AI computing contracts are reaching the income statement at scale, but the returns available to shareholders remain markedly different. In the second quarter of 2026, CoreWeave recognized $2.575 billion of revenue, Nebius reported $582.3 million at group level, and Bitdeer generated $14 million from AI Cloud. The useful comparison is how equipment, customers and financing interact to determine earnings.[1][2][3]
That interaction is easily lost in neocloud valuations. A long contract can support borrowing; an advance payment can finance equipment; an annualized revenue figure can indicate the latest sales pace. Each is valuable. None alone establishes that a fleet of GPUs will earn an adequate return after depreciation, interest and replacement spending. These results offer three views of that problem at different stages of development.
In this article6 sections
CoreWeave has scale—and a substantial financing bill
CoreWeave reported second-quarter adjusted EBITDA of $1.510 billion, or roughly 59% of revenue. Its GAAP operating loss was $49 million and net interest expense reached $640 million. Operating earnings must absorb equipment consumption and financing costs before shareholders receive a return. Treating adjusted EBITDA, which excludes several expenses, as distributable profit would substantially overstate the maturity of the business.[1]
Net interest expense was approximately 24.9% of quarterly revenue, calculated from the reported figures. This does not mean every incremental revenue dollar carries the same interest burden. It shows how much weight the existing capital structure already carries. Growth can improve earnings rapidly if revenue rises faster than relatively fixed financing charges. If each new contract requires comparably intensive borrowing and equipment purchases, the interest bill can expand alongside the business.
The company reported 1.5 GW of active power, approximately 3.7 GW of contracted power and about $104 billion of revenue backlog. Its backlog definition includes remaining performance obligations and other amounts expected from committed contracts, subject to delivery and service availability.[1] The gap between contracted and active power cannot simply be multiplied by today's revenue density: hardware, contract terms, start dates and power definitions may differ.
The more revealing comparison follows the same cluster through commissioning and into service, tracking revenue against depreciation and financing expense. If new clusters improve operating earnings without comparable growth in receivables and interest, scale begins to benefit shareholders. If contract growth mainly supports a larger balance sheet, valuation remains unusually dependent on funding conditions and future utilization.
Nebius receives cash before completing the service
Nebius's cash flow looks stronger than its income statement. Second-quarter group operating cash flow was $2.2461 billion. Changes in deferred revenue contributed $1.1970 billion and changes in receivables contributed $1.1868 billion. Purchases of property, equipment and intangible assets were $5.6574 billion.[2] These figures help explain why operating cash can greatly exceed current revenue while the business continues to need financing.
Operating cash flow less those asset purchases was negative $3.4113 billion. This is a subtraction of two reported lines, excludes other investing activities and is not a company-defined free-cash-flow measure. Separately removing the deferred-revenue and receivables movements from operating cash gives negative $137.7 million. That should not be labeled normalized operating cash flow either: other working-capital movements, taxes and non-cash adjustments remain. The exercise identifies the sources of cash generation.
Customer advances provide a real financing advantage. They reduce the external capital an operator must supply during construction. They also create an obligation to deliver future service. When the associated revenue is recognized, the cash may already have arrived in an earlier period, while electricity, operations and replacement equipment still require payment. Advances cannot therefore be extrapolated indefinitely at the same growth rate or treated as profit without delivery costs.
Management said on its earnings call that four new transactions averaged more than $1 billion in total contract value and that certain medium-term contracts represented annual value of around $20 million–$25 million per MW, with advance-payment arrangements. These are terms for a particular contract mix, not a universal price for capacity. The company also reported approximately $3 billion of AI Cloud annualized revenue run-rate at June-end. That point-in-time pace is a different measure from full-quarter group revenue.[4]
Bitdeer is growing AI revenue below gross break-even
Bitdeer's disclosure makes the costs of an early transition particularly visible. AI Cloud generated $14 million of quarterly revenue against $16.3 million of revenue costs, producing a $2.3 million gross loss. Revenue had grown from $1.3 million a year earlier, demonstrating a paying customer base. The increase was not yet sufficient to cover the business's reported cost of revenue.[3]
Using the rounded segment figures, gross margin was approximately negative 16.4%. Even if every cost were temporarily treated as fixed, revenue would need to increase by about 16.4% from its present level to reach gross break-even. In practice, additional revenue can bring additional power, cloud-service and staffing costs. This is a simplified lower bound, not a forecast of next quarter's earnings.
Bitdeer also operates mining, AI Cloud and colocation businesses. An operator that owns GPUs carries procurement and replacement risk. A colocation provider mainly sells usable facilities, power and related services. Revenue per MW, capital requirements and risks differ. Applying GPU-cloud margins to a colocation contract, or treating the entire power portfolio as an AI revenue asset, would obscure the economics.
The project schedule distinguishes these activities. Tydal phase-one colocation and the Cyberjaya expansion were listed for service readiness in the fourth quarter of 2026, while other projects extended into later periods. The table also explicitly separates those dates from tenant equipment delivery.[3] Fourth-quarter readiness therefore applies to specified projects and engineering scope, not a commitment that the entire AI Cloud business will bill at full capacity then.
What does a quarter of delayed delivery cost?
A transparent scenario illustrates the importance of timing. Assume a new cluster can recognize $100 million a year after acceptance and satisfaction of service conditions. Billing is uniform, there are no construction-period fees and delayed revenue cannot be recovered later. A January start produces $100 million that year. A three-month delay reduces the contribution to $75 million, a $25 million shortfall. No company's contract price is assumed here.
Now assume the lost revenue would have earned a 60% cash contribution margin after cash costs that vary with usage. The lost contribution is $15 million. That still excludes construction overruns, fixed personnel expenses, financing costs and contractual compensation. If equipment has already been paid for and borrowing already accrues interest, cash pressure can emerge before the revenue shortfall appears.
The same delay affects business models differently. For a dedicated cluster whose minimum payments begin after acceptance, the acceptance date matters most. For usage-based sales, the speed of customer adoption after commissioning matters as well. For colocation, the relevant terms include rent commencement, minimum power commitments and construction responsibility. Strong customer demand does not answer these contractual questions.
Conversely, a non-refundable capacity reservation fee or phased acceptance and billing can reduce the loss relative to this scenario. The model should change with the terms. Giving every delayed project the same discount would miss a major source of differentiation. Precise project dates and payment commencement clauses can be more informative for valuation than a headline GPU count.
Premium short contracts and long commitments carry different costs
Long contracts improve revenue visibility and may support lower-cost asset financing. Strong counterparties and clearer payment obligations make debt-service cash flows easier to estimate. Operators may surrender some upside in exchange: when new hardware becomes scarce and spot prices rise, a locked contract cannot immediately capture the increase.
Short contracts preserve repricing opportunities but require more frequent customer acquisition. Suppose a long contract secures billable time equal to 90% of the year, while a short-contract strategy achieves 70%. Holding costs and service terms equal, the short-term price must be at least 90 divided by 70 minus one, or approximately 28.6%, higher to generate the same revenue. If facility and financing bills continue during idle periods, the profit comparison requires further care.
These are not estimates of Nebius or CoreWeave utilization. They explain why a higher quoted price does not alone establish a better strategy. Advance payments, termination rights, counterparty credit and hardware generation also matter. Premium pricing supported by durable software, service and delivery advantages can improve returns. Pricing supported mainly by temporary scarcity is harder to sustain when additional equipment becomes available.
Customer concentration has a similar trade-off. A large buyer can fill a cluster quickly, reduce selling costs and support financing. The same buyer can affect revenue, collections and utilization simultaneously by delaying acceptance or changing procurement plans. Diversification reduces exposure to an individual counterparty but requires additional commercial and operating capabilities.
Valuation should follow sustainable returns
A single revenue multiple is an inadequate way to rank these businesses. For CoreWeave, the question is whether service earnings can grow faster than depreciation and interest. For Nebius, it is whether construction funded partly by advances becomes recurring profit, and how much additional equity future delivery requires. For Bitdeer, it is whether AI Cloud gross margins and phased colocation commissioning provide a repeatable source of cash for the transition.
The optimistic interpretation has a sound basis. Assets require investment before they produce revenue, and recognition follows acceptance. Judging an entire project's life by cash consumption at the construction peak can understate the return from mature assets. Reliable customer commitments, timely delivery and recovery of investment within the equipment's economic life can justify negative cash flow during expansion.
The opposing case deserves equal precision. GPU rental prices can fall, equipment can lose competitiveness before its accounting depreciation period ends, and new construction can continue consuming cash generated by established sites. Repeated equity issuance can leave growth in group revenue well ahead of growth in value per share. Industry demand alone cannot settle those questions without asset-level returns and financing terms.
The next results should clarify three developments: whether commissioned clusters contribute a full quarter of revenue; whether underlying operations improve after considering working-capital timing; and whether incremental investment begins growing more slowly than revenue and cash earnings. Consistent progress on those measures would provide a stronger foundation for long-term valuation.
Sources
This website is for information and research only and does not constitute investment advice.