Who pays for the AI boom? $CRWV, $NBIS, $IREN and $ORCL from contracts to cash

01The missing link between an AI contract and shareholder value

A customer can need more computing power, sign a substantial contract and still contribute very little revenue to its supplier in the current quarter. The supplier may first need to arrange financing, complete a building, connect electricity, install accelerators, test the cluster and obtain customer acceptance. Cash can move before any of those steps, through a deposit, or considerably later, after an invoice becomes due. A convincing analysis follows all three timelines: the commercial agreement, the physical delivery and the money.

That distinction makes CoreWeave, Nebius, IREN and Oracle a useful comparison. They participate in the same expansion of AI infrastructure, but they do not offer identical exposures. CoreWeave makes financing and the conversion of a large contracted book central to the story. Nebius combines a rapidly expanding AI cloud with a software and engineering proposition. IREN brings an electricity and data-center development base to its transition toward AI services. Oracle adds a substantial existing software and cloud business, so consolidated cash generation cannot automatically be attributed to newly built AI capacity.

The immediate news hook is CoreWeave’s financing completed on September 22, 2026. Its final size was $4.2 billion, including the fully exercised purchaser option. The proposed transaction and the subsequently priced base offering were earlier states of the same financing. A reader comparing a headline from one day with a balance-sheet discussion from another can therefore reach the wrong conclusion without either headline having been false when published. The completed transaction is documented in the September 22 Form 8-K.

The broader question is how additional capital changes the ability to serve customers and the claim that existing shareholders retain on the result. A new financing can remove a near-term construction constraint while adding interest, refinancing obligations or possible dilution. A customer prepayment can reduce the cash needed from lenders while committing the operator to provide future service. The value of these arrangements depends on their terms and on execution; a large announced amount alone does not settle the question.

02A translation guide for the numbers that circulate on social media

The first step in comparing these companies is to preserve each metric’s definition. Revenue measures activity recognized over a stated accounting period. A contractual balance measures a set of commitments at a point in time. An annualized run rate describes the pace of a selected recent period or an expected future period. These quantities answer different questions, even when all are expressed in dollars.

This is an analytical comparison framework, not a claim that the four companies define these measures identically. Their specific disclosures appear in the company sections. If a reporting definition changes, the historical series needs reconciliation before a growth rate is calculated. A larger number can reflect new contracts, a longer measurement horizon or a broader perimeter rather than a comparable increase in near-term production.

Timing also changes the meaning of a seemingly simple ratio. Dividing a multiyear contractual balance by one quarter of revenue annualized produces a rough scale comparison, not a contract duration or an assured number of years of sales. Customers can ramp at different speeds; agreements can start in different years. A useful comparison therefore places the anticipated recognition schedule alongside the funding and commissioning schedule instead of turning all three into a single multiple.

03Electricity is a delivery condition, not just a capacity headline

The financial model starts operating only when the physical system works. A site with a prospective power allocation is different from a building with an energized connection. An energized building is different from a commissioned cluster, and a commissioned cluster is different from a customer-accepted service earning revenue. The weakest step can hold up the whole chain even when the other steps are complete.

The IEA’s Electricity 2026 report explains the mismatch in infrastructure timelines: planning and completing grid infrastructure can require five to fifteen years, whereas new data centers can be built in one to three years. These are broad ranges, not completion forecasts for any of the four companies. Their relevance is that obtaining a site does not remove the need to examine the actual connection and reinforcement work. IEA, Electricity 2026: Grids.

For a project-level comparison, the useful questions are concrete. Which entity controls the connection? Is the capacity firm or subject to curtailment? Which construction milestones remain? Who pays if energization slips? Does customer billing begin when equipment is delivered, when the service is accepted, or when reserved capacity becomes available under the agreement? Where terms are not public, that uncertainty belongs in the analysis rather than being filled with an industry average.

An announced power portfolio may still be valuable because it provides development options. The analytical mistake is to price every option as a completed project. Conversely, ignoring all undeveloped capacity would miss the possibility that a well-positioned site can shorten the next expansion. The appropriate distinction is between operating capacity, funded construction and development opportunities, with a different degree of execution uncertainty attached to each.

04CoreWeave: the funding bridge behind a $104 billion backlog

CoreWeave’s investment case increasingly depends on the relationship between three calendars: customer deployments, infrastructure payments and financing maturities. A larger order book helps secure funding, but shareholders benefit only if the resulting services earn enough after the infrastructure and financing costs. September’s financing announcements make that distinction especially relevant: the company has increased both its resources and its potential claims on future cash flow.

The operating business is growing; financing still changes the result

CoreWeave reported the following figures for the quarter ended June 30, 2026. Adjusted EBITDA excludes expenses that remain relevant to shareholders, so its margin should be read alongside the GAAP figures. CoreWeave results, August 11, 2026.

This combination explains why revenue acceleration alone cannot settle the equity debate. The company can improve the economics of delivering compute while financing costs absorb much of that progress below operating income. Conversely, a GAAP loss does not reveal which newly deployed clusters are attractive or unattractive: consolidated results combine installations at different stages and capital raised at different costs. The useful question is whether mature deployments ultimately finance their own replacement and leave a residual for shareholders.

CoreWeave’s 2025 annual report assigns technology equipment a six-year useful life. That accounting estimate allocates historical cost; it does not promise six years of unchanged rental prices. An older GPU can remain useful while earning less than a newer configuration. Whether that is acceptable depends on the purchase price, remaining debt and replacement demand. Depreciation and commercial obsolescence therefore need to be examined together, without assuming that either a newer chip launch or a long accounting life automatically establishes an impairment. Annual-report PP&E schedule.

A contract book with a delivery timetable and concentrated customers

At June 30, remaining performance obligations were $103.7 billion: 41% was expected to become revenue within 24 months, another 39% during months 25–48. Customers labeled A, B and C represented 36%, 26% and 10% of quarterly revenue. The filing does not identify those labels here. CoreWeave Q2 Form 10-Q.

That timetable is more useful than treating the entire balance as near-term sales. Equipment ordered today may support a contract whose revenue continues well beyond the initial installation. The financing bridge must survive the interval between those dates. Customer concentration also has two sides: large commitments can support underwriting and deployment planning, while a dispute or delayed rollout involving one major customer can affect several parts of the business simultaneously.

The approximately $104 billion headline backlog uses CoreWeave’s broader company definition, including estimated future amounts under committed contracts subject to delivery and availability. Its proximity to reported RPO does not make the labels interchangeable. Readers should track both definitions consistently across reporting periods. Q2 backlog definition, August 11.

Even the 24-month recognition bucket is not a quarterly revenue schedule. It does not specify equal installments or disclose how much depends on infrastructure that still needs to be completed. A credible financial model would preserve that uncertainty instead of dividing the amount by eight quarters and presenting the result as contracted quarterly revenue. The same discipline applies to customer diversification: several new customer announcements can broaden the pipeline without immediately reducing the revenue share of the largest existing accounts. The relevant comparison is the subsequent financial disclosure, using the same revenue basis and period.

CoreWeave subsequently disclosed more than $25 billion of net new customer commitments early in Q3. Simply adding those commitments to June backlog would omit intervening recognized revenue and any other adjustments. It would manufacture a September balance that the company did not report. September 17 commercial update.

Active capacity and short contracts answer different questions

June’s active power was 1.5 GW. Contracted power subsequently reached approximately 4.2 GW as of August 11. The latter is a development pipeline, not a statement that every site already contains revenue-producing equipment. The dates also differ, preventing a precise contemporaneous utilization calculation. June operating figures; September update confirming the August figure.

The September update additionally describes three-to-six-month contracts priced at approximately $40 million per MW on an annualized basis. That denominator measures power required for the relevant clusters. It is neither a portfolio-wide average nor the revenue collected over each short contract’s life. Multiplying it by all contracted power would combine selected pricing, different hardware and unfinished capacity into an unsupported forecast.

Commercially, short contracts can monetize urgent demand and retain repricing flexibility. They also require another sale sooner. A high annualized price is most valuable when the operator can sustain occupancy between customers without expensive reconfiguration or idle periods. The next evidence to seek is repeat demand for the same deployed equipment, not merely another attractive quotation for scarce new capacity.

Cash generation does not yet cover infrastructure payments

At June 30, cash was $5.524 billion; restricted cash was separate. Reported liquidity included $10.014 billion of undrawn facilities. Debt carrying amounts totaled $35.068 billion. First-half operating cash flow was $3.663 billion, against $14.117 billion of cash PP&E and capitalized-software purchases. Q2 Form 10-Q, balance sheet and cash-flow statement.

The difference is a $10.454 billion cash investment gap, a Merlintrader subtraction rather than company-defined free cash flow. It excludes other investing activities and should not be mistaken for a complete financing forecast. It nevertheless explains why a business generating positive operating cash flow keeps returning to funding markets.

Undrawn lending capacity has conditions and eventually creates debt when used. Restricted cash may secure specific obligations. Neither deserves the same treatment as unrestricted cash available for any corporate purpose. Similarly, non-recourse borrowing remains a claim against the financed project even when creditors’ access to the wider group is limited. Equity ultimately receives what remains after those project obligations, rather than the project’s gross operating surplus.

A particularly useful test comes from the $2.6 billion DDTL 5.5 facility announced August 10: approximately five-year financing, priced at SOFR plus 5.50%, supports contracts averaging approximately three years. The facility permits renewals and re-leasing under its conditions. Its economics therefore require attention to the period after the first customer contract ends; successful initial deployment does not remove renewal risk. Closing a delayed-draw facility also does not establish that every dollar has already been borrowed. CoreWeave financing announcement.

September’s convertible and ATM change the per-share equation

The convertible offering closed September 22 at $4.2 billion, including the fully exercised $500 million option. It carries a 2.875% coupon and matures April 1, 2033. Initial conversion terms imply approximately $97.85 per share; conversion may be settled in cash, shares or a combination under the terms. That conversion price is a financing parameter, not a valuation target. Closing Form 8-K.

Net proceeds after purchaser discounts, before other expenses, were $4.137 billion. Capped calls cost $566.2 million, leaving approximately $3.571 billion before further expenses by subtraction. The hedge has a $199.70 cap. It can reduce dilution or excess cash settlement within its structure; it does not make the financing free or protect shareholders without limit. Same closing filing.

Separately, the September 17 ATM program authorizes up to 35 million shares. Authorization is not evidence that the shares have been sold. Its possible collared-forward structure also separates market sales by a forward seller from the company’s later receipt of settlement proceeds. ATM announcement; distribution-agreement Form 8-K.

These mechanisms create different tradeoffs. Equity issuance can reduce reliance on borrowing, but increases the ownership denominator. A convertible postpones the choice between repayment and potential dilution, while its coupon still consumes cash. The meaningful assessment is whether the assets financed generate a sufficient return for the remaining ownership interest. Higher enterprise revenue does not automatically mean higher value per existing share.

The September hedge expenditure also makes gross proceeds a poor shorthand for deployable construction funding. Part of the financing purchases protection for the equity structure rather than another GPU cluster. That may be a sensible allocation, but a funding model must preserve the distinction. Counting the entire principal as infrastructure purchasing power would overstate the amount left for the operating plan before any other transaction expenses.

For the next reporting period, the clearest evidence would combine delivered capacity, the pace of revenue recognition, cash investment, new borrowings and actual share issuance. A larger backlog with slower delivery or deteriorating funding terms would weaken the conversion from demand to shareholder economics. Faster delivery with controlled costs and transparent financing would strengthen it. None of those outcomes can be established from the gross amount of September’s fundraising alone.

05Nebius: delivery progress, customer advances and the terms beneath megadeals

Nebius presents a different analytical challenge: rapid cloud expansion sits alongside a changing mix of dedicated customer infrastructure, shared cloud capacity and partner-owned assets. These arrangements can all support growth, but they distribute capital requirements and commercial risks differently. Reading them as one undifferentiated collection of multibillion-dollar contracts would obscure the part that matters to an NBIS shareholder.

Microsoft delivery is a milestone, not five years of immediate revenue

The original Microsoft arrangement disclosed in September 2025 specified $17.4 billion, potentially $19.4 billion with additional capacity, subject to delivery and service availability. It included remedies for deployment delays. In its August 12, 2026 shareholder letter, Nebius reported delivery of all capacity tranches. Original Microsoft Form 6-K; Q2 shareholder letter.

That is meaningful progress from a construction commitment toward a service business. It does not make the full contract value revenue in the delivery quarter. The relevant follow-through is reliable operation over the service period, collection under the payment schedule and sufficient cash generation after the associated infrastructure obligations. The delivery statement is Nebius’s disclosure; a separate public Microsoft confirmation of every tranche was not established in this research.

For the equity thesis, delivery also changes the risk being evaluated. Before deployment, attention centers on schedule and construction funding. Afterwards, attention shifts toward uptime, service economics and the amount of capital needed elsewhere in the expansion plan. A completed customer rollout can lower one project’s uncertainty without making the whole company’s future construction budget self-financing.

Meta’s $27 billion headline contains two different arrangements

The March 16 agreement includes $12 billion of dedicated capacity, scheduled to begin coming online in early 2027, plus up to $15 billion for remaining capacity on certain upcoming clusters. Nebius intends to sell that second pool to third parties, with Meta purchasing the residual. Nebius–Meta agreement, March 16, 2026.

The distinction is economically important. The dedicated component supports a specified customer buildout. The residual-capacity arrangement supports the broader cloud business by providing another outlet for qualifying capacity. Those are different routes to monetization, with different implications for customer mix and sales flexibility.

A model should not count a third-party sale and Meta’s purchase of the same residual capacity as two independent revenues. Nor should it treat the maximum combined headline as cash already collected. The practical questions concern which clusters qualify, when they become available and how much of their capacity is ultimately sold through each channel. The announcement does not publish enough contractual detail to calculate the final margin for either route.

This structure can be attractive even when not all revenue ultimately comes from Meta: selling capacity through the ordinary cloud business may itself be the intended outcome. The important test is total realized economics for the shared pool, rather than insisting that the maximum headline must appear as recognized sales to one customer.

It also changes how diversification should be interpreted. A larger third-party customer base can reduce dependence on one buyer in the eventual revenue mix while the underlying development still benefits from that buyer’s contractual support. Conversely, a greater residual purchase by Meta may support utilization while leaving sales more concentrated. Those are alternative outcomes within the arrangement, not evidence that its maximum value must be incremental to every other customer commitment announced for the same infrastructure.

The June financial baseline is smaller than the headline contract book

The latest completed-quarter figures provide the baseline against which those future arrangements should be judged. Nebius Q2 results, August 12, 2026.

Core AI cloud revenue was $575 million in Q2, while the reported ARR was $3 billion. Here ARR annualizes the final month’s cloud revenue; it is not annual recurring revenue or a completed year’s sales. Q2 shareholder letter.

For a company adding capacity quickly, the exit rate can help reveal momentum that a quarterly average understates. It still assumes that the latest month’s activity is a useful starting point. It does not prove a full year of future utilization, unchanging prices or immediate access to the cash implied by the annualization. Comparing ARR directly with another issuer’s realized annual revenue would disguise that difference.

Customer advances finance construction but also create obligations

Nebius’s first-half cash-flow statement records a $4.395 billion increase in deferred revenue. Subtracting its cash capital purchases from operating cash flow produces a $3.626 billion investment gap, using Merlintrader’s simple calculation rather than an issuer-defined free-cash-flow measure. Q2 financial statements.

Customer advances can materially improve funding economics: capital arrives before the full service obligation is discharged. But their appearance in operating cash flow does not mean that the corresponding service has already been delivered or that the funding can recur indefinitely at the same pace. An expanding contract book can generate substantial advances during construction, followed by periods when revenue is recognized against amounts collected earlier.

The cash-flow analysis therefore needs to follow each deployment beyond the initial advance. If subsequent collections fund operations, debt service and replacement investment, the arrangement can produce a durable residual. If construction consumes the advance and the remaining receipts are thin relative to the obligations, strong initial operating cash flow will have overstated the ease of financing continued growth. The disclosed consolidated figures do not provide enough detail to calculate that return project by project.

The August notes cost more than their coupons suggest

Nebius closed $5.75 billion of convertible notes on August 24: $3.45 billion carrying a 0.50% coupon and $2.30 billion carrying 4.50%. The respective maturities are in February 2030 and February 2034. Crucially, principal accretes to 110% and 125% of its original amount if outstanding to maturity. Closing announcement; August 19 terms.

Applying those percentages gives a combined $6.670 billion repayment amount across the two maturities, before coupons, assuming no earlier conversion, redemption or repurchase. The $920 million increase is a financing obligation that a comparison of coupon rates alone would miss. It is not an annual interest rate and should not be represented as one.

Nebius also disclosed agreements to exchange $800 million of older notes for approximately 15.8 million shares. That is a separate exchange arrangement, not fresh cash raised. August 24 disclosure.

A July secured facility adds another funding model: approximately $775 million against deployed GPUs and contracted investment-grade customer cash flows, at SOFR plus 2.50%, maturing October 31, 2030. Nebius attributes capital coverage to the facility together with customer cash flows, not the loan alone. July 17 financing announcement.

These post-quarter transactions cannot simply be added to June cash to declare a current balance. Construction spending, drawdowns, repayments and transaction costs also move. The next financial statements must reconcile them. The distinction matters particularly when a funding announcement arrives while several large deployments are consuming cash simultaneously.

Partner-owned infrastructure and Palantir offer different forms of expansion

In July, Nebius introduced arrangements in which partners finance and own facilities and hardware, while Nebius provides architecture, software, commercial access and service responsibilities. Initial arrangements were disclosed, with several possible economic structures rather than one published standard margin. Infrastructure-partner model, July 15.

This can reduce the incremental capital required for the affected capacity. It does not erase the investment budget of Nebius’s own developments. Nor does it eliminate execution risk: a partner that cannot finance or operate its facility reliably can still damage the customer experience delivered under the Nebius platform. The evidence needed is the amount of capacity deployed through partners, its contribution to earnings and the obligations retained by Nebius.

September’s Palantir partnership is another distinct channel. Nebius was designated a preferred sovereign-AI infrastructure partner, with eligible customer access following integration. The joint announcement did not disclose a minimum purchase, contract value or recognized revenue. Palantir–Nebius announcement, September 8.

That opens a route to customers with demanding infrastructure and governance requirements. Commercial success still requires integration, actual purchases and adequate infrastructure in the relevant locations. Preferred status should not be converted into exclusivity or an immediate addition to contracted sales.

Nebius’s stated year-end target of 5 GW of contracted power remains a development target. August shareholder letter. For the next assessment, the strongest evidence would link that pipeline to delivered service, subsequent collections and a manageable funding burden per share. Delayed installations, dependence on repeated advances or rising ownership dilution could weaken the economics even while the headline pipeline keeps growing.

06IREN: the AI transition is real, but its financial history still includes Bitcoin

IREN presents a particularly useful test of the neocloud story because two businesses coexist in its accounts. Bitcoin mining monetizes computing power through a cryptocurrency network; an AI cloud sells access to infrastructure under customer agreements. Both require electricity, equipment and operational discipline. They differ in hardware, revenue formation, customer relationships and the obligations attached to delivery. Calling the entire company an AI revenue stream erases precisely the transition investors need to understand.

The fiscal year ended June 30, 2026 generated $707.0 million of revenue: $128.8 million from AI Cloud Services and $578.2 million from Bitcoin mining. The final quarter tells a different story about the direction of travel: AI revenue reached $70.5 million, against $66.7 million from mining. These are reported results for different periods, not competing estimates of the same annual business. IREN FY26 results, August 27, 2026.

The analytical challenge is the gap between the business producing cash during construction and the business expected to produce cash afterward. Converting a site can reduce mining output before the replacement AI cluster becomes available to its customer. Land and electricity connections may be useful across both activities, but a mining machine does not become an AI accelerator through a software update. Different equipment and cooling requirements create a fresh investment cycle.

That also changes how to interpret losses. IREN reported a $702.6 million fiscal-year net loss, including substantial noncash asset impairments associated primarily with decommissioning mining hardware. An impairment does not consume that amount of cash again when recorded. It nevertheless signals that previously purchased assets will not deliver their earlier accounting value. Conversely, removing impairments from an adjusted measure does not fund the replacement equipment.

The proper question is therefore whether the new AI assets produce sufficient revenue and cash after delivery, operating costs and financing. Neither the historical mining cash margin nor an annualized AI sales target answers it. A credible transition analysis follows individual deployments from construction through customer acceptance and then watches the resulting revenue mix over successive reporting periods.

07IREN’s three contract numbers measure three different things

The headline most easily misread is IREN’s $4 billion of contracted annualized run-rate revenue for 2026 capacity. The August update targeted that run rate becoming operational by December 31, 2026, subject to commissioning, testing and customer acceptance. Its approximately $1 billion operating ARR was dated August 26. ARR annualizes GPU hourly pricing and associated services; it does not report revenue already earned during twelve months. The company’s own footnotes warn that recognized revenue may be materially lower. IREN’s ARR definitions and assumptions.

Imagine a hypothetical cluster that reaches its full contracted rate near year-end. Its exit run rate can be impressive while its contribution to that year’s reported revenue remains small. This is a timing distinction, not evidence that the contract lacks value. A reader needs both figures: one describes the operating level reached or targeted; the other records the services actually delivered during the accounting period.

The annual filing adds a second distinction. At June 30, IREN disclosed approximately $5.1 billion of unsatisfied remaining performance obligations under revenue accounting, plus approximately $11.4 billion of contracted lease arrangements. It reported an approximate combined total of $16.6 billion. The component amounts are rounded; their displayed sum should not be used to overwrite the company’s stated aggregate. Of the $5.1 billion RPO component, $0.9 billion was expected within twelve months, $1.3 billion in months 13–24 and the balance in months 25–60. That schedule should not automatically be applied to the lease component. IREN 10-K, Note 4.

Why can an AI contract contain a lease? A customer may control the use of identified GPUs and dedicated data-center space. IREN then acts as lessor, while the equipment remains among its assets and continues to depreciate. No lease revenue had been recognized in the periods presented through June 30, 2026; the reported AI revenue for those periods arose under the applicable service-revenue accounting. This is different from IREN leasing equipment from someone else to finance its fleet. One arrangement creates customer revenue; the other creates a payment obligation.

The distinction prevents several errors. ARR is not added to contract value. A named customer agreement is not added again to a portfolio total that already includes it. Customer prepayments are not an extra layer of revenue on top of the services they fund. And “contracted” does not mean that all deployment, service-credit and collection risks have disappeared. IREN’s accounting classification is a reason to inspect the underlying agreement, rather than assume every company’s use of backlog describes the same economic object.

08IREN’s capacity must cross the acceptance line

There is a concrete operational milestone behind the numbers. On August 13, IREN announced that Horizon 1 had been delivered to and accepted by Microsoft. It is the first of four 50 MW IT liquid-cooled deployments at Childress associated with the five-year Microsoft agreement announced in November 2025. Acceptance is more informative than a rendering, equipment order or construction photograph because it marks a change in the customer’s relationship with the completed capacity. IREN’s Horizon 1 announcement.

It does not confer the same status on the remaining deployments. The August 27 update described Horizon 2 as undergoing commissioning and Horizon 3–4 as in late-stage construction, targeting fourth-quarter 2026 delivery. Likewise, a global development pipeline exceeding 5 GW is not the same as the company’s targeted cumulative delivery of approximately 0.3 GW IT in 2026 and 0.8 GW IT in 2027. Pipeline and delivered IT load have different boundaries and different dates.

The physical sequence matters financially. A power connection must support the required load; the building needs appropriate distribution and cooling; equipment must arrive and communicate correctly; a cluster must sustain the contracted performance; the customer must accept it. Progress at one stage may reduce risk without completing the next. The commercial clock can therefore lag behind the construction clock, even when workers and suppliers have been paid.

Software also belongs in this sequence. IREN completed its Mirantis acquisition on August 4, adding capabilities in workload orchestration, monitoring and customer support. The consideration included approximately 12.6 million ordinary shares and approximately $40 million of cash, restricted stock units and other consideration at closing. That is both a software investment and a financing choice. It does not, by itself, demonstrate that every acquired capability has already produced incremental revenue. IREN’s Mirantis closing announcement.

For the next operational update, the useful checklist is short: which deployment was accepted, how much capacity started billing, how quickly utilization increased and whether delivery credits reduced the expected receipts. A larger pipeline is encouraging only conditionally: its economic value still depends on the capital, equipment and execution needed to turn it into a functioning service.

09IREN: financing the GPUs does not finance the entire company

IREN’s Microsoft-related financing demonstrates why the customer matters to the cost of capital. The June 1 announcement described a $3.65 billion facility, split between a $2.10 billion private placement and a $1.55 billion delayed-draw term loan. Together with customer prepayments, it covered approximately 96% of the associated GPU capital expenditure. The announced blended debt cost was 6.00%, excluding fees. This does not mean that 96% of all future corporate spending was funded or that every subsequent borrowing carries the same rate. IREN financing announcement, June 1.

The later Mackenzie transaction illustrates the difference. Blue Owl confirmed a $2.4 billion package comprising a $1.2 billion secured term loan and $1.2 billion of secured notes, drawn in stages alongside equipment delivery and commissioning. IREN described a 9% fixed rate for that financing in its August results. A committed facility provides access under specified conditions; it is not automatically cash already in the bank. Blue Owl, August 28.

The June 30 balance sheet and annual cash-flow statement make the funding bridge visible. Cash and equivalents were approximately $5.896 billion, with another $1.724 billion restricted. Fiscal-year operating cash flow of $2.100 billion included a $1.842 billion increase in deferred revenue. Payments for property and equipment excluding computer hardware were $2.998 billion; computer hardware payments added $1.335 billion. Debt principal was approximately $7.706 billion, distinct from its carrying value after issuance costs, and financial-lease liabilities were additional. IREN audited financial statements.

Prepayments genuinely help fund construction. They also mean that some future service will be delivered against money collected earlier. A growing prepaid order book can produce strong operating cash flow before mature operations have demonstrated their cash profitability. The test is what remains after delivery costs, financing payments and replacement investment, not whether the accounting cash-flow total has a positive sign.

Equity remains part of this story. Ordinary-share issuance brought in approximately $4.743 billion before issuance costs during FY26. Convertible debt can create future dilution, while capped calls mitigate only within their contractual limits. Neither asset-backed lending nor an acquisition paid partly in shares justifies a blanket claim that future dilution has ended.

IREN’s economics will be strongest if accepted capacity earns the contracted receipts while debt amortizes and equipment remains useful. The adverse scenario is a timing mismatch: acceptance slips, a customer struggles to pay or a renewal occurs at a weaker price while financing obligations continue. Those are conditional mechanisms to monitor, not predictions that any named customer will default.

10Oracle: a diversified software company building a much larger infrastructure business

Oracle belongs in the comparison because its infrastructure expansion competes for the same scarce inputs: accelerators, power, buildings and financing. It should not be treated as economically identical to a specialist GPU cloud. Its installed software base, applications, support relationships, hardware and services provide a much broader commercial platform. The question is how the infrastructure build changes that platform’s capital requirements.

The latest reported period is the first quarter of fiscal 2027, ended August 31, 2026. Oracle reported $19.345 billion in total revenue, including $7.388 billion of cloud infrastructure revenue and $4.219 billion of cloud applications revenue. Infrastructure grew 121% year over year; applications grew 10%. Software licenses and support together contributed approximately $5.550 billion, with hardware and services making up the remainder. These are quarterly results, unlike the full-year IREN figures above. Oracle Q1 FY27 results, September 10.

The distinction between infrastructure and applications is economically useful. Infrastructure provides computing, storage and networking resources; applications package business functionality for customers. Selling more of the former can require large additions of physical equipment and data-center capacity before the associated revenue arrives. Growth percentages alone do not show whether that investment is producing an adequate return.

Nor is all infrastructure revenue synonymous with AI accelerator rental. Oracle’s cloud supports a broader range of workloads. A consolidated cloud-and-software margin cannot be assigned to a new AI campus without additional disclosure, because the mix includes businesses with different cost structures and investment needs.

The mature businesses matter because they create relationships and cash-generating capacity from which expansion can be supported. They do not remove the need to finance that expansion. For readers comparing four tickers, Oracle provides a useful counterexample to the idea that only young neoclouds need outside capital: scale and diversification can improve funding access while leaving a very substantial construction bill.

11Oracle’s $664 billion RPO has a delivery calendar

Oracle reported $664 billion of remaining performance obligations at August 31. Its recognition schedule placed approximately 13% in the following twelve months, 37% in months 13–36, 34% in months 37–60 and the remainder afterward. These are separate time bands, not cumulative percentages. The headline therefore contains commitments stretching well beyond the next reporting year. Oracle 10-Q, Note 1.

RPO includes contracted revenue not yet recognized, spanning amounts already deferred, invoices awaiting collection and amounts still to be invoiced. The category does not mean every dollar has arrived or every required data center is operational. Conversely, the company’s disclosure excludes certain variable consideration under an accounting exemption, so it is not a complete estimate of every possible future sale. Oracle FY26 10-K, revenue policies.

The schedule is more useful than dividing the total by an arbitrary number of years. A contract that begins later demands a different funding bridge from a service already being consumed. Construction, equipment procurement and lease commitments may precede recognition. If that gap lengthens, the provider can need more financing even when the ultimate contracted value remains unchanged.

There is also a concentration distinction. Oracle’s annual filing said no individual customer represented 10% or more of total revenue in fiscal 2026, 2025 or 2024, while warning that certain OCI offerings are more concentrated among large customers. These statements describe different denominators. Diversification in historical company-wide revenue does not establish diversification in future AI commitments or at an individual campus.

A practical reading separates three questions. How many buyers support the future capacity? When do their payment obligations start? What happens if the provider or customer fails to meet a condition? A large order book is evidence of contractual demand, but those questions determine how that demand translates into a financeable operating business. Without a disclosed customer-by-customer RPO allocation, assigning a precise share to one buyer would create certainty the filings do not provide.

12Oracle’s cash flow shows why customer financing must be counted once

For the quarter ended August 31, Oracle generated $23.103 billion of operating cash flow and paid $28.499 billion of capital expenditure. Its reported non-GAAP free cash flow was therefore negative $5.396 billion. Operating cash flow included $11.363 billion from customer prepayments with a significant financing component. Those receipts support construction before the related performance obligations have been fulfilled. Oracle Q1 financial statements and supplemental cash-flow analysis.

Oracle additionally reported $17.966 billion of net cash outlay for capital expenditures, a supplemental measure reflecting customer prepayments and relevant short-term financing movements. This is useful for understanding funding needs, but it cannot simply replace gross cash capex in the usual free-cash-flow calculation. Subtracting the smaller net outlay from operating cash flow would count the benefit of those customer prepayments twice: once inside operating cash flow and again in the reduction of investment spending.

The lesson is broader than Oracle. Two individually correct company numbers can produce a misleading result when combined with incompatible definitions. A reader should follow each cash receipt through the statement before deciding which adjustment belongs in a comparison.

Financing also included the full use of Oracle’s $20 billion at-the-market common-stock program. The company issued approximately 141 million shares, receiving approximately $19.9 billion net. This was capital raised during the quarter, not an unused facility waiting to be tapped. At August 31, cash and equivalents were $36.369 billion, with $708 million of marketable securities; restricted cash was disclosed separately. Oracle 10-Q, liquidity and equity.

Debt and leases require equally careful labels. Current and noncurrent notes and other borrowings totaled approximately $125.337 billion, a calculation from the balance-sheet lines. Operating-lease liabilities were $34.621 billion and finance-lease liabilities $9.185 billion. Separately, Oracle disclosed $288 billion of additional lease commitments, mostly for data centers, expected to begin between Q2 FY27 and FY29 with terms of fifteen to nineteen years. These future commitments are not all immediate cash maturities, and their undiscounted value should not be casually added to recognized liabilities as though the measurement bases matched.

They still matter. A building commitment can outlast the initial customer or accelerator generation. The funding analysis must therefore connect payments, useful life and renewal assumptions. Strong demand during construction does not establish the economics of every subsequent year of a long lease.

13Oracle’s power announcements need the same discipline as its contracts

Power supply is another place where one headline can hide several states of progress. Oracle’s September 8 request for proposals sought 2 GW of renewable generation in New Mexico, with deliveries between 2027 and 2031. Seeking proposals is a procurement step, not proof that all contracts have been awarded or that the power is available today. Oracle’s New Mexico RFP.

On September 15, Oracle announced investments in more than 1.7 GW of Texas wind generation. Electricity supplied into ERCOT and progress toward carbon-free energy matching should not be relabeled as 1.7 GW of continuously available IT load at a particular campus. Annual energy generation, hourly availability, transmission and a site’s connection limits answer different questions. Oracle’s Texas energy announcement, September 15.

Project Jupiter illustrates the permitting distinction. Oracle’s September 14 statement separated the data-center buildings from the adjacent microgrid and its air-permit process. A September 10 response from New Mexico agencies described requirements attached to the proposed permit. These documents establish separate workstreams; neither a company statement nor a proposed permit should be presented as proof that the entire campus is already authorized, energized and delivering customer workloads. Oracle permitting statement, New Mexico agencies’ response.

For financial analysis, the implications are concrete. A completed building awaiting energy cannot earn the same receipts as an accepted operating cluster. Equipment arriving too early can tie up capital; equipment arriving too late can postpone delivery. Local generation can change the dependency on the grid while creating its own fuel, permitting and maintenance requirements.

The next evidence to watch is therefore a combination: accepted capacity, recognized infrastructure revenue, customer cash receipts, gross capex, new financing and the start of lease obligations. A favorable outcome brings these timelines into alignment. A difficult outcome stretches the interval between spending and revenue. The size of the backlog makes that interval more consequential; it does not make it disappear.

14Why a strong operating margin can coexist with tight cash

The most useful bridge starts with money collected from customers, subtracts cash operating costs, then identifies interest, equipment purchases, lease payments, debt principal and other obligations. EBITDA can help describe operating performance, but it does not by itself show the cash available to common shareholders. Depreciation is noncash in the current period; the equipment that is depreciating nevertheless had to be financed and may eventually need replacement.

Consider a deliberately simplified Merlintrader illustration, unrelated to any company’s guidance. An asset costs 100 monetary units. The initial funding is 80 of debt and 20 of equity. Assume five equal annual principal repayments of 16, an 8% annual interest rate, and interest of 6.4 in the first year on the opening debt balance. Assume cash operating costs of 18, customer collections equal to revenue, no working-capital changes, no taxes and no additional capital spending. These assumptions isolate the mechanics; they are not realistic enough to value a company.

All figures are illustrative calculations by Merlintrader. The reference case has a substantial operating surplus, yet that surplus does not quite cover the first year’s assumed debt service. The stronger case covers it before the excluded items. Neither result is a prediction for CoreWeave, Nebius, IREN or Oracle. The exercise shows why a margin headline and a liquidity headline can describe different aspects of the same business.

The construction period creates another distinction. If customers start paying six months late while some financing and site costs continue, the problem is not merely a lower annual growth rate. Cash has to fund a longer interval before the project begins contributing. The actual impact depends on draw schedules, capitalized interest, contractual protections and available liquidity. A model that assumes all financing is drawn at inception can exaggerate some costs; one that assumes no cash outflow until revenue begins can understate them.

Replacement spending adds a further layer. The value of an older accelerator is not automatically zero when a new architecture appears. It can still serve appropriate workloads. But the price, utilization, power efficiency and support costs obtainable on that older equipment determine its economic life. A depreciation policy and an eventual resale value are therefore separate assumptions. Extending an accounting useful life does not itself create new customer collections.

15Read financing as a set of obligations, not a single interest rate

For secured equipment debt, assess the collateral, repayment schedule, draw conditions, covenants and recourse. For a lease, identify the fixed cash commitments and whether the associated capacity has already entered service. For a convertible, consider the coupon together with the conversion and settlement terms. For an equity issue, examine how much new capital arrives and the increased share count against which future earnings must be measured. These are different funding tools, not interchangeable labels for financial strength.

The distinction between a facility and its drawn balance is especially important during expansion. An available commitment may help finance a project but still depend on eligible equipment, customer contracts, construction progress or other conditions. Adding an undrawn commitment to cash without those qualifications gives an exaggerated impression of immediate liquidity. Conversely, a funded project need not require all of its expected lifetime capital expenditure to be sitting in cash on the reporting date.

Customer prepayments deserve the same discipline. They may fund part of a build and provide useful evidence of customer commitment. They are also linked to future service obligations. Their accounting classification, application against future invoices and any refund or termination provisions affect the economic interpretation. The relevant comparison is the funding gap after considering the project timetable and the agreement’s terms, not simply which company reports the largest cash balance.

16NVIDIA’s involvement: support, exposure and the question of final demand

NVIDIA’s role in the infrastructure ecosystem includes equipment supply, technical cooperation and capital investment. Its January 26 announcement with CoreWeave confirmed a $2 billion investment in CoreWeave shares. NVIDIA’s separate $2 billion investment in Nebius was completed through a pre-funded warrant and is reflected in Nebius’s first-half financing cash flows and equity. The funding was received; the associated infrastructure expansion still requires execution. Completion of the investment does not establish that the warrant has been exercised into all underlying shares. NVIDIA–CoreWeave announcement; Nebius H1 financial statements; NVIDIA ownership filing.

The analytical issue is how these relationships affect the reliability and diversity of cash flows. Industrial support can make deployment easier and align product development. It does not make every customer equally creditworthy, guarantee utilization beyond a contract’s term, or remove the operator’s funding obligations. Equally, overlapping commercial and investment relationships are not, by themselves, evidence of artificial revenue or misconduct.

A useful test traces the transactions separately: who provides capital, who sells hardware, who contracts for compute and who ultimately pays for the resulting service. A diversification claim is stronger when it is supported by a broader base of paying customers rather than only a longer list of counterparties in the same financing network. Where public information cannot establish the end customer’s economics, a claim that demand is either fully validated or entirely circular would go beyond the evidence.

17What the social debate reveals, and what it cannot establish

Two accessible Reddit discussions illustrate the information gap. One links directly to CoreWeave’s completed convertible financing and includes questions about the purpose of the capital. Another debates whether neocloud economics depend mainly on scarce compute or on durable capabilities such as electricity access, engineering and software. These are observations about the questions being asked, not a representative sample of investors or a verified measure of bullishness. CoreWeave financing discussion; neocloud business-model discussion.

The comments are views of social-media users whose professional status and positions have not been verified. They are not treated as institutional research. Claims about liabilities, GPU lives, financing terms or customer contracts need checking against primary documents. A repeated number is not more authoritative because it appears in several communities.

The useful editorial takeaway is the tension between two hypotheses. One says scarcity provides temporary pricing power that competition will erode. The other says reliable operation, software integration and access to usable power can sustain differentiation. Neither can be settled by counting optimistic posts. Evidence would come from renewals, realized pricing, utilization, customer retention and returns after the capital required to deliver the service.

18The comparison that matters at the next results release

The strongest new evidence would connect several stages in the same direction: more capacity enters service, the intended customer accepts it, recognized revenue follows, cash collections remain sound and financing needs develop as disclosed. A new contract is useful information, but a quarter that demonstrates conversion of earlier commitments can resolve a different and sometimes larger uncertainty.

A weaker outcome would be a widening gap between construction commitments and service commencement, especially if it is accompanied by additional financing needs or a change in expected cash generation. That does not automatically invalidate the demand thesis. It can instead shift the question toward timing, financing terms and how much capital must be committed before the demand becomes economically useful to the provider.

For CoreWeave, follow the relationship between contracted work, deployment and the obligations attached to its financing. For Nebius, connect cloud growth and operating run rates with actual service capacity and the spending needed to produce them. For IREN, track the conversion from its power and development base to accepted AI services while retaining the distinction between AI and bitcoin operations. For Oracle, separate the cloud-infrastructure expansion from consolidated software cash generation and compare RPO conversion with the funding burden. These are company-specific tests, not a ranking based on one headline number.

The long-term question is whether incremental projects earn enough over their economic lives to cover their full costs and compensate the capital that finances them. The near-term question is whether the operator can reach the next delivery milestone with the funding already available or credibly arranged. Both questions matter. Keeping them separate makes it possible to recognize substantial commercial progress without treating future shareholder returns as a completed fact.