A Stanford geologist trained AI on a century of accidental discoveries to find the first commercially viable blind geothermal system in over three decades - proving the American West still holds gigawatts of hidden clean energy.
Carl Hoiland had a theory most of the energy industry considered naive: that the American West was sitting on vast reservoirs of geothermal energy that nobody had found simply because nobody knew how to look. The Stanford-trained geologist co-founded Zanskar Geothermal and Minerals in 2019 with that conviction at its core, and spent the next six years building artificial intelligence models that could do what human geologists had failed to do for three decades - locate commercially viable geothermal systems with no surface signs, no historical drilling data, and no roadmap except the patterns buried in a century of accidental discoveries. In December 2025, those models delivered their most compelling proof yet: a 250-degree Fahrenheit reservoir hidden beneath the Nevada desert, confirmed by drill bits, and named with characteristic understatement "Big Blind."
The geothermal industry's stagnation had a simple, demoralizing explanation. Finding a viable conventional geothermal site required a specific geology - underground reservoirs of hot water or steam sitting in permeable rock, close enough to the surface that standard drilling could reach them - and for most of the twentieth century, geologists found those sites by looking for surface clues. Hot springs. Fumaroles. Steam vents. The places where the earth announced itself. By the 1980s, the most obvious sites in the western United States had been mapped and, in many cases, developed. The implicit conclusion hardened into orthodoxy: conventional geothermal was tapped out.
That orthodoxy rested on a fundamental gap in the data. The surface indicators that had guided the industry were absent from roughly 95% of geothermal systems, Hoiland estimates. These "blind" systems - reservoirs with no surface expression at all - were not rare anomalies. They were the norm. The industry had been finding and developing the visible minority while writing off everything else as too difficult to locate.
The late 1970s and early 1980s had seen a brief attempt to crack this problem. With oil shocks reshaping energy economics, major oil and gas companies deployed serious capital to explore for blind geothermal systems across the intermountain west. They drilled deep, expensive wells on a grid pattern, hoping to stumble onto something viable. The cost was enormous - Zanskar estimates those early explorers spent more than $100,000 per megawatt of capacity in discovery costs - and the success rate was low enough to end the experiment. By the mid-1980s, the majors had moved on, and the industry settled into a development-only mode, extracting value from sites already discovered while the frontier of new discovery went cold.
What those early explorers lacked was a systematic way to read the subsurface. The signals that indicate a blind geothermal system - heat flow anomalies, fault patterns, rock composition, magnetic and gravitational signatures - exist, but no single dataset makes them legible. Integrating them requires modeling millions of possible subsurface configurations simultaneously, which is exactly the kind of problem that modern machine learning handles well and human geologists cannot do at scale.
Hoiland and co-founder Joel Edwards, both with backgrounds in geology, spent years assembling the training data that would make Zanskar's AI models possible. The foundational insight was that the world already had substantial evidence of where blind geothermal systems exist - it had simply accumulated accidentally. Over a century of drilling for oil, gas, minerals, and agricultural water, drillers had punched into unexpected geothermal reservoirs without intending to. Those discoveries were scattered across records, filings, and analog datasets that had never been digitized or systematically analyzed. Hoiland and Edwards tracked them down, digitized them, and fed them to machine learning models as training examples of what a blind system looks like from the available surface and near-surface data.
The models then scour multiple datasets simultaneously - rock composition, magnetic field variations, heat flow measurements, fault geometry, electrical conductivity, seismic signatures - looking for the same combination of factors that characterize known blind systems. "There's no one type of data that tells you that a system is below you, even if you're right on top of it," Edwards told CNN in December 2025. The AI's advantage is that it can hold all of those signals in tension and identify the compound signature that indicates a viable reservoir.
A critical threshold came in 2024, when the models crossed a significant milestone: they had become more accurate at predicting geothermal locations than human geologists. The AI was not just processing more data faster - it was finding patterns that exceeded expert intuition. That shift created the confidence to move aggressively into unexplored territory. And the economics had been transformed: Zanskar estimates it can identify and confirm a geothermal prospect at less than $15,000 per megawatt - an order-of-magnitude reduction from the $100,000-plus costs of the 1980s wildcatters.
The site that would become Big Blind started as a geothermal anomaly - a cluster of signals in Zanskar's regional AI models that suggested elevated heat flow in a basin in western Nevada, near Tonopah. The basin was entirely unremarkable on the surface: scrubby desert, no springs, no venting, no history of geothermal exploration or any other drilling. Nobody had looked there because nobody had any reason to.
Zanskar acquired Bureau of Land Management leases at auction, then deployed field teams to collect high-resolution ground-level datasets that sharpened the AI's predictions about where to target exploration wells. In July and August of 2025, the company drilled two intermediate-depth wells, pushing roughly 2,700 feet into the earth. Both hit the same result: a permeable geothermal reservoir at 250 degrees Fahrenheit. That temperature, at that depth, cleared the minimum threshold for utility-scale power generation. Reaching comparable temperatures in the surrounding regional geology without the benefit of a localized geothermal system would typically require drilling to around 10,000 feet and likely expensive hydraulic stimulation techniques to create the permeability that Big Blind already possesses naturally.
The announcement on December 4, 2025 marked the first time the U.S. geothermal industry had confirmed a commercially viable blind system in more than 30 years. James Faulds, a professor of geosciences at the Nevada Bureau of Mines and Geology - whose research on characterizing geothermal systems informed Zanskar's work, and whose former graduate students are among the company's geologists - called the discovery "very significant." His assessment of the broader stakes was direct: estimates suggest over three-quarters of U.S. geothermal resources are blind systems, meaning the industry's true resource base is far larger than its development history suggests.
Big Blind was not Zanskar's first success, but it was its most consequential. The company's track record coming into the December announcement had already established a pattern that investors were watching closely.
The first proof of concept came at Lightning Dock, a struggling power plant in New Mexico that Zanskar acquired from Cyrq Energy. The site had been underperforming, and conventional wisdom held that the resource had been essentially exhausted at shallow depths. Zanskar's models identified a deeper, hotter zone that prior operators had missed. The company drilled to roughly 8,000 feet and found temperatures several tens of degrees Fahrenheit higher, with better and more stable production. By May 2025, Zanskar had tripled Lightning Dock's output. The plant became the most productive pumped geothermal well in the United States.
The second discovery, at Pumpernickel in northern Nevada, validated the exploration model on a previously known but undeveloped system. The site had attracted interest from potential developers for decades given its cluster of hot springs, and both oil and gas companies and the federal government had drilled exploratory wells since the 1970s. None had found commercially viable temperatures. Zanskar's AI identified the precise drilling targets that prior exploration had missed. Pumpernickel is now on a development path toward a 20-megawatt plant.
Big Blind completed a sequence the company's backers had been waiting for: three for three, with each discovery more technically challenging and more convincing than the last. "What does it look like when you try 10?" Edwards asked publicly after the December announcement - a question framed less as rhetoric than as a research agenda.
On January 21, 2026, Zanskar announced a $115 million Series C round led by Spring Lane Capital, with participation from a broad investor syndicate that included Obvious Ventures, Union Square Ventures, Lowercarbon Capital, Munich Re Ventures, Susquehanna Sustainable Investments, and more than a dozen other energy and technology-focused funds. The round brought Zanskar's total equity funding to $180 million and was the largest venture investment to date in AI-enabled geothermal resource discovery.
The commercial logic was described plainly by Spring Lane's Jason Scott, who joined Zanskar's board: the company had identified more geothermal anomalies in North America than any other company in decades, and the Series C existed to convert that discovery pipeline into generating capacity. Hoiland's own framing was characteristically expansive: "We started Zanskar with the belief that AI would have as profound an impact on geothermal cost and scalability as modern drilling technologies have. The result is a terawatt-scale opportunity."
The company is planning six power plants of approximately 20 megawatts each across its portfolio, targeting commercial operations before 2030. Grid interconnection is the identified bottleneck - queues across the western U.S. are measured in years - and Zanskar is structuring its plants in 20-megawatt increments specifically to manage that constraint, allowing individual projects to move through permitting and interconnection processes independently.
The most contested claim in Hoiland's worldview is also the most important: that the undiscovered geothermal resource base in the U.S. is vastly larger than official estimates suggest.
The most widely cited government analysis, a 2008 U.S. Geological Survey study, estimated undiscovered conventional geothermal reserves at roughly 30 gigawatts of capacity - enough to power approximately 25 million homes. That figure had become the working ceiling for the industry. Hoiland believes it is off by at least an order of magnitude, for two compounding reasons: the 2008 analysis systematically undercounted blind systems because the methods available at the time could not reliably identify them, and modern drilling techniques can extract substantially more power from each system than the estimates assumed. If both adjustments hold, the true resource base could approach 300 gigawatts or more.
The timing works in Zanskar's favor politically. While the Trump administration's 2025 tax legislation stripped incentives from wind and solar, it preserved geothermal tax credits. Energy Secretary Chris Wright has called geothermal "an awesome resource that's under our feet" with the potential to pow…