Daily Brief, August 11th 2026

Editorial

The vendor finances the buyer, again

Nvidia announcing $500 billion of third-party financing for purchases of Nvidia products is a structure with a long pedigree, and the pedigree is not reassuring. Lucent and Nortel spent the late 1990s lending their customers the money to buy their own switches; the sales were real, the revenue was booked, and when the borrowers failed, the equipment came back worth a fraction of the receivable. The 1840s railway promoters sold bonds against traffic projections on lines not yet built. In every version, the trade works until the asset's residual value is tested, and the test always arrives.

Here is what separates this from the lazy bubble call, and it deserves stating precisely. Nvidia is not lending its own balance sheet. It has convinced Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to raise dedicated vehicles, which means professional credit investors will be pricing the risk rather than a vendor's sales desk. That is a genuinely better structure than Lucent's. The capital is also walking in with open eyes: the BIS named AI capex and circular financing among its top systemic risks in June. Nobody buying this paper can claim surprise.

But look at the load-bearing claim in Huang's own words: Nvidia compute offers the "longest life" and is "fungible and transferable." The entire structure prices off that sentence. A data center shell is infrastructure and amortizes over decades. The chips inside it are inventory that Nvidia itself makes obsolete on a roughly annual cadence, and Nvidia is the single most motivated party on earth to keep that cadence fast. The company arranging the financing controls the depreciation schedule of the collateral. That is not a conflict Lucent had.

Anthropic's side of the trade is the rational response to the same fact. Theseus and the reported Riot lease convert what would be depreciating assets on a pre-IPO balance sheet into rent, and the grid-cost commitments buy down the political risk that yesterday's brief flagged as electoral. If you believe your compute needs will keep compounding, renting from capital that priced a 20-year tenor is cheap. If they do not compound, the leases become the overhang. Either way the lab's prospectus looks cleaner than the economics underneath it, which is worth remembering when the S-1s surface.

What would change my mind: published terms. If the definitive vehicles carry short tenors, meaningful spreads over infrastructure paper, or explicit residual-value support from Nvidia, then the risk was priced and this is just project finance doing its job. If $500 billion of GPUs get funded at infrastructure-grade terms on a "fungible and transferable" narrative, the mispricing is structural, and F4 on the ledger says credit is where it shows first. I have logged 2026-08-11-F1 to hold the headline number itself to account: MOUs are press releases until a vehicle closes.

CoreWeave reports in a few hours carrying the whole argument in miniature: $99 billion of contracted backlog, a nine-figure quarterly loss, and a depreciation schedule its critics already call generous. Watch the cost of debt line, not the revenue beat.

Elias Marchetti