Energy and Civilization

Smil’s opening move is to refuse metaphor. “All natural processes and all human actions are, in the most fundamental physical sense, transformations of energy.” Every advance a civilization records — bigger harvests, more materials, more goods, more mobility, more information — is a claim on more conversion. That framing sounds obvious until you notice how rarely anyone prices it: the book’s running complaint is that historians write the story of civilization with the energy ledger left out. Braudel defined a civilization across a full page of houses, roofing, arrow-feathering, dialects, and tastes — and never mentioned energy in any form.

What makes Smil worth shelving next to the market books is that he polices his own thesis with the same severity. A late section titled “The Limits of Energy Explanations” concedes that the master variable explains less than its fans want it to. Energy constrains what a society can do; it does not choose what a society does. That double discipline — insist on the ledger, refuse the monocause — is the posture this desk wants toward its own data. The coverage record constrains the story. It does not write it.

The unit of account throughout is the prime mover: whatever converts stored energy into useful work. Muscles, then waterwheels and windmills, then engines and turbines. The book’s periodization runs on two axes — dominant fuel and leading prime mover — and rejects any clean two-era split of history, because waterwheels were doing critical work centuries before anything modern. Keep that frame: when the desk argues about compute, it is arguing about the newest prime mover’s fuel bill.

The most tradeable idea in the book is a ratio: power density, the rate at which energy is produced or consumed per unit of area. Smil calls it “a critical structural determinant of energy systems,” and his worked example is medieval urbanism. A traditional city of 500,000 needed roughly 150,000 hectares of cropland for food and another ~100,000 hectares of sustained-yield woodland for fuel — while occupying as little as 2,500 hectares itself. A city consumed at ~25 W/m² and its fuel supply produced at ~0.25 W/m², so every city dragged behind it a supply hinterland up to a hundred times its own size. City size was capped not by food or water but by the power density of wood.

Read that paragraph with 2026 eyes and it is a datacenter story. A gigawatt campus on a few hundred acres is the highest-power-density settlement humans have ever built, and it drags the same invisible hinterland behind it — generation, transmission, gas supply — that the medieval city dragged in forest. The Power Wall is this book’s box 6.11 with substations in place of fuelwood groves: the constraint never sits inside the fence line; it sits in the hundred-times-larger area that feeds it.

The phytomass numbers explain why coal changed everything: photosynthesis converts under a percent of sunlight, so wood’s power density was pinned near 2% of what traditional urban heating and manufacturing demanded. Fossil fuels collapsed the hinterland into a mine shaft. Every subsequent energy era is a story about raising the density of supply — and the AI era’s bet is that the demand side can now outrun a grid built for the old ratio.

Smil singles out electrification as unlike every prior energy innovation. Earlier prime movers slotted into existing arrangements — a waterwheel turned the millstones the animals had turned. Electricity “required the invention, development, and installation of a whole system”: generation, transmission, and conversion to heat, light, motion, and chemistry, all at once. In 1870, four decades after Faraday’s induction discovery, electric everything was still science fiction, because no one could yet generate it at scale.

The adoption ledger is the useful part. Less than 2% of all fuel became electricity in 1900; still under 10% in 1945; about 25% by century’s end. Global generation grew ~11% a year from 1900 to 1935, better than 9% annually into the early 1970s, then settled near 3.5%. That is the base rate for “the century’s defining technology”: two generations of double-digit growth, then maturity. Anyone modeling AI-driven load growth is implicitly claiming we have re-entered the 11% regime after fifty years at 3.5%. Maybe — but say it out loud, against this baseline.

The system-not-component lesson compounds the density lesson. What broke in 1870 was not the generator; it was everything around it. What binds in 2026 is not the chip — the arithmetic of the buildout keeps landing on turbine order books, interconnect queues, transformer lead times. Component innovation moves at the speed of a product cycle; system innovation moves at the speed of the slowest install.

The book’s most quoted finding: energy transitions are evolutionary, not revolutionary. “Established sources and prime movers can be surprisingly persistent, and new supplies or techniques may become dominant only after long periods of gradual diffusion. A combination of functionality, accessibility, and cost explains most of this inertia.” Sailing ships persisted into the steam era; draft animals worked fields decades after tractors existed; water power still makes roughly a sixth of the world’s electricity.

The one great counterexample proves the rule’s shape rather than breaking it. China quadrupled per-capita primary energy use between 1980 and 2010, became the world’s largest energy consumer in 2009, and by 2015 ran about 30% ahead of the United States — compressing into one generation what earlier industrializers took three to do. Speed like that is possible, but it required a state rebuilding its entire supply system at once, which is precisely the whole-system move Smil says transitions demand. It is also the context the China AI stack sits inside: a country that has already run one compressed transition attempting a second.

Even after that sprint, China’s 2015 per-capita use (~95 GJ) matched France in the early 1970s — and billions in India, Southeast Asia, and Africa intend to replicate the climb. Smil’s point is that the demand curve underneath every other story only goes one way, and none of us can foresee how it will be met. The desk’s translation: power demand is the one input in the AI thesis with no bear case.

A small box in the book does more work than most chapters. Grid reliability is counted in nines: 99.99% availability sounds immaculate and still means 53 minutes dark per year. Five nines cuts it to about 5 minutes; the industry’s aspiration of six nines leaves 32 seconds. Actual U.S. performance runs near 99.98% — weather, vandalism, fuel interruptions.

The gap between 99.98% and what a frontier training run tolerates is the entire behind-the-fence story. When the buildout reaches past the grid — reciprocating engines, cargo-ship gensets, turbines bought on deposit — it is not buying kilowatt-hours, which are cheap. It is buying nines, which are not. Smil’s box prices the difference between energy as commodity and energy as guarantee, and the guarantee is where the margin lives.

The book closes with a gallery of confident energy forecasts, kept honest by hindsight. Glenn Seaborg, chairman of the Atomic Energy Commission, forecast in 1972 that half of America’s generating capacity would be nuclear by 2000, with nuclear-powered ferries running to Mars. The 1980s then ended Western reactor orders almost completely. Smil’s label for the genre: “electricity’s promise that never ends.”

He applies the same audit to the current favorites. PV capacity factors run 11–15% in cloudy climates and about 25% even in Arizona — and in 2015, cloudy Germany produced nearly three times the solar electricity of sunny Spain, which tells you the driver was subsidy design, not sunlight. The desk-relevant discipline is not cynicism about any one technology; it is that capacity announcements are not generation, and generation is not reliability. Every megawatt claim in the coverage record deserves the Smil treatment: what capacity factor, what baseline, who pays.

Set this shelf-mate against the 2026 power book and the pairing is deliberate: one book is the long ledger, the other is the current trade. Smil supplies the base rates — transitions take generations, density is destiny, forecasts overshoot — and the trade is a bet on which of those rates the AI era actually breaks. The record will say. The ledger says: probably fewer than the tape assumes.