Crusoe: The energy-first AI factory company

The Energy-First AI Factory Company

Everybody’s a builder in this AI era, Crusoe is not trying to be another general-purpose cloud with a GPU sticker slapped on it, It is trying to own the stack that builders actually hit when they leave the demo and enter production. Two customers, one bottleneck:

1. Infrastructure engineers who want clusters, scale, and every last unit of performance.

2. Developers who want agents that solve a business problem and need the metal abstracted.

The company line that matters is this: training is still essential, but inference is becoming the dominant service. Value shows up when a query is served and an agent runs. That workload is always on, always scaling, always under latency and reliability pressure. Inference is where AI becomes a product instead of a science project.

Crusoe’s claim is vertical, not incremental: purpose-built data centers designed around how AI is actually used, plus in-house capability to source and produce energy, build the campus, and run the cloud on top. Their phrase is the whole thesis — electrons to tokens. That is not marketing fluff in 2026. Grid interconnection, firm power, and time-to-MW are the binding constraints of the AI buildout. Whoever connects power to reliable inference at scale owns the scarce layer.

What Crusoe Is:Crusoe started in 2018 converting stranded / flared oilfield gas into compute. First use case was Bitcoin. That chapter is over, The mining and Digital Flare Mitigation business was sold to NYDIG so the company could go all-in on AI factories and Crusoe Cloud. Today the company sits in three stacked businesses:

• Energy-first campus development for hyperscalers and frontier labs

• Purpose-built AI data centers (gigawatt campuses and modular Spark units for distributed inference)

• Crusoe Cloud — NVIDIA-native training and managed inference, sold to labs, enterprises, and now large financial buyers

Customers and contracted work already include OpenAI / Stargate in Abilene (~1.2 GW), a separate ~900 MW Abilene campus for Microsoft, Meta sites in Texas and Missouri, Oracle, and a reported ~$13 billion, five-year cloud contract with Jane Street.

Management has cited roughly 4.9 GW contracted against a 40–45 GW pipeline. NVIDIA has been in the story from the beginning as partner and investor. The stack is being built for when systems fail, not if they fail: fault-tolerant orchestration, fleet management, AI-expert-to-AI-expert support. Recent GTC-cycle work extends that into Vera / Rubin generation hardware, Omniverse DSX factory design, and managed inference (Nemotron, Dynamo, proprietary tokenizer / MemoryAlloy claims).

They are also hedging the energy wall beyond gas: UPS systems sized for AI load swings, modular Spark factories in Colorado, SMR work with Aalo at Idaho National Lab, and gas-to-nuclear bridging with Blue Energy in Texas. That is the founder mindset the Gauntlet screens for — they keep pushing the constraint, not milking last year’s narrative.

Capital Structure Snapshot (Private)

Valuation went $2.8B → $10B+ → $30B in under two years. That is the AI-infra bid in one line. Sacra’s last clean print is ~$276M for 2024 (+82% YoY), with cloud bookings later described as growing ~5x through 3Q25. Other private-market models have 2025 anywhere from high-$200Ms to a near-$1B forecast once Stargate / campus revenue is recognized. Treat operating numbers as directional and unaudited until an S-1 drops. What is not sloppy: contracted GW, named counterparties, and a $13B financial-buyer cloud deal. That is how private marks jump.

Why Crusoe Fits the G.L.A. Lens  

Coach Kyle’s filter is simple. We want companies building must-have infrastructure at the start of a multi-decade S-curve, run by people who reinvest in the next problem and will continue to innovate. Crusoe’s next problem is not “can we rent H100s.” Everyone can try that. The next problem is:

• firm power at GW scale

• campuses designed for inference density, not 2015 enterprise IT

• always-on reliability when agents run 24/7

• abstraction for developers and raw performance for cluster engineers

If inference is where AI “comes to life,” then the winner is the operator who can keep tokens flowing when the grid, the cooling loop, and the orchestration layer are all under stress. That is an energy company wearing a cloud costume, or a cloud company that learned energy first. Either way, Wall Street’s old “data center REIT vs. AWS clone” box does not hold it.

THE G.L.A GAUNTLET

Rule #25 — G.L.A. Proprietary Scoring System (Used to Identify Great Longterm Investments)

Horizon: 5-year and 10-year risk vs. reward
Scoring: 1–10 per rule. 80+ = serious compounder watch. 90+ = rare.

#1 Size & Predictability — 8.5 / 10 Minimums: $2B market cap, $100M revenue.
Private mark ~$30B.

Revenue already cleared $100M years ago and is scaling with contracted campuses. Rare-exception clause not needed. Predictability: large contracted MW, concentrated counterparties, project-finance complexity.Talks of IPO ongoing.

#2 Visionary Founders Who Keep Innovating — 8.0 / 10 Not household names like Musk, Jobs, Karp, Bezos, or Saylor. The behavior matches the rule.Lochmiller (MIT math/physics, Stanford CS/AI) and Cavness (energy/geology background) started on stranded electrons, jumped the S-curve into AI factories, sold the old cash cow, and are now stacking nuclear, modular inference boxes, and NVIDIA-native cloud. No dividend/buyback culture. Capital is being poured into the next constraint. Founder brand is still under-recognized relative to the asset they control. That can be a feature.

#3 Must-Have Product / Service — 9.0 / 10 AI training clusters and, increasingly, always-on inference are not nice-to-haves. If the iPhone was the consumer must-have of the last cycle, tokens on demand are the industrial must-have of this one. Crusoe is selling the factory, the power, and the runtime. That is closer to “Amazon delivery + Tesla energy” than to a feature app.

#4 Hyper-Growth + Fortress Balance Sheet + Path to FCF — 7.0 / 10 Growth is real: cloud bookings, GW pipeline, named hyperscaler and Jane Street contracts. Balance sheet was just recapitalized with $3B+ of new equity on top of earlier equity and project debt. This is not a fortress FCF machine today. It is a capital-intensity business that prints cash after campuses energize and reserved GPU contracts season. High degree of future FCF if utilization and power costs hold. Until then, the risk is dilution, leverage, and timing slippage. Score reflects growth yes, fortress-FCF not yet.

#5 Huge TAM — 9.5 / 10 AI data-center capex, GPU-as-a-service, and dedicated generation for compute are a multi-trillion stack over a decade. Inference share of that stack should rise as agents leave the lab. Crusoe’s TAM is not “GPU rental.” It is power + purpose-built capacity + cloud software in the same motion. Few private companies sit across all three.

#6 Impenetrable Moat — 7.5 / 10 Real moat pieces:

• energy origination and on-site / behind-the-meter optionality

• speed at GW-scale construction (Stargate / Abilene proof points)

• NVIDIA-day-one architecture and software integration

• vertical ops from electrons to tokens

 CoreWeave is public and scaled. Hyperscalers can build themselves. Other neoclouds and utilities are running the same race. Switching costs exist once a training cluster or inference fleet is live, but the industry is still in land-grab mode. Moat is advantaged, not fortress-monopoly yet.

#7 Risk vs. Reward We Can Live With — 7.0 / 10 Bull case (5–10 years): inference becomes the cash engine; 40 GW pipeline converts; Crusoe is a top independent AI factory + cloud; IPO or strategic outcome compounds from a $30B base if execution matches the mark.

#8 Solving a Major Societal Problem — 9.0 / 10 Rule: companies that solve the largest problems produce the largest long-term returns.
Crusoe sits on two: wasted / constrained energy and insufficient AI compute. Those are not boutique ESG talking points. They are the reason model labs, governments, and trading firms are writing nine- and ten-figure checks. Foundation-of-the-next-era language in the transcript is earned only if the campuses actually turn on.

#9 Proven Investors — 9.0 / 10 Cap table quality is high:Founders Fund, NVIDIA, Valor Equity, Mubadala, Fidelity, Franklin, T. Rowe, Tiger, Salesforce Ventures, Ribbit, Winklevoss, Lowercarbon, and a long list of crossover names. ARK-related capital appears on private databases. This is not a lonely seed round looking for a mark. It is the same class of capital that backed the last generation of category owners.

#10 Wall Street Does Not Understand It Yet — 8.5 / 10 Still private, so “doesn’t understand” is almost automatic. More important: the category is mis-boxed. Analysts want to model Crusoe as Core Weave-lite or as a contractor on Stargate. Crusoe’s model is energy company + AI factory + inference cloud.

GAUNTLET TOTAL

Classification: High-conviction private-market compounder . Clears the Gauntlet. Not a 90+ “own it at any price” name because capital intensity, concentration, and the $30B mark already pull forward a lot of the 5-year dream.

Crusoe reportedly met with four of Wall Street’s largest underwriters — Goldman Sachs, Morgan Stanley, JPMorgan Chase and Bank of America — about a potential near-term IPO, the same week it closed a private round tripling its valuation to $30 billion.

AI infrastructure companies are reaching public-market conversations on a dramatically faster timeline than AI model-layer companies: OpenAI and Anthropic are still confidentially filed with no public S-1, while infrastructure names like Crusoe and Cerebras are already public or in banker talks within a few years of founding.

COACH KYLE 2 CENTS:Crusoe has the right backers and are providing must have services to the AI Build out, also with talks of a near term IPO all sounds good to me.

G.L.A INVE$T RESEARCH
Private / educational analysis. Not an offer to buy or sell securities. Crusoe is private; figures are compiled from public reporting and third-party estimates and will change when primary filings appear.