every dollar it bills) and 28% of guided 2026 revenue going to interest it is the machine with the least room to be wrong. Moderate

01What a neocloud actually sells

A customer who signs with CoreWeave, IREN or Nebius is reserving a physical thing, not a software subscription: a specified number of Nvidia acceleratorsThe specialised chips — GPUs, in the common shorthand — that AI models are trained and run on. They do a narrow kind of arithmetic thousands of times in parallel, which is what a neural network is made of. Nvidia designs the ones every company in this study buys, and sets their price., wired into a specified network fabricThe high-speed wiring that lets thousands of accelerators behave as one machine. A large AI job is split across the whole cluster, so every chip must exchange results with every other chip constantly; if the fabric is slow, the chips sit idle waiting. It is architecture rather than plumbing — the fabric is a large part of what a customer is actually renting., sitting in a specified building drawing a specified number of megawatts, for a specified number of years. Almost everything about the economics follows from that sentence. The asset is bought up front with borrowed or raised money. The revenue arrives monthly over the term. The gap between those two facts is the business.

Take-or-pay, backlog and RPO are the terms every table in this report is built from, and none of them means quite what it sounds like.

The chain from a megawatt to a monthly invoice has five links, and each of the three companies owns a different subset of them. The constraint moves down the chain as the market tightens. In 2023 it was GPU allocation. By 2026 it is power and the time it takes to energiseTo bring a site's electrical connection into live service — the substation built, the utility's approval given, current actually flowing to the racks. A site can be bought, permitted and half-built and still not be energised, which is why "secured" power and live power are different numbers, often by a factor of a hundred. it — which is why every one of the three now leads its disclosure with gigawatts rather than chip counts.

Pricing sits on top of that chain and is quoted, in this market, per megawatt of IT load per year. Nebius told shareholders in August that its second-quarter deals averaged "more than $20 million in revenue per MW"⁴ and that it sees "a price opportunity in the $40–50 million per MW range" for short-duration capacity. IREN's August results deck puts recent three-year contracts at ">$20m revenue per MW (IT)" with "active discussions at ~$25m per MW"⁵. CoreWeave does not disclose a per-megawatt price at all. The largest operator in the group is the one that has chosen not to publish its unit price.

Three contract shapes now coexist, and they are converging across all three companies rather than distinguishing them. Long-dated deals with investment-grade counterparties run typically five years, priced low and used to raise cheap secured debt. Mid-length deals of one to three years go to AI labs at the market price. Short deals of three to six months carry a large premium, and have appeared at both Nebius and CoreWeave within the last quarter. CoreWeave's chief executive was explicit about why the short book matters: enterprises "tend to think in shorter cycles" than five years, and the company needed a financing structure that could support two- and three-year contracts before it could sell them. That structure arrived on 10 August 2026⁶ as a $2.6 billion facility.

02The stakes, sized

All three can be set against each other on a single date. The table below is dated 3 September 2026 for market values and uses each company's most recent reported balance sheet — 30 June 2026 for all three, which happens to be a common date because IREN's fiscal year ends then.

Two rows in that table deserve a warning label. The contracted-book row is the one most often quoted and most often misused: CoreWeave's $104 billion and IREN's $5.1 billion are not the same measure, because IREN's largest arrangements are not service contracts at all. Its FY2026 revenue note states that the company "determined that these arrangements contain leases of specified GPU equipment and dedicated data center space," that the lease component "is predominant," and that the aggregate contracted value of lease arrangements was approximately $11.4 billion¹³. That value sits outside RPO entirely. Any comparison that puts $104 billion next to $5.1 billion is comparing a total to a fragment.

The power row carries the same hazard in a different form. CoreWeave's "1.5 GW active" is facility power across its footprint; IREN's "40 MW" is AI Cloud IT load specifically, alongside roughly 380 MW still drawing power for bitcoin miners it is decommissioning; Nebius does not publish a live figure at all and reports only contracted power secured by land and grid commitments. These three numbers are not measuring the same thing and should not be ranked against each other. One comparison does survive: CoreWeave has built and energised more than the other two combined, by a wide margin, and has therefore taken and survived more delivery risk.

Who owns what is the last of these comparisons. The same GPU sits in three quite different corporate structures, and the differences show up in the balance sheet long before they reach the income statement.

03CoreWeave — the leveraged contract machine

CoreWeave's second quarter of 2026 contains the whole argument about the company in six numbers. Revenue of $2,575 million, up 112%. Adjusted EBITDA of $1,510 million, a 59% margin. Adjusted operating income of $128 million, a 5% margin. A GAAPGenerally Accepted Accounting Principles — the audited US rulebook a company must report under. Anything labelled "adjusted" is the company's own alternative presentation, with items it considers unrepresentative taken out. Both can be honest. Only one is standardised, which is why a GAAP number and an adjusted number for the same quarter can differ by a great deal and neither is a mistake. operating loss of $49 million. Net interest expense of $640 million. A net loss of $626 million. Those are one quarter seen through six filters, and the distance between the top and the bottom is the argument. EBITDA is profit measured before the cost of the chips (depreciation) and the cost of the money (interest) are taken out. In most industries that is a fair stand-in for the cash a business throws off. Here the chips and the money are the business, so EBITDA is the one measure built to leave out what separates these three companies. The business converts most of its revenue into EBITDA and then hands nearly all of it to depreciation and lenders.

Sources: CoreWeave 10-Q statement of operations¹⁶, Q2 2026 results release¹, debt schedule⁸ and debt narrative²¹, all as at 30 June 2026.

What it sells and to whom

CoreWeave sells a managed AI cloud rather than bare metalRenting the physical machines and nothing else: the customer gets root access to the hardware and brings its own software for everything above it. The opposite of a managed cloud, where the supplier also runs the scheduling, storage and monitoring layers. Bare metal is cheaper per chip and shifts the work, and the blame, to the customer. — networking backbone, GPU clusters, purpose-built storage with GPU-local caching, and an orchestration layer it calls Mission Control. Its Q2 2026 10-Q²² states plainly that revenue "primarily comes from committed contracts" rather than on-demand consumption. The company builds against signed commitments, not against a demand forecast.

In FY2025, a single unnamed customer was 67% of revenue²³, up from 62% in 2024. By the second quarter of 2026 that customer was 36%²⁴, with a second at 26% and a third at 10%. Concentration has genuinely fallen, and it fell because the denominator more than doubled rather than because the anchor shrank.

The filings do not name these customers. CoreWeave's S-1²⁵ lists Cohere, IBM, Meta, Microsoft, Mistral and Nvidia among the "leading AI labs and AI enterprises" using the platform, but attaches no percentage to any of them. The widespread attribution of the 67% figure to Microsoft is trade-press inference, not disclosure. What is disclosed, from both sides, is the contract book itself.

The financing, which is the business

CoreWeave has raised, by its own count on the August call, "over $32 billion of debt and equity capital secured to date." The debt schedule at 30 June 2026 runs to more than a dozen separate instruments: delayed-draw term loansA loan agreed in full up front but taken down in instalments as the borrower needs the cash — typically as each building fills with equipment. It suits a company spending money in stages, and it means the headline facility size and the amount actually borrowed are two different numbers. numbered 1.0 through 5.0, dollar senior notes with stated couponsThe interest rate printed on a bond — the cash the borrower must hand over every year for as long as the bond is outstanding, whatever the business is doing that year. Unlike a dividend, it is not optional. of 9.00%, 9.625%, 9.75% and 10.00%, euro notes at 8.50%, convertible issues at 1.75% and 2.00%, OEM and software licence financings at 9.00% and 11.00%, and a legacy Magnetar loan at 12.00%. Gross long-term debt is $35,551 million⁸. The disclosed weighted-average rate is 9.00%²¹.

Management presents the trajectory as an improving one, and on its own terms it is. The company says it has reduced its weighted-average cost of debt by "almost 300 basis pointsHundredths of a percentage point. A hundred basis points is one percent. Lenders count in them because on a balance sheet of this size a quarter of a point is real money, and saying "twenty-five basis points" leaves no room to mishear "a quarter percent" as "point two five percent."" over the past year, worth about $1.1 billion of annualised interest saving on the end-Q2 debt load. The centrepiece of that repricing closed on 31 March 2026³⁰: an $8.5 billion delayed-draw term loan rated A3A rung on Moody's credit-rating ladder, comfortably inside investment grade — roughly the standing of a solid large corporate. The ladder runs Aaa, Aa, A, Baa and then down into the speculative grades. DBRS's "A (low)" sits at about the same height on a different agency's scale. by Moody's and A (low) by DBRS, priced at SOFRThe Secured Overnight Financing Rate — the benchmark US interest rate that floating-rate loans are quoted against. "SOFR + 2.25%" means the borrower pays whatever short-term rates happen to be, plus 2.25 percentage points. The spread is the lender's judgment of the borrower; the SOFR part is the weather. + 2.25% floating or roughly 5.9% fixed, maturing March 2032 and secured by substantially all assets of a bankruptcy-remote subsidiary together with the associated customer contract. It was the first investment-grade-rated financing secured by high-performance computing infrastructure, and it is a genuine capital-markets achievement.

Ring-fencing is how the cheap debt in this sector gets raised. A company sets up a separate subsidiary, puts one customer contract and the hardware that serves it inside, and borrows against that subsidiary alone. The lenders can seize what is inside the fence and nothing beyond it; if the subsidiary fails, the parent is not on the hook. What such a lender is really buying is one customer's promise to pay, insulated from everything else the parent is doing. The A3 rating attaches to a ring-fenced vehicle — "substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC" plus the associated customer contract — and not to CoreWeave. What the parent itself pays is visible in the same debt schedule: unsecured senior notes issued at 9.00%, 9.625%, 9.75% and 10.00%, which is high-yieldThe polite name for debt rated below investment grade, and the older name is junk. Lenders charge these rates when they think default is a live possibility. The label describes the price the market demands, not a verdict on the company — but the price is the market's verdict. pricing by any reading. The structure demonstrates that a strong counterparty's payment obligation, isolated from the operating company, can be financed at investment-grade cost. That is a statement about the customer's credit and the enforceability of the contract. It is not, by itself, a statement about CoreWeave's.

Then there are the leases, which are debt in everything but name. At 30 June 2026 CoreWeave carried $16,595 million of operating lease right-of-use assets against $16,319 million of operating lease liabilities. The note discloses a further block of leases signed but not yet commenced — approximately $35.5 billion of data centre and office building obligations over sixteen-year terms¹⁷, including a single site of about $14.7 billion. None of that appears on the balance sheet until the landlord delivers. Add the on-balance-sheet leases to the debt and CoreWeave has roughly $52 billion of fixed obligations today, with another $35 billion queued behind them.

A sixteen-year lease against a six-year depreciation schedule is a bet that the building outlives four fleets of chips.

The structural question inside CoreWeave's lease book

Growth, and what has to go right

The build plan is the most aggressive of the three by a distance. Capex guidance for 2026 is $35–39 billion against revenue guidance of $12.4–13.2 billion — roughly 2.9 times revenue. Active power is guided to exceed 1.85 GW by year end, with contracted power at 4.2 GW as of 11 August and a stated target of "at least 8 gigawatts by 2030." International capacity has passed 1 GW, including 360 MW in Indonesia due online in about eighteen months.

Against that, the conversion evidence is better than the sceptics allow and worse than the headline suggests. Management disclosed that "more than 50%" of the Q2 backlog is attached to a contract where customer delivery has commenced, and expects that to reach two-thirds by year end. The RPO note is more precise: 41% of the $103.7 billion is expected to be recognised within 24 months from 1 July 2026, and a further 39% within the 48 months from 1 July 2028. So roughly $42.5 billion converts by mid-2028 and the balance stretches into the 2030s. A backlog with a weighted life that long is a portfolio of long-dated obligations, not a forward order book in the ordinary sense. Its value depends on counterparties still wanting the capacity in 2031.

CoreWeave raised prices roughly 25% across SKUsStock-keeping units — the distinct products on a price list. Here each SKU is a particular configuration of chips, memory and networking rented by the hour. A price rise "across SKUs" means the whole menu moved, not one line of it. in July 2026 "in response to the current demand environment," and says it remains "largely sold out of prior generations of NVIDIA GPUs in addition to the current SKUs." If that holds through a cycle it is the strongest evidence yet against the residual-valueWhat a GPU is still worth once its first contract has run out — whether it can be re-let, and at what price, or sold second-hand. Nobody knows yet: no operator has taken a large modern fleet through a full contract term and remarketed it. Lenders nonetheless have to put a number on it, because residual value is a large part of what secures the loans this sector runs on. bear case, and it belongs in the ledger alongside the leverage.

04IREN — the power company that learned to sell compute

IREN spent $186.6 million on electricity in the year to 30 June 2026 to run bitcoin miners that produced $578.2 million of revenue. In the same year it spent $4.5 million on electricity to run an AI cloud that produced $128.8 million. Power was 32% of the cost of one product and 3.5% of the other. That single pair of numbers, from the cost of revenue note³¹ in its FY2026 10-KA US-listed company's annual report to the Securities and Exchange Commission — the audited, legally binding account of the year, and the most reliable document any issuer produces. Its relatives appear throughout this brief: the 10-Q is the quarterly version, the 8-K a one-off disclosure of a material event, and the 20-F and 6-K the equivalents filed by foreign issuers such as Nebius. Press releases and investor decks are none of these., explains the entire strategic pivot and most of what makes IREN interesting.

Sources: IREN FY2026 10-K statement of operations², FY26 results release⁵, 10-K Item 1¹⁸, and rate computed from the debt note¹².

The asset nobody was bidding for

Between 2021 and 2024 IREN acquired land and grid connections in places chosen for cheap renewable power and indifferent competition: Canal Flats, Prince George and Mackenzie in British Columbia; Childress in the Texas panhandle; later Sweetwater, Kiowa in Oklahoma, Bundey in South Australia. Bitcoin mining paid for the option. As of 30 June 2026 the company reports "executed grid connection agreements, letters of agreement or equivalents representing approximately 5GW of total power capacity¹⁸" across the United States, Canada, Spain and Australia, with announced site capacities of 1,600 MW at Kiowa, 1,400 MW at Sweetwater 1, 800 MW at Bundey, 750 MW at Childress and 600 MW at Sweetwater 2.

The word "equivalents" is doing real work in that sentence and should not be waved past. A letter of agreement is not an energised substation. IREN's own operating footprint tells the truth about the gap: roughly 40 MW of AI Cloud capacity and about 380 MW still running miners at year end, against a 5 GW paper portfolio. The distance between those numbers is the company's entire execution risk.

Where the power already flows, the numbers are clear. Strip out depreciation and IREN's AI Cloud segment ran at an 86.9% gross margin in FY2026 — $128.8 million of revenue against $16.9 million of electricity, employee and other direct costs. Nebius, on the same pre-depreciation basis, ran at 77.1% in the June quarter. CoreWeave's 65.9% is not directly comparable because its cost of revenue line includes depreciation of power installation and distribution systems, which the other two report separately. The ranking of the three on this measure gives the order, not the size of the gaps, and it points the same way as everything else about IREN's power.

The contract book, and what it is made of

IREN's commercial breakthrough was the Microsoft agreement announced on 3 November 2025³²: approximately $9.7 billion over five years, including a 20% prepayment, for NVIDIA GB300 capacity deployed in four 50 MW liquid-cooledCooling the chips by running coolant directly over them rather than blowing cold air through the room. The newest Nvidia systems draw too much power per rack for air to carry the heat away, so liquid cooling is what lets a hall be packed densely. It also means a different building, and a costlier one, than an air-cooled site. halls at Childress. IREN simultaneously agreed to buy the GPUs and ancillary equipment from Dell for about $5.8 billion. The first hall, Horizon 1, was delivered to and accepted by Microsoft in August 2026³³, with the remaining three targeted for the fourth quarter.

Then, on 7 May 2026³⁴, a second anchor: a $3.4 billion five-year AI cloud contract with NVIDIA itself for air-cooled Blackwell systems inside about 60 MW at Childress, alongside a strategic partnership contemplating up to 5 GW of NVIDIA DSX-aligned infrastructure and a five-year right for NVIDIA to purchase up to 30 million IREN shares for about $2.1 billion, conditioned on delivery of up to 600,000 GPUs. By the August results, IREN listed Cohere, Prometheus, Perplexity, Figure AI, Fal AI, Higgsfield AI, Together AI, Fluidstack, Fireworks AI and Hume AI among newer counterparties, plus an unnamed "leading frontier AI lab," and said 2026 capacity was "largely sold out."

IREN's revenue note states that these arrangements "contain leases of specified GPU equipment and dedicated data center space," that the lease component "is predominant," and that they are therefore classified as operating leases under ASC 842. Only the non-lease services element runs through ASC 606. The consequence is that IREN's disclosed RPO of $5.1 billion excludes its largest contracts, whose aggregate contracted value of approximately $11.4 billion is disclosed separately.

This is not an accounting quibble. A contract classified as a lease of specified equipment is, economically, a lease of specified equipment. The customer has taken the hardware; the differentiation IREN claims from owning "all three layers of the stack" does not obviously reach the layer the anchor customer is paying for. IREN acquired Mirantis in August 2026 for approximately $544 million precisely to thicken that software layer — an admission, in the form of a cheque, that it was thin.

The price of an investment-grade customer

Two disclosures belong side by side. The Microsoft contract is $9.7 billion over five years across 200 MW of IT load. That is $9.7 million of revenue per MW per year. The August results deck says recent three-year contracts clear ">$20m revenue per MW (IT)" with discussions "at ~$25m per MW." Nebius reported the same range for its own second-quarter deals.

The cost side of the same contract is derivable, which is rare in this sector. IREN agreed to buy the GPUs and ancillary equipment for the Microsoft deployment from Dell for approximately $5.8 billion, against 200 MW of IT load — $29.0 million of GPU capital cost per megawatt. Set that against $9.7 million of annual revenue per megawatt and the GPU spend alone takes three years of contract revenue to return, before the data centre, before operating costs and before interest, on a five-year contract. At the $20 million per megawatt that newer three-year contracts clear, the same GPU cost returns in about eighteen months. That arithmetic is the whole reason IREN's management describes recent deals as carrying roughly a two-year payback and Nebius describes its own second-quarter deals at "1 year and 10 months." It is also why the Microsoft contract, taken alone, is a much thinner piece of business than its headline value suggests. A further Dell purchase of approximately $1.6 billion was agreed in May 2026³³; the filing does not state the megawatts it serves, so no second per-megawatt figure is computed here.

The trade is explicit in IREN's own financing disclosure. The Microsoft deployment is funded by "$3.6bn investment grade GPU financing for Microsoft contract at 6.0%," which together with prepayments covers 96% of the associated GPU capex. The non-investment-grade deployments at Mackenzie are funded by "$2.4bn financing led by Blue Owl and Pacific Investment Management Company LLC ... at a 9.0% fixed rate," covering 90% of capex. The credit spread IREN saves on Microsoft is 300 basis points on $3.6 billion — about $108 million a year.

Set that against the revenue gap. Two hundred megawatts at $9.7 million versus a market that now clears above $20 million is roughly $2 billion a year of foregone revenue at today's prices. Even discounting heavily for the nine-month gap and the longer tenor, the funding saving recovers a small fraction of it. The lesson generalises beyond IREN. An investment-grade anchor is a financing instrument, and the market charges for it in price. That cuts against the standard reading of these deals as unambiguous validation.

The balance sheet

IREN's debt at 30 June 2026 was $7,705.7 million of principal in two clearly separated buckets. The split matters more than the total. Recourse debt is borrowed by the parent company itself, and a lender who is not repaid can pursue everything the company owns. Non-recourse debt is borrowed inside one of the ring-fenced vehicles described above, and that lender's claim stops at the equipment and the contract sitting behind the fence. A balance sheet adds the two together. They behave nothing alike in a bad year. Recourse debt of $6,745.7 million is almost entirely convertible notes: $3,000m at 1.00% due December 2033, $1,150m at 1.00% due June 2033, $1,150m at 0.25% due 2032, $1,000m at 0.00% due 2031, $233.4m at 3.50% due 2029 and $212.3m at 3.25% due 2030. Principal-weighted, that is a 0.88% cash coupon. Non-recourse debt of $938 million takes the blended figure to 1.64%. It is a delayed-draw facility at 7.12% effective and US private placement notesBonds sold directly to a handful of institutions — insurers and pension funds, mostly — instead of being offered publicly. The buyers hold to maturity and negotiate tighter terms; the borrower avoids a public issue and the disclosure that comes with it. There is no traded price, so the market never marks the debt in public. at 7.05% effective, both secured by GPUs and infrastructure with debt-service-coverage covenantsA promise, written into the loan, that the cash the financed assets generate will stay above an agreed multiple of the payments owed on them. Fall below it and the lender gains rights: cash can be trapped inside the vehicle, distributions to the parent stopped, and the loan declared in default. Covenants are how a lender acts on a deteriorating position before a payment is actually missed.. Total interest expense for the whole fiscal year was $46.6 million¹².

The price of that cheapness is dilution, and it has been severe. Dilution is what it costs to raise money by printing shares instead of borrowing. The company is worth what it was worth a moment earlier, but that value is now split among more owners, so each existing share is a claim on a thinner slice. Nothing about it appears as an expense on the income statement. The bill goes straight to the shareholder, in full, and it does not reverse. Weighted-average shares went from 99.6 million in FY2024 to 214.6 million in FY2025 to 316.1 million in FY2026, with 394.1 million outstanding by the August filing date — very nearly four times the FY2024 base in two years. Convertible holders sit above that with claims on a further several hundred million shares. IREN's shareholders have funded this build; they have simply done it through the share count rather than the interest line.

Nearly all of IREN's committed spending comes due inside a year. Capital commitments at 30 June 2026 were $13,810.0 million, of which $13,611.0 million falls within twelve months³⁵, against $7,619.5 million of cash and restricted cash. IREN's own framing is that it is "well capitalized with existing cash and committed GPU financing and prepayments of $14bn," which closes the gap arithmetically — provided the committed financings draw and the prepayments arrive on schedule. Both are conditional on delivering capacity. The fault line in this machine is execution, not solvency.

05Nebius — the software company with a balance sheet

Nebius sold 12.7 million of its own shares into the market in the first half of 2026 at a weighted-average price of $223.60, raising about $2.8 billion, and issued $4.34 billion of convertible notes at coupons of 1.25% and 2.625%. It also expects to collect more than $9 billion of customer prepayments during 2026. That is how a company with $582 million of quarterly revenue funds $20–25 billion of capital spending without a syndicated loan book. It is elegant, and it works precisely as long as the share price and the prepayment culture both hold.

Sources: Nebius Q2 2026 results release³, interim financial statements¹⁴ and shareholder letter⁴, all 12 August 2026.

A different product, sold to a different mix

Nebius came out of the Yandex separation with engineers rather than land. Its 2025 Form 20-F²⁰ describes designing servers and racks in-house, building the orchestration software, and running "one of the few global, at scale, multi-tenant clouds purpose built for AI." The distinction from bare-metal rental is real: multi-tenantOne physical cluster serving many separate customers at once, each isolated from the others and each able to take as much or as little as it needs. The alternative is single-tenant: a hall of machines handed over whole to one customer. Multi-tenancy is harder to build and is what allows a supplier to sell to hundreds of small buyers rather than to five large ones. provisioning "from a single node to thousands of nodes" is a different engineering problem from handing a customer a dedicated hall. It is why Nebius can serve hundreds of small customers alongside its anchors.

That shows up in the concentration data. In the June 2026 quarter Nebius's three disclosed customers were 24%, 21% and 14% of revenue, leaving 41% outside the top three. A year earlier two customers had been 39% and 15%. The identity of the top customer changed entirely between the two periods — the filing's Customer A and B of 2025 do not appear in 2026, replaced by C, D and E. A customer base that rotates at the top is a different risk profile from one anchored on a single name, and on balance a healthier one.

The two anchors, and their terms

On 8 September 2025³⁶ Nebius disclosed a five-year Microsoft agreement with "total contract value ... about $17.4 billion through 2031," rising to about $19.4 billion if Microsoft takes optional additional services, with GPU services "deployed in several tranches during 2025 and 2026" at a new data centre in Vineland, New Jersey. The same 6-K discloses the termination mechanics, and they are the most informative sentences either side has published about how these contracts actually work: Microsoft may terminate individual GPU services if Nebius fails to meet agreed delivery dates, after a grace period, and cannot provide alternative capacity. Either party may terminate for material breach with a sixty-day cure, or on insolvency.

CoreWeave's 2025 delivery failures showed what a clause like that is for. These are delivery-conditioned commitments rather than unconditional take-or-pay obligations, and the condition runs against the supplier. A neocloud that misses a hall does not merely lose margin; it hands the customer an exit. By August 2026 Nebius reported it had "delivered all capacity tranches to Microsoft under the contract, including two tranches over the past month" — the condition has been met so far, and the fact that management chose to say so tells you how closely it is watched.

The second anchor is larger. On 16 March 2026³⁷ Nebius announced a five-year Meta agreement: $12 billion of dedicated capacity committed, plus a Meta commitment to purchase additional capacity across upcoming clusters "up to a total of $15 billion," for a contract value of up to approximately $27 billion, on the NVIDIA Vera Rubin platform, delivering from early 2027. It followed a roughly $3 billion Meta agreement signed the previous autumn. As of August the buildout supporting the second Meta agreement was "on track to come online in early 2027."

Total disclosed customer commitments are "more than $40 billion," against $37.49 billion of RPO at 30 June 2026, of which 36% is expected to be recognised in the 24 months to June 2028 and 40% in months 25 to 48. The book is younger than CoreWeave's and converts on a similar clock.

Where the profit comes from, and where it doesn't

Nebius's income statement needs reading twice. Group adjusted EBITDA of $236.2 million in the June quarter was a 41% margin, against a $21.0 million loss a year earlier; the AI cloud business on its own delivered $285.7 million and a 49.7% margin. Cost of revenue fell from 29% to 23% of revenue as colocation and lease costs were spread over more capacity. That is real operating leverageWhat happens when a business's costs are largely fixed: once revenue passes the level that covers them, most of each additional dollar drops through to profit. It works ferociously in both directions, which is why a business with heavy fixed costs looks brilliant while it is filling and brutal while it is emptying. in a business that had none twelve months ago.

The headline net income of $430.8 million for the half is a different animal. Nebius's 2025 accounts show the same pattern at annual scale: a $611.7 million loss from operations rescued by a $598.9 million "gain from revaluation of investments." The equity stakes in ClickHouse — which raised at a reported ~$15 billion valuation in January 2026 — and Toloka are genuine assets and genuine optionality. They are not compute earnings, and a reader who takes the net income line as evidence that renting GPUs is profitable has been misled by the arithmetic rather than by the company.

The capital plan is the largest relative to revenue in the group. Capex guidance of $20–25 billion against revenue guidance of $3.0–3.4 billion is roughly seven times revenue, against CoreWeave's 2.9 times. Nebius has three answers. Prepayments: roughly 70% of Q2 deals included them, covering 50–60% of associated capex, with more than $9 billion expected in 2026. Secured debt: a first $775 million facility in July 2026 at SOFR + 2.50%, backed by deployed GPU infrastructure and the contracted cash flows of an investment-grade customer. And equity, in every available form — the ATM, the converts, and a $2 billion prefunded warrantA right to buy shares where the buyer pays almost the whole price now and a token amount later on exercise. In practice the company has the cash immediately and the shares are issued at a time of the holder's choosing. It is equity raised today, counted in the share tally tomorrow. issued to NVIDIA. In August 2026 it priced a further $5.0 billion of convertible notes, partly to exchange the 2029 and 2031 issues.

The fourth answer is the one Nebius says least about. Nebius introduced an "asset-light partnership model" in the June quarter under which partners build and own the data centres to Nebius specifications while Nebius supplies the software, systems architecture and demand. If that works at scale it changes the company's capital intensity fundamentally — it would be selling the platform layer while someone else carries the fleet. It is early, unquantified, and the largest source of upside optionality in this study.

06The funding stacks compared

Three companies bought roughly the same asset with money that costs three very different amounts. CoreWeave's disclosed weighted-average interest rate at 30 June 2026 was 9.00%. Computed on the same principal-weighted basis from their own debt notes and offering documents, Nebius pays 1.91% and IREN 1.64%. On the balances outstanding that is about $3.6 billion of annual cash interest at CoreWeave against roughly $155 million and $126 million.

The gap is a currency story, and reading it as one about credit quality misses the mechanism. CoreWeave raised the money as debt because it had contracts and no equity market to speak of when it needed the capital most; it went public in March 2025 with the debt already in place. IREN and Nebius raised the money as equity and equity-linked paper because their share prices allowed it. The result is that CoreWeave's shareholders own a levered claim on the contracts and its lenders take the first loss, while IREN's and Nebius's shareholders have already taken the loss in dilution and hold an unlevered claim on the same assets.

Non-recourse structures move risk; they do not delete it. CoreWeave's $3,663 million of non-recourse debt and IREN's $938 million sit in vehicles whose lenders can take the GPUs and the contract but cannot reach the parent. That is genuinely protective for equity holders in a bad scenario, and it is exactly why lenders demand debt-service-coverage covenants and, in IREN's case, "restrictions on indebtedness and distributions" at the vehicle level. The parent keeps the downside optionality and gives up the cash-flow flexibility. In a normal cycle nobody notices. In a downcycle, cash trapped in a covenanted vehicle is cash the parent cannot use to fund the next build.

07The depreciation clock

How many years a GPU earns is the number underneath every argument about neocloud profitability. A company that buys a chip does not charge the whole purchase to one year. It spreads the cost over the years it expects the chip to earn, and each year's slice is depreciation. No cheque is written for it; the cash left when the hardware was bought. What makes it worth arguing about is that the assumed life is an estimate made by management. Stretch it by a year and every future depreciation charge shrinks and every future reported profit grows, with nothing at all happening inside the business. The three companies have picked different answers, and two of the three have moved their answer in the direction that raises reported profit.

CoreWeave depreciates technology equipment over six years. Its FY2025 10-K records that "effective January 1, 2023, the Company changed its estimate of the useful life for its computing equipment utilized in data centers from five to six years," citing "continuous advancements in hardware performance, software optimization, and data center design improvements." IREN depreciates GPU hardware over five years on a straight-line basisCharging the same amount every year of an asset's assumed life — a fifth of the cost a year over five years. The simplest method, and the one all three companies here use. It assumes a chip loses value at an even rate, which is a convention rather than an observation.. Nebius uses five years for server and network equipment, "up from the four-year useful life used prior to 2026" — a change made effective January 2026 that reduced first-half depreciation by $86.1 million and increased net income by $75.7 million.

The obvious reading is that CoreWeave is the aggressive one. That reading is too quick. CoreWeave made its change in January 2023, before the current cycle and before it was a public company with an earnings line to defend; Nebius made its change in January 2026, in the middle of the boom, and disclosed a benefit worth 18% of its half-year net income. Neither is evidence of bad faith. Both are evidence that the useful life of an AI accelerator is a judgment rather than a fact, and that three companies looking at the same hardware and the same customers reached three different judgments within a three-point range.

The argument that matters is physical rather than accounting. Michael Burry's November 2025 claim — that the realistic economic life is two to three years and that the industry's schedules understate depreciation by more than $176 billion across 2026–2028⁵⁰ — rests on the premise that Nvidia's annual product cadence makes older silicon uneconomic well before six years. The supporting evidence offered is resale: H100 systems reportedly trading at less than half the price of new ones by year three. Nvidia's stated counter is that "customers consistently use four-to-six-year depreciable lives based on observed utilization and longevity." CoreWeave's management gave the counter-evidence directly on the August call: the company remains "largely sold out of prior generations of NVIDIA GPUs in addition to the current SKUs," and describes a re-contractingLetting the same hardware again once its first contract expires. It is the moment the whole sector's economics get tested: the chips are older, the debt that bought them may be repaid, and the price is whatever the market will pay that day rather than what it paid years earlier. No large fleet has been through it yet. economics in which "when an initial contract ends, the cluster no longer has any leverage and we are free to recontract." That is a testable claim, and so far the test is passing — an operator with 1.5 GW live and a July price increase of about 25% is not a company struggling to fill old racks.

The answer sits between the two. The bear case is about the second half of the asset's life, and no operator has yet reached it at scale with Hopper- or Blackwell-class hardware. What can be said now is narrower: reported operating margins across these three are not comparable without the adjustment in Figure 15, and the group is one restatement away from re-ranking on GAAP profitability. Anyone ranking these businesses on reported operating income is measuring an accounting choice.

08The customer's alternative, and everyone's incentives

In March 2025 Microsoft chose not to exercise a nearly $12 billion option to buy more data-centre capacity from CoreWeave. Semafor reported it⁵¹ as "a sign that big tech companies are starting to right-size and tailor their AI budgets," and noted that OpenAI picked the contract up. CoreWeave said publicly that there had been no cancellations. Both can be true — an unexercised option is not a cancelled contract — and the episode shows what the customer's alternative actually is. Microsoft's alternative was to not buy. That is always available, it costs nothing, and it is the ceiling on every neocloud's pricing power at renewal.

The demand side has three tiers and they behave differently. HyperscalersThe handful of companies that operate computing at planetary scale and build their own data centres to do it — Amazon, Microsoft, Google, Meta. They are simultaneously the neoclouds' largest customers, their most credible competitors, and the benchmark their prices are measured against. Every contract in this brief is signed with a counterparty that could, given time, build the same capacity itself. buy neocloud capacity to cover a shortfall against their own build, which makes the contract a hedge on their own construction schedule. AI labs buy because they cannot build fast enough and are capital-constrained in a different way — OpenAI's roughly $22.4 billion of CoreWeave commitments exist because compute arrives faster from a specialist than from a construction programme. Enterprises are the newest tier and, on the disclosed pricing, the most attractive: shorter contracts at higher rates. CoreWeave's ability to sell into that tier depended on a financing structure that only arrived in August 2026.

Oracle closed its fiscal 2026 with remaining performance obligations that "ended the quarter at $638 billion, up 363% USD year-over-year"⁵² on cloud infrastructure revenue of $18.1 billion. That is six times CoreWeave's backlog, at a company with an investment-grade balance sheet and a decades-old enterprise sales channel. Any moat claim for a neocloud has to survive the observation that a general-purpose incumbent booked more AI infrastructure commitments in one year than all three of these companies have booked in their existence.

Street prices are the discipline on all of this, and what they show is dispersion rather than a single price. A street-price index dated 10 July 2026⁵⁴ puts H100 on-demand rental at a median of $4.29 per GPU-hour across seven providers, against $1.33 at the cheapest marketplace and $6.88 to $12.29 at the hyperscalers; B200 runs $2.71 at the cheapest marketplace, $3.50–$7.15 at specialist clouds and $14.00–$27.04 at hyperscalers. CoreWeave's own GB200 NVL72 pricing appears in the same index at $10.50 per GPU-hour for inference and $42.00 on-demand. Two things follow. Prior-generation hardware has not become worthless — a two-year-old H100 still clears above a dollar an hour, and CoreWeave says it is sold out of prior generations. And the spread between marketplace and hyperscaler pricing for the identical chip is roughly ten to one, which means the price of compute is mostly the price of everything around the chip.

The private peers show how contestable this is. Lambda reached $505 million of annualised revenue by May 2025 at a 61% gross margin and hired Morgan Stanley, J.P. Morgan and Citi to prepare a US listing⁵⁵; Crusoe has raised billions on the same story. Neither owns anything the listed three do not, and both buy the same GPUs. The barrier to entering this business is capital and power, in that order — and capital has been abundant.

09The criteria, and why they are weighted this way

The criteria come before the answers. Five of them carry weights, and the reasoning for each weight sits on the page where a reader can disagree with it.

10Head to head

Funding-model resilience — IREN, then Nebius, then CoreWeave

How much each company still needs from outside decides the ranking inside that rate gap. IREN's recourse debt is $6.75 billion of convertible notes at a 0.88% weighted coupon with maturities from 2029 to 2033; its cash interest bill last year was $46.6 million; its GPU financings are non-recourse and 90–96% funded by a combination of facility and prepayment. It could stop building tomorrow and service everything from the mining business alone. Nebius holds $8.0 billion of cash against $155 million of annual coupon and has demonstrated repeated access to the convertible market — but its $20–25 billion capex plan has no committed loan book behind it, and its equity-linked funding is priced off a share price that has ranged from $63.26 to $299.86 over the past year. CoreWeave's $3.6 billion of annual interest is contractual, senior, and does not shrink if a customer defers.

CoreWeave's answer to the leverage charge is stronger than the headline number suggests. Its debt is matched to contracts: the DDTL structures amortise against specific customer cash flows, and management describes each deployment as "fully repaying asset level debt used to fund the CapEx" so that "when an initial contract ends, the cluster no longer has any leverage." If that is true across the book, then the leverage is not a permanent claim on the enterprise but a self-liquidating claim on individual contracts — which is a fundamentally different risk. The evidence for it is the A3 rating on the DDTL 4.0 vehicle and the fact that CoreWeave has refinanced at successively lower rates for a year. The evidence against it is that gross debt rose from $20.7 billion to $31.4 billion of recourse debt in six months while the amortisation was supposedly happening. Both can be true in a company growing this fast. The resolution is arithmetic: the structure is self-liquidating only if new builds stop, and stopping is the one thing a company with $35.5 billion of uncommenced leases cannot easily do.

Contracted economics — CoreWeave, then Nebius, then IREN

On this criterion the order reverses cleanly. CoreWeave has $104 billion of backlog with more than half attached to capacity where delivery has commenced, a customer set that has broadened from one name at 67% to three names at 72%, a Meta contract running to December 2032, and a $6.3 billion NVIDIA obligation to buy residual unsold capacity through April 2032 — a floor no competitor has. Nebius has $37.5 billion of RPO plus more than $40 billion of stated customer commitments, two hyperscale anchors and 41% of revenue outside the top three. IREN has roughly $16.5 billion across two accounting regimes, one dominant customer, and — on the disclosed per-megawatt figures — the worst price in the group on its largest contract.

Size is the easy half. The harder question is when the book turns into revenue, and all three disclose enough to answer it.

CoreWeave's book is the longest-dated, which is protection and exposure at once: 39% of its RPO converts in the four years beginning July 2028, so a material fraction of the value depends on demand conditions in 2030 and beyond. Nebius's Microsoft contract carries a disclosed customer termination right on missed delivery, and its Meta capacity does not come online until early 2027. IREN's book is smallest but has the highest disclosed prepayment intensity — 45–55% of GPU capex arriving as cash before the service is delivered, plus 20% of the Microsoft contract paid up front — which is the most protective contract feature available in this market and the one least visible in a backlog table.

Unit economics and cost position — IREN, then Nebius, then CoreWeave

IREN's AI Cloud segment ran at 86.9% gross margin before depreciation with electricity at 3.5% of segment revenue. Nebius's group figure on the same basis was 77.1% in the June quarter with an AI-cloud adjusted EBITDA margin of 49.7%. CoreWeave reported 59% adjusted EBITDA and a 5% adjusted operating margin, on a cost-of-revenue definition that includes depreciation of power infrastructure and is therefore not directly comparable at the gross line.

Capex intensity sharpens the same ordering differently. CoreWeave is spending 2.9 times its guided 2026 revenue; Nebius roughly 7 times. IREN gives no calendar capex guidance. Its $13.8 billion of commitments against a >$4 billion exit-ARR target implies a similar multiple to CoreWeave's. High capex intensity is what a land-grab looks like. It also means every one of these companies is valued on capacity it has not yet built, and Nebius most of all.

Moat mechanism and optionality

Both are argued in full in the next section. In summary: IREN holds the only asset in the group whose replacement cost has risen materially since it was acquired; Nebius holds the only software layer with independent evidence of quality and the only equity stakes worth a material fraction of its market value; CoreWeave holds the only scale advantage, which is real, and the only supplier-backstopped revenue floor, which is a contract rather than a moat.

11The moat question

The strongest case against everything in this report goes roughly as follows, and it is not a straw man.

There is no moat here and there was never going to be one. Three companies buy identical hardware from a monopoly supplier at prices that supplier sets. They sell it to four or five buyers who could build the same capacity themselves and are visibly doing so. The contracts that look like moats are hedges the buyers take against their own construction schedules, priced by the buyers, terminable by the buyers if delivery slips. The financing that looks sophisticated is a mechanism for transferring the buyers' credit quality to the seller's balance sheet, which means the value being financed belongs to Microsoft and Meta rather than to CoreWeave or IREN. Power is a genuine constraint but it is a constraint on everyone equally, and permitting queues clear; the same argument was made about fibre in 1999 and about data-centre land in 2007. On this reading the correct number of neoclouds in 2032 is one or two, they will be owned by their largest customers or by an incumbent with a real balance sheet, and today's contracted backlogs are the high-water mark of a period when compute was scarce.

That case has real evidence behind it. Microsoft did decline a $12 billion expansion. Oracle booked $638 billion of RPO in a single fiscal year without owning a single differentiated thing about GPU rental. On-demand H100 prices have fallen through the asset's life exactly as a commoditising input should. Lambda and Crusoe can raise billions on the same story. And the most-quoted proof of the sector's maturity — the A3-rated GPU financing — is a rating on a ring-fenced vehicle holding one customer's contract.

Where the case breaks down is on the word "equally." Power is a constraint on everyone, but not at the same price, and the difference is not small.

What each company would actually make a competitor replicate

CoreWeave: scale, and the financing structures scale unlocked. A competitor would have to reach 1.5 GW of energised capacity, more than a dozen live financing instruments, an investment-grade-rated GPU securitisationPackaging a stream of contracted payments, and the assets that produce them, into a separate vehicle and selling bonds backed by that stream. It is how mortgages, car loans and aircraft leases have long been financed. Applying it to GPUs is new, and it works only for as long as lenders trust their estimate of what the hardware will be worth at the end., and a relationship with NVIDIA that produced both a $4.7 billion equity position and a $6.3 billion residual-capacity backstop. Time to replicate: three to four years and roughly $35 billion, assuming the capital markets stay open. That is a formidable barrier and it is a barrier of capital, not of right. It compounds while the sector is short of capacity and erodes the moment it is not. On the strict test — could a competitor buy this at today's prices — the answer is yes, expensively. Crusoe is trying.

Nebius: the software layer, and the multi-tenant architecture underneath it. A competitor would have to build an orchestration and inference platform that hundreds of small customers choose over a hyperscaler's. There is independent evidence this exists: Nebius reports full accuracy parity on GLM-5.2 in Artificial Analysis's Endpoint Accuracy Index while ranking among the fastest providers, and inference workloads on its Token Factory more than tripled in a quarter. Time to replicate: two to three years of engineering, which is cheap relative to $20 billion of capex. The harder question is whether the layer can hold margin against AWS, Azure and Google, who have the same software and better distribution. Nebius's own answer — that it serves "the full workload lifecycle on one platform" where standalone inference providers do not — is a product argument, not yet a pricing-power argument. The 41% of revenue outside its top three customers is the best available evidence that it is winning small customers on something other than price, and it is indirect.

IREN: grid positions bought when nobody wanted them. This is the one that survives the test. IREN's Childress, Mackenzie, Prince George and Canal Flats sites were acquired to mine bitcoin, in an era when a data-centre developer would not have bid for them. The company now runs AI capacity at those sites where electricity is 3.5% of revenue. A competitor starting today cannot buy 2022 grid positions at 2022 prices; it can only join an interconnection queue, and queue position is the one input in this business that money does not accelerate. The queue is a physical fact rather than a figure of speech. Anyone wanting to draw large power from a grid applies to the utility, which studies whether the local network can carry the load and generally has to reinforce it before the connection can be granted. Those studies and that construction run on their own clock, and applications are worked in the order they arrive. A buyer can pay more for the land. There is no price at which the utility moves the project up the line. That is a genuine, mechanism-level advantage, and it is the only one in the group.

Three qualifications belong with it. IREN does not own generation — it buys power under regulated tariffs in British Columbia and Oklahoma and at ERCOT market prices in Texas, so the advantage is a position in a queue and a low-cost region rather than a physical hedge. The advantage is a spread, and it narrows if wholesale power prices fall or if AI compute prices fall faster than power prices. And it applies to a portfolio that is still overwhelmingly unbuilt: 5 GW secured against roughly 40 MW of AI capacity operating. An advantage in an asset you have not yet developed is an option, not an annuity.

The debates, argued on their merits

Are take-or-pay contracts from self-building customers durable demand or a bridge? The evidence says bridge, with a long deck. Every anchor is expanding its own capacity; Microsoft has already declined one expansion; the Nebius disclosure shows the customer holds the termination right on delivery failure, not the reverse. But a bridge with a 2032 end date and a prepayment attached is a very good asset, and the renewal question is genuinely open rather than settled. What would resolve it: a single disclosed renewal, at term, at a stated price, by any of the three. None exists yet. That absence is the largest unanswered question in the sector and it will start answering itself around 2028.

Is owned power a moat or only a cost advantage that matters at commodity prices? Both, and the distinction is about when. At today's prices, where IREN sells a megawatt-year for above $20 million and pays 3.5% of that for electricity, power cost is nearly irrelevant to margin. If contract prices halved, power cost would become the difference between a viable and an unviable megawatt. Owned cheap power is therefore insurance that pays out precisely in the scenario that threatens the sector — which is the definition of a valuable but currently invisible asset.

Can a software layer hold margin above rented hardware? Nobody has shown it yet at scale. Nebius's 49.7% AI-cloud EBITDA margin is real but is earned on scarce hardware in a supply-constrained market, and CoreWeave earns 59% with a different software story. The test comes when capacity is abundant and the two companies' prices diverge or do not. Nebius's own strongest evidence is not its margin but its customer mix: 41% of revenue from outside its three largest accounts, in a business where CoreWeave's equivalent is 28%.

Is GPU-backed leverage intelligent structuring or the sector's fault line? It is intelligent structuring of an asset whose residual value nobody has yet observed. The A3 rating is real; so is the fact that it depends on collateral valuations for hardware that has never been sold second-hand at scale after a full contract term. The private-credit question is not whether the first-lien lender gets paid — the contract cash flow covers it — but what happens to the market's appetite the first time an SPV has to remarket a fleet mid-life. That has not happened.

Alternatives considered and rejected

Three framings were tested and set aside. Ranking by backlog was rejected because the three books are measured under different accounting regimes and the scope error is larger than the differences it would purport to show. Ranking by reported operating margin was rejected because the depreciation policies differ by up to two years of useful life, worth roughly $950 million a year of pre-tax income to CoreWeave alone. Ranking by revenue per megawatt was rejected as a primary measure because CoreWeave does not disclose it, because rack density differs by generation and cooling type — IREN's own air-cooled NVIDIA contract prices at $11.3 million per MW against liquid-cooled deals above $20 million — and because contract vintage dominates the variation. Each of these reopens if the underlying disclosure changes: a common useful life across the three, or per-megawatt pricing from CoreWeave, would make the second and third framings usable.

Second- and third-order effects

A GPU glut does not hit these three equally, and the transmission mechanism is not the one people expect. The channel is not utilisation — all three are substantially contracted for two years — it is the renewal price against a fixed capital cost. CoreWeave's fixed cost is $3.6 billion of cash interest plus lease payments on sixteen-year buildings; a re-contracting at 70% of current pricing leaves it covering interest and little else. IREN's fixed cost is $126 million of interest today and roughly $534 million once its GPU facilities fully draw; the same 30% price cut leaves it comfortably covered. Nebius sits closer to IREN. The order in which these machines break under price pressure is the reverse of the order of their contracted books.

An NVIDIA architecture shift produces a subtler asymmetry. Fleets age on different clocks: CoreWeave's is the oldest and largest, IREN's is almost entirely 2026 vintage, Nebius's is in between and it has already taken delivery of Vera Rubin NVL72 systems for validation. A generational reset therefore lands hardest on the company with the most depreciated-but-not-written-down silicon — and that company also has the longest assumed useful life. Those two exposures compound rather than offset. The third-order effect is on the lenders: an architecture shift that impairs residual values would reprice GPU-backed credit for everyone, including the two companies that barely use it, because the sector's marginal financing cost is set by the marginal lender's view of collateral.

If Nebius's asset-light model works, the most capital-efficient way to be a neocloud stops being to own the fleet. That would devalue exactly the thing CoreWeave has spent $35 billion assembling and would leave IREN's position — owning the scarce input rather than the depreciating one — looking the most durable of the three. It is early, it is unquantified, and it is the scenario least priced into any of the three narratives.

12The downcycle test — who breaks first

The useful stress on this sector is a price shock at re-contracting, against a capital structure that does not reprice. A demand collapse is the wrong test: all three are contracted two years deep, and a demand shock takes that long to reach them. So the test below holds each company's disclosed adjusted-EBITDA margin constant, cuts revenue to 75% and then 50% of its current run rate, and measures the result against contractual cash interest. What that produces is an interest-coverage ratio: earnings divided by the interest owed on the debt. Cover of two times means a business earns twice what it must hand to its lenders. Below one it is not earning its interest bill at all and has to find the shortfall somewhere else — cash reserves, fresh borrowing, or new shares — which is the point at which a bad year stops being a profit problem.

Two caveats keep this from being a verdict. The first is timing: CoreWeave's book converts over years, so a price shock reaches it gradually, and the company has already demonstrated it can refinance downward — 300 basis points in a year, worth $1.1 billion annualised. Repeat that twice more and the base-case coverage improves materially without any change in revenue. The second is that the scenario tested is severe. A 50% cut in re-contracting prices has not happened in this market; the observed move in the last twelve months has been upward, with CoreWeave raising list prices about 25% in July 2026 and Nebius reporting more than 30% higher pricing on older-generation GPUs than in the prior quarter.

The table does not establish that CoreWeave will break. It establishes that CoreWeave is the only one of the three for which a price shock is a solvency question rather than a profitability one, and that the difference comes from the funding decision, not from the quality of the business.

13The ranking

The ranking, stated plainly: IREN first, Nebius a very close second, CoreWeave third. IREN and Nebius should be read as jointly ranked; the evidence does not separate them and the tiebreak is a judgment about what kind of advantage survives a bad year. IREN's cheap grid positions do not depend on anyone's willingness to buy its shares. Nebius's funding menu does. That is the whole of the difference between first and second, and a reader who thinks equity-market access is more durable than an interconnection queue should swap them.

CoreWeave's third place needs defending, because on the criterion most people use — contracted revenue — it wins outright, and this report has argued that contracted cash flow is the asset. The defence is that the ranking grades durability, and CoreWeave's contract book is financed in a way that converts a bad year into a solvency problem rather than a bad year. It has the largest book, the most delivered capacity, the strongest supplier relationship and the only investment-grade GPU financing in existence. It also pays 9.00% on $35.6 billion, carries $16.3 billion of lease liabilities with another $35.5 billion queued, and covers cash interest 1.69 times at its current run rate before any lease payment. Rank it first on scale and execution; the evidence does not support ranking it first on durability.

Verdicts are structural assessments of business durability. They are not investment recommendations and carry no view on any security's price. Confidence is sized to the disclosure available: all three are rated Moderate because none of the three has yet been observed through a full contract renewal cycle.

14Predictions, falsifiers and what to watch

Nine claims below are dated and checkable. Each can be scored against a disclosure that will exist by the stated date, and each is wrong if the stated observable does not occur.

Dates as disclosed in each issuer's most recent results materials: IREN FY26 release⁵, CoreWeave Q2 2026 call¹⁰, Nebius Q2 2026 shareholder letter⁴, CoreWeave–Meta release²⁸.

Every argument in this report treats the three as competitors, and on any given contract they are. But the constraint they share is more important than the differences that separate them: none of the three sets the price of its input, none controls the architecture cadence that resets its fleet, and none has a customer whose alternative is worse than building. What they are competing for is the right to carry other people's capital risk in exchange for a spread. The best business in that trade is the one with the cheapest capital and the scarcest input. On today's evidence that is IREN, narrowly, and the case is not yet proven.

§Evidence register

Every source below was opened and read in the preparation of this brief. Tier 1 = issuer filings and regulator submissions. Tier 2 = issuer-hosted transcripts, shareholder letters and press releases. Tier 3 = single-source, aggregated or estimated material, used only where flagged in the text. Dates are filing or publication dates; event dates are stated separately in the prose where they differ.

§Not verified, and not load-bearing

The following could not be verified from a primary source in this work and therefore carries no conclusion anywhere in this brief. Each is listed with what was tried.

• The identity of CoreWeave's largest customer. Its filings identify customers only by letter. The S-1 names Microsoft among its customers without a percentage. The widely repeated attribution of the 67% FY2025 concentration to Microsoft is not disclosure and is not relied on here; every concentration statement in this brief uses the letters the filings use.
• CoreWeave's corporate issuer credit ratings. S&P's rating page returned HTTP 403 and no primary rating action from Fitch or S&P could be opened. The DDTL 4.0 vehicle rating is now directly supported by a retrievable Morningstar DBRS action plus CoreWeave's own release. Statements about CoreWeave's cost of unsecured capital rest on coupons in its debt schedule, not on a corporate rating.
• Microsoft's aggregate spending with neocloud providers. The trade report of a figure above $60 billion could not be opened (HTTP 403) and no counterparty disclosure quantifies it. No claim in this brief depends on it.
• The upfront payment inside the Nebius–Microsoft contract. Secondary sources report roughly $6.96 billion of aggregate upfront payments. The 6-K disclosing the contract does not state an upfront amount and the figure is not used. Nebius's prepayment intensity is instead taken from its own shareholder letter.
• Nebius's live, energised capacity in megawatts. Nebius discloses contracted power only. The "live power" row in Figure 2 is marked undisclosed rather than estimated, and no comparison of live capacity across the three is attempted.
• The gap between Nebius's convertible principal and its carrying value. The six disclosed series total $8,087.5m of principal against $8,545.7m of debt on the balance sheet at 30 June 2026. The interim statements do not itemise the difference. The weighted-average coupon is computed on principal and labelled as such.
• Revenue per GPU per hour at any of the three. None discloses GPU counts alongside revenue, so a per-GPU-hour rate cannot be derived from the filings. The observable rental prices in section 08 come from a third-party street-price index and describe the market, not these three companies' realised rates. CoreWeave's own GB200 list prices appear in that index and are the only issuer-attributable hourly figures used anywhere here.
• Cost per contracted megawatt at CoreWeave and Nebius. Neither publishes capacity delivered against capital spent in a way that supports the calculation; CoreWeave's 2026 capex funds capacity delivering well beyond 2026. The only cost-per-megawatt figure computed in this brief is IREN's $29.0m of GPU capital cost per MW on the Microsoft deployment, which is derivable because both the Dell purchase price and the megawatts are disclosed in the same release.
• Third-party estimates of IREN's funding gap. A consultancy estimate of roughly $21bn circulates in trade coverage; no primary source for it was opened and it is not used. The funding-gap discussion in section 04 uses IREN's own commitments note instead.
• Crusoe's revenue and valuation. Discussed only in aggregate as evidence that private competitors are well capitalised; no figure is used.
• Termination and renewal mechanics in the CoreWeave–OpenAI, CoreWeave–Meta and IREN–Microsoft contracts. None is publicly disclosed. Where this brief discusses termination rights it cites the one contract whose mechanics are disclosed — Nebius's Microsoft agreement — and does not generalise beyond it without saying so.
• The economic useful life of an AI accelerator. No operator has yet run a fleet through a full contract term and remarketed it. Every statement about depreciation in this brief is about accounting policy and its arithmetic effect, not about physical life.

AIR Research. This document is an internal structural assessment of business quality and competitive position. It is not investment advice, not a recommendation, and not a solicitation. The ranking grades the durability of three business models on the evidence available at the cutoff date; it expresses no view on any security's price. Derived figures are computed from the cited disclosures and their method is shown in the relevant figure caption.