Neocloud economics: contract first, hardware second

Summary

• The GPU cloud operators studied sign multi-year take-or-pay contracts before they order hardware. Committed contracts produced over 98% of CoreWeave's 2025 revenue [1], and Nscale typically issues purchase orders once a committed contract is signed [14].
• Customers prepay part of each contract. Weighted averages were 15% to 25% of contract value at CoreWeave [1] and 23% at Nscale [14]. Deals signed in 2026 at Nebius and IREN carried prepayments covering 45% to 60% of the related capital spending [13, 16].
• Debt is secured by the hardware and the contract, repaid from the contract and priced on the customer's credit. Within the same borrower and year, the stronger credit cost two to three percentage points less: SOFR + 2.25% on CoreWeave's A3-rated facility against SOFR + 4.50% on one backed by two non-investment-grade customers [7, 8], and 6.0% against 9.0% at IREN [16].
• Operators report contract paybacks between one year and ten months and about 2.5 years, each on its own definition [2, 13, 16].
• Net results were losses at CoreWeave, Nscale and IREN in the periods reviewed [3, 14, 16]. At CoreWeave, depreciation and net interest together exceeded adjusted EBITDA in the second quarter of 2026 [3].
• Where no investment-grade customer exists, credit support has come from NVIDIA, through capacity purchase and revenue-share agreements, and in leases from hyperscalers [11, 19, 21, 24].

Background

A neocloud is a cloud provider built around GPU capacity for AI. Its main assets are GPU servers and the networking around them, housed in owned or leased data centers. The servers wear out far sooner than the buildings: Nscale depreciates buildings over 25 to 50 years and technology equipment over five to six years [14]. CoreWeave depreciates technology equipment over six years [2]. Nebius uses five years for server and network equipment, up from four years before 2026 [13].

The amounts involved are large. CoreWeave spent $6,422 million on property and equipment in the second quarter of 2026 alone [3], and its total indebtedness was $21.6 billion at the end of 2025 [1]. Nebius reported capital expenditures of about $5.7 billion in the same quarter [13].

An operator that buys hardware first and looks for customers later carries two risks for the life of the equipment: the price it can charge and the share of hours it can sell. Both have moved sharply. SemiAnalysis reported that one-year H100 rental contracts rose almost 40%, from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March 2026 [22]. A contract signed before the hardware is ordered moves both risks to the customer for the contract term. This report examines how the public operators use that structure, what it does to their cost of debt and what it leaves unresolved.

Data and method

What was measured. For each operator we recorded:

• the share of revenue from committed contracts and their duration
• the prepayment customers make, as each operator defines it
• the order of contract signing, financing and hardware purchase
• the terms of GPU debt facilities: size, security, recourse, repayment, rating and pricing
• payback statements and the definition behind each one
• company-level results for recent reporting periods

Sources. Primary documents come first: annual and quarterly reports, registration statements and current reports filed with the U.S. Securities and Exchange Commission, and the operators' own press releases and shareholder letters. Operators covered are CoreWeave [1, 2, 3, 4, 5, 6, 7, 8, 10, 11], Nebius [12, 13, 25], Nscale [14], IREN [15, 16], Lambda [17, 18] and SharonAI [19], with NVIDIA's quarterly report [21] for the supplier side and a TeraWulf release [24] for a hyperscaler lease backstop. Research and reporting fill gaps the issuers leave: SemiAnalysis on lender requirements and market conditions [20, 22], Octus on a 2026 loan syndication [9] and Semafor on a lapsed negotiation [23]. SOFR comes from the Federal Reserve Bank of New York [26].

Period. From CoreWeave's first delayed draw term loan in July 2023 to Lambda's fixed-rate facility of 1 October 2026.

How. Each figure is recorded as the issuer states it, with its date and definition. We did not convert one company's metric into another's. Where a figure comes from reporting rather than from the issuer, the text says so. The report uses no Rillor data.

Findings

1. The order: contract, financing, hardware, acceptance

The filings describe the same sequence.

• Contract. CoreWeave customers "purchase a specified amount of capacity on a take-or-pay basis over the contract term" [1]. Committed contracts were over 98%, 96% and 88% of revenue in 2025, 2024 and 2023 [1]. Nscale says it generates the vast majority of its revenue from multi-year take-or-pay contracts with a weighted-average life of about 5.7 years [14].
• Financing. Nebius signed a Microsoft agreement on 8 September 2025 with fees estimated at up to $17.4 billion through 2031. It expected to fund the capital spending through cash flow under the agreement and "the issuance of debt secured against the Agreement", and the parties' obligations began only once Nebius confirmed it had secured that financing [12].
• Hardware. Nscale describes a contract-to-cash cycle in three phases: signing, infrastructure purchase and installation, then go-live. "Upon executing a committed contract, we typically issue purchase orders for the infrastructure components" [14]. IREN announced a $9.7 billion, five-year Microsoft contract and an agreement to buy the GPUs and ancillary equipment from Dell for about $5.8 billion in the same release [15].
• Acceptance. Nscale's contract terms begin on delivery of the GPU clusters [14]. IREN says revenue ramps after commissioning, testing and customer acceptance [16]. Debt follows the build: CoreWeave's facilities are drawn as it builds for customer requirements [2], and Lambda's October 2026 facility releases funds at cluster commissioning milestones [18].

The sequence has exceptions. Nscale says that in certain markets it may deploy capital before finalizing customer contracts to shorten project timelines [14].

2. Prepayment funds part of the build

Sources: [1, 13, 14, 15, 16].

Nebius expected more than $9 billion of customer prepayments in 2026 and described prepayment as "the market standard for securing capacity in a supply-constrained environment" [13]. SemiAnalysis reported that smaller AI labs and AI-native companies had agreed to prepayments above 20% on contracts of four years or more, which it described as previously atypical [22]. The measures differ. Three are shares of contract value and two are shares of capital spending, so the figures show direction more reliably than level.

3. How the debt is built

Security and recourse. CoreWeave's delayed draw term loans "are collateralized by contractual cash flows and infrastructure assets" [1]. Its March 2026 facility sits in a subsidiary. It is secured by substantially all of that subsidiary's assets and a pledge of its equity, and the parent's guarantee is limited to specified "bad acts" [6]. Nscale's 2026 GPU facilities are secured by first-priority interests in the borrower's assets and equity, with limited guarantees from the parent [14].

Repayment. CoreWeave's first two facilities required quarterly repayment to reduce principal to an amount based on projected contracted cash flows or on the depreciated purchase price of the GPU servers, whichever required more [2]. Nscale's facilities "mature on the earlier of the date on which the customer makes its final scheduled payment" and a stated date [14]. Lambda's August and October 2026 facilities amortize fully, the first on a schedule aligned with the contracted cash flows and the useful life of the GPUs [17, 18].

Covenants. CoreWeave's March 2026 facility requires a debt service coverage ratio of at least 1.15x and includes events of default tied to adverse events under certain material contracts [6]. Both Nscale facilities require a historical debt service coverage ratio [14].

Leverage. IREN's $2.4 billion facility for non-investment-grade deployments funds 90% of the associated GPU spending, and its investment-grade financing, together with prepayments, funds 96% [16]. SemiAnalysis reported that lenders to backstopped operators look for a debt service coverage ratio of at least 1.3x, which typically corresponds to a loan-to-value ratio of 70% to 80% [20].

Pricing. The table lists the GPU facilities in the period whose pricing was disclosed or reported.

Sources: [2, 4, 5, 6, 7, 8, 9, 13, 14, 16, 17, 18].

Two effects run through the table. The first is time and track record: CoreWeave's spreads fell from SOFR + 9.62% to SOFR + 2.25% between 2024 and 2026, which CoreWeave describes as progress in reducing its cost of capital and enhancing its credit profile [2, 7]. The second is customer credit, which shows clearly when time is held fixed. In 2026 CoreWeave borrowed at SOFR + 2.25% in an A3-rated facility backed by one customer contract, and at SOFR + 4.50% against two non-investment-grade customers [7, 8]. According to Octus, a third CoreWeave loan, backed by a basket of unrated and speculative-grade lessees, struggled to attract commitments and was pushed to SOFR + 5.50% with a discount [9]. CoreWeave had not confirmed that pricing when Octus reported it [9]. IREN's two GPU financings reported in August 2026 differ by three percentage points along the same line [16].

The CoreWeave registration statement states the mechanism directly: the DDTL 2.0 spread was set "based on the credit quality of the associated customer contracts" [2]. For scale, overnight SOFR was 3.87% on 8 October 2026 [26]. SemiAnalysis modeled the effect of debt cost on margin and found that raising a neocloud's all-in cost of debt from 5.62% to 10% cut pre-tax margin from 14.8% to 5.4% [20].

4. Payback, as each operator defines it

Sources: [2, 13, 14, 16].

The definitions differ in what they count. CoreWeave measures against adjusted EBITDA and includes prepayments [2]. IREN measures GPU spending against contracted revenue net of direct costs [16]. Nebius excludes prepayment [13]. None of the definitions given mentions interest. A simple check on disclosed totals gives a slower figure for an earlier contract. IREN's Microsoft contract is $9.7 billion over five years, or about $1.94 billion a year, against about $5.8 billion of servers, networking, deployment services and software bought from Dell [15]. Equipment cost divided by annual contract revenue is about three years before any operating cost. IREN reports faster paybacks on later contracts at higher prices per megawatt [16].

5. Company-level results

Sources: [1, 3, 14, 16]. Entries marked n/a were not taken from the cited document.

Net results were negative at each company in the table. At CoreWeave, depreciation and amortization equaled 54% of second-quarter revenue and net interest expense 25%, against an adjusted EBITDA margin of 59% [3]. IREN's net loss includes $638.8 million of non-cash impairments, mainly from decommissioning Bitcoin mining hardware as sites convert to AI cloud [16]. Nebius reported depreciation and amortization equal to 45% of second-quarter revenue, down from 72% a year earlier [13]. Debt service is a large claim on cash: in 2024 about 32% of CoreWeave's operating cash flow before interest went to principal and interest [2].

Backlog shows where the revenue is meant to come from. CoreWeave reported about $104 billion of revenue backlog at 30 June 2026 [3]. Of its $60.7 billion of remaining performance obligations at the end of 2025, 43% was expected to be recognized within 24 months and 38% in months 25 to 48 [1]. Nscale had $2.6 billion of active and $103.4 billion of active and contracted contract value at 31 August 2026 [14].

6. Short contracts and list prices

Capacity outside long contracts priced well above long-contract rates in 2026.

• CoreWeave said that since 30 June 2026 it had continued to contract new capacity at higher prices, including three-to-six-month contracts at about $40 million per megawatt, calculated as annualized revenue divided by the power required for the clusters [10].
• Nebius reported annual contract value per megawatt of about $12 million on its 2026 base, over $20 million on second-quarter deals and over $40 million on third-quarter short-term deals [13]. It also reported more than 30% higher pricing on older-generation GPUs than in the first quarter [13].
• Nebius's on-demand price list from 1 October 2026 raised the HGX H100 from $3.85 to $4.50 per GPU-hour, the H200 from $4.50 to $5.40, the B200 from $7.15 to $8.50 and the B300 from $7.85 to $9.50 [25].

SemiAnalysis reported in April 2026 that on-demand rental capacity was sold out across GPU types and that some H100 contracts were renewing at the rates signed two to three years earlier [22]. It found that large neoclouds wanted terms of at least one year, preferred two or three years, and that five years was the common tenor for long-term offtake [22]. It also noted that in the on-demand market, utilization is a better high-frequency indicator of demand than price [22]. Under take-or-pay, the customer pays for the reserved capacity whether or not it uses it [1], so the operator's utilization risk sits mainly in capacity outside such contracts.

These prices describe a period of shortage. Rental prices in late 2025 were lower [22], and a contract signed in one market condition is repaid in whatever condition follows.

7. Credit support where no investment-grade customer exists

SemiAnalysis reported that lenders generally require an offtake contract or a backstop from an investment-grade hyperscaler before they lend for large GPU builds [20]. Several structures fill the gap.

• Supplier capacity purchase. CoreWeave disclosed a $6.3 billion order under which NVIDIA must buy residual unsold capacity through 13 April 2032, when CoreWeave's own customers do not fully use it [11].
• Supplier revenue share. NVIDIA's quarterly report for the period ended 26 July 2026 lists $36 billion of commitments under agreements with AI clouds. The clouds buy NVIDIA infrastructure, NVIDIA commits to cloud service agreements that the clouds may stop providing and sell to third parties instead, and NVIDIA participates in that third-party revenue if certain criteria are met [21]. The same report lists land, power and shell guarantees for AI cloud partners' lease obligations with maximum gross exposure of $3.5 billion. It also describes memoranda of understanding with capital providers for financing platforms designed to mobilize more than $500 billion, with limited residual-value support at NVIDIA's option [21].
• An example. SharonAI disclosed a six-year agreement with NVIDIA of up to $4.88 billion, structured through "a revenue-sharing and credit-support model", for up to 40,000 GB300 GPUs [19]. SemiAnalysis described such backstops as a minimum revenue guarantee, typically six years long, and estimated SharonAI's implied floor at about $2.33 per GPU-hour on average over the term [20].
• Hyperscaler lease backstop. For a data-center lease rather than a GPU contract, Google agreed to backstop $1.8 billion of Fluidstack's lease obligations to support project debt at TeraWulf's campus, and received warrants for about 8% of TeraWulf [24].

Each of these substitutes a stronger credit for the operator's customer. The lender still underwrites the party that pays when demand falls short.

8. Letters of intent and delivery risk

A letter of intent precedes a contract and does not replace it. Nscale says committed contract execution "typically follows the conversion of a letter of intent" [14]. Semafor reported that Microsoft and Google had both been in talks to rent capacity at Nscale's West Virginia campus. Microsoft walked away over the summer of 2026 amid a review of its data-center portfolio, and the capacity was leased to another customer [23].

Signed contracts carry their own risk until delivery. Nscale discloses that certain customers may terminate a tranche for delivery delays, and its contract terms begin only when the clusters are delivered [14]. IREN's annualized revenue targets depend on commissioning, testing and customer acceptance [16].

Implications

For buyers of compute. Long-term capacity is priced and financed around the buyer's credit. In the cases above, a stronger customer credit lowered the operator's debt cost by two to three percentage points, which gives the buyer room to negotiate on price. A prepayment funds part of the operator's hardware before service starts, so the buyer carries operator and delivery risk until acceptance. Delivery dates, termination rights for delay and acceptance tests deserve the same attention as the hourly rate.

For builders. The order of steps decides the cost of capital. In the filings reviewed, spreads of 2.25% to 3.00% over SOFR went to debt secured by GPUs and a signed contract, where the customer or the facility carried an investment-grade rating [7, 13, 14, 17]. An operator without such a customer pays more, or brings in a backstop from a supplier or a hyperscaler. A letter of intent supports neither.

For lenders. The filings show lenders underwriting two things: the contract's cash flows and the depreciated value of the hardware [2]. Coverage covenants, amortization to the contract's final payment and limited parent recourse are common [6, 14]. Two questions remain open in each deal: customer concentration (Microsoft was about 67% of CoreWeave's 2025 revenue [1]) and what the hardware earns after the contract ends.

For researchers. Payback, annualized revenue and revenue per megawatt are defined differently by each company. IREN computes annualized revenue from contracted GPU-hour prices times 8,760 hours [16]; Nebius multiplies the last month's revenue by 12 [13]. Comparisons need the definitions alongside the numbers. Contract-level and company-level results answer different questions and should be reported separately.

Limits of this analysis

• The study covers operators that publish filings or detailed releases. Private operators disclose less, and their terms may differ.
• Figures are taken as each issuer states them. Definitions of prepayment, payback and annualized revenue differ across companies and were not restated onto a common basis.
• Some facts come from reporting and research rather than from issuers: the July 2026 CoreWeave loan pricing (Octus), the Nscale negotiation (Semafor), and lender requirements and market conditions (SemiAnalysis). CoreWeave had not confirmed the July 2026 pricing when it was reported.
• Spread differences reflect the operator's track record, market conditions and deal structure as well as customer credit. The within-year comparisons reduce those other effects without removing them.
• Market conditions in 2026 included a reported shortage of rental capacity. Contract economics signed in other conditions may look different.
• The report uses no Rillor data. It is research, not investment advice, and it recommends no purchase or sale of any security.

Sources

1. CoreWeave, Inc. Annual Report on Form 10-K for the year ended 31 December 2025. U.S. Securities and Exchange Commission, filed 2 March 2026. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm
2. CoreWeave, Inc. Registration Statement on Form S-1. U.S. Securities and Exchange Commission, filed 3 March 2025. https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm
3. CoreWeave Reports Strong Second Quarter 2026 Results. CoreWeave, 11 August 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx
4. CoreWeave Closes $2.6 Billion Secured Debt Financing Facility (Exhibit 99.1 to Form 8-K). CoreWeave, 31 July 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000033/ddtl30pressrelease-ex991x6.htm
5. CoreWeave, Inc. Current Report on Form 8-K (Fifth Amendment to the DDTL 2.0 credit agreement). U.S. Securities and Exchange Commission, filed 2 October 2025. https://www.sec.gov/Archives/edgar/data/1769628/000119312525227562/d910811d8k.htm
6. CoreWeave, Inc. Current Report on Form 8-K (DDTL 4.0 Facility). U.S. Securities and Exchange Commission, filed 31 March 2026. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/crwv-20260330.htm
7. CoreWeave Closes Landmark $8.5 Billion Financing Facility (Exhibit 99.1 to Form 8-K). CoreWeave, 31 March 2026. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/ex991.htm
8. CoreWeave Closes $3.1 Billion Loan Facility, Expanding Access to Public Markets for GPU-Backed Financing. CoreWeave, 18 May 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-3-1-Billion-Loan-Facility-Expanding-Access-to-Public-Markets-for-GPU-Backed-Financing/default.aspx
9. CoreWeave Gets Caught in an AI Implosion; FOMC's Decision Unleashes Steeper Curve. Octus, 31 July 2026. https://octus.com/resources/articles/coreweave-gets-caught-in-an-ai-implosion-fomcs-decision-unleashes-steeper-curve/
10. CoreWeave Continues to Contract New Compute Capacity at Higher Prices. CoreWeave, 17 September 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx
11. CoreWeave, Inc. Current Report on Form 8-K (NVIDIA order form). U.S. Securities and Exchange Commission, filed 15 September 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm
12. Nebius Group N.V. Report on Form 6-K, Exhibit 99.2 (subsequent events: Microsoft agreement). U.S. Securities and Exchange Commission, filed 10 September 2025. https://www.sec.gov/Archives/edgar/data/1513845/000110465925088860/tm2525580d2_ex99-2.htm
13. Nebius Group Letter to Shareholders, Q2 2026. Nebius Group, 12 August 2026. https://assets.nebius.com/assets/4462517b-ce83-41f2-96ed-f2ac1bc06a05/SHLQ226%20(1).pdf-0
14. Nscale Limited Registration Statement on Form S-1. U.S. Securities and Exchange Commission, filed 18 September 2026. https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm
15. IREN Secures $9.7bn AI Cloud Contract with Microsoft (Exhibit 99.1 to Form 8-K). IREN, 3 November 2025. https://www.sec.gov/Archives/edgar/data/1878848/000114036125040072/ef20058139_ex99-1.htm
16. IREN Reports FY26 Results (Form 8-K). IREN, 27 August 2026. https://www.sec.gov/Archives/edgar/data/0001878848/000187884826000051/irenreportsfy26results.htm
17. Lambda closes $926 million senior secured term loan B facility. Lambda, 27 August 2026. https://lambda.ai/blog/lambda-closes-926-million-senior-secured-term-loan-b-facility
18. Lambda closes $1 billion senior secured fixed rate financing. Lambda, 1 October 2026. https://lambda.ai/blog/lambda-closes-1-billion-senior-secured-fixed-rate-financing
19. SharonAI Holdings Inc. Registration Statement on Form S-1. U.S. Securities and Exchange Commission, filed 12 August 2026. https://www.sec.gov/Archives/edgar/data/2068385/000149315226037453/forms-1.htm
20. Nvidia GPU Debt Backstop Unleashes the AI Project Trinity: Capital, Offtake and Datacenters. SemiAnalysis, 6 July 2026. https://newsletter.semianalysis.com/p/nvidia-gpu-debt-backstop-unleashes
21. NVIDIA Corporation Quarterly Report on Form 10-Q for the quarter ended 26 July 2026. U.S. Securities and Exchange Commission, filed 26 August 2026. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm
22. The Great GPU Shortage: Rental Capacity. SemiAnalysis, 2 April 2026. https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
23. Google and Microsoft were in talks for Nscale compute deal that went to Anthropic. Semafor, 28 August 2026. https://www.semafor.com/article/08/28/2026/google-and-microsoft-were-in-talks-for-nscale-compute-deal-that-went-to-anthropic
24. TeraWulf Signs 200+ MW, 10-Year AI Hosting Agreements with Fluidstack (Exhibit 99.1 to Form 8-K). TeraWulf, 14 August 2025. https://www.sec.gov/Archives/edgar/data/1083301/000110465925078084/tm2523008d2_ex99-1.htm
25. Nebius AI Cloud pricing. Nebius, viewed 9 October 2026. https://nebius.com/prices
26. Secured Overnight Financing Rate, most recent observations. Federal Reserve Bank of New York Markets Data API, retrieved 9 October 2026. https://markets.newyorkfed.org/api/rates/secured/sofr/last/3.json