Extreme heat, overloaded grids and the Tumbler Ridge tragedy reveal what happens when powerful technologies expand faster than the systems meant to sustain and govern them.
Adapted from the Susbstack article of the same name.
The phrase “near-death experience” is often used loosely to describe a crisis that forces an institution to reconsider its assumptions. In the summer of 2026, however, the convergence of artificial intelligence, extreme heat and electricity scarcity has given the phrase an unsettling literal dimension.
In British Columbia, the Tumbler Ridge tragedy raised profound questions about what an AI company must do when its systems detect signs of possible real-world violence. Across Europe and North America, heat waves have pushed electricity demand upward while simultaneously reducing the performance of some thermal power plants. In the United States, emergency orders have authorized grid operators to call on backup generation at data centres and other large facilities when the public system is approaching its most serious emergency conditions.
These events are different and should not be forced into a single causal story. Yet they expose the same weakness: digital technology is expanding faster than the physical infrastructure, public institutions and legal standards required to support it.
AI is often presented as weightless intelligence in “the cloud.” There is no cloud. There are data centres, transformers, transmission lines, cooling systems, power plants, water supplies and human beings making consequential decisions. AI is therefore not merely a software revolution; it is an enlargement of society’s physical and institutional metabolism.
Tumbler Ridge and the Limits of Private Moderation
On February 10, 2026, eight people were killed in Tumbler Ridge, British Columbia, and many others were injured. For their families and community, “near-death experience” is not a metaphor. The tragedy itself must remain central, rather than becoming merely an example in a debate about technology.
The subsequent disclosures nevertheless created an unavoidable public-policy question. OpenAI had identified and banned an account associated with the perpetrator in June 2025 after its abuse-detection systems flagged activity involving violence. The company considered contacting the Royal Canadian Mounted Police but concluded that the material did not meet its threshold for referral. After the attack, OpenAI CEO Sam Altman apologized for the company’s failure to alert law enforcement. British Columbia Premier David Eby described the apology as necessary but grossly insufficient.
It is important to state the evidence carefully. The public record supports the conclusion that the account contained disturbing discussions of violent scenarios and that OpenAI’s systems escalated the material internally. It does not justify the broader claim that ChatGPT definitively designed or operationally planned the eventual attack.
The failure was not that an algorithm possessed perfect knowledge of the future and ignored it. The failure was that a private company had warning signals serious enough to ban an account, yet no accepted public framework determined whether, when or how those signals should reach Canadian authorities.
OpenAI now says that cases indicating potentially serious real-world harm receive deeper investigation and that conversations presenting an imminent and credible risk of harm to others are referred to law enforcement. The company also says behavioural and mental-health experts assist with difficult assessments.
That is an improvement, but it does not resolve the democratic problem. A commercial AI company is being asked to perform threat assessment, legal interpretation and public-safety triage. Reporting too little can leave a community exposed; reporting too much could create mass surveillance, overwhelm police and violate privacy.
The answer cannot simply be “trust the company,” nor can every disturbing private conversation be sent automatically to police. Canada needs a transparent framework defining levels of concern, evidence-preservation duties, human review, emergency disclosure and independent auditing. Software engineers and content moderators cannot be left to invent national public-safety policy inside a corporate escalation channel.
Tumbler Ridge was therefore a near-death experience for the belief that AI safety can remain principally a matter of voluntary corporate policy. The technology may be private, but the consequences are public.
The Thermodynamic Trap Facing Nuclear Power
The second warning is physical rather than institutional.
Every thermal power plant must reject waste heat. Nuclear fission supplies the heat that produces steam, the steam drives a turbine, and the turbine produces electricity. Afterward, the steam must be condensed so that the cycle can continue. That requires a colder sink, commonly a river, lake, ocean or cooling tower connected to a water supply.
As the temperature of the cooling water rises, the temperature difference available for heat rejection shrinks. Condensation becomes less effective, plant efficiency can decline and the environmental margin for returning warmed water to an ecosystem becomes narrower.
Nuclear plants are not normally discharging “boiling water” into rivers. The actual constraint is that the discharge must remain within regulated temperature limits designed to protect aquatic ecosystems.
This creates a brutal summer paradox. The hotter the air becomes, the more electricity people need for cooling. Yet the same heat can warm rivers, lower water levels and impair the heat-rejection systems on which nuclear and other thermal plants depend.
During the June 2026 European heat wave, French nuclear output fell by about 4.1 gigawatts as several reactors reduced production in response to high river temperatures and environmental restrictions. Further reductions occurred during the July heat wave.
These were not reactor-safety failures; they were controlled operational responses to cooling and ecological limits. Nevertheless, they removed substantial electricity at precisely the moment power systems were being stressed by cooling demand, weak wind output and expensive replacement generation.
The vulnerability differs by location, cooling design, water availability and access to cooling towers. Nuclear energy remains a major source of low-carbon firm power. The lesson is not that nuclear has failed, but that “firm” does not mean independent of climate.
Electricity planning has traditionally treated heat waves mainly as demand events. They must now be understood as simultaneous demand-and-supply events. Extreme heat raises air-conditioning loads while affecting thermal generation, hydropower, transmission equipment and sometimes wind and solar performance.
The International Energy Agency reports that air conditioning and data centres are among the structural forces accelerating electricity demand and that grid investment has lagged generation investment in many systems.
A power system designed around historical temperature and water conditions may therefore become progressively less reliable even when its nameplate capacity appears adequate.
When the Grid Tells AI to Bring Its Own Power
The third warning comes from the extraordinary growth of AI infrastructure.
Global electricity use by data centres rose by 17 percent in 2025, according to the International Energy Agency. Consumption by AI-focused data centres rose approximately 50 percent. The IEA projects that total data-centre electricity consumption could reach roughly 945 terawatt-hours by 2030, more than double its recent level.
Globally, that demand remains a minority of electricity use. Locally, however, data centres arrive as concentrated loads requiring large blocks of power and rapid interconnection.
PJM, which serves 67 million people in the eastern United States, projects that data-centre growth could add roughly 30 gigawatts between 2025 and 2030. It describes a transition gap in which demand growth has outpaced infrastructure development.
During the July 2026 East Coast heat wave, this company briefly paid nearly US$28,000 per megawatt for regulation resources capable of rapidly balancing supply and demand
In 2026, the U.S. Department of Energy issued a series of emergency orders authorizing PJM to direct backup-generation resources at data centres and other large customers to operate as a last resort before, or during, the grid’s most serious emergency alert level.
A May order was issued amid a heat wave and elevated generation and transmission outages. Later orders extended similar authority during additional summer emergencies.
It would be inaccurate to say that every AI data centre was simply “booted off the grid.” The orders created emergency authority to call on behind-the-meter resources, which can include generators, batteries and other on-site systems.
Yet the precedent is remarkable. One of the most advanced industries in history was effectively told that, under extreme conditions, it might have to reduce its reliance on the public grid and supply more of its own electricity.
This raises a question that engineering alone cannot answer: who receives priority when electricity becomes scarce?
During a severe heat wave, electricity can mean survivable indoor temperatures, refrigerated medicine, functioning hospitals and safe water systems. For an AI company, interruption usually means lost computing time and financial cost.
Those uses are not morally equivalent in an emergency. Data centres should therefore be flexible grid participants rather than untouchable loads. New facilities should disclose realistic demand, finance required upgrades, maintain cleaner backup resources, provide demand response and accept curtailment rules established before crises occur.
Big Tech’s Nuclear Turn
Technology companies understand that electricity availability is becoming a strategic constraint.
Microsoft’s 20-year power-purchase agreement with Constellation is intended to support the restart of Three Mile Island Unit 1, renamed the Crane Clean Energy Center. The proposed project would return about 835 megawatts of carbon-free generation to the PJM grid, subject to regulatory approval, with operation targeted for 2028.
AI firms are no longer merely buying electricity from an existing market. They are helping determine which major generating assets are financed and restarted.
Nuclear power can provide low-carbon, around-the-clock electricity that complements variable wind and solar generation. But the nuclear turn does not eliminate the underlying thermodynamic constraint.
All thermal plants must reject heat, and many require dependable access to cooling water. A climate-resilient AI energy strategy must therefore consider not only carbon emissions and annual megawatt-hours, but also peak demand, water availability, heat rejection, transmission congestion, local ecosystems and performance under future temperature extremes.
The danger is not that AI alone causes heat waves. The danger is that AI demand is being added to systems already stressed by climate-driven cooling loads and aging infrastructure.
Where additional electricity is supplied by fossil fuels, AI growth can also increase the emissions that intensify long-term warming. The result is a feedback loop: hotter conditions increase cooling demand; strained grids rely more heavily on marginal generation; expanding AI adds further load; and the infrastructure needed to manage both heat and computation struggles to keep pace.
From Artificial Intelligence to Thermodynamic Intelligence
AI does not have to remain merely another claimant on a failing energy system. It could become one of the tools used to redesign that system.
But doing so requires a different definition of intelligence. A truly intelligent energy strategy would not evaluate technologies only by how much electricity they produce or how little carbon they emit at the point of generation. It would also ask what happens to heat, water, ecosystems, materials and grid stability.
This is where Thermodynamic Geoengineering deserves rigorous investigation.
TG proposes using the temperature difference between warm surface-ocean water and colder water at depth to operate a heat engine, converting a fraction of transferred heat into useful work while moving the remainder away from the surface layer.
It is not a licence for deployment. Major environmental, engineering, economic and governance questions remain. It is a research proposition built on a simple insight: civilization should investigate energy systems capable of converting part of the heat accumulating in the climate system into useful work.
AI is well suited to evaluating such complexity. It can integrate oceanographic data, improve climate and fluid models, optimize equipment, identify ecological risks, model grid integration and test thousands of operating scenarios. It can also expose weak assumptions.
The convergence of AI and TG would reverse the present relationship. Instead of AI simply demanding ever-larger quantities of electricity from a heating world, AI would help develop energy infrastructure intended to reduce the thermal burden contributing to that world’s instability.
The Warning We Should Not Waste
The events of 2026 do not prove that AI is doomed, nuclear power has failed or the grid faces universal collapse. They warn that these systems are becoming tightly coupled.
An AI platform’s moderation decision can become a matter of public safety. A river’s temperature can influence the availability of electricity across national borders. A data centre’s load can affect the reserve margin available to millions of households. A technology company’s power contract can determine whether a nuclear plant is restarted.
Decisions once confined to separate industries now interact within one thermodynamic and social system.
The near-death experience is therefore not AI’s alone. It belongs to a development model that treats computation as immaterial, electricity as automatically available, climate as an externality and corporate judgment as a substitute for public governance.
We should respond neither with panic nor technological worship. We should establish enforceable AI safety standards, build grids for extreme rather than historical weather, require large digital loads to contribute to reliability, adapt power plants to water and heat constraints, and direct artificial intelligence toward the discovery of energy systems that address accumulated heat as well as emissions.
AI may become one of humanity’s most powerful tools. But intelligence is not measured only by the ability to generate words, images or predictions.
It is measured by whether a system can recognize the conditions required for life and organize its power accordingly.
 The warning from these near-death experiences is that we cannot keep responding to a warming world simply by building ever more capacity to produce energy while ignoring the heat accumulating around us. Rita Mae Brown captured the problem in her 1983 novel Sudden Death: “Insanity is doing the same thing over and over again and expecting different results.” Artificial intelligence gives us unprecedented ability to model, test and challenge alternatives; extreme heat gives us increasing reason to use that ability. Thermodynamic Geoengineering is such an alternative. Its hypothesis, preliminary assessment, and research agenda does not claim it is ready for deployment. It asks a more fundamental question: if the ocean has absorbed most of the excess heat driving climate change, should we at least investigate whether some of that heat can be converted into useful work while reducing the thermal stress at the surface? Continuing to add generation, cooling equipment and grid capacity without confronting the accumulated heat itself is the experiment we have already been running. The results are increasingly visible. Perhaps the genuinely intelligent response—artificially and otherwise—is to try something different.