Nvidia pauses the rent-back guarantee that made neocloud debt bankable

The line underneath the 50 percent was where the interesting term lived. Nvidia and each provider were to set a base hourly rate covering the provider's costs, and those costs explicitly included depreciation on the Nvidia chips, the data-center expense and the staffing, with Nvidia taking half of every dollar earned above that line ^[4]. Which makes Nvidia's second revenue stream subordinate to the price of its first: the more the GPU costs, the higher the base rate the provider needs to clear, and the later the revenue share starts paying ^[17].

Then the size of the thing. Nvidia's quarterly filing put $36 billion of commitments under agreements that typically run six years ^[9], or about $6 billion a year of promised rent ^[15]. Capacity rents for a few dollars an hour per GPU ^[11], and a chip billed continuously for six years brings in roughly $105,000 at two dollars an hour and $158,000 at three, which puts the commitment somewhere between 230,000 and 340,000 GPUs of capacity Nvidia offered to absorb if the provider's customers never appeared ^[16]. Idle hours push that implied count up, not down.

That absorption promise was the product ^[6]. A provider spending billions on chips and shells before it holds firm contracts could show a lender a guaranteed tenant, and the first two names attached to the AI Compute Partnership were Sharon AI and Firmus Technologies ^[10][5]. The condition, according to the people the Journal spoke to, was that the tenant also picked the other tenants: approved customers only, and preferably several small AI companies rather than one large one, which some providers refused ^[7]. The control is where the antitrust discomfort came from and it is also where the credit enhancement came from, and you do not get to keep one of those without the other.

The sequence is worth noting, since the reporting is dated August 27 ^[20]: the step-back came the week before the scale showed up in the filing, and the CFO was still describing the stream to investors on the Wednesday of that same filing week ^[14][12]. That order isn't improper; it just means the program was disclosed at its widest only after it had already been narrowed.

The pause opens a few readings, and they price differently. The first is a terms dispute, revamped or folded into something else within a quarter with commitments intact ^[2]. The second is legal, and it is the expensive one, because stripping out the control that made the guarantee bankable leaves something closer to a volume discount than a credit enhancement ^[1]. The third is balance sheet, of a piece with the scrutiny Nvidia already draws for using its own balance sheet to support projects that create demand for its chips ^[13]. This is probably wrong, but I weight the second: a guarantee that can be paused inside two months prices differently to a lender than one that cannot ^[19], and the $36 billion is designed to shrink as providers resell capacity anyway ^[9]. I am wrong if the next filing carries a bigger commitment number under a revamped structure, or if one of these providers borrows against the same buildout, at the same spread, with no rent-back behind it.