Cheap AI Compute from Methane: Crusoe's Business Model

Waste gas creates cheap electrons for busy GPUs

Welcome to the latest edition of Buy the Rumor; Sell the News. We’re closing in on 2,000 subscribers, including institutional allocators, venture capitalists, litigators, senior executives and entrepreneurs. Thank you to all who have subscribed!

In today’s post, I take a look at Crusoe, the company which builds cheap AI compute infrastructure powered by methane gas.

If you want to connect with me directly, my contact information is at the end of this post.

---

A few years ago, if you stood on a ridge outside Midland, Texas on a clear night, you’d see the horizon pulsing with orange fire. It looked almost beautiful. In reality, it was an enormous waste. Shale drillers, interested in oil, had no cost-effective way to deal with the gas that came along with the oil. Pipelines were full or nonexistent, and building infrastructure for small or remote wells didn’t pencil out. So they did what oilmen have done for decades: they burned it right at the wellhead, flaring off billions of cubic feet of natural gas every day.

At its peak around 2019, the U.S. was flaring or venting enough gas to power every home in Texas and then some. This was terrible for climate optics and bad economics. But it also created a strange opportunity

Flared methane gas. Image sourced from here.

Crusoe’s Founding Insight: Arbitrage Stranded Energy

In 2018, two entrepreneurs, Chase Lochmiller and Cully Cavness, launched Crusoe Energy with a simple but radical idea. If gas was worthless at the point of extraction, why not capture it on-site, convert it to electricity with mobile generators, and use that ultra-cheap power to run energy-hungry data workloads?

They started with Bitcoin mining. Each containerized unit had a generator hooked into the flare stack and rows of ASIC miners humming inside. The economics were spectacular when BTC was north of $50,000. Crusoe paid next to nothing for the gas, while Bitcoin’s price meant huge margins.

But after the crypto winter of 2022, Crusoe saw the fragility in depending on a single volatile commodity. They pivoted hard: same gas, same modular data centers, but now filled with racks of NVIDIA GPUs for AI workloads.

Crusoe’s Business Model

Today, Crusoe’s business model is deceptively simple: They take stranded or flared gas that has almost zero market value, convert it on-site into electricty using mobile generators, power modular data centers filled with GPUs, and sell that compute capacity to companies training or running AI models at a steep discount to hyperscaler rates.

It’s an arbitrage play. Crusoe buys for pennies energy that would otherwise be wasted. They turn it into rentable GPU-hours, a commodity that’s scarce and expensive in today’s AI gold rush.

This means they effectively transform wasted hydrocarbons into one of the most valuable resources on Earth right now: affordable AI compute.

Who buys from Crusoe, and why?

Crusoe has two sets of customers:

1. The Supply Side: Oil & Gas Companies

These are exploration & production (E&P) operators across the Rockies and Permian basin. These include public companies like EOG and Pioneer (acquired by ExxonMobil in 2024), and hundreds of smaller private equity-baced drillers. They have three incentives:

• Regulatory: Many states impose strict limits on flaring. If you exceed them, your drilling permits are at risk.

• Economic: Instead of wasting the gas, they can now monetize it or at least avoid paying penalties.

• ESG optics: Investors increasingly care about emissions. Being able to say “we captured our flaring” matters.

2. The Demand Side: AI Labs, SaaS Startups, and HPC Researchers

On the demand side, Crusoe sells compute capacity to:

• LLM startups and AI research labs who need thousands of GPUs to train massive models. The catch? They don’t always care about latency or location. They just want cheap, reliable access.

• SaaS firms doing occasional big training runs, for example re-training a recommendation engine or fraud model.

• High performance computing (HPC) scientists running genomics or fluid dynamics models that aren’t time-sensitive.

Why do they choose Crusoe?

• Price: Training on Crusoe’s clusters is often 50-70% cheaper than on AWS, Azure, or GCP. For startups burning millions on compute, that’s game-changing.

• Availability: When hyperscalers have waitlists or capacity constraints (as happened with H100s in 2023-24), Crusoe can fulfill orders immediately.

• Marketing: Buying from Crusoe means your compute is net carbon negative relative to normal data centers. It lets green-conscious VCs or customers feel better.

The Economics Under the Hood

Crusoe’s economics rely on three stacked layers:

1. Energy Arbitrage: They acquire gas for 10–30 cents per MCF (thousand cubic feet) versus a market price often over $3. Their fully loaded power costs are under 1 cent per kWh, which is far cheaper than grid electricity.

2. Compute Margin: They charge ~$8–12 per H100 hour (vs $25–30 on AWS), achieving solid gross margins especially since the infrastructure is mobile and purpose-built for GPU cooling and density.

3. Carbon Revenue: They also sell carbon offset credits on voluntary markets. Not massive, but it sweetens returns.

Capex is significant (each 2MW mobile module costs ~$7–8 million fully installed), but they depreciate over 4 years. With average utilization climbing past 60%, Crusoe’s payback periods hover around 18–24 months¹.

What is its Moat?

What Protects Crusoe?

1. Regulatory and Permitting Know-How: Crusoe has multi-year contracts and compliance systems with over two dozen oil & gas operators. Navigating local permitting, emissions tracking, and interconnect approvals is not trivial.

2. Engineering IP: They designed their own ducted immersion cooling, waste heat recovery systems, and sound-dampening containers. Competitors typically buy generic OEM containers and retro-fit.

3. First-Mover Climate Credibility: Big-name AI labs (some even Anthropic-scale) have chosen them specifically for the ESG narrative, which helps those labs raise money. That brand story is hard to replicate overnight.

4. Balance Sheet: They’ve raised over $500 million, allowing them to buy GPUs directly from NVIDIA on allocation (avoiding the spot market) and lock in gas supply with pre-paid agreements. Capital intensity itself becomes a moat.

But It’s Not Invincible

• Regulatory Shocks: If the U.S. bans flaring outright, the gas could be forced through pipelines or used onsite for chemicals, drying up Crusoe’s ultra-cheap feedstock. If there’s a big carbon tax, oil firms might want to use the gas for their own internal offsets instead of sharing it.

• Hardware Risk: Crusoe’s business assumes the current GPU generation holds value for 36 months. If new Blackwell B200s or custom AI ASICs make H100s obsolete faster, resale value tanks.

• Network Bottlenecks: Many sites use microwave or Starlink because there’s no fiber. That’s fine for training big models you upload once. But for tasks needing constant data in and out (like human feedback tuning), it’s a constraint.

• Capital Markets: Rising interest rates or an AI slowdown could choke new funding for building out more clusters.

• Copycats: Exxon or even Amazon could theoretically replicate the model. So far, oil majors face investor pressure to divest these side ventures rather than build them. But it’s a lurking threat.

Crusoe’s Next Moves, and What Investors Should Watch

Short-Term (1–2 years)

• Fill out the Permian footprint: More generators on more well pads, denser local networks to get utilization even higher.

• Lay major fiber into Midland: So they can attract enterprises that demand lower-latency or want to run inference, not just big batch training.

Medium-Term (3–5 years)

• Diversify Energy: Crusoe is piloting landfill gas in the Midwest, where the same flare-avoidance model applies, and studying modular nuclear. When small-scale advanced reactors become licensable (NRC Part 53 projected 2026+), they could power GPUs directly.

• Become a Compute Exchange: They’re building a software platform to let other stranded-energy compute operators list GPU capacity, positioning Crusoe as the ICE of off-grid FLOPs².

Long-Term (5–8 years)

• Potential to spin out the carbon credits platform as a standalone marketplace or even repackage the data centers as an infrastructure REIT.

Strategic Takeaways for Allocators, Investors, and Entrepreneurs

For Allocators & Institutional Investors

• Treat Crusoe-like assets as a hybrid between midstream energy and hyperscale cloud. They yield cash flows like pipelines but refresh hardware like Silicon Valley. Allocate from infra-tech buckets, not pure real assets.

• Underwriting shouldn’t fixate on ESG headlines. Instead, dig deep on the durability of flare-gas supply contracts and GPU leasing counterparty risk.

For Growth Equity & VCs

• Expect modest IRRs (8–12%) unless carbon pricing spikes or leverage gets creative. Exits likely look like infra YieldCos or SPAC-style roll-ups.

• Keep tabs on others exploring landfill methane. Over 1,100 U.S. landfills flare more than 200 MMCF/year. A portfolio of these niches could hedge regulatory shifts.

For Entrepreneurs & Operators

• If you’re building an AI-heavy product, design your workload scheduler to treat Crusoe as a cheap, pre-emptible pool of GPU hours, then failover to AWS. The cost savings can extend your runway dramatically.

• For hardware startups, there’s huge opportunity in serving these edge data centers: immersion pumps, portable SMRs, microwave bandwidth optimizers.

The Broader Lesson

At the end of the day, the cheapest joule wins. As AI’s hunger for compute keeps doubling every 6–12 months, it will pull in progressively more exotic energy niches, from flare gas to landfill methane to offshore wind barges. Crusoe is just the first high-visibility example. For anyone investing in or building AI infrastructure, it pays to think not only like a technologist, but also like an energy merchant banker.

---

Coda

If you enjoy this newsletter, consider sharing it with a colleague.

Most posts are public. Some are paywalled.

I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

• follow me on Twitter,

• connect with me on LinkedIn, or

• send an email to dave [at] davefriedman dot co. (Not .com!)

1. A rough estimate: (1) early ramp periods mean 20-40% utilization for the first few months; (2) occasional downtime for maintenance or site gas disruptions; (3) effective average price often dips due to negotiated discounts on large blocks.

So if you blend ramp periods + spot discounts + partial downtime over the life of the cluster, average effective gross cash flow might be $400-500k/month. So payback becomes $7.5 million / $450k ~= 17 months.

2. There are similar platforms being built, including Compute.exchange. The financialization of GPU compute, possibly with GPU futures, akin to the invention of oil futures in the early ‘80s, seems like a possibility.