Cloud Computing
PrivateThe first vertically-integrated AI merchant utility — owns power + real estate + compute, not a neocloud — and the best-financed private in the buildout; but it is a levered, tenant-concentrated bet on hyperscaler demand staying elastic, and the Stargate expansion collapse + Project Jade exit are the first cracks in that assumption. WATCHING into an IPO that is 18–36 months out.
Research
The verdict
The first vertically-integrated AI merchant utility — owns power + real estate + compute, not a neocloud — and the best-financed private in the buildout; but it is a levered, tenant-concentrated bet on hyperscaler demand staying elastic, and the Stargate expansion collapse + Project Jade exit are the first cracks in that assumption. WATCHING into an IPO that is 18–36 months out.
Crusoe is best understood not as a cloud company but as America's first vertically-integrated AI "merchant utility" — it originates the land, generates the power behind-the-meter, builds the data-center shell, and (at some sites) also operates the GPU cloud on top. That full-stack ownership is the whole thesis; competitors lease power from the grid, Crusoe makes its own.
Origin and the pivot. Founded August 2018 in Denver by Chase Lochmiller (CEO) and Cully Cavness (COO/President) to capture flared methane from oil wells and convert it into on-site power for computing — "Digital Flare Mitigation" (DFM). From 2018–2022 that power ran bitcoin ASICs. The inflection: in 2023 Lochmiller took a $200M loan to buy NVIDIA H100/A100s and stand up Crusoe Cloud; in June 2024 the company entered the hyperscale data-center business with the Abilene, Texas announcement; and on 25 March 2025 it sold the entire bitcoin-mining + flare-gas business to NYDIG (425+ modular datacenters, 135 employees, seven US states + Argentina) to go 100% AI. Crusoe today is a pure AI-infrastructure company that kept the energy-first DNA but shed the crypto.
What it sells now (three lines):
Key economics headline: ~4.9 GW of AI infrastructure contracted as of June 2026, against a 40GW+ development pipeline. Contract structure at the flagship is a ~15-year Oracle lease (asset-backed) — recurring, take-or-pay-like, utility-grade cash-flow quality.
Map upstream → Crusoe → end customer, named at every node:
Chain verdict: the differentiated node is power generation + electrical (owned/insourced), not GPUs (bought like everyone else). The single most dangerous dependency is gas turbines + gas supply + emissions permits — the thing that lets Crusoe move fast is also the thing regulators and lead-times can choke.
Crusoe's moat is structural cost + speed from vertical integration, and it is real but capital-gated:
Bargaining power: mixed. Over its cloud customers (startups) Crusoe has pricing power via cost advantage. Over its anchor tenants (Oracle/Microsoft/Meta) it has little — they are the credit that makes the debt bankable (Lens 5), so the tenant holds the whip. Over suppliers, NVIDIA and the gas-turbine OEMs hold the power (allocation + lead times). Net: the moat is against other developers, not against its own customers or suppliers.
Moat durability risk: it is a capital moat, and capital moats erode if the cost of capital rises or hyperscaler demand softens — exactly the pre-mortem in Lens 12.
No audited segment reporting exists (private). Reconstructed revenue-mix by year, all ``, unaudited:
| Year | Total rev (approx) | Bitcoin/DFM | AI cloud + infra | Note |
|---|---|---|---|---|
| 2023 | ~$276M | ~14% AI cloud + ~6% other | Pre-pivot; crypto-dominant | |
| 2024 | ~$276M | Crossover year | ||
| 2025E | ~$998M–$1B (+~262% YoY) | ~0% (divested Mar 2025) | ~100% AI | First pure-AI year |
| 2026E | ~$2B | 0% | 100% AI |
Segment reading: the story is a clean, violent business-model swap — bitcoin fell from ~80% to zero in ~two years while AI infra went from a rounding error to the entire company, and total revenue is compounding ~2–2.6x annually off the AI base. Within the AI base, the mix is shifting from Crusoe Cloud (GPU rental, higher-velocity/lower-margin, price-compressing) toward hyperscale leasing (Abilene/Microsoft — asset-backed, long-dated, utility-margin). That mix shift toward leased-infrastructure is the bull's "recurring quality" argument and the bear's "capital-intensity + leverage" argument — same fact, two readings.
Crusoe is one of the best-financed private companies in the entire AI buildout, running two parallel capital stacks: corporate equity (the venture rounds) and project-level debt (the campus financings). Keeping them separate is essential to reading the risk correctly.
Corporate equity rounds (all `, unaudited):**
| Round | Date | Amount | Post-money val | Lead(s) |
|---|---|---|---|---|
| Series C | Apr 2022 | $350M | — | G2 / Valor era |
| Series D | Dec 2024 | $600M | ~$2.8B | Founders Fund |
| Series E | Oct 2025 | $1.375B (oversubscribed) | >$10B | Valor Equity + Mubadala Capital |
| Series F (in talks) | ~Jul 2026 | ~$3B (reported, not closed) | ~$30B (≈3x in ~8mo) | TBD — |
Series E syndicate is a who's-who: NVIDIA, Fidelity, T. Rowe Price, Franklin Templeton, Altimeter, Tiger Global, Founders Fund, Salesforce Ventures, Ribbit, Winklevoss, Supermicro, Blue Owl, + Saquon Barkley. The valuation slope is the headline: ~$2.8B (Dec-24) → >$10B (Oct-25) → ~$30B (Jul-26) — roughly 10x in ~19 months, one of the steepest private mark-ups in the cycle.
Project-level debt (Abilene, the flagship) — this is separate from corporate leverage and largely asset-backed / non-recourse-style, wrapped by tenant credit:
Burn / cost signals: annual interest expense projected ~$300M by end-2025, and debt service is utilization-dependent — the model works only if the leased capacity stays full. Revenue ~$1B (2025E) against that interest load is serviceable while tenants pay; it is the pressure point if a tenant walks.
Traction / unit economics (Phase B add): Crusoe Cloud ~946K GPU capacity, 3.4GW, 99.5% SLA, "100% customer-satisfaction" (self-reported Oct-25). Spark unit economics disclosed via the Energy Vault framework: expected EBITDA of ~$1.5–2.0M per MW per year — a rare hard margin data-point, implying a 25MW Spark deployment ≈ ~$37–50M annual EBITDA.
Verdict: capital is not the constraint — Crusoe has arguably solved financeability better than any private peer. The constraint is demand certainty against a fixed, levered cost base.
No earnings calls (private). Substitute: founder posture across interviews/podcasts (Not Boring, McKinsey, Bain Capital Ventures) and the arc of company messaging.
Consistent throughline (2018→2026): "energy-first." Lochmiller's framing has been remarkably stable — the belief that power, not chips, is the binding constraint on AI, and that whoever controls stranded/behind-the-meter generation controls the buildout. The tone shift worth flagging: 2023–24 messaging leaned on climate/flare-mitigation ("we're cleaning up methane"); post-NYDIG-divestiture 2025–26 messaging is unabashedly "AI factory company" and gas-forward (behind-the-meter gas turbines, SMR optionality) — the ESG halo has faded as the gas footprint grew (see Lens 10/13). Recent tell: the deliberate move to de-risk single-tenant exposure narratively (multi-tenant hub framing, Microsoft + Meta added) after the Oracle/OpenAI expansion collapse — management is visibly managing the concentration story.
"Stopped saying": the flare-mitigation / carbon-avoidance stats (21 Bcf captured, 2.7M tons CO2 avoided) that dominated 2024 decks are now legacy — the company sold that business.
Team-quality signal: senior hires (Erwan Menard ex-Google Cloud SVP → Crusoe Cloud; Matt Field, Chief Real Estate Officer) show a deliberate build-out from founder-led startup toward an institutional operator — appropriate for a company signing 15-year, multi-billion leases.
Syndicate quality — the IPO-proximity tell. Crusoe's cap table is dense with crossover / public-market funds, the classic pre-IPO signal: Fidelity, T. Rowe Price, Franklin Templeton, Altimeter, Tiger Global all entered at Series E. Add sovereign capital (Mubadala Capital, co-lead) and strategic NVIDIA + Supermicro + Salesforce. Structured-credit/real-assets giants Blue Owl (~$192B AUM real-assets) and JPMorgan anchor the project debt. This is not an early-stage cap table — it is a pre-IPO one.
Secondary marks (private-market price discovery):
Public-comp valuation frame (for the eventual IPO — all ``, NTM EV/Sales, sourced to Contrary Research's Oct-2025 comp set unless noted):
| Comp | EV/Sales (NTM) | Note |
|---|---|---|
| CoreWeave | ~9.3x (Contrary) / ~12–13x fwd (electroneconomics) | Closest listed analog; IPO'd Mar-25 ~$35B on ~$5B rev |
| Nebius | ~24.3x | Growth-premium neocloud |
| Equinix | ~10.2x | Colocation incumbent |
| Digital Realty | ~12.6x | REIT DC |
| Vertiv | ~6.1x | DC power/cooling equipment |
| Crusoe (private) | ~10x on ~$1B 2025E rev → ~$10B (Oct-25); ~$30B in-talks ≈ ~15x on ~$2B 2026E |
Read: at ~$10B on ~$1B revenue Crusoe was priced in line with CoreWeave/Equinix (~10x). The jump to ~$30B on ~$2B (~15x) re-rates it toward the Nebius growth-premium band — the market is paying for the pipeline/optionality (40GW+) and the "utility, not neocloud" recurring-revenue framing, not for delivered financials. Whether that premium is earned is the Lens 12/13 fight. No fabricated multiples — every figure above carries an outlet + date or shows the arithmetic; where a precise private multiple is genuinely unknowable it is derived and labeled ``.
For a private, "what moved the mark / narrative" replaces "what moved the stock." Key >material events, all ``:
Pattern: the mark reacts to (1) mega-financings, (2) named anchor tenants, and (3) tenant/pipeline defection (Oracle expansion scrap, Jade exit). The 2026 cracks did not stop the valuation tripling — the private mark is being driven by pipeline narrative + capital access, and is lagging / insulated from the operational stumbles a public tape would have punished. That gap is the single most important thing to watch.
Management verdict: high-quality, well-credentialed founder team with a demonstrated ability to read the market and pivot hard — the biggest asset in the file. The watch-item is capital-allocation discipline at 40GW-pipeline scale: the Jade exit shows they will (rightly) walk from bad projects, but it also shows the pipeline is softer than the headline number implies.
No audited statements exist, so classic forensic-accounting analysis is not possible — that absence is itself the first flag (a ~$30B company with no public financials, servicing ~$300M/yr interest on utilization-dependent leases). What can be assessed:
n/a — private, not disclosed.Regulatory findings (required sub-section) — from regulatory/regulatory-findings.md + web:
Private-watch grounding: stage late, ipo_readiness 3/5 (late-stage; scale: 4=pre-IPO/secondary-active, 5=S-1/imminent), leads Founders Fund + Mubadala, catalyst "AI datacenter + power (Abilene/Stargate)." This dossier is the write-back trigger — the entry currently has dossier: null and, on the evidence below, its ipo_readiness should be bumped 3 → 4 (crossover funds are in, secondaries are active at ~$123/sh, a ~$30B round is in market). (Registry edit is out of wave scope — flagged for Connor, not executed here.)
Milestones that unlock an S-1:
Estimated window: ~2027–2028 (12–36 months). A CoreWeave-style direct IPO is the base case; a secondary-heavy/continuation path (Mubadala/sovereign takeouts, structured secondaries) is a live alternative given how much sovereign + crossover money is already in.
Path-to-tradeable comps precedent: CoreWeave (the closest analog) went private-to-public in ~18 months and listed ~$35B on ~$5B revenue. Crusoe at ~$30B on ~$2B (2026E) is ahead on multiple, behind on delivered revenue vs CoreWeave-at-IPO — consistent with a listing 12–24 months out once revenue catches the mark.
(No forecast.ts create — this is a --watchlist unattended run; per SKILL the Brier forecast is only logged on a genuine committed base case, and for a private the tracked binary would be an IPO-timing/round-close event, not an EPS line. Deferred to a human-gated /thesis pass.)
Bull case. Crusoe is the only company that owns the entire scarce-asset stack of the AI buildout — power generation, interconnect rights, real estate, and compute — at a moment when power, not chips, is the binding constraint. It is 3–5 years faster to energized capacity than any grid-dependent competitor, has the best financeability of any private peer (JPMorgan + Blue Owl + Mubadala + a crossover-stacked cap table), insources its worst supply-chain bottleneck (electrical), and holds the lead-developer position on the single largest AI-infra program in America (Stargate). Revenue is compounding ~2–2.6x/yr off a pure-AI base with a utility-quality long-dated lease backbone forming underneath. If AI compute demand stays anywhere near consensus, Crusoe is a generational infrastructure franchise — and the ~$30B mark could look cheap against a 40GW pipeline. Contrarian earnings-surprise lever: the Spark modular line ($1.5–2M EBITDA/MW/yr) is an under-appreciated, higher-margin, faster-cycle second act that hedges campus-obsolescence risk.
Bear case (permanent-impairment risks).
Pre-mortem (18 months out, thesis broke): AI-capex growth decelerated from consensus; two anchor tenants pushed out or renegotiated leases citing "changing demand forecasts" (Oracle already showed the template); utilization on 6GW of simultaneously-constructing capacity fell below debt-service breakeven; the IPO window shut on rates; the ~$30B mark reset toward ~$12–15B (a CoreWeave-multiple on delivered ~$2B revenue). The pipeline, revealed as optionality not backlog, stopped supporting the premium.
Are multiples too high? At ~15x forward sales for a levered, capital-intensive, tenant-concentrated developer, the mark prices the pipeline as if it were backlog. It is not (Jade proves it). The premium over CoreWeave (~10x) is only justified if the "utility recurring-revenue" framing is real and the pipeline converts — an open question the 2026 stumbles argue against, not for.
Contrarian view (what the market refuses to see): the crowd is debating "neocloud vs utility." The thing being underweighted is that Crusoe's greatest asset — vertical integration into gas generation — is also its greatest regulatory and demand-cyclicality liability, and that the private mark has decoupled from operational reality (tripled through a cooling failure, a tenant defection, and a flagship-project exit). The public market, when it finally prices Crusoe, will re-couple them — and that re-coupling is the risk the ~$30B round is papering over.
Dismantling the bull case:
A real, cash-generating neocloud retrofitter trading at ~18x trailing sales on a single $865M Nscale contract and a still-71%-Bit-Digital-controlled cap table — the build is genuine, but the multiple already prices the NC-1 inflection that hasn't happened yet.
A merchant-power balance sheet wearing a regulated-utility's contracted growth — long-dated nuclear PPAs to AWS/Meta de-risk the AI-demand story, but the GAAP P&L is hostage to hedge mark-to-market and the equity carries ~3.4x the net debt of Constellation. Cheapest large-cap way to own the data-center power trade if (and only if) ERCOT/PJM load growth shows up; bull at ~10x forward EBITDA, but leverage + commodity beta make it the high-volatility expression, not the safe one.
The default arms dealer of the AI buildout — a real moat compounding a $15B backlog into 30% organic growth, but priced at 82x for perfection while insiders sell 65:0 and EMEA orders are already cracking.