What an eight-second geothermal breakthrough reveals about software reliability, engineering, and the future of energy
An extraordinary achievement. An extraordinarily short run
On September 27, roughly 5,000 feet below the surface of the northeast Pacific, a three-ton machine sat on an active underwater volcano and did something its builders say no machine had done before: it produced more electricity than it consumed.
It did this for 8.7 seconds.
Then the battery’s overcurrent protection stopped it. About ten minutes later, during a restart attempt, a power cable feeding one of its pumps failed, and the dive was over.
The machine is Adélie, a test plant built by Seattle startup Endurance Energy. The volcano is Axial Seamount, about 300 miles off the Oregon coast. The team drilled a well into the seafloor, landed a power plant nearby, and ran it in real time from Seattle. And the run ended the way so many of our own deployments end: not with a volcanic catastrophe, but with a protection threshold and a bad cable. Humanity can now park a power plant on an underwater volcano, and it can still be stopped by the same category of problem that takes down a homelab.
It would be easy to read “8.7 seconds” as a punchline. But proof-of-concept tests exist to answer one question cheaply before you spend a fortune answering a hundred others. Endurance’s question was blunt: can a seafloor geothermal plant generate more power than it uses? Briefly, yes. Everything that went wrong afterward is data, arguably the more valuable kind.
How do you generate electricity beneath the ocean?
Axial Seamount sits on a mid-ocean ridge, where tectonic plates pull apart and magma rises close to the seafloor. Seawater seeps into the crust, gets superheated, and returns through hydrothermal vents. The heat is free. Collecting it is not.
According to Canary Media, Adélie uses an organic Rankine cycle, the approach behind many land-based “binary cycle” geothermal plants. The hot geothermal fluid never touches the turbine. Instead, it boils a separate working fluid with a lower boiling point than water inside a heat exchanger; that vapor spins a turbine, which drives a generator, and is then condensed and pumped around the loop again. Endurance mentions a refrigerant pump, which fits: something has to keep that fluid moving, and it runs on electricity.
Which brings us to “net-positive.” Every power plant spends part of its output running itself. Adélie’s 8.7 seconds were the window in which generation exceeded that internal consumption, at a reported median of about 2 kilowatts net and a peak of 16.5 kilowatts net. Canary reports that the turbine ran for about 29 seconds in total and produced roughly 60 watt-hours, with the surplus charging the battery system. For scale, that would run a 2-kilowatt electric kettle for a little under two minutes.
That is calibration, not criticism. “Net-positive” here means the plant briefly out-produced its own electrical appetite, not that the project recovered the energy spent building hardware, sailing a ship, or drilling. Nor does it say anything about cost. Even the well was a mixed result: Endurance’s 52-foot, roughly two-inch-wide well delivered fluid at 260°C but was too restrictive to run the plant, so the team connected Adélie to a natural vent at about 281°C instead. The company itself calls Adélie a test article, not a commercial plant.
The software lesson beneath the sea
Here is where developers should lean in, because the most interesting word in this story is “protection.”
Endurance says the run ended when battery overcurrent protection stopped the system. Its account doesn’t say whether that protection lives in a hardware breaker, battery-management firmware, or control software, and for this argument it doesn’t need to. The shape of the event is the same: a monitored value crossed a limit, and the system stopped rather than continued.
That is not, by itself, a bug. It may be the system doing exactly what it was designed to do.
Developers know the circuit breaker pattern: when a dependency starts failing, stop calling it so the failure doesn’t cascade. It borrows its name from the electrical device, which exists because excessive current can overheat wiring, damage cells, or start a fire. Tripping is the feature.
The questions that matter are subtler. Was the overcurrent a genuine electrical event or a measurement artifact? If genuine, was it a harmless transient a better-tuned limit could have tolerated, or a real hazard? Could the system have degraded gracefully, shedding a load rather than stopping outright? Those are questions about sensors, thresholds, power electronics, and integration, not about whether someone typed > when they meant >=. The public accounts don’t answer them, and pretending otherwise would be fiction.
It also helps to separate two verdicts that are easy to blur. The protection may have worked exactly as intended while the test as a whole fell short of its operating goal. Those statements don’t contradict each other. Anyone who has watched a health check correctly evict a failing node and still ended up with an outage knows the feeling: each component did its job, and the system still didn’t do what you needed.
This is the difference between fail-safe and fault-tolerant design. A fail-safe system, when in doubt, moves to a safe state, usually “off.” A fault-tolerant system keeps operating through a fault, often through redundancy or degraded modes. On a first end-to-end dive, leaning fail-safe is a defensible choice. You don’t get clever with a battery pack you can only reach with a ship and a robot.
This world differs from the web. When a web service misbehaves, you roll back, retry with exponential backoff, page someone. Nobody’s 500 error has ever corroded. In machinery, energy is stored in batteries, spinning turbines, and pressurized loops, and “just restart it” is itself a physical operation with its own risks.
Observability becomes literal here. Endurance’s engineers received temperature and flow data in Seattle, and sent commands back, over the Regional Cabled Array, a seafloor fiber-optic network the University of Washington operates for the National Science Foundation. That link made the prototype debuggable at all. In software we say you can’t fix what you can’t see. A mile underwater, you can’t even touch what you can see.
When the physical world refuses to cooperate
The second failure was not a decision. During the restart attempt, the cable powering Adélie’s refrigerant pump failed. Endurance hasn’t detailed the cause, but whatever the exact mechanism, the episode is a compact summary of deep-sea engineering. Seawater is conductive and corrosive, and at roughly 1,500 meters it presses on everything at about 150 times atmospheric pressure. Every connector and seal is a place where water is looking for a way in.
And nobody can swim down with electrical tape. Retrieval means a ship, a crane, and cooperative weather. Endurance says it once recovered, modified, recharged, and redeployed Adélie in under 16 hours, a genuinely fast turnaround, but a commercial plant can’t depend on that kind of intervention. A system that works is the beginning. A system that keeps working for years, at a maintenance cost its electricity can pay for, is the actual product.
The connection to AI, data centers, and our energy-hungry future
Why should anyone shipping code care about a vent field off Oregon? Because our industry has become one of the fastest-growing electricity customers around.
The International Energy Agency estimated that data centers used about 415 terawatt-hours in 2024, roughly 1.5% of global electricity, and projected this could more than double to around 945 TWh by 2030, with AI as the main driver. In the United States, a Lawrence Berkeley National Laboratory report found data centers used about 4.4% of national electricity in 2023 and could reach 6.7% to 12% by 2028.
Data centers need power continuously and predictably, which makes “firm” clean power, available regardless of weather, especially valuable. Geothermal is one of the few renewables that is firm by nature. The U.S. Department of Energy has estimated that next-generation geothermal could supply around 90 gigawatts in the United States by 2050.
That projection concerns land-based technologies. Subsea geothermal is a separate and much earlier bet. Endurance lists data centers among potential customers, but a 16.5-kilowatt peak powers no GPU cluster, and nothing in this test shows it eventually will.
The difference between possible and practical
The open questions are substantial: running for months rather than seconds, deployment and maintenance costs, getting power to shore, and the effect of years of operation on vent ecosystems. Endurance ran voluntary environmental monitoring and reported no significant changes, with samples still at independent laboratories. A responsible start, but three weeks is not a lifetime.
Skeptics exist. Speaking to Canary Media, geologist Emily Pope of the Center for Climate and Energy Solutions questioned whether offshore geothermal can pay off given the harsh environment and transmission distances, suggesting island nations could more easily adopt solar and batteries.
Endurance’s roadmap includes an upgraded unit called Gentoo this winter, deeper wells, a 2-megawatt generator, and a first commercial project in Tonga targeting power in 2027. Those are plans, not results, and they will be judged the way hardware always is: one deployment at a time, each failure shaping the next design.
Conclusion: Eight seconds can tell us something important
Endurance’s CEO compared the run to Robert Goddard’s first liquid-fueled rocket, which flew for about two and a half seconds. The comparison is flattering, but it carries a useful lesson: what mattered most was everything learned afterward to make flights routine.
Demonstrating that something works once is a milestone. The engineering lives in what follows: systems that keep working, notice when something is wrong, stop safely when they must, and come back up without breaking something else on the way.
Adélie’s first stop may well have been a protection system doing its job. Its second was a cable failing in one of the least forgiving places to have one. The interesting work sits in the narrow space between them, where code meets current and the decision to restart is also a physical event.
So the next time your system decides to stop something, it’s worth asking: are you sure “off” is the safe state, and do you know what it will cost to turn it back on?
Sources and Further Reading
Ars Technica. “Startup taps underwater volcano as energy source in proof-point test.” October 2026. https://arstechnica.com/science/2026/10/startup-taps-underwater-volcano-as-energy-source-in-proof-point-test/
Endurance Energy. “Endurance Energy Drills World’s First Subsea Geothermal Well and Runs First Net-Positive Power Plant.” PR Newswire, October 19, 2026. https://www.prnewswire.com/news-releases/endurance-energy-drills-worlds-first-subsea-geothermal-well-and-runs-first-net-positive-power-plant-302912079.html
Gallucci, Maria. “This geothermal startup produced a tiny bit of power from the seafloor.” Canary Media, October 19, 2026. https://www.canarymedia.com/articles/geothermal/geothermal-startup-power-seafloor-drilling
World Oil. “Endurance Energy drills subsea geothermal well, demonstrates net-positive power.” October 20, 2026. https://www.worldoil.com/news/2026/10/20/endurance-energy-drills-subsea-geothermal-well-demonstrates-net-positive-power/
ThinkGeoEnergy. “Endurance Energy demonstrates subsea geothermal power in Pacific Ocean.” October 2026. https://www.thinkgeoenergy.com/endurance-energy-demonstrates-subsea-geothermal-power-in-pacific-ocean/
Ocean Observatories Initiative. Regional Cabled Array. https://oceanobservatories.org/regional-cabled-array/
National Renewable Energy Laboratory. “Geothermal Electricity Production Basics.” https://www.nrel.gov/research/re-geo-elec-production.html
International Energy Agency. Energy and AI, Executive Summary. 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
Lawrence Berkeley National Laboratory. “Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers” (2024 Report on U.S. Data Center Energy Use). https://energyanalysis.lbl.gov/news/berkeley-lab-report-evaluates-increase-electricity-demand-data-centers
U.S. Department of Energy. Pathways to Commercial Liftoff: Next-Generation Geothermal Power. March 2024. https://liftoff.energy.gov/next-generation-geothermal-power/
