Why this is in the vault
A live, numbers-heavy case study of how the AI-compute buildout is actually financed (take-or-pay contracts, customer prepayments, vendor financing circularity) — the fourth entry in Mostly Metrics' recurring S1-breakdown format and directly useful as a data point for the capital-cycle investing thesis.
The core argument
Nscale, a two-year-old UK "neocloud" (rents Nvidia GPUs by the hour, same category as CoreWeave/Nebius/Crusoe/Lambda), filed to list on the NYSE (ticker NSCL) targeting a reported $35B valuation against a $14.6B March private round. The headline number — $103.4B in total contracted value, mostly from Microsoft ($43.8B) and Anthropic (~$44.6B) — is a ceiling, not revenue: only $2.6B is "live" because Nscale's take-or-pay contracts don't start billing until a cluster is built, powered, and accepted. Gross margin is negative (-159% after depreciation), net loss is $1.02B (much of it non-cash warrant markups), yet operating cash flow is +$1.7B because customers prepay 23% of contract value up front — deferred revenue went from $2.0B to $6.5B in six months. Nvidia sits on every side of the deal: investor ($2.2B), guarantor of $860M of data-center rent, and a $1.2B GPU-capacity customer — paid in warrants that have since appreciated, feeding the non-cash loss. CJ flags five red flags: going-concern language, material weaknesses in accounting controls (new auditor, KPMG, disclosed its own independence issue), 85% customer concentration in two logos, ~$7.3B of growth paid for in stock, and local community opposition to sites. At $35B the deal prices at ~90x current run-rate and ~10x where run-rate could plausibly land a year out — CoreWeave, the direct comp, priced its IPO at ~12x trailing revenue and still traded at 4.5x what its first public year actually delivered.
Mapping against Ray Data Co
Directly feeds the Markov capital-cycle investing thesis (project_investing_markov_capital_cycle): the founder's placement of the current chip-fab/memory cycle in "Phase 2" (capacity announcements, pre-glut) needs exactly this kind of primary-source data — contract structure, financing terms, customer-concentration math — to calibrate when capacity commitments outrun real demand. Nscale's S1 makes concrete what the memory-cycle thesis treats abstractly: (1) take-or-pay backlog is a promise, not revenue — only 2.5% of Nscale's contracted value is live; (2) the Nvidia-as-investor/guarantor/customer loop is the vendor-financing circularity the thesis should be pricing as a risk factor, not treating chip-maker equity stakes in AI-infra customers as a clean demand signal; (3) CoreWeave is already a named comp in the model — Nscale is a second neocloud data point on how the market prices unbuilt-capacity backlog (Nscale trading ~7x CoreWeave's backlog multiple on a twelfth the revenue), which sharpens the "when does the market stop rewarding backlog over delivered revenue" signal the Phase 2→3 transition depends on.
Related
- [[2026-09-06-mostlymetrics-oura-ipo-s1-breakdown]]
- [[2026-05-12-mostlymetrics-cerebras-ipo-s1-breakdown]]
- [[2026-02-28-ark-invest-stock-commentary-lunr-crcl-bfly-crwv-xyz]]
- [[project_investing_markov_capital_cycle]]
⚠️ Sponsorship
Brex — top-of-issue paid placement ("Mostly metrics is proudly powered by Brex"), pitching corporate cards + AI-native expense workflows, unrelated to the Nscale content itself. Standard CFO-audience ad slot, not a conflict with the S1 analysis. House-promo — a London happy-hour RSVP (CJ's own event, October 15) is self-promotion, not a paid third party; flagged for completeness, no bias implication for the editorial content. No Abacum/Koyfin/Rillet/Mostly Talent presence in this issue.