Energy · Deep dive
Crusoe
Started by burning stranded gas to mine bitcoin at the wellhead. Now builds the gigawatt campuses OpenAI, Oracle and Microsoft run on — and is reportedly being priced at $30B.
emerging
The question that decides it: Crusoe's equity is a residual claim behind roughly $10B+ of project debt on campuses leased to a handful of counterparties. Do the 15-year NNN leases at Abilene and Childress hold their credit quality if OpenAI's or Meta's compute demand curve bends — or does the levered developer, not the tenant, absorb the shortfall?
- HQ
- Denver, CO
- Founded
- 2018
- Ownership
- VC-backed (Series E; pre-IPO round reported in talks)
- Funding
- ~$3.9B equity raised; multiples of that in project and corporate debt
- Valuation
- Over $10B (Oct 2025 Series E); reportedly ~$30B in talks (July 2026, Bloomberg)
- Revenue
- ~$276M in 2024 (Sacra estimate, +82% YoY); ~$1B reported for 2025; company projects ~$2B for 2026
- Headcount
- ~900 (roughly 912 as of 2026, per Tracxn); ~135 staff transferred to NYDIG with the bitcoin unit in 2025
- Screen
- Scaled private — raised well over $100M
- Published
- 2026-07-14
- Web
- www.crusoe.ai
- Elsewhere
- LinkedIn · Crunchbase
Founders and leadership
-
Chase Lochmiller Co-founder and CEO
Math and physics at MIT, then a quantitative trader at Jump Trading and Getco; later ran crypto investing at Polychain Capital and took a Stanford CS master's focused on AI. Took six months off to climb five of the world's seven summits. The mountaineering trips with his childhood friend Cavness are where the flared-gas idea took shape.
-
Cully Cavness Co-founder and President
Third-generation Colorado oil and gas — grew up in the business and studied geology at Middlebury, then energy economics in Europe on a Watson Fellowship, then an Oxford MBA. Worked in upstream oil and gas before Crusoe. He supplied the field reality: stranded gas at the wellhead has negative economic value to the producer.
Snapshot
Crusoe began in 2018 as an oddity: shipping-container data centers dropped next to oil wells in North Dakota, burning gas that producers were flaring anyway, and mining bitcoin with the electricity. Eight years later it is one of the largest AI data centre developers in the United States, the builder of the Abilene, Texas campus that anchors OpenAI and Oracle’s Stargate programme, and a contractor to Meta and Microsoft. Sacra estimated 2024 revenue at $276M (up 82% year on year); 2025 came in near $1B on reported figures, and the company projects roughly $2B for 2026. In July 2026 Bloomberg reported it was in talks to raise about $3B at a valuation near $30B — triple the mark set in its October 2025 Series E. The bitcoin business that funded all of it was sold off in 2025.
Founding story
Chase Lochmiller and Cully Cavness met at high school in Denver. Lochmiller went the quant route — MIT physics, then Jump Trading and Getco, then crypto investing at Polychain and a Stanford CS master’s in AI, with a six-month detour to climb five of the seven summits. Cavness went the opposite way: geology at Middlebury, a Watson Fellowship studying European energy economics, an Oxford MBA, and a job in upstream oil and gas, in a family that had been in Colorado energy for three generations.
The company came out of the collision of those two resumes on hiking trips in the Colorado 14ers around 2018. Lochmiller understood that bitcoin mining’s only real input is cheap electricity. Cavness understood that at a remote wellhead, associated gas has negative value — there is no pipeline to take it, so the producer burns it into the sky. Roughly 150 billion cubic metres is flared globally each year, per World Bank figures. The insight was arbitrage in its purest form: move the load to the fuel rather than the fuel to the load.
This is the part of the story that gets sanitised. Crusoe was not founded as a climate company that later discovered AI. It was founded as a bitcoin miner with an unusually good cost-of-power story and a defensible emissions argument attached to it.
How it works
The original mechanism, which Crusoe patented as Digital Flare Mitigation, was physical and unglamorous. Crusoe would sign a deal with a producer — Exxon and Devon were among the partners — truck in a skid of reciprocating gas gensets and a stack of modular data centre containers, and tie them into the flare stack. Gas that was going to be combusted inefficiently in an open flare instead gets combusted in an engine, which turns a generator, which powers compute sitting metres away. The claimed environmental benefit is real but specific: engine combustion destroys methane far more completely than a wind-blown flare, so Crusoe’s case has always been about methane destruction efficiency, not about the CO2, which still happens. At its peak the fleet was 425+ modular units across roughly 30 sites in seven US states and Argentina, processing over 10 million cubic feet of gas a day and around 250-270 MW of generation.
That business taught Crusoe two things that mattered more than bitcoin: how to commission generation fast, and how to build compute in a box. When H100s became the scarcest asset in technology in early 2023 — hyperscaler waitlists running six months, spot GPU-hours touching $8 — Crusoe already had the hardest ingredient. It launched Crusoe Cloud, then went further and began developing the campuses itself.
The modern mechanism is different in kind. At Abilene, Crusoe is the developer: it holds the land, builds the shells, installs the on-site generation (about 360 MW of gas turbines plus backup diesel at the Stargate site), and signs a long-term triple-net lease — Oracle’s is 15 years — with the tenant, who sells the compute onward to OpenAI. Power is no longer scavenged from flares; it is contracted at industrial scale, including roughly 1 GW of GE Vernova aeroderivative turbines and 650 MW from ProEnergy for delivery through summer 2027, plus forward deals for iron-air storage with Form Energy (12 GWh from 2027) and a nuclear-plus-gas-bridge plan with Blue Energy. Crusoe also builds its own modular units at a 352,000 sq ft “Spark Factory” in Brighton, Colorado — a $200M+ investment, first modules targeted for Q3 2026.
Product and business overview
Three businesses sit under one roof, and they are not equally important.
Data centre development. The real engine. Crusoe originates the site, builds the power, and leases the facility. Abilene is 1.2 GW across eight buildings for the Oracle/OpenAI Stargate deal; a separate 900 MW campus in Abilene is being built for Microsoft (first building mid-2027), taking the town’s total Crusoe footprint toward ~2.1 GW. Meta is under contract for roughly 1.6 GW across Childress, Texas and Warrenton, Missouri.
Crusoe Cloud. The GPU cloud: bare-metal and virtualised H100/H200/B200/GB200 and AMD MI300X/MI355X instances with InfiniBand fabric, sold to AI labs and enterprises. Customers named publicly include Sony, Databricks, Together AI and Luma AI.
The software layer above it. AutoClusters for distributed-training orchestration, Managed Inference (Crusoe claims up to 9.9x faster time-to-first-token versus vLLM on its own benchmarks, via its MemoryAlloy KV-cache), Command Center for observability, and Edge Zones on Crusoe Spark modules. Crusoe acquired Israeli GPU-management startup Atero to staff this, making it the Tel Aviv R&D centre.
Business model and pricing
Revenue books three ways. Development and leasing produces long-dated NNN rent and PPA revenue — the most bond-like income, and increasingly the majority. Cloud produces on-demand and reserved GPU-hours. Software services are still immaterial.
Published Crusoe Cloud rates as of mid-2026: H100 at $3.90/GPU-hour on-demand, H200 at $4.29, billed per minute. B200 and GB200 require a sales conversation — a tell that the newest silicon is being sold on committed contracts, not on the credit card. Sacra’s model puts reserved pricing nearer $2-3/hour with up to ~81% off on-demand for three-year terms, and sketches a 100-GPU enterprise on a one-year commit at roughly $2.2M a year including storage, networking and support.
The unit economics diverge sharply by segment. Cloud is exposed to a GPU rental price that has fallen from roughly $8/hour in 2023 to $2-3 today. Development is not: it is a spread between a levered construction cost and a 15-year lease from an investment-grade-ish tenant, with Sacra sketching something like a 3.7x MOIC on a ten-year hold. That is a real-estate business wearing a technology company’s valuation multiple, and it is worth being honest that this is what has been financed.
Traction over time
| Date | Metric |
|---|---|
| 2021 | ~$30M revenue era; DFM fleet scaling across the Bakken |
| 2023 | $152M revenue (per company projections cited by Sacra); Crusoe Cloud launched |
| 2024 | $276M revenue, +82% YoY (Sacra estimate); 425+ modular units, ~250-270 MW |
| Dec 2024 | Series D at $2.8B — 10.2x trailing revenue |
| Mar 2025 | Bitcoin mining unit sold to NYDIG; ~135 staff transfer |
| Jun 2025 | First GB200 racks delivered at Abilene; first two buildings (980,000 sq ft, 200MW+ IT) live |
| 2025 | ~$1B revenue (reported); cloud bookings up ~5x over first three quarters vs 2024; cloud ARR +150% |
| Oct 2025 | Series E at over $10B; power pipeline stated above 45 GW |
| Feb 2026 | Company reports ~17x YoY growth in added total contract value, ~70% growth in new logos for 2025 |
| Jun 2026 | ~4.9 GW of contracted capacity; development pipeline reported above 40 GW |
| 2026E | ~$2B revenue projected; ~900 employees |
Market analysis
The demand side is not in question. Dell’Oro raised its global data centre capex outlook past $1 trillion for 2026 and projects $1.7 trillion by 2030, with the top four US hyperscalers alone approaching $600B of combined 2026 capex. Synergy Research forecasts the neocloud segment specifically approaching $400B by 2031 from roughly $20B in 2026. Goldman Sachs has data centre power demand growing at a 15% CAGR to 2030, taking data centres to around 8% of US electricity.
The supply side is where Crusoe’s thesis lives. Over 2,000 GW sits in US interconnection queues, with median time to commercial operation approaching five years and reported waits of four to seven years in Northern Virginia, Phoenix and Dallas. The binding constraint on AI capacity in 2026 is not GPUs; it is transformers, turbines and an energised substation. A developer that brings its own generation behind the meter compresses a seven-year timeline into eighteen months. That is the entire product.
Competitive intel
See the competitor table above for the named set. The structural point: Crusoe now competes in two markets with two sets of rivals and is not obviously the favourite in either. Against CoreWeave and Nebius on cloud, it has fewer GPUs and faces listed rivals with cheaper equity. Against Vantage, QTS and Switch on development, it has less land-banking history and a shorter track record with lenders. Its one hard-to-copy asset is energy origination — turbine slots, gas contracts, self-built modules — which is exactly what the incumbents are now buying too.
History and evolution
- 2018 — Founded in Denver by Lochmiller and Cavness.
- 2019 — $4.5M seed (March), $30M Series A led by Bain Capital Ventures (December). Digital Flare Mitigation units go live in the Bakken.
- 2021 — $128M Series B led by Valor. Fleet scales across seven states.
- April 2022 — $505M package ($350M Series C equity led by G2 Venture Partners plus debt) at ~$1.75B; Crusoe Cloud announced.
- 2023 — GPU scarcity peaks. Crusoe pivots hard to AI compute.
- October 2024 — $3.4B JV with Blue Owl and Primary Digital for the Abilene campus.
- December 2024 — $600M Series D led by Founders Fund at $2.8B, with NVIDIA participating.
- January 2025 — Stargate announced; Abilene named Site 1.
- March 2025 — Bitcoin mining and the DFM business sold to NYDIG. The founding business is gone.
- May-June 2025 — $11.6B of project financing for Abilene phase 2; $750M Brookfield credit facility; first GB200 racks arrive.
- October 2025 — $1.375B Series E at over $10B.
- December 2025 — Michael Gordon, ex-COO/CFO of MongoDB, hired as COO and CFO. An IPO hire, plainly.
- March 2026 — 900 MW Microsoft campus announced in Abilene; Form Energy storage deal.
- June 2026 — Bloomberg reports Crusoe was displaced from a Wyoming campus after failing to land an anchor tenant. 4.9 GW contracted.
- July 2026 — Reported talks for ~$3B at ~$30B.
What people say
The case for. The praise is remarkably consistent on one point: speed. Crusoe energised the first two Abilene buildings within roughly a year of breaking ground in June 2024, a pace hyperscale developers regard as close to impossible, and Oracle began racking GB200s by June 2025. Customers of Crusoe Cloud cite the InfiniBand fabric and bare-metal access; Sony’s GT Sophy research programme and Databricks are reference accounts. Investors have voted with unusual conviction — NVIDIA, Founders Fund, Fidelity, Mubadala and Tiger all in the cap table, an oversubscribed Series E, and a secondary market that reportedly repriced the shares well above the October 2025 round within months. On Glassdoor, the recurring positive is the calibre of colleagues; employees repeatedly describe the team as the best part of the job.
The complaints. Three clusters, all serious.
Environmental. The Texas Tribune reported in July 2026 that Stargate Abilene’s developers built ten gas turbines and 62 backup diesel generators under “permit by rule” and standard permits — minor-source authorisations more commonly used by dry cleaners, requiring no environmental study, public notice or comment period. Those permits allow over 1.6 million tons of greenhouse gases and 1,000 tons of criteria pollutants a year; applications for 41 more turbines and 18 more generators are pending, which would make the site one of the largest fossil-fuel power plants in Texas. A neighbouring rancher told reporters she learned of the project only when construction began. The older critique still stands too: capturing flared gas reduces methane venting, but the carbon still enters the atmosphere, and critics argue the model creates a paying customer for gas that would otherwise be a liability.
Employees. Glassdoor sits around 3.4/5 with 58% recommending the company. The negative reviews are specific and repeated: constant pivots, meeting overload, overwork, and — pointedly — one review describing unrealistic performance and reliability SLAs on the cloud side colliding with operational missteps, producing customer pain and revenue reversals. That is an employee, not a competitor, saying it.
Financial. The bear case writes itself and several analysts have written it. Revenue is concentrated in a handful of counterparties — OpenAI via Oracle, Meta, Microsoft. The capital stack is layered with billions in debt including a ~$7.1B J.P. Morgan-led construction loan on Abilene phase 2. The valuation rests on roughly a 2x revenue jump in 2026. GPU rental prices have already fallen 60-75% from their peak. And the Wyoming setback in June 2026 — reportedly losing a campus after failing to secure an anchor customer — is the first public evidence that Crusoe’s pipeline is not the same thing as Crusoe’s backlog.
Outlook: the open question
For this to work, the 15-year leases at Abilene, Childress and Warrenton have to behave like bonds; for it to fail, they only have to behave like contracts with counterparties whose own AI revenue does not arrive. Crusoe has built something genuinely scarce: the ability to put a gigawatt of behind-the-meter generation next to a data hall in eighteen months, when the grid quotes seven years. That is why Blue Owl and J.P. Morgan lend against it and why NVIDIA sits on the cap table.
But the equity that is being marked at $30B sits behind ten-plus billion dollars of project debt on assets whose value is almost entirely a function of tenant credit. Crusoe is not taking technology risk any more; it sold the business that made it interesting. It is taking counterparty risk on the AI capex cycle, financed with leverage, and being paid a venture multiple to do it. The company that gets a good IPO is the one whose backlog converts to rent on schedule and whose cloud margin survives GPU price decay. The one that does not is a levered real-estate developer with a very good story about turbines.
Sources and further reading
- Crusoe revenue, valuation & funding — Sacra, updated April 2026. Revenue estimates, capital stack, unit economics, project detail.
- Crusoe in Talks to Raise $3 Billion in Round That May Triple Firm’s Value — Bloomberg, 2 July 2026.
- Planned Texas data centers could emit significant pollution — The Texas Tribune, 9 July 2026. The permitting-loophole investigation.
- Crusoe closes $1.375B Series E, reaches $10B valuation — Crusoe newsroom, October 2025.
- Crusoe Energy sells bitcoin mining unit to NYDIG to focus on AI — CNBC, 25 March 2025.
- Crusoe’s Abilene data center campus officially live, serving Oracle and OpenAI’s Stargate — Data Center Dynamics, 2025.
- Crusoe secures $750 million credit facility from Brookfield — Crusoe newsroom, June 2025.
- AI Boom Drives Data Center Capex to $1.7 Trillion by 2030 — Dell’Oro Group, 2026.
- Neocloud Market Forecast to Approach $400B by 2031 — Synergy Research Group, 2026.
- Crusoe Business Breakdown & Founding Story — Contrary Research.
- Crusoe Cloud pricing — company pricing page, accessed July 2026.
Capital history
| Date | Round | Amount | Valuation | Lead(s) |
|---|---|---|---|---|
| March 2019 | Seed | $4.5M | Undisclosed | Bain Capital Ventures, Winklevoss Capital |
| December 2019 | Series A | $30M | Undisclosed | Bain Capital Ventures |
| April 2021 | Series B | $128M | Undisclosed | Valor Equity Partners |
| April 2022 | Series C | $350M equity (part of $505M with debt) | ~$1.75B | G2 Venture Partners |
| October 2024 | Project JV (Abilene phase 1) | $3.4B | n/a (project-level) | Blue Owl Capital Real Assets, Primary Digital Infrastructure |
| December 2024 | Series D | $600M | $2.8B post ($29.17184/share, per Sacra) | Founders Fund (with NVIDIA, Fidelity, Mubadala, Valor, Ribbit) |
| Early 2025 | Credit facility | $225M | n/a (debt) | Upper90 |
| May 2025 | Project debt + equity (Abilene phase 2) | $11.6B (incl. ~$7.1B J.P. Morgan-led construction loan) | n/a (project-level; ~$15B total committed to Abilene) | Blue Owl Capital, Primary Digital Infrastructure, J.P. Morgan |
| June 2025 | Credit facility | $750M | n/a (debt) | Brookfield Asset Management (infrastructure debt platform) |
| 2025 | Credit facility | $175M | n/a (debt) | Victory Park Capital |
| October 2025 | Series E | $1.375B | Over $10B | Valor Equity Partners and Mubadala Capital (co-led) |
| July 2026 (reported, in talks) | Pre-IPO | ~$3B (reported) | ~$30B (reported by Bloomberg; not confirmed by the company) | Undisclosed |
Investors / owners: Valor Equity Partners, Mubadala Capital, Founders Fund, Bain Capital Ventures, G2 Venture Partners, NVIDIA, Fidelity Management & Research, Tiger Global Management, Franklin Templeton, T. Rowe Price, Salesforce Ventures, Ribbit Capital, Lowercarbon Capital, Polychain Capital, Winklevoss Capital, DRW Venture Capital, Upper90, Brookfield Asset Management (debt), Blue Owl Capital (project)
Competitive set
- CoreWeave — The public comp and the scale leader. Passed $5B of annual revenue faster than any cloud platform, and ended 2025 with roughly $66.8B of contracted backlog. Attacks Crusoe on GPU fleet size, Nvidia relationship depth and the credibility of a listed balance sheet; Crusoe's counter is that it owns the power, not just the racks.
- Nebius — Listed neocloud spun out of Yandex. Guided 2026 capex to $20-25B and is on a path analysts see reaching well past $1B of revenue. Vertically integrated on hardware and software, but buys grid power like everyone else — no self-generation story.
- Lambda — The developer-experience neocloud (~$893M raised as of 2025, per Sacra). Wins bottom-up with researchers on price and simplicity — publishes H100/B200 rates openly and undercuts on-demand. Not a data-center developer; competes only for the cloud dollar.
- Together AI — Inference and fine-tuning platform sitting a layer above raw GPUs — and, notably, a Crusoe customer as well as a rival. Attacks the part of the stack Crusoe is climbing into with Managed Inference and AutoClusters.
- Vantage, QTS (Blackstone), Switch, Aligned — The incumbent hyperscale developers, several PE-owned, with decades of land, permitting and leasing muscle. They compete for the same 15-year NNN anchor leases and have cheaper capital; what they lack is Crusoe's willingness to build its own generation.
- Hyperscalers building in-house (Meta, Microsoft, Amazon, Google) — Combined 2026 data-center capex approaching $600B (Dell'Oro, early 2026). Every campus they self-develop is a lease Crusoe does not sign. Meta is simultaneously a ~1.6 GW Crusoe customer and its most credible substitute.
- GE Vernova, ProEnergy, Bloom Energy — Not rivals so much as the chokepoint. Crusoe has contracted ~1 GW of GE Vernova aeroderivative turbines and 650 MW from ProEnergy. Whoever controls turbine and fuel-cell slots controls who gets to build behind-the-meter — and they sell to everyone.