Divide the $19.4bn by the more than 100,000 GB300s it covers and you get something under $194,000 a chip of contracted spend [2][13], which is fleet-purchase money moving through a services agreement: Microsoft gets the capacity, Nebius keeps the asset, and the lenders behind the asset get a Microsoft contract plus an Nvidia guarantee as collateral [6].
That one contract is about 59% of the more than $33bn Microsoft has committed across neocloud providers [1][14], and roughly 1.9 times what CoreWeave's most recent reported quarter of $2.575bn annualises to [3][16]. CoreWeave's 112% year-on-year growth implies a prior-year quarter near $1.215bn, so about $1.36bn of new quarterly revenue arrived in twelve months [15]. That is real cash rather than a booking.
Cryptobriefing reports neocloud GPU instances running 60-70% below equivalent AWS, Azure and Google Cloud offerings [4], which inverted means the hyperscalers are charging something like 2.5 to 3.3 times the neocloud rate for comparable silicon [17]. A vendor winning volume at a third of the incumbent's price is not usually the vendor dictating terms; the vendor's structural advantage here is that it carries none of the general-purpose overhead, and none of the legacy architecture, that the incumbents amortise across everything else [11]. The same publisher frames all this as neoclouds holding an increasingly powerful negotiating position and the industry's power dynamics flipping [18]. What the material actually documents is volume and a discount. It does not show contract minimums, prepayment structures, take-or-pay language, gross margins on these specific agreements, or which side conceded on price, and without those the leverage claim is a reading rather than a finding.
The party whose position visibly improved is Nvidia, which supplies priority hardware access, takes equity in neocloud companies, and offers revenue backstops on unsold capacity [5]. That last item is the financing mechanism: it converts an idle-GPU risk into a guaranteed cash flow a lender will underwrite [6], and it buys Nvidia a customer base broader than the handful of hyperscalers that currently absorb most of its output, along with competitive tension among those buyers [7].
The residual-value arithmetic is where this gets decided. Architectures now turn over on roughly annual cycles, and a fleet bought today could be worth meaningfully less in 18 months on price-per-performance grounds even though nothing has broken [8]. Whoever owns the metal eats that, funds the replacement out of capital markets, and does it while carrying significant existing debt [9] and bidding against its own customers for power in the locations that have any [12]. My read, and it rests on one publisher's reporting: the 60-70% discount is partly the price of taking depreciation and power procurement off a hyperscaler's books, and the concentration of Microsoft and OpenAI revenue in a few neoclouds [10] is what makes it a transfer of risk rather than a shift in bargaining power.
Two things would show that read to be wrong. If these contracts turn out to carry hard minimums and upfront prepayments, the neoclouds are financing fleets with customer money and the leverage is real. If renewals reprice up rather than down as GB300 gives way to the next part, the scarcity is structural rather than a window. The debt schedules will not wait for either answer.
