Home / Insights / Compute

Research report

Neocloud economics: contract first, hardware second

Filings show GPU cloud operators sign customers before buying hardware, then borrow against those contracts at interest rates set by the customer's credit.

Published by
Rillor
Published
Period covered
July 2023 to October 2026
Reading time
5 min read

GPU cloud operators that build at scale work in a fixed order. They sign a customer to a multi-year take-or-pay contract, collect part of its value up front and borrow against the rest. The hardware is ordered after that. The filings of the public operators describe this order in their own words, and the terms of their debt show how much the customer's credit decides.

This article reads the filings and financing announcements of CoreWeave, Nebius, Nscale, IREN and Lambda from July 2023 to October 2026, along with the credit support NVIDIA now gives some operators.

Contracts come before purchase orders

CoreWeave says it finances its infrastructure primarily through asset-level debt supported by take-or-pay customer contracts [1]. Committed contracts produced over 98% of its 2025 revenue and had a weighted-average duration of about five years [1]. Nscale's registration statement describes "building against contracted customer demand" and says it typically issues purchase orders once a committed contract is signed [14].

Nebius tied its Microsoft agreement to its financing. The parties' obligations were to begin only once Nebius confirmed it had secured the financing for the capital spending, which it expected to raise partly as debt secured against the agreement [12].

Revenue starts later again. Nscale's contract terms begin when the GPU clusters are delivered [14]. IREN says revenue ramps after commissioning, testing and customer acceptance [16].

Prepayment carries part of the build

Customers pay part of the contract before service starts. CoreWeave's weighted-average prepayment across active contracts was 15% to 25% of total contract value at the end of 2025 [1]. Nscale reported 23% as of 31 August 2026 [14].

The share rose in 2026. About 70% of the deals Nebius closed in the second quarter included prepayments, covering 50% to 60% of the related capital spending [13]. IREN reports recent prepayments of 45% to 55% of GPU capital spending [16], against 20% of contract value on its November 2025 Microsoft contract [15].

The customer sets the interest rate

CoreWeave's 2024 facility priced each loan by the credit of the contract behind it: 6.00% to 6.50% over SOFR for investment-grade customers and 13.00% over SOFR for the rest [2]. Facilities signed since then follow the same line, and two borrowers show the split inside a single year.

Facility Closed Credit behind it Pricing
CoreWeave DDTL 4.0 Mar 2026 One customer contract; facility rated A3 SOFR + 2.25%, or about 5.9% fixed
CoreWeave DDTL 5.0 May 2026 Two non-investment-grade customers; rated Ba2 SOFR + 4.50%
Nebius secured facility Jul 2026 Investment-grade customer SOFR + 2.50%
Nscale GPU facilities Aug 2026 Facilities rated investment grade SOFR + 2.375%
Lambda term loan B Aug 2026 Investment-grade offtaker; rated Baa2 SOFR + 3.00%
IREN GPU financing Reported Aug 2026 Investment-grade customer 6.0% average
IREN GPU financing Reported Aug 2026 Non-investment-grade customers 9.0% fixed

Sources: [6, 7, 8, 13, 14, 16, 17].

For scale, overnight SOFR was 3.87% on 8 October 2026 [26]. Within one borrower and one year, the credit split is worth two to three percentage points a year, on debt that can fund 90% of the GPU spending [16]. In a SemiAnalysis model, raising a neocloud's all-in debt cost from 5.62% to 10% cut pre-tax margin from 14.8% to 5.4% [20].

Lenders also tie repayment to the contract. CoreWeave's earlier facilities must be paid down to a level based on projected contracted cash flows or on the depreciated purchase price of the servers, whichever requires more [2]. Nscale's 2026 facilities mature on the date the customer makes its final scheduled payment, if that comes before the stated maturity [14].

Payback sits inside the contract

Each operator reports a payback period, and each defines it differently. CoreWeave estimated about 2.5 years, including prepayments, from contracts in effect at the end of 2024 [2]. IREN puts recent three-year contracts at about two years, measured as GPU capital spending divided by contracted revenue less direct costs [16]. Nebius reports one year and ten months for deals signed in the second quarter of 2026, down from two to three years, on a revenue basis that excludes prepayment [13].

The figures do not compare one for one. They point the same way: the contracts are priced to return the hardware's cost before they end.

Company results still show losses

Contract payback and company profit are different measures. CoreWeave's second quarter of 2026 shows the gap. Revenue was $2,575 million and adjusted EBITDA $1,510 million. Depreciation and amortization took $1,393 million, leaving an operating loss of $49 million. After net interest expense of $640 million, other income and taxes, the net loss was $626 million [3].

Nscale had $2.6 billion of active and $103.4 billion of active and contracted contract value at 31 August 2026. It reported a net loss of $1,020.1 million on revenue of $140.6 million for the first half of 2026 [14]. Large contracts sit on the books well before their revenue does.

Short contracts and list prices move with the market

Capacity outside long contracts sells at different prices. CoreWeave said in September 2026 that it had signed three-to-six-month contracts at about $40 million per megawatt of annualized revenue [10]. Nebius reported $20 million to $25 million per megawatt on its large second-quarter deals [13]. Nebius also raised its on-demand list prices from 1 October 2026, with the HGX H100 moving from $3.85 to $4.50 per GPU-hour [25]. SemiAnalysis reported in April 2026 that on-demand capacity was sold out across GPU types [22]. Prices set in those conditions describe those conditions.

Credit support where no investment-grade customer exists

Operators without such a customer have found credit elsewhere. NVIDIA agreed in September 2025 to buy CoreWeave's residual unsold capacity through April 2032 under a $6.3 billion order [11]. Its quarterly report for the period ended 26 July 2026 lists $36 billion of commitments under agreements with AI clouds, with a share of the clouds' third-party revenue when certain criteria are met [21]. SharonAI disclosed a six-year agreement with NVIDIA of up to $4.88 billion under a revenue-sharing and credit-support model [19]. For data-center leases, Google backstopped $1.8 billion of Fluidstack's lease obligations at a TeraWulf campus to support project debt [24].

Letters of intent do not finance a build

Nscale says its committed contracts typically follow the conversion of a letter of intent [14]. Semafor reported in August 2026 that Microsoft had been in talks to rent capacity at Nscale's West Virginia campus and walked away during a review of its data-center portfolio [23]. Signed contracts carry delivery risk too: some Nscale customers may terminate for delays [14].

Read the full report

The full report, with each facility's terms, the payback definitions side by side and the complete source list, is available as a PDF.

The full report

Neocloud economics: contract first, hardware second

The research behind this article, with every source listed. 16 pages.

  1. Summary
  2. Background
  3. Data and method
  4. Findings
  5. Implications
  6. Limits of this analysis
  7. Sources

Notices

Rillor is not a registered investment adviser or commodity trading advisor and does not provide investment or trading advice. Research, forecasts, data and software described on this site are for research and engineering use. Nothing here is an offer or recommendation to buy or sell any security, commodity interest or digital asset. Past or simulated results do not indicate future results.

NVIDIA, AMD and related product names are trademarks of their owners and are used only to identify products. No affiliation or endorsement is implied.

Ask about the data behind this report.

Talk to us