The neocloud divide: cloud growth versus hosting payback

US neocloud revenue is growing, but the label covers two different businesses. One buys GPUs and sells compute; the other builds sites, delivers power and facilities, and earns long-term rent. The first bears more hardware-refresh and utilization risk. The second depends on matching delivery and leases with financing maturities.

CoreWeave and Nebius already generate substantial direct-cloud revenue. IREN, Bitdeer and mining-to-hosting peers sit at different points in their expansion. This analysis retains its August 12, 2026 Beijing-time cutoff and compares revenue, contracts and billing status separately. Future MW are not assets already collecting cash.

Revenue, run rates and future contracts measure different things

CoreWeave reported second-quarter revenue of $2.575 billion, versus $1.212 billion a year earlier, and adjusted EBITDA of $1.510 billion. Its $104 billion backlog, about 1.5GW of active power and about 3.7GW of contracted power describe future business and supply at different stages. None can replace the measure of services already delivered.^[1]

Nebius reported Q2 group revenue of $582.3 million, up 454% year over year; AI Cloud was $575 million, up 514%. The company reported 50% adjusted EBITDA margin for AI Cloud and $3 billion of ARR. This is among the closest comparisons to an operating GPU-cloud revenue model, but ARR remains an annualized run-rate rather than recognized quarterly revenue.

The remaining companies occupy earlier or different conversion stages. IREN reported FY2026 Q3 AI Cloud revenue of $33.6 million, up from $17.3 million in the prior quarter. Bitdeer reported Q2 AI Cloud revenue of $14 million and total revenue of $228.8 million. Core Scientific reported $136.7 million of Q2 colocation revenue, with 395MW billing at quarter-end and 437MW by mid-July; billed MW is closer to delivered service than contracted capacity, but it still does not make every signed project current revenue.

CoreWeave’s $1.510 billion of adjusted EBITDA divided by $2.575 billion of revenue is approximately 58.6%. This company-defined measure precedes depreciation, interest and capital expenditure. Comparing it with Nebius AI Cloud’s 50% should prompt reconciliation of adjustments, not a conclusion that one delivers higher lifetime equity returns. The measures also cover different business scopes.^[1][3]

Growth rates depend on the deployment starting point. A small business can accelerate sharply when its first customers go live; a mature business requires a much larger absolute increment for the same percentage. IREN and Bitdeer’s early cloud revenue should be examined for newly operating equipment versus utilization of existing assets. A small denominator does not make their scale comparable with established operators.

Revenue per MW reflects different assets and obligations

Direct GPU cloud bundles equipment, networking, software and customer usage. Nebius said new core-cloud contracts exceeded $20 million in annualized value per MW. That reflects the scope of a cloud service, not average recognized revenue or long-term margin.^[3] A host responsible for fewer assets and services has a commercial reason to report a much lower contract value per MW.

Hosting contracts can show lower per-MW values while assigning risk differently. Bitdeer disclosed a 121 IT MW, 16-year, $4.7 billion Tydal contract. Dividing $4.7 billion by 16 years and 121 IT MW produces roughly $2.4 million per IT MW per year. This is only a conversion of disclosed total contract value; it excludes discounting, operating cost, financing cost and revenue-recognition timing.

TeraWulf / Justified disclosed 401MW, 20 years and $19 billion. The same arithmetic yields roughly $2.37 million per IT MW per year. It must not be represented as all TeraWulf campuses, all company revenue or operating capacity; the official release confines it to the Justified Atlas campus. Contract value indicates what long-duration capital will pay for deliverable IT load, not that racks, GPUs or cash flow already exist.

Higher cloud revenue per MW does not by itself justify a higher valuation. A cloud operator may buy GPUs and networking and bear refresh and service obligations; a host may provide only specified space, power and cooling. Equipment ownership, included services and electricity pass-through need to be aligned before comparing net returns on invested capital.

Dividing TCV by years also hides ramp-up, rent escalators, construction reimbursements and end-of-term arrangements. Partial first-year delivery can leave revenue below the average annual value, while later escalators can lift nominal rent above it. Without an annual payment schedule, the calculation should remain a normalization of contract size, not an input directly substituted into debt-service coverage.

Large contracts also require large upfront funding

CoreWeave’s $104 billion backlog provides demand visibility; Nebius’s $3 billion ARR describes an annualized business pace. Both still require equipment, electricity and funding. Direct cloud can report revenue earlier while bearing refresh, pricing and idle-capacity risks. Long contracts can reduce some exposure, but cancellation, acceptance and payment provisions determine how much.

IREN announced $2.8 billion of new TCV in July and a $4 billion ARR target, improving forward visibility while increasing the burden of customer activation and capital deployment. Bitdeer's Tydal contract, Hut 8's 949MW of contracted IT capacity and $26.6 billion of base TCV are inputs to future hosting economics, not recognized revenue. Capacity must retain its state boundary: contracted, under construction, energized, accepted and billing.

Financing is not a side variable. GPU clouds need rapid deployment of servers, networks and working capital. Campus hosts must complete land, power, civil works, transformers, cooling and networking before service revenue. Larger contracts often increase the coordination and front-loaded-capital problem. The best subsequent disclosure puts customer commitment, funding source, energization milestone and actual billing on one timeline.

An independent scenario illustrates the financing cost of delay. Assume a project has $1 billion of outstanding debt at 8% while awaiting acceptance. A three-month delay with no principal repayment adds about $20 million of interest: $1 billion × 8% × 3/12. This is not any sample company’s loan and excludes lost revenue, retained staff and contractor costs. Even a contract that remains in place can consume the equity cushion when delivery slips.

Prepayments ease construction funding while creating future service obligations. Cash restricted to specified equipment, or refundable after delay, is not available for arbitrary use. A valuation should distinguish unrestricted cash, restricted funding, customer commitments and unpaid equipment costs. The same advance cannot be treated both as a free cash asset and as the full profit on future services.

Initial rent receipts begin to test the mining transition

Applied Digital recognized $44.1 million of base HPC rent in FY2026 Q4. Its $208.2 million of service revenue included $152.4 million associated with construction and equipment deployment, while signed leases covered 1,410MW over 15 years with $36 billion of value. Contracts describe future business; rent begins to demonstrate operations. Construction revenue does not have the same recurring character.^[10]

TeraWulf reported $31.9 million of Q2 HPC lease revenue and said it had 102MW of revenue-critical IT capacity in early July. Cipher reported $25 million of Q2 revenue and negative $30 million of adjusted EBITDA, with first Black Pearl capacity delivered in August. Both make the same point: entering AI/HPC does not instantly transform the current income statement; initial energization, installation and customer acceptance decide when revenue crosses over.

Hut 8's $72.5 million of Compute revenue combines ASIC, AI and traditional computing services, so it is not a pure GPU-cloud measure. Chronoscale has only been independent since May. For these companies, the most revealing changes are billed IT MW, first-project timing, recurring-rent mix and post-financing dilution—not total power or distant TCV forced into a common valuation multiple.

A revenue multiple still needs equipment renewal and debt

Direct-cloud revenue includes services supplied by GPUs and other equipment. Hosting rent may principally compensate for facilities. Assigning both the same sales multiple assumes each dollar requires equivalent capital, upkeep and renewal. A more useful equity analysis begins with receipts, subtracts service costs, investment needed to maintain existing capability, interest and required principal payments, and then assesses new construction.

Growth spending also differs from maintenance spending. New GPUs and sites serve additional customers; replacement equipment and power or cooling repairs preserve existing capability. Real investments can do both, and public reports may not fully separate them. Treating all equipment purchases as indefinitely removable growth capex would overstate cash available to owners.

Long-term funding eases immediate maturities but can leave debt outstanding beyond an asset’s economic life. If a new hardware generation depresses old-equipment pricing or utilization, the loan remains due. A hosting asset similarly needs an assessment of reletting and conversion costs after a customer exits. Long maturity is an advantage only while assets continue to earn cash.

This comparison therefore does not assign a common target multiple. Stable rent can be assessed against contractual cash and remaining obligations. Ramping direct cloud requires utilization, pricing and renewal estimates. Undelivered campuses still need funding to complete. At the same enterprise value, different net debt, unpaid equipment obligations and future equity issuance leave different value per share.

Valuation must follow billing, collection and reinvestment

Over the next twelve to twenty-four months, direct-cloud value depends on realized pricing, retention and paid utilization covering equipment replacement. Hosting value depends on contracted IT capacity beginning to bill on schedule. Both must fit delivery and revenue into their financing terms. Grid, transformer, cooling, networking or commissioning delays can defer cash even when the customer remains committed.

Positive evidence should be visible in quarterly results: recurring service or rental revenue rises, more capacity becomes billable, collections keep pace with revenue and the remaining construction funding gap narrows. Together these changes can reduce dependence on external capital during expansion. Contract growth alone cannot establish the same result.

The adverse case combines slower customer activation, continuing fixed payments and more expensive financing. Existing contracts may remain valid while interest and dilution reduce the residual per share. Delivery dates, replacement spending and funding costs should then change in the valuation model; an unchanged ARR target cannot resolve the deterioration.

CoreWeave and Nebius have demonstrated direct-cloud revenue, while AI/HPC landlords occupy different stages of rental ramp-up. The fastest revenue growth need not earn the highest capital return, and the largest contract need not produce shareholder cash first. Distinguishing the businesses and reconciling their investment, billing and collections is the basis for valuing that growth.