On September 2, Broadcom's CFO disclosed — verbally, in an earnings call — that the company had provided potential "residual value guarantees" to AI labs, designed "to bridge the gap between their current cash flow and the significant upfront investments." It was the clearest single marker yet that the AI capex buildout was developing a vendor-financing layer, the mechanism that defined the telecom bubble of 1999. But it was still a disclosure without a number.
Seventeen days later, the number exists — in an SEC filing.
The filing
Nscale, the British AI datacentre group that supplies compute capacity to Anthropic, filed publicly on Friday for a New York IPO under the ticker NSCL, seeking to raise up to $3 billion with Goldman Sachs, JPMorgan and Morgan Stanley as lead bookrunners. The financials are stark: a $1.02 billion net loss on $140.6 million of first-half revenue — a burn multiple above 7x — against $56.4 billion in remaining performance obligations and more than $8 billion of debt. A March funding round valued the company at $14.6 billion.
The line that matters is this one: Nvidia has agreed to guarantee up to $860 million of Nscale's obligations on a Texas data-centre lease.
Why the guarantee layer matters
A vendor guaranteeing its customer's obligations is not a detail. It is the structure through which supplier financing becomes demand creation. When Nvidia sells chips to a neocloud, and simultaneously stands behind that neocloud's lease commitments, the revenue is partially self-referential: the demand that justifies the capex exists partly because the supplier has underwritten it. This is the exact mechanism — vendor financing bridging cash-flow gaps until utilization materialized — that froze telecom capex for a decade when credit markets stopped absorbing the paper in 2001.
The AVGO disclosure suggested the pattern existed. The Nscale filing documents it, at the neocloud layer, with a number attached. The evidence standard for the funding-spiral thesis has moved from inference to documentation.
The RPO quality question
$56.4 billion of contracted backlog against $140.6 million of recognized revenue is a 400x ratio. Backlog of that scale is only as good as the solvency of the counterparties and the guarantees standing behind the contracts. The question the market should now ask of every AI-infrastructure RPO: how much is underpinned by circular flows — customers backed by the supplier, commitments guaranteed by the chipmaker, revenue loops that net out to vendor financing? AVGO's disclosure promised quantification in future filings. Nscale delivered the first one. Others will follow.
The IPO window is bifurcating, not closing
The same week this filing landed, the labs layer stayed shut: Anthropic's mega-IPO remains deferred and OpenAI's listing timeline has slipped, raising the stakes for SoftBank's $50 billion SB Energy data-centre IPO. Holtec pulled its offering days before Nscale filed. The pattern is not a window closing — it is a window sorting. Public capital will still fund compute infrastructure at 7x burn multiples, but it wants the layer closest to contracted demand, not the layer closest to model risk. That is not froth deflating. It is froth reallocating — which is how late-cycle funding spirals typically extend before they break.
What we are watching
Three markers decide whether the guarantee layer stays benign or becomes the stress point: (1) IG spreads on AI issuers versus the 250bps stress threshold — the cost of the debt that funds the buildout; (2) NSCL's pricing and first trade — whether the market absorbs a $3 billion raise against $8 billion of existing debt at a $14.6 billion private mark; (3) the Anthropic S-1 whenever it lands — capex commitments and guarantee analogs at the layer where the money actually concentrates.
The thesis is unchanged: funding-stack strain is the watch item, not yet the break. The break condition remains IG spreads blowing through 250bps alongside hyperscaler capex running above 1.8x operating cash flow. What changed on Friday is narrower and more important: we no longer need to argue that the guarantee layer exists. We can read it.