Nebius's customers fund 50-60% of capex upfront on four $1bn+ deals

The prepayment lands in the cash flow statement well before it lands in revenue: $2.3bn of operating cash flow in a quarter that recognised $582.3m of it ^[8][1], just under four times the revenue actually booked ^[2]. Four contracts at roughly a billion apiece, with half to six-tenths of the associated capital expenditure paid on signature ^[3], puts somewhere between $2.0bn and $2.4bn of customer money in the door ahead of delivery ^[1]. Call that the cheapest tranche of financing available to a company that spent $5.7bn on capex in the same three months ^[9] and therefore had to find $3.4bn somewhere other than operations ^[3].

The arithmetic on the rest of the year is where prefunding stops being a talking point. First-half capex above $8.1bn against a reaffirmed full-year guide of $20bn to $25bn leaves $11.9bn to $16.9bn for the second half ^[9][10][4], and the $8bn of cash on hand covers between 47% and 67% of that ^[8][5]. The remainder comes from debt, equity, or more customers agreeing to fund GPUs they will never own.

Which is why I read the auction as the weaker of the two pricing signals. In the Q2 material summarised by Quartr ^[15], management says a capacity auction cleared at a record 15% above previous highs on Blackwell parts ^[4], and that is one auction, self-reported, with no volume attached, from the party that gains most if you believe compute is getting dearer. The contract terms are sturdier evidence: a buyer funding 50% to 60% of somebody else's hardware is telling you it values a 2027 delivery slot above its own cash ^[3].

It could go otherwise in a few ways. The four counterparties may be fewer than four names, in which case the prepayment is credit concentration wearing a working-capital costume. The raised 5 GW contracted-power target and the plan to deploy more than 1 GW a year from 2027 ^[6] is new supply, and if enough operators do the same then whoever locks 2027 capacity at today's cleared price has overpaid. And the 2026 run-rate guide of $7bn to $9bn against $3bn at June-end needs 2.3 to 3 times growth in six months ^[11][10][6], so higher pricing is already carried inside the guidance rather than tested against it.

This is probably wrong, but the view I would take is that buyers should contract early, because a seller collecting cash up front can afford to wait for a better price. The counter sits in the same numbers: capex of $20bn to $25bn against group revenue of $3bn to $3.4bn is six to eight times revenue ^[10][7], which describes an obligation more than it describes pricing power, and obligations get discounted when the deployment wave lands.

The cash is going into hardware already spoken for, not into acquisitions or a price war further up the software stack, at that multiple of revenue ^[7]; the $692.9m swing from a $502.5m profit to a $190.4m loss on continuing operations ^[5][8] is the depreciation of a decision already taken.