Two emerging cloud players in the AI infrastructure space, CoreWeave (CRWV) and Nebius (NBIS), delivered impressive Q2 2026 earnings reports. Their quarterly revenue surged by 112% and 454% year-over-year, respectively. However, beneath these striking growth curves, the capital market is witnessing an unprecedented business drama: a capacity bet fueled by extreme leverage, profit pains from timing mismatches, and an eventual evolution from "computing landlords" toward "asset pricing power." The logic within the AI infrastructure sector is shifting, as revealed by these information-rich earnings reports.

01 Order Frenzy: Clients Prepay for Computing Power

If power grids and servers form the foundation of computing giants, then massive backlogs represent their strongest financial moats. In Q2, both companies demonstrated exceptional order-locking capabilities. CoreWeave reported a record backlog of $1,042 billion, up 246% year-over-year, excluding over $250 billion in net new customer commitments added early in Q3. Critically, more than 50% of this backlog has entered the delivery phase. Nebius saw explosive growth in customer prepayments, with full-year 2026 prepayments expected to surpass $9 billion—nearly three times its full-year revenue guidance of $3.0 to $3.4 billion. Furthermore, for new medium-term contracts (1-3 years), customer prepayments cover 50% to 60% of CapEx spending. The order structure shows AI demand rapidly spreading from pure large model training to enterprise applications and inference. The entry of industrial giant Caterpillar (physical AI), life sciences firm Isomorphic Labs, and quantitative financial institutions signals that computing power is no longer exclusive to tech giants, but is accelerating into a core production tool for traditional high-wage industries. This prepayment model not only eases the cloud providers' own financial pressure but also proves to the capital market that downstream clients are willing to place long-term orders and even prepay massive sums to secure future AI computing capacity.

02 ARR Surge: From Hardware Rentals to High-Margin Subscriptions

More than total revenue, the capital market is excited about the explosive growth in annualized recurring revenue (ARR) and high-margin derivative services. For CoreWeave, its Managed Inference platform's ARR soared from $1 million to over $100 million within months of its launch, targeting $250 million by year-end. Meanwhile, its non-GPU add-on services (storage, CPU, networking, software) have surpassed $400 million in ARR. Since software and add-on services boast significantly higher margins than pure GPU hardware leasing, this is becoming a key driver of overall profitability. Nebius reported ARR of $3.0 billion as of end of June (up 598% year-over-year, 58% quarter-over-quarter), and has anchored its year-end 2026 ARR guidance at an ambitious $7.0 to $9.0 billion. This metric surge signals a major shift: AI cloud providers are shedding their "one-time hardware rental" label. Whether through AI agents and fine-tuned models moving rapidly from experiment to production, or integrated subscriptions combining compute, storage, and software, customer usage of high-stickiness, high-margin services is climbing exponentially. This ARR explosion is transforming one-time capital expenditure into long-term, stable, high-margin cash flow.

03 Timing Mismatch: Why Revenue Doubles but Margins Halve?

CoreWeave's Q2 revenue approached $2.6 billion, with adjusted operating profit jumping from $21 million in Q1 to $128 million. While operating leverage seems apparent, its adjusted operating margin plummeted from 16% in the same period last year to 5% this quarter. Why do margins shrink even as business scales double? The answer lies in the data on "timing mismatch." The company added about 500 MW of active power in Q2, with over 300 MW coming online in June alone—surpassing any previous full quarter's addition. This front-loaded delivery drove Construction in Progress (CIP) to $11.9 billion. With massive server deployments finalized in the last month of the quarter, depreciation and operational costs were fully booked, while the corresponding service revenue will be recognized later in Q3 and Q4. This is the inevitable "financial time lag" of heavy-asset models during extreme expansion: capital is racing ahead, and cash is on its way back.

04 Debt Devours: Interest Becomes the Heaviest Tax

Supporting this supercharged expansion is wild financial leverage. CoreWeave's quarterly interest expense hit $640 million, up 140% year-over-year, representing 24.9% of quarterly revenue. For every $4 earned, nearly $1 goes to creditors. Q3 guidance projects interest expenses rising further to $860 million to $940 million. Nebius reported a quarterly CapEx of $5.7 billion, while operating cash flow was only $2.3 billion. Even with $8 billion in cash, the gap from high-intensity construction remains enormous. Under the weight of massive interest, CoreWeave maintained a GAAP net loss of $626 million for the quarter. This "max leverage" game is essentially a race against time: the company must ensure the slope of AI computing demand growth outpaces the slope of debt interest growth.

05 Narrative Shift: From GPU Rentals to Scarce Energy Operators

Facing market anxiety over hardware depreciation and debt burdens, management collectively reframed their narrative this earnings season. First, they challenged hardware depreciation logic by "annuitizing" older GPUs. CoreWeave disclosed that its five-year-old A100 architecture secured new contracts extending to 2029, with pricing flat or even higher than years ago. Management cleverly defused asset devaluation fears: clusters already connected to power grids, cooled, and plug-and-play are no longer fast-depreciating consumables, but "unlevered pure free cash flow generators." Second, they shifted from "production to order" to "pricing by production." Nebius management adopted a seller's stance: "Under current conditions, we could sell all of our 2027 capacity now, but we choose not to." By reserving some capacity to match short-term spot demand (up to 6 months), they can price single MW at $40 million to $50 million, nearly double the medium-term contract price of $20 million to $25 million. Through innovative auctions and arbitrage, cloud providers are attempting to gain pricing power for basic resources akin to "oil" or "electricity."

06 Business Evolution: Lightening the Heavy-Asset Model

Constrained by their own balance sheets for building data centers, both companies are pursuing dual leaps in financial engineering and software platforms. Through tool innovation, CoreWeave launched a delayed draw term loan (DDTL 5.5), successfully financing shorter-term client contracts (2-3 years) and opening up the enterprise market reluctant to sign 5-year deals. Nebius completed a $775 million asset-backed debt financing (ABS) based on investment-grade client contracts, fully financializing GPUs as assets. In an asset-light approach, Nebius even introduced a model where partners fund construction while it provides the software platform and clients, aiming to upgrade from "computing landlord" to "AI computing operating system."

07 Endgame Thoughts: The Next Turning Point for Computing Prosperity

Based on Q2 and full-year 2026 guidance, the massive CapEx of the five major cloud providers ($785 billion to $810 billion) continues to inject ample liquidity into the industry. CoreWeave expects Q4 adjusted operating margin to enter low teens, while Nebius reaffirms a full-year 40% EBITDA margin. However, for investors, the perspective has fundamentally changed. The capital market is no longer simply rewarding "revenue growth," but is now rigorously evaluating the certainty of margin improvement and the specific timeline for positive free cash flow (FCF). This fast-paced AI infrastructure bet is entering a second half that tests refined operations and financial engineering.