Inside SERIO: geothermal's hardest problem is the decision, not the drill.

Every major renewable has come down the cost curve. Geothermal has gone the other way.

IRENA's Renewable Power Generation Costs in 2025 puts it starkly: since 2010 the levelised cost of solar PV has fallen 89%, onshore wind 71%, offshore wind 63%. Over the same period the levelised cost of geothermal rose 53%, to around USD 89/MWh — against USD 44/MWh for solar PV and USD 33/MWh for onshore wind.

The instinct is to blame the technology. The instinct is wrong, and IRENA says so in the same document: the increases in dispatchable renewables were "driven by lower output, rather than changes in technology costs."

That sentence relocates the entire problem. Geothermal did not get more expensive to build. It delivered less than it was expected to deliver. That is not a drilling failure. It is an evaluation failure — a gap between what the subsurface was predicted to do and what it actually did.

It is worth being precise here, because the two facts sit awkwardly together and a careful reader will notice. Geothermal's operating fleet runs hard: US plants averaged a 65.9% capacity factor in 2025, roughly twice wind and nearly three times solar. IRENA's series measures something different — the output of newly commissioned capacity worldwide, against what that capacity was expected to produce. The two are not in conflict. Geothermal plants do not run badly once built; too many of them simply arrive at commercial operation having promised more than the reservoir delivers. Reliability is the product. The resource forecast is the risk.

Where the money goes, and where the risk goes

One number needs correcting before anything is built on it. It is widely repeated that drilling is 75–80% of a geothermal project's capital cost. It is not. The IEA puts well drilling and steamfield development at 30–45% of conventional geothermal project investment — the second-largest line, behind the power plant itself. NREL's 2025 cost-curve update gives a range of 30–57%, depending on whether the project is conventional or next-generation.

Drilling is not most of the cost. But it is most of the risk, and the risk is front-loaded: a substantial share of a project's capital must be committed to exploration and confirmation drilling before anyone knows whether the resource is commercially viable at all. The World Bank's ESMAP describes the resource risk at that early stage as "a unique and critical barrier that can effectively stall geothermal development at its inception", and notes that exploration drilling is typically funded with owner equity — an exposure many developers struggle to carry, and so do not take on.

The odds on that spend are sobering, though not in the way oil-and-gas intuition suggests. The largest study of its kind — the IFC and World Bank's survey of 2,613 wells across 57 fields, covering fields that supply some 71% of world geothermal capacity — found that the first well drilled in a new field succeeds about half the time. Success climbs to 74% during development and 83% in operation.

And note what failure means here. A geothermal "dry hole" is a rarity; almost every geothermal well flows to some extent. The failure mode is not an absence of fluid but insufficient commercial capacity. The authors are also honest that fields abandoned after a few bad wells never made it into their database — which skews their own numbers optimistic.

So: a capital-intensive business in which the largest uncertainty is resolved last, on a resource that reveals itself one expensive well at a time.

The constraint is decision quality

Here is the part the industry finds least comfortable. The evidence suggests the binding constraint is not the rock, and not the rig. It is what people do with what they already know.

The cleanest demonstration comes from petroleum geoscience. Studying 25 exploration wells drilled on the Norwegian Continental Shelf between 2011 and 2015, Alexei Milkov found that explorers "consistently and significantly underestimated the geological probability of success and overestimated the success case volumes relative to the available base rate information." His diagnosis is not that geoscientists are careless. It is that they do not treat their own historical track record as relevant to the specific prospect in front of them — and so, in his words, "exploration portfolios generally fail to deliver on promise."

The cost of that is quantifiable, at least in modelling. McVay and Dossary, in SPE Economics & Management, modelled what overconfidence does to a portfolio: at moderate levels, expected disappointment runs at 30–35% of estimated net present value; at greater levels it approaches the whole of it.

Geothermal has its own version. A 2025 paper co-authored by an operator and a Stanford decision-theory group found that the standard practice of loading unsystematic risk into discount rates "significantly reduces Expected Monetary Value, increases exploration costs, and impairs the ability to choose the most appropriate development scenario or to walk away when warranted" — while failing to actually reduce expected losses. Their proposed remedy is a framework that integrates physical numerical models with probabilistic economic evaluations.

That is worth dwelling on. It is a peer-reviewed call for precisely the thing SERIO is built to be: one place where the technical, the commercial, the operational and the financial meet, at the moment a decision is taken.

What AI in the subsurface can and cannot do

It would be easy, and dishonest, to leap from there to a claim that artificial intelligence solves this. It does not, and the published record is unusually clear about why.

Machine learning's demonstrated wins in the subsurface are overwhelmingly in perception, not decision. At the Utah FORGE site, a deep-learning model applied to fibre-optic sensing data detected more than 5,700 microearthquakes where a conventional method found 1,307. That is a real result. It is also a pattern-detection task — finding signal in an enormous sensor stream — and not an act of capital allocation.

Where the field is weakest is exactly where the money is. A review published in Energies this month concludes that machine learning's advantages in geothermal "are not absolute; rather, they are highly contingent upon the quality and scale of the dataset, the rigorousness of the validation method, and the appropriateness of the selected algorithm", and names the open problems as dataset scale and quality, model generalisation, physical consistency, interpretability and deployment feasibility. Its summary of the direction of travel is the telling part: such approaches "aim to shift machine learning from high-accuracy prediction toward trustworthy and deployable decision support." An aim. Not an achievement.

The generalisation problem is not theoretical. Models trained to forecast induced seismicity at one geothermal site have been shown to fail at another — none could forecast all of the larger events, and the results did not transfer between fields.

Then there is data. The largest geothermal well database ever assembled holds 2,613 wells; oil and gas has millions. Geothermal machine learning is trained almost entirely on successes, because failures are rarely published. A model cannot manufacture data that was never recorded, and no amount of architecture repairs that.

Finally, the cautionary tale the sector ought to tell about itself. The most dramatic cost reduction in modern geothermal — the collapse in drilling times at Cape Station — is not an AI story, and the people who achieved it do not claim it is. Fervo attributes it to learning curves and successive generations of well design; NREL attributes it to drill-bit technology and physics-based optimisation. Meanwhile the sector's most prominent AI-driven exploration claims remain company-reported, with no published method, no published data and no published hit rate.

A sober summary comes from a US Department of Energy-funded consortium that includes working geothermal operators, writing about the operational end of the problem: machine learning in geothermal operations "is still a nascent technology that has not seen wide-reaching application in power plant optimization." The scale of the research effort tells the same story — the US federal machine-learning-for-geothermal programme has spent under USD 10 million since 2018. This is a young field, and anyone implying otherwise is selling something.

What SERIO is — and is not

SERIO — Screening, Evaluation, Risk, Investment, Operations and Optimization — is Catoxy's decision-support platform. It is developed in-house and expands with the company. It is built to integrate technical, commercial, operational and financial information so that a decision can be taken with all four in view.

It is not a finished product, and we will not describe it as one. It is not for sale. And it does not claim to improve drilling or exploration outcomes, because we have no outcome data — and, as above, the literature is explicit that nobody credibly does yet.

What it is built to do is narrower, and we think more defensible: to ensure the next decision is taken with everything the last one taught us. Not to preserve expertise — nothing captures tacit expertise, and claims to the contrary do not survive scrutiny — but to preserve the record, the context and the rationale of decisions, and to make them reachable at the moment the next decision is made.

That is a real problem, and a documented one. The Construction Industry Institute finds that 94% of organisations already run some form of lessons-learned process. The recurring finding is that the lessons are never captured, or are captured but unfindable, or are available but ignored. It is not a filing problem. It is a retrieval-and-context problem.

It is also getting harder. The IEA's World Energy Employment 2025 reports that in advanced economies there are 2.4 energy workers nearing retirement for every worker under 25 — against roughly one-to-one in emerging and developing economies — and that two of every three new hires between now and 2035 will be needed simply to replace those retiring; around 60% of companies already report labour shortages. Experience is leaving this industry faster than it arrives. And organisations genuinely do forget: the economics literature on learning curves has long documented that knowledge gained in production depreciates, and that it does not fully carry from one generation of a project to the next.

The obvious objection

There is a rejoinder that deserves a straight answer rather than silence.

MIT's 2025 study of enterprise generative AI found that roughly 95% of pilots produced no measurable return, and that internally built systems succeeded about a third as often as vendor partnerships. SERIO is an internal build. By that base rate, it should fail.

We take the finding seriously — and note what MIT identified as the reason those systems fail: they "do not retain feedback, adapt to context, or improve over time." That is not an argument against building SERIO. It is close to a specification for it. A platform whose entire purpose is to carry context and feedback forward is a response to that diagnosis, not a counterexample to it.

We would rather be measured against that standard than against a claim we have not earned.

The ground we are standing on

Geothermal's cost curve went the wrong way because projects under-delivered against what was expected of the subsurface, and because each project relearned the subsurface from the beginning. The physics is not the obstacle. The drill is not the obstacle. The obstacle is an industry making consequential, expensive, irreversible decisions without systematically using what it already knows.

That is the problem SERIO is being built to work on. Not solved — worked on. In this field, at this moment, that is the only honest thing to say. It is also the more useful one.

---

Sources: IRENA, Renewable Power Generation Costs in 2025 (2026); International Energy Agency, The Future of Geothermal Energy (December 2024) and World Energy Employment 2025 (December 2025); NREL, 2025 Geothermal Drilling Cost Curves Update, NREL/CP-5700-92793 (July 2025); IFC / World Bank, Success of Geothermal Wells: A Global Study (2013); ESMAP, Comparative Analysis of Approaches to Geothermal Resource Risk Mitigation, World Bank (2016); A.V. Milkov, "Integrate instead of ignoring: base rate neglect as a common fallacy of petroleum explorers", AAPG Bulletin 101(12) (2017); D.A. McVay & M. Dossary, "The Value of Assessing Uncertainty", SPE Economics & Management 6(2) (2014); T. Clemens et al., "Incorporating risk in geothermal field development planning: applying decision-theoretic agents", Geothermal Energy (December 2025); Q. Zhang, L. Gou & L. Xu, "Machine Learning for Geothermal Energy Systems", Energies 19(13):3193 (July 2026); S. Karimpouli et al., "Forecasting induced seismicity in enhanced geothermal systems using machine learning", Geophysical Journal International 242(2) (June 2025); P. Yu et al., "DASEventNet: AI-Based Microseismic Detection on DAS Data From the Utah FORGE Well 16A(78)-32 Hydraulic Stimulation", JGR: Solid Earth (2024); P. Siratovich et al., "GOOML: Real World Applications of Machine Learning in Geothermal Operations", GRC Transactions (US DOE award DE-EE0008766); US Department of Energy, Geothermal Technologies Office — Machine Learning; Construction Industry Institute, research on lessons-learned programmes (RS230-1); C. Benkard, "Learning and Forgetting: The Dynamics of Aircraft Production", American Economic Review 90(4) (2000); MIT Project NANDA, The GenAI Divide: State of AI in Business (2025).