The geothermal industry drilled blind for a century. AI found the sites in one year.

For a century, finding geothermal power meant drilling blind. Companies drilled where hot springs surfaced, or where a neighbor's well got lucky, and wrote off most of what they hit. In 2025 a Salt Lake City startup used AI to locate more overlooked geothermal sites than the industry had found in a decade, including one with no surface expression at all. The question now is whether the capital follows.

The International Energy Agency's Future of Geothermal report projects the resource could supply close to 4% of global electricity by 2050. The nearer-term number matters more for investors. U.S. enhanced geothermal capacity is expected to climb from roughly 4GW at the end of 2024 to 8-12GW by 2030, with Fervo, Sage, and a few new entrants splitting the hyperscaler load.

This is a sector that moved from pilot to pipeline between 2024 and 2026. What changed is who pays for it.

The blind-drilling problem

Geothermal's problem was never the heat. The planet stores more of it than the economy will use in centuries. The problem was finding it cheaply enough to justify a well that costs tens of millions of dollars.

For most of the industry's history, exploration meant surface signals. Geysers. Hot springs. A known fault line with a temperature anomaly. The industry got good at the obvious spots and then stalled, because the obvious spots were mostly taken. Drilling a wildcat well to test an unproven formation was the financing equivalent of buying a lottery ticket with a rig attached.

The result was a sector that stayed tiny. In the United States, geothermal still produces around 0.4% of electricity, nearly 70 years after the first plant opened in Northern California.

The AI bet is that the resource isn't rare. It's hidden.

Finding heat with AI

Zanskar's method inverts the drill-first logic. The company trains regional AI models on known hot spots and on its own simulations, then feeds in geological, satellite, gravity, and tectonic data across wide areas. The models flag where a reservoir might sit. Field crews then walk the candidate zone, roughly 100 square miles, drill shallow holes, and check whether the ground temperature matches the prediction.

In December 2025 Zanskar announced the payoff: a blind geothermal system in western Nevada, meaning no surface expression at all, confirmed as a commercial prospect. The company says it is the first such system identified and confirmed in over 30 years. The site, called Big Blind, holds 250°F water at about 2,700 feet, shallow enough to drill with conventional equipment.

"This problem had been unsolvable for decades," CEO Carl Hoiland told MIT Technology Review. CTO Joel Edwards was blunter: "We have dozens of sites that look just like this."

The wildcatter framing the Wall Street Journal applied to Zanskar is not wrong. The company behaves like an oil-patch explorer, only the map is a model. It runs the kind of subsurface maps, gravity surveys, and stochastic simulations a shale operator would recognize, then bets a drilling program on the output.

There is a second proof point beyond greenfield discovery. In 2023 it bought the Lightning Dock geothermal plant in New Mexico after its prior operator put it up for sale for underperformance. Less than eight months later, the company's first step-out production well produced a flow test that its own team called a world-beating gusher, capable of feeding the entire 15MW nameplate plant. By May 2025 production had tripled.

That sequence matters. It separates the models from the marketing, because the same toolset that finds new sites also revived one the industry had given up on.

From discovery to dispatch

Exploration is only half the pipeline. The other half is turning a confirmed reservoir into dispatchable megawatts, and that is where Fervo Energy became the sector's industrial anchor.

It took the oil and gas playbook: horizontal wells, multistage stimulation, and geothermal reservoirs instead of hydrocarbon ones. Its Project Red pilot in Nevada came online in November 2023, producing 3.5MW net for Google's data center load, the first commercial enhanced geothermal system in the United States. Cape Station in Utah then scaled that pilot into a 400MW buildout, the largest enhanced geothermal project under construction anywhere. The first 100MW is contracted to Southern California Edison under a 15-year power purchase agreement and is targeted for the fourth quarter of 2026. The remaining 300MW stages through 2028.

Public research capital de-risked the early innings. The Department of Energy's (DOE) FORGE program in Utah spent a decade defining what a commercial enhanced geothermal well needed to prove, and both Fervo and the national labs built on that baseline. The private capital that followed came in once the drilling data made the resource bankable.

Drilling economics are improving faster than the sector's own projections. Per-lateral drilling time at Cape Station fell from 21 days in 2022 to 6 days in 2024, a 70% reduction. An appraisal well reached 15,765 feet in 16 drilling days, roughly five times the penetration rate of the FORGE site in 2017. In July 2026 it reported that its third-generation well design lifted drilling rates 143% versus the first Cape Station well, with a lateral reaching 7,500 feet in 21 days.

Each step change compounds. Each one also has to keep compounding for the 2030 forecasts to hold.

Hyperscalers are the buyer

Demand is pulling this market, not pushing it. AI data centers need clean, firm, dispatchable power on a 24/7 schedule, and the conventional answers do not fit the timeline. New nuclear sits on a 2030-plus schedule. Solar with four-hour batteries cannot match the temporal profile of a data center load. That leaves geothermal, which runs at a 90%+ capacity factor without storage, as one of the few firm options that can be built inside this decade.

The contracts are the tell. Google holds 115MW of dedicated offtake from Cape Station Phase II under an expanded master agreement signed in September 2023. Southern California Edison buys the first 100MW of Phase I. Meta announced a 150MW agreement with Sage Geosystems in October 2024 for a Texas site, the first hyperscaler geothermal contract east of the Rockies. Microsoft has run a solicitation since late 2024 and, as of mid-2026, had not closed a geothermal deal.

The money gets boring

The technology story is easier to tell than the capital story, and the capital story is the one that changed.

It has been venture-backed since inception, which is itself a statement about the sector. Early geothermal exploration used to be funded like a research project. The company raised $12M in a 2022 Series A, $30M in a 2024 Series B led by Obvious Ventures, then $115M in a January 2026 Series C led by Spring Lane Capital, taking total equity to $180M. In April 2026 it added a $40M development capital facility, structured to scale to $100M, to finance the earliest, riskiest phase of geothermal development: drilling, permitting, and field development. Zanskar describes it as one of the first facilities of its kind.

That facility is the quiet shift. Development-stage geothermal has historically run on expensive equity because lenders had no way to price subsurface risk. A debt structure that scales to $100M treats the resource as a project-finance asset, not a science project. As we wrote in August, green hydrogen reached its final investment decision (FID) moment when the Barrow project found a way to fund the hydrogen economy. Geothermal is working through the same transition, one facility at a time.

Its path shows the equity side maturing in parallel. The company closed an oversubscribed $462M Series E, then went public on Nasdaq under the ticker FRVO, giving the sector a traded equity benchmark for the first time. In June 2026 it announced a partnership with NVIDIA and Pacific Northwest National Laboratory to build EGS-Twin, a digital twin platform that trains AI models on subsurface field data to forecast reservoir behavior in real time. Chief technology officer Jack Norbeck framed it plainly: digital twins could expedite the learning curve of geothermal development the way they have in manufacturing.

The tax code helped at the margin. The Inflation Reduction Act's (IRA) production credit, at $27.5 per MWh for a decade, pushes the post-credit cost of a new enhanced geothermal plant close to competitive with new firm gas in the western interconnection. Credits do not make a project real on their own, but they shorten the distance from the cost curve to the market.

The cost math is where an investor's attention should land. It targets a levelized cost of energy (LCOE), the average cost per unit of power over a plant's lifetime, of roughly $70-80 per MWh at Cape Station Phase I, falling toward $45 at scale. The Energy Information Administration's 2024 unsubsidized benchmarks put combined-cycle gas at $37, utility solar at $40, and onshore wind at $32. Geothermal is not the cheapest electron. It is the most valuable one, because a 90%+ capacity factor with no storage is what a data center actually needs.

What keeps it small

Every asset class has a floor, and geothermal's floor is the wellbore. Drilling, completion, and stimulation account for 50-70% of enhanced geothermal capex. The cost curve is improving, but it is improving from a small base, and the sector will not reach the scale that compounds the way shale did. The IEA modeled an 80% reduction in drilling cost per kilometer by 2035 if the industry replicates the learning rates of unconventional oil and gas. That assumption is not guaranteed. Shale was running about 15,000 horizontal wells a year by 2014. Geothermal will not approach that volume this decade.

Geography is the second constraint. Conventional geothermal is concentrated west of the Rockies, where the crust is thin and the anomalies sit close to the surface. Across most of the United States, heat sufficient for power generation lies seven kilometers or more down, and the deepest geothermal well ever drilled reaches just over six kilometers. That gap is what the exotic drilling technologies are aiming at, and it is why they matter.

Permitting has started to loosen. The House passed bipartisan measures in mid-2026 to speed federal geothermal lease sales and extend the same streamlined environmental reviews the oil and gas industry gets. The effect will show up in project count, not headlines.

The skepticism is fair

The honest counterargument is that a breakout built on three companies is not yet a sector. Fervo, Zanskar, and Sage carry most of the momentum. The rest of the pipeline is thinner than the marketing suggests.

There is also a genuine technical debate about which geothermal wins. Fervo and Zanskar bet on conventional and enhanced systems that pump hot water from fractures in the rock. Eavor chose a closed loop, circulating working fluid through sealed wellbores with no fluid touching the formation, which sidesteps the induced seismicity that derailed projects in Basel in 2009 and Pohang in 2017. Its first commercial loop came online at Geretsried, Germany in late 2024, delivering 8MW thermal and scaling toward 64MW thermal plus 8.2MW electric by 2027. The trade-off is lower power density per well, offset by extreme lateral reach.

Quaise, an MIT spinout, sits at the speculative end. Its gyrotron concept aims to vaporize rock rather than drill it, targeting 10-20 kilometer depths and 400-500°C superhot rock anywhere on the map, including under retired coal plant interconnects. The company has raised roughly $95M through 2024 from Breakthrough Energy Ventures, Mitsubishi, and Standard Investments, with a pre-pilot vertical hole targeted for 2026 and a full pilot for 2028. It is the call option on the whole asset class. It either breaks the geographic constraint or stays a physics demonstration.

Outside the United States the field is smaller but real. Iceland, Indonesia, Kenya, and Turkey already run conventional geothermal at meaningful scale, and Germany is testing both closed-loop and superhot paths. The IEA's 4%-by-2050 projection depends on every one of those paths compounding, not just the American ones.

And the demand side is real but not universal. Microsoft has been soliciting proposals since late 2024 and still has not signed. If hyperscalers hesitate, the premium that makes today's economics work softens.

Signals that would change the call

The difference between a sector and a story is what happens next. These are the markers that separate them.

Geothermal spent decades being the power source that was always a decade away. The 2026 version is different: confirmed sites, signed contracts, public equity, and debt that finally understands the resource. The investment thesis is no longer about whether the heat is there. It is about whether the wellbore economics keep falling, and who is positioned on the other side of that curve.

Sources