Nebius backs $775m of senior secured debt with GPUs and customer contract cash flows

In ordinary procurement the buyer checks whether the supplier has the resources to perform, and in the neocloud version the contract itself may be what gives the supplier those resources, which is the reversal PYMNTS puts at the centre of its argument ^[8]. Nebius has published the cleanest example of it: a first senior secured financing of roughly $775m backed by GPU infrastructure and contracted customer cash flows ^[6], with the company saying the debt and the cash generated by the associated customer contract together cover more than 100% of the capital expenditure for the underlying deployment ^[7]. Coverage above 100% means the equity cheque for those GPUs is arithmetically zero, replaced by a lender's read on one customer's willingness to keep paying.

CoreWeave shows the same machine at larger scale, with $5.1bn of equipment and software financing outstanding as of June 30 alongside a set of delayed-draw term loans ^[4], and a May facility that added $3.1bn of borrowing capacity ^[5], so about $8.2bn between the disclosed stock and the newest availability ^[14]. Availability is not drawn debt, and the gap between the two is precisely what a capacity buyer now has to price, because the metric that matters has moved from cost per GPU-hour to the probability that the promised GPU-hour will exist when the enterprise needs it ^[10].

The demand backdrop makes the scarcity real: the Bank for International Settlements estimated the five largest global technology companies will invest more than $1 trillion in AI across 2025 and 2026 ^[3], which is above $500bn a year, or roughly $1.4bn a day ^[15], and the spending lands on GPUs, networking, data centres and electricity before much of the revenue can be recognised ^[17]. Cost per GPU-hour is a price; delivery probability is what converts it into a value. Put arbitrary numbers on that to see the shape of it: a rate ten per cent under a hyperscaler quote that shows up nine times in ten is not a ten per cent saving, it is a saving contingent on never needing the tenth deployment, and the replacement cost in a scarce market is the part nobody models.

In the benign reading, contract-backed financing does what it does in other capital-intensive industries and lowers the cost of capital, which PYMNTS notes does not inherently make the model riskier ^[12], and the buyer gets cheaper compute because its own credit was bankable. A second reading has a deployment milestone slip, and the buyer discovers which entity owns the GPUs, which leases the facility, which holds the power contract, which owes the debt and what rights lenders have over the assets ^[13]. A third has compute getting cheaper faster than the multiyear commitment amortises, which is the duration mismatch of buying hardware that can depreciate economically faster than the contracts financing it ^[11].

My view is that entity mapping is the item to underwrite before price, and the checklist of whether GPUs are installed, ordered or still dependent on financing is a treasury exercise rather than a procurement one ^[9]. The capital-structure evidence is all provider disclosure, and no enterprise is named as having run this diligence ^[16]. The thesis fails if contracted deployments land on schedule and GPU resale values hold above the amortisation of the debt secured on them, in which case the cheaper financing shows up as cheaper compute and the extra diligence was a cost with no claim attached.