PART II The Energy Budget Why Cheap Power Is a Medical Intervention

PART II

The Energy Budget

Why Cheap Power Is a Medical Intervention

Arthur Gazaryants, DOM

Part I ended on a claim about bodies. Aging is entropy winning. Life is what energy buys, and when a body can no longer afford to maintain itself, it starts shutting things down in a predictable order.

What I did not say is that the same sentence describes the society your body lives in.

I want to make an argument that will sound like it belongs to someone else’s field. Cheap, reliable, abundant energy is not an economic preference or a political position. It is a precondition for health. Not health in the abstract, civilizational sense. Health in the sense of your blood pressure, your sleep, the rate at which your cells accumulate damage they cannot repair.

This is not an energy policy essay written by a doctor. It is a medical argument that happens to be about the grid.

What an Energy Budget Actually Is

Start with you.

Right now, sitting still, you are drawing about a hundred watts. Roughly what an old incandescent bulb pulled. You are running that draw continuously, awake and asleep, and you have never once stopped since the day you were born.

That power goes somewhere. Most of it goes to maintenance. Your cells are constantly repairing DNA, refolding proteins that have come undone, clearing out debris, replacing components that have worn past usefulness. None of this is optional and none of it is free. It is the cost of staying organized in a universe that has no particular interest in keeping you that way.

So you are running a budget. Income is what you eat and breathe. Expenses are everything required to hold you together.

When income clears expenses, you build. You lay down muscle, consolidate memory, invest in tissue you will not need for years. When income falls short, you triage. You defer the long-term repairs, fund immediate survival, and hope the deferred maintenance does not come due too soon. Anyone who has run a household through a bad year understands the logic exactly.

Now scale it up without changing anything else.

A country keeps the same ledger. Its income is generation capacity. Its expenses are everything that has to stay running for the place to remain coherent: water treatment, refrigeration, hospitals, heating, the cold chain that keeps insulin and vaccines viable, and now the computation that increasingly mediates all of it. Same rule applies. Surplus lets you build. Deficit forces you to triage, and the things you defer are the ones whose absence will not be noticed until later.

I want to be precise about something, because this is where the argument either holds or falls apart. I am not saying a country is like a body. I am saying they are subject to the same accounting, and that the accounting is not metaphorical. Order costs energy. Take the energy away and the order degrades, at whatever scale you happen to be looking at.

The ledger does not care whether the shortfall starts inside the cell or outside the house.

How I Got Here

For most of my career I treated energy as an interior problem. Everything I worked on lived inside the patient. Mitochondrial function. NAD+. Redox balance. Hormonal signaling. When someone came to me depleted, I went looking for where the production line had broken down, and I nearly always found it somewhere in the machinery of the cell.

What changed my thinking was not something I saw in clinic. It was reading about grid failures.

I was going through the reporting on the Texas freeze, mostly out of general interest, and I kept having a strange sensation of recognition. The sequence was familiar. Not similar. Familiar. Systems shutting down in a specific order, the least urgent first. Reserves burning off faster than anyone had modeled. A cascade where each failure raised the cost of the next one. I had watched that exact pattern a hundred times, on a much smaller stage, in the metabolic panels of people whose cells had run out of money.

Then it occurred to me that the resemblance might not be a resemblance.

Consider what a body actually experiences during a multi-day outage in February. Core temperature drops. The cardiovascular system compensates by constricting peripheral vessels, which raises blood pressure and increases cardiac workload. Sleep fragments, because you cannot get warm. Cortisol rises and stays risen. Meals get skipped or become whatever is shelf-stable. And underneath all of it, the cell’s own energy production is being taxed by a thermoregulatory demand it did not budget for.

Every one of those is something I treat. Every one of them, in my clinic, has an internal cause I can name and address.

But the cell has no way of knowing where the deficit came from. A mitochondrion that cannot make enough ATP and a household that cannot keep the power on produce the same downstream problem, and the body responds identically, because the body is only ever responding to the shortfall itself.

Which means the energy budget I have been treating one patient at a time is a small subsystem of a much larger budget. And I have no ability to prescribe for that one.

When the Grid Fails, Bodies Fail

I want to walk through three events. Not to catalog disasters, and not because the death tolls are the interesting part. They are not. What interests me is the mechanism, because the mechanism is the part that gets left out when these events are covered as infrastructure stories or political stories.

Texas, February 2021.

Winter Storm Uri took most of the state’s grid down for several days. The official state count was 246 deaths. Independent excess-mortality analyses put the real figure closer to 700, and that gap is itself worth noticing: the difference between deaths a coroner attributes to a storm and deaths a population statistician can see in the data.

Here is what a clinician notices. Cold does not primarily kill by making people cold.

Hypothermia was a factor, and it was not the whole picture, or even most of it. When you get cold, your body protects your core by constricting the vessels in your extremities. That raises blood pressure and makes the heart work harder against increased resistance. In a person whose coronary arteries are already narrowed, that is exactly the wrong load at exactly the wrong moment. Cold snaps reliably produce a spike in heart attacks and strokes, and they produce it in people who were managing fine the week before.

Then there is the category almost nobody predicts. Carbon monoxide killed people too, from generators run in garages, grills brought indoors, cars left running in attached spaces. Portable generators were implicated in at least ten Texas deaths, and Harris County alone fielded more than three hundred carbon monoxide calls. Those deaths are not from cold. They are from the absence of electricity, and every one of them was preventable if the grid had held.

And underneath the deaths, there is a much larger number that no dataset contains. Millions of people were simply cold and frightened for a week. Their cortisol rose. Their sleep broke apart. Their glucose control degraded, as it always does under sustained stress and disrupted sleep. None of that killed anyone. All of it left a mark, and nobody drew labs on four million Texans, so as far as the record is concerned, it never happened.

Europe, winter 2022.

After Russian gas supply collapsed, European energy prices went vertical, and across the continent people turned their heating down or off. The Economist’s excess-mortality modelling for that winter suggested the toll ran into the tens of thousands, possibly exceeding the continent’s COVID deaths over the same months.

The mechanism here is different from Texas, and the difference matters more than it might seem.

Nobody in those figures froze to death. A cold home is not an acute exposure. It is a chronic one, delivered continuously for four months. The World Health Organization puts the minimum healthy indoor temperature around 18°C for a reason: sustained cold below that raises blood pressure, increases blood viscosity, suppresses immune function, and makes respiratory infection substantially more likely. Sleep in a cold room is worse sleep, night after night, for a season.

That is a dose problem. In pharmacology we would never confuse a single large exposure with a small one sustained over months. They are different events with different physiology, and the chronic version is frequently the larger one. The Texas story is acute toxicity. The Europe story is chronic exposure, and it is the one that more closely resembles what I actually see in practice.

India, heat and load-shedding.

Recent Indian summers have produced a pattern I think of as the heat-power trap. Extreme heat drives electricity demand up sharply. The demand surge stresses the grid, which produces outages. The outages make the heat considerably more dangerous, because they remove the equipment people rely on to survive it. And when power returns, demand spikes again.

The physiology here is worth spelling out, because popular coverage usually gets it wrong.

Humans shed heat mainly by evaporating sweat. That works well in dry air and poorly in humid air, because humid air is already close to saturated and will not accept much more water. There is a widely repeated claim that the human survivability limit sits at a wet-bulb temperature of 35°C. More recent work suggests the real limit is meaningfully lower, closer to 30-31°C for young healthy adults under experimental conditions, and lower still for older people and anyone with cardiovascular disease.

Air movement is part of how that equation resolves. In humid heat below those thresholds, a fan meaningfully improves evaporative cooling, which is why a fan is not a comfort appliance in a Delhi summer. It is part of the physiological mechanism. Cut the power and you have not made someone uncomfortable, you have removed a component of their thermoregulation.

The second-order effects are the ones that stay with me. Refrigerated medication spoils. Insulin denatures. Dialysis stops. A patient who dies three weeks after an outage because their insulin degraded in a warm refrigerator is recorded as a death from diabetes. Nothing in the record connects them to the grid.

The death counts, in all three cases, are the visible portion, and they are the smallest portion. For every person who died in Texas, millions spent a week in the cold, and their stress hormones and their sleep and their rate of biological aging moved anyway. Nobody counted that. Nobody was measuring.

What I’m Claiming and What I’m Not

I need to be careful here, because I have just made a large leap and I would rather point at it myself than have someone else point at it for me.

Here is the argument, with each piece labeled by whose evidence it actually is.

The first piece is mine. I know what happens when a body’s energy budget fails. I know the order things break in, I know which markers move first, and I know what it takes to restore them, because that is the work I do and have done for years. That part I will defend.

The second piece is established physiology, and I did not discover any of it. Chronic stress degrades mitochondrial function. Sustained cortisol elevation disrupts sleep architecture, impairs glucose handling, and promotes inflammation. Cold raises cardiovascular load. Heat kills through mechanisms we understand well. This is textbook material.

The third piece is documented, and it is also not mine. Energy insecurity produces chronic physiological stress across whole populations. Texas, Europe, and India are the examples I have used, and there are many more.

The fourth piece is an inference. If the first three hold, then energy insecurity should be producing, across millions of people at once, the same failure I treat one patient at a time. That is what I believe, and I think the reasoning is sound.

But it is reasoning, not data. Nobody has run the study. Nobody has followed a cohort through a blackout with serial cortisol, epigenetic clocks, and continuous glucose monitoring, which is exactly the study I would want and exactly the study that does not exist. I am arguing from a mechanism I know well to a scale I cannot personally observe.

I would rather say that plainly than dress it up as something firmer. If someone runs that study and the effect turns out to be small, I will have been wrong about the magnitude. I do not think I am wrong about the direction.

Scarcity Is Mostly Administrative

Which brings me to the part of this that made me want to write it in the first place.

Most of what we experience as energy scarcity is not physical. It is a scheduling problem we created.

Start with a piece of vocabulary, because the whole argument depends on it and it almost never appears in general coverage. Capacity factor is the fraction of the time a power source actually delivers its rated output over a year. On EIA’s 2024 figures, nuclear runs above 92%. Natural gas combined cycle sits just under 60%. Wind manages 34%. Solar comes in at 23%.

Those numbers are not opinions about renewables. They are measurements. And they have a consequence that follows immediately: replacing one unit of reliable generation requires substantially more than one unit of intermittent generation, plus storage to cover the gaps, plus something dispatchable for when both the storage and the weather fail together. I am not against solar and wind. I am for them wherever the arithmetic works, and the arithmetic works in a lot of places. But the arithmetic is the arithmetic.

Now, the part I actually have standing to say.

In medicine there is a rule so basic that we do not usually bother to state it. You do not discontinue a working therapy before the replacement is proven to hold. If a patient is stable on something, you bridge. You start the new agent, you overlap them, you confirm the new one is carrying the load, and only then do you taper the old one off. Stopping first and hoping the replacement works is not a difference of philosophy. It is malpractice, and it is malpractice regardless of how good your reasons were.

Germany shut down its last three reactors in April 2023.

I want to be fair about what happened next, because the popular version of this story is wrong in both directions. In the immediate aftermath, with Russian gas gone and nuclear gone at the same time, coal filled the gap and emissions moved in the wrong direction. Then, over the following two years, renewables scaled and demand fell, and German coal generation dropped to its lowest level in something like seventy years. The destination was reached.

But look at the sequence. A therapy running above 90% capacity factor with almost no carbon output was discontinued before the replacement could carry the load, during a period when the other major input to the system had just been cut off by a war. The buffer was removed at precisely the moment the shock arrived. Several years of higher emissions and a brutal price spike were the cost of arriving at the right destination in the wrong order.

I want to be careful not to overclaim here, because this matters. The 2022 European price spike was driven primarily by the loss of Russian gas, not by German nuclear policy. Anyone who tells you otherwise is selling something. The honest version is narrower and, I think, more damning: the phase-out removed resilience from a system that was about to need all the resilience it had, and the decision was made without any apparent modeling of what would happen if something went wrong at the same time.

California scheduled the closure of Diablo Canyon and then, facing reliability projections it did not like, reversed course and extended the plant’s life. That is the same error, caught mid-taper. It is also, to be clear, the correct response to catching it: you stop the taper and you keep the patient stable. Spain is currently working through a phase-out plan for all seven of its reactors with, as far as I can tell, the same chart in front of it.

None of this is left or right. A grid does not hold political opinions. It holds load. What I am describing is a sequencing error, and sequencing errors are the kind of thing clinicians recognize immediately, because we are trained to, and because we have all seen what happens when someone gets impatient with a taper.

The Demand Nobody Budgeted For

There is a complication in my own argument that I should raise before someone else does.

I am asking for two things that are in tension.

In Part I, I argued that AI is the mechanism that will make longevity medicine affordable. I believe that. The reason comprehensive longevity care currently costs a fortune and reaches almost nobody is that it requires a physician to sit with each patient and integrate dozens of variables continuously over years. AI collapses that cost. It is the difference between a practice that serves a few hundred people and a platform that serves millions.

But AI is also the largest new load anyone has put on the grid in a generation. Training a frontier model consumes electricity on the order of what tens of thousands of homes use in a month. Individual queries are cheap and getting cheaper, though estimates vary widely and older figures are now badly out of date. Multiply even a small number by billions of daily interactions and the aggregate is substantial and growing fast.

Meanwhile generation capacity grows a couple of percent a year.

So I am arguing for abundant energy partly because I want the thing that is about to consume it. That is a real bill, and I would rather name it than pretend the two arguments sit comfortably together. They do not. The resolution, if there is one, is that we have to build faster, and I am aware that saying so is easier than doing it.

Although the more I looked at where capacity has actually been coming from, the more I found that a piece of the answer had already been built, by an industry almost nobody credits for it.

Where the Capacity Came From

If you want to know whether new generation gets built, do not follow the policy debate. Follow whoever is willing to sign a contract for power that nobody else wants.

For most of the past decade, that was Bitcoin miners.

I am aware of how that sentence reads, and I want to be straightforward about my own position. I spent years regarding mining as an elaborate mechanism for converting electricity into nothing, and I have not found much of the industry’s cultural output persuasive. What changed my mind was narrower than that, and it has nothing to do with what Bitcoin is worth. It has to do with a property that mining load has and almost no other large industrial customer does.

It can stop instantly.

A mining operation can shed its entire load in minutes and pick it back up just as fast. Nothing spoils. No process has to restart. No customer is waiting on the other end of it. That makes mining the only kind of very large load a grid operator is actually glad to see, because it behaves as a paying customer during normal conditions and as reserve capacity during bad ones.

In Texas this stopped being theoretical. In August 2023, during a brutal heat wave, Riot Platforms earned $31.7 million in power and demand response credits from ERCOT. In the same month it mined 333 bitcoin, worth about $8.9 million. The grid paid the company roughly three and a half times more to stop than the company could earn by running, and Riot curtailed more than 95% of its load during peak hours to collect it. That is a strange sentence until you see what it describes: a grid operator buying capacity back from a customer who was glad to sell it, at the exact hour the system needed it most.

What that flexibility unlocked was power that previously had no path to market. Remote hydro with no transmission. Wind that was being curtailed because it generated at the wrong hours. Gas being flared at wellheads because moving it was uneconomic. Plants running well below capacity because nobody nearby needed the output. In each case the electricity existed and the buyer did not, and a buyer who can co-locate and does not care about latency or proximity to anything is a buyer who makes the project pencil out.

Then there is the part almost nobody noticed at the time.

Connecting a large new load to the American grid takes years. Interconnection studies, transmission upgrades, queue position. Several years is normal and longer is routine. You cannot pay to skip it, because the constraint is engineering review and physical construction rather than money.

Which means an energized site with a large existing interconnection is not really a real estate asset. It is a time asset. It is the only way to have serious power now instead of at the end of the decade.

Miners spent ten years accumulating exactly that, for reasons that had nothing to do with artificial intelligence.

And then artificial intelligence arrived needing enormous, reliable power immediately, from an industry with no patience and a great deal of capital. The result is that a significant share of the American mining fleet has spent the past two years converting to AI and high-performance computing, or leasing their sites to companies that do. Core Scientific’s agreement with CoreWeave runs past ten billion dollars. TeraWulf leased a 401 MW campus in Kentucky to Anthropic on a twenty-year deal with expected revenues near nineteen billion. Hut 8, IREN and Cipher have signed comparable arrangements with Google-backed partners, Microsoft and AWS respectively. The balance sheets and power portfolios that were built around hash rate are being rebuilt around FLOPS.

Now the honest complication, because without it this is just a good story.

The flexibility does not transfer cleanly. Mining is perfectly interruptible. A large training run is not. Interrupting one is expensive, and the whole reason those sites are valuable is a grid-friendly load profile that a training cluster does not have. If every mining site simply becomes an AI site, the grid has traded its most cooperative customer for one of its least cooperative, on the same interconnection.

The most interesting response to this comes from Fred Thiel at MARA, and it is more sophisticated than the pivot story. His argument is not that AI is as interruptible as mining. It is that you keep the mining. You retain mining capacity on the same sites, both as a bridge during buildout and as permanent flexible ballast, so that the combined load can still curtail even when the AI portion cannot. Pair that with generation you own or control behind the meter, and the site as a whole stays grid-compatible even though one component of it is inflexible.

MARA demonstrated the principle during Winter Storm Fern in February 2026, voluntarily curtailing roughly 770 MW across three major power markets and powering down close to 70% of its global hashrate ahead of peak stress. It was not alone. Across the whole bitcoin network something like 12 GW of load came off the grid during that storm, enough that block times stretched to fourteen minutes. An entire industry stepped back from the meter during the hours a cold snap was threatening to break the system.

That is a genuinely different proposition from a clean swap. It treats mining’s interruptibility as a permanent engineering feature rather than a phase to be grown out of.

I do not think this vindicates everything about the industry, and I am not trying to. What I think it establishes is narrower and, to me, more interesting. A speculative business that a great many serious people regarded as pure waste turned out to be one of the mechanisms that financed and de-risked a meaningful slice of American generation capacity, and secured the grid positions that the next technological era now depends on. I do not have to admire the reason to notice the result.

What the Data Center Panic Gets Wrong

Most of the alarm about data centers is either overstated or aimed at a version of the technology that is already being engineered out. Let me take the objections in order, and concede what should be conceded.

Start with the grid, because this is where the framing is most wrong.

The common claim is that we do not have enough power for AI. That is not quite what the data shows. Researchers at Duke have looked at how often the American grid is actually at its limit, and the answer is: rarely. Peak conditions occupy a small fraction of the hours in a year, and the system is built to survive those hours, which means it is substantially underused during most of the others.

Their finding is specific and worth stating precisely. Across the twenty-two largest balancing authority areas, which carry about 95% of American load, 76 GW of new demand could be absorbed by the existing system if that demand agreed to curtail 0.25% of the time, which works out to roughly twenty-two hours a year. At half a percent the figure is 98 GW. At one percent it is 126 GW. The study’s own term for this is curtailment-enabled headroom.

Set that against projected AI demand. EPRI and Epoch AI put US AI power capacity around 5 GW today and above 50 GW by 2030. The shortage looks very different once you put those two numbers beside each other. We are not short of capacity. We are short of flexible capacity, which is a considerably easier problem and a much cheaper one.

And here is the catch, which matters as much as the statistic. The headroom only exists if the new loads actually curtail during the genuine peak stress hours. Not on average. On the specific afternoons when the grid is in trouble. A training run with a deadline and a large capital cost attached is precisely the load least inclined to volunteer, and an argument that assumes otherwise is doing arithmetic on a promise. This is exactly why the hybrid approach matters: something on the site has to be willing to go dark, and if the AI will not, something else has to.

On water.

The image of a data center evaporating a town’s water supply comes from a real design, and that design is becoming a legacy one. Traditional cooling relies on evaporation, which consumes water by definition. Closed-loop systems and direct-to-chip or immersion liquid cooling recirculate instead, and cut on-site consumption dramatically. The trajectory for new builds is clear enough that the objection is aging out faster than the debate about it is.

I had assumed there was a catch here, that saving the water simply meant spending more electricity, and I was wrong. Closed-loop liquid cooling runs at a power usage effectiveness around 1.05 to 1.2, against roughly 1.4 to 1.5 for air cooling, which is a ten to fifty percent energy saving. The old tradeoff was between evaporative cooling, cheap on power and expensive on water, and air cooling, which uses no water and a great deal of power. Liquid cooling largely escapes both horns.

There is a real catch, and it is a policy one rather than an engineering one. Water-efficiency targets written without reference to energy efficiency can push operators off evaporative cooling and back onto air, which saves the water and burns considerably more power. A rule aimed at one number can quietly worsen the other.

And the water consumed upstream at the generating station does not disappear because you changed the design on site. Thermoelectric generation is itself water-intensive. The honest claim is that on-site water use is a solved or solvable engineering problem, not that water is a non-issue.

On siting and local impact.

These facilities are moving toward power, land, and fiber, which usually means industrial and remote locations rather than neighborhoods. That genuinely reduces friction, and a well-sited facility is a large contained industrial load rather than something the surrounding community has to live inside.

But I would not tell you the problem is solved, because the record says otherwise. Loudoun County and Memphis are live disputes, not hypotheticals, and people who ended up with a bad neighbor and a constant low-frequency hum are not being irrational. The industry’s track record on community relations is uneven, and the correct posture is that siting and design can address this rather than that they already have.

On the strategic question, which is the one that actually decides this.

Compute is becoming a foundational national capability in the way that steel and oil once were. It sits underneath military capacity, industrial capacity, scientific output, and the information environment simultaneously. A country that cannot generate its own frontier compute does not merely lose an industry. It becomes dependent across all of those domains at once, and it does so without anyone having to fight a war over it.

The constraint on that capability is electrical. Not talent, not capital, not silicon in the long run. Power and the ability to connect to it.

China is adding generation at a pace the United States is not remotely matching, across nuclear, solar, wind, and coal at the same time, and it is doing so under a planning regime that does not spend a decade on permitting. If the American build-out stalls on local environmental anxiety and procedural delay while that continues, the resulting gap does not stay an energy gap. It becomes a compute gap, then a capability gap, and the timelines involved mean it is not the sort of thing you correct in a single administration. Interconnection queues and construction schedules do not accelerate because the situation became urgent.

Underbuilding is a decision. It is simply one that gets made by default, through a thousand separate delays that each look reasonable in isolation.

What Abundance Actually Buys, Medically

Here is the part that motivates all of it for me.

Energy costs propagate into computation costs, and computation costs are currently the thing standing between preventive medicine and everybody who needs it.

Consider what changes when continuous monitoring stops being expensive. Right now, most people’s metabolic health is assessed with a fasting glucose drawn once a year, which is roughly like judging a city’s traffic by photographing one intersection at 6 a.m. every January. A continuous glucose monitor produces a reading every few minutes, which reveals variability, post-meal response, the effect of a bad night’s sleep, the effect of a stressful week. It is a different kind of information, and it turns treatment from guesswork into something closer to engineering.

The barrier has never really been the sensor. It is the cost of analyzing millions of data points per patient per year and turning them into a decision. That is a compute problem, and compute problems are energy problems.

Concierge longevity medicine, done comprehensively, runs somewhere near a hundred thousand dollars a year. That is not a market. It is a rounding error serving a fraction of a percent. The trajectory I am building toward, and I think it is achievable this decade, is a few hundred dollars a year for the analytical layer. Not because the medicine gets worse but because the marginal cost of the analysis approaches zero, and it approaches zero largely because power gets cheap.

At that price, prevention becomes cheaper than treatment for essentially everyone, and the entire economics of medicine inverts. Disease gets caught five years earlier, while it is still reversible. Interventions get dosed on your response rather than a population average.

This is what Interface Doctor is for. It calculates your biological age, tracks your markers over time, and shows you where entropy is gaining and where you are pushing it back. The reason I keep writing about power plants is that the platform is only affordable for most people if electricity is.

Where I Could Be Wrong

Two things.

The first is that nuclear has real problems and I have been describing it favorably. Vogtle Units 3 and 4 in Georgia came in around thirty-six billion dollars against an original estimate of $14.8 billion, and took roughly fifteen years from construction start to Unit 4 entering service in April 2024. In pharmaceutical terms, that is a therapy with an outstanding efficacy profile and an unacceptable time-to-onset. Time-to-onset matters when the patient is deteriorating now. Small modular reactors are the proposed answer to exactly this, and I should be honest that at the moment they are mostly a promise with a small number of units actually operating. Waste is manageable and it is not nothing.

The second is more basic, and it applies with extra force to the two sections above. I am a physician reading energy data. I am not a grid engineer, an economist, or an energy analyst. I can speak with real authority about what energy deficits do to a human body, and about the order in which biological systems fail when the budget runs out. Everything upstream of that, I am reasoning about from public sources like anyone else. Where those two domains meet, I think I see something that people working in only one of them tend to miss. That is the entire claim I am making, and I would not want it mistaken for a broader one.

What This Means Practically

Nobody reading this can fix a grid, so let me split it into what you can do and what you can support.

Personally, treat energy resilience as a health intervention, because that is what it is. If you or anyone in your household depends on a CPAP, a refrigerated medication, insulin, or a powered device, you need a plan for the days when the power is out, and battery backup for medical equipment is cheaper than it used to be. Think about thermal resilience in your home before you need it: heat is the exposure people most consistently underestimate, and it is the one with the clearest metabolic signature. And if you already wear something that tracks sleep or glucose, look at what your own data does during the next heat wave or outage. Most people are surprised. You are running the experiment on yourself regardless; you may as well read the result.

Civically, the useful things are boring. Permitting reform so that building anything does not take a decade. License extensions for plants that are working. Grid hardening, which nobody campaigns on because nobody notices when it succeeds. I wrote in an earlier essay that energy should be politically boring, and I meant it as a goal rather than a complaint. Boring infrastructure is infrastructure that holds.

The Same Argument at Two Scales

Part I made a claim about a person: when your cells can no longer afford to maintain you, your future contracts, and it contracts before you notice it happening. The woman I described in that essay had lost the ability to plan, and what she had actually lost was the metabolic capacity that planning requires.

A country does the same thing. When the energy budget tightens, maintenance gets deferred, long projects get abandoned, and the horizon shortens until nobody in charge is thinking past the current crisis. The mechanism is not similar to the biological one. It is the same mechanism, and the bill comes due in the same currency.

That is why I have written a piece about power plants and put it in a series about longevity. The two are not adjacent topics. They are the same topic at different magnifications.

In Part III I want to take up what happens next: if energy is what a civilization can produce, human capital is what it does with the people producing it. And AI is in the process of breaking the connection between the two.

Glossary of Key Terms

Capacity factor: The fraction of the time a power source actually delivers its rated output over a year. Nuclear runs above 90%, wind around 35%, solar around 25%. The single most useful number for understanding why replacing one source with another is not a one-for-one swap.

Baseload: The minimum continuous demand a grid must always meet. Baseload sources run constantly rather than following demand up and down.

Dispatchable vs. intermittent: A dispatchable source can be turned up or down when you need it. An intermittent source produces when conditions allow, on its own schedule rather than yours.

Levelized cost of energy (LCOE): The average cost per unit of electricity over a plant’s lifetime, including construction, fuel, and operation. Useful for comparison, but frequently misleading because it does not price reliability.

Small modular reactor (SMR): A factory-built nuclear reactor, typically around 300 MW, designed to be assembled from standard components rather than custom-built on site. The intended solution to nuclear’s construction-time problem.

Grid inertia: The stabilizing effect of large spinning generators, which resist sudden frequency changes simply by virtue of their mass. Traditional plants provide it automatically; solar and wind do not, which is a real engineering challenge with real engineering answers.

Curtailment: Shutting down generation the grid cannot absorb, or reducing a load on request. Generation curtailment usually happens when intermittent sources overproduce during low demand. Load curtailment is a customer agreeing to stop drawing power when the system is stressed, generally in exchange for payment.

Interconnection queue: The waiting list to connect a new generator or a new large load to the grid. Requires engineering studies and physical upgrades, routinely takes years, and cannot be shortened with money. Currently the binding constraint on how fast anything gets built in the United States.

Behind-the-meter generation: Power generated on site and consumed on site, without passing through the public grid. Bypasses the interconnection queue, which is why large loads increasingly want it.

Flexible load: A customer that can reduce or stop consumption on demand. Valuable to a grid operator because it functions as reserve capacity without anyone having to build a plant.

Wet-bulb temperature: A combined measure of heat and humidity that reflects how effectively sweat can evaporate. It matters medically because human cooling depends on evaporation, and above certain wet-bulb values that mechanism fails regardless of how healthy you are.

Terawatt-hour (TWh): A unit of energy equal to a billion kilowatt-hours. Used for describing national-scale consumption.

References

1. Texas Department of State Health Services. February 2021 Winter Storm-Related Deaths. December 30, 2021. https://www.dshs.texas.gov/sites/default/files/news/updates/SMOC_FebWinterStorm_MortalitySurvReport_12-30-21.pdf
2. Texas Tribune. “Texas winter storm official death toll now put at 246.” January 2, 2022. https://www.texastribune.org/2022/01/02/texas-winter-storm-final-death-toll-246/
3. Aldhous, Peter. “Texas Is Still Not Recognizing The Full Death Toll Of Last Year’s Devastating Winter Storm.” BuzzFeed News. Excess-mortality estimate of 702 deaths, range 426–978. https://www.buzzfeednews.com/article/peteraldhous/texas-winter-storm-death-toll
4. Texas Tribune. “Congress investigates portable generator manufacturers following carbon monoxide deaths in Texas and other states.” June 29, 2022. https://www.texastribune.org/2022/06/29/texas-winter-storm-carbon-monoxide-deaths-generators/
5. The Economist. “Expensive energy may have killed more Europeans than covid-19 last winter.” May 10, 2023. https://www.economist.com/graphic-detail/2023/05/10/expensive-energy-may-have-killed-more-europeans-than-covid-19-last-winter
6. World Health Organization. WHO Housing and Health Guidelines. 2018. Minimum indoor temperature of 18°C.
7. Marmot Review Team. The Health Impacts of Cold Homes and Fuel Poverty. Friends of the Earth and the Marmot Review Team, 2011.
8. Vecellio, D.J., Wolf, S.T., Cottle, R.M., and Kenney, W.L. “Evaluating the 35°C wet-bulb temperature adaptability threshold for young, healthy subjects (PSU HEAT Project).” Journal of Applied Physiology, 2022. Mean critical wet-bulb temperature 30.55 ± 0.98°C. https://journals.physiology.org/doi/full/10.1152/japplphysiol.00738.2021
9. US Energy Information Administration. Capacity factor data by generation source, 2024. Electric Power Monthly.
10. Clean Energy Wire. “Germany’s coal power production drops to lowest level in 60 years in 2023.” https://www.cleanenergywire.org/news/germanys-coal-power-production-drops-lowest-level-60-years-2023
11. Clean Energy Wire. “German coal use continues downward trend in 2024.” https://www.cleanenergywire.org/news/german-coal-use-continues-downward-trend-2024
12. NucNet. “Germany’s Nuclear Phaseout Has Increased CO2 Emissions – And Prices, Analysis Suggests.” January 3, 2025. https://www.nucnet.org/news/germany-s-nuclear-phaseout-has-increased-co2-emissions-and-prices-analysis-suggests-3-1-2025
13. Our World in Data. Germany CO₂ and Greenhouse Gas Emissions Country Profile. https://ourworldindata.org/profile/co2/germany
14. California Senate Bill 846 (2022), extending Diablo Canyon operations to 2030.
15. NucNet. “Vogtle-4 / US Nuclear Power Plant Connected To Grid For First Time.” January 3, 2024. https://www.nucnet.org/news/us-nuclear-power-plant-connected-to-grid-for-first-time-3-1-2024
16. POWER Magazine. “What Was Learned from Building New Nuclear Reactors?” https://www.powermag.com/what-was-learned-from-building-new-nuclear-reactors/
17. Norris, T.H., Profeta, T., Patino-Echeverri, D., and Cowie-Haskell, A. Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems. Nicholas Institute for Energy, Environment & Sustainability, Duke University. February 13, 2025. https://nicholasinstitute.duke.edu/publications/rethinking-load-growth
18. Utility Dive. “Existing US grid can handle ‘significant’ new flexible load: report.” https://www.utilitydive.com/news/us-grid-headroom-flexible-load-data-center-ai-ev-duke-report/739767/
19. S&P Global Commodity Insights. “Artificial intelligence power demand in US could top 50 GW by 2030: EPRI.” https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/081325-artificial-intelligence-power-demand-in-us-could-top-50-gw-by-2030-epri
20. Lawrence Berkeley National Laboratory. 2024 United States Data Center Energy Usage Report.
21. Lawrence Berkeley National Laboratory. Queued Up: 2025 Edition — Characteristics of Power Plants Seeking Transmission Interconnection. https://emp.lbl.gov/queues
22. CNBC. “Texas paid bitcoin miner Riot $31.7 million to shut down during heat wave in August.” September 6, 2023. https://www.cnbc.com/2023/09/06/texas-paid-bitcoin-miner-riot-31point7-million-to-shut-down-in-august.html
23. Riot Platforms. August 2023 Production and Operations Update. SEC filing exhibit. https://www.sec.gov/Archives/edgar/data/1167419/000155837023015517/riot-20230906xex99d1.htm
24. MARA Holdings. “Case Study: Balancing the Grid During Winter Storm Fern.” https://www.mara.com/posts/balancing-the-grid-during-winter-storm-fern
25. Blockspace. “Bitcoin hashrate drops 8% as US miners curtail during Winter Storm Fern.” https://blockspace.media/insight/bitcoin-hashrate-drops-8-as-us-miners-curtail-during-winter-storm-fern/
26. MARA Holdings. “The Future of Bitcoin Mining with Fred Thiel.” https://www.mara.com/posts/the-future-of-bitcoin-mining-with-fred-thiel
27. Eaton. “Energy consumption in data centers: air versus liquid cooling.” https://www.eaton.com/us/en-us/markets/data-centers/data-center-cooling/efficiency/energy-consumption-in-data-centers-air-versus-liquid-cooling.html
28. Equinix. “What Is Water Usage Effectiveness (WUE) in Data Centers?” November 13, 2024. https://blog.equinix.com/blog/2024/11/13/what-is-water-usage-effectiveness-wue-in-data-centers/
29. Environmental and Energy Study Institute. “Data Centers and Water Consumption.” https://www.eesi.org/articles/view/data-centers-and-water-consumption
30. CNBC. “Elon Musk’s Memphis AI empire is the epicenter of the data center backlash.” July 16, 2026. https://www.cnbc.com/2026/07/16/elon-musk-memphis-ai-colossus-data-center.html
31. CBS News. “Mississippi homeowners blame a noisy data center plant for sleepless nights.” https://www.cbsnews.com/news/mississippi-elon-musk-xai-data-center-power-plant-noise/
32. Tom’s Hardware. “Virginia county with 250 data centers begins to rein in building.” https://www.tomshardware.com/tech-industry/data-centers/virginia-county-with-250-data-centers-begins-to-rein-in-building-loudouns-more-than-250-data-centers-made-it-one-of-the-richest-counties-in-the-us-but-residents-are-pushing-back
33. Enerdata. “China installs record capacity for solar (+45%) and wind (+18%) in 2024.” https://www.enerdata.net/publications/daily-energy-news/china-installs-record-capacity-solar-45-and-wind-18-2024.html
34. US Energy Information Administration. “China continues rapid growth of nuclear power capacity.” https://www.eia.gov/todayinenergy/detail.php?id=61927