Energy and Civilization

Smil’s last chapter opens by refusing metaphor. “All natural processes and all human actions are, in the most fundamental physical sense, transformations of energy.” On that view each gain a society makes, from food to travel to knowledge, has to be paid for in converted energy. 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. His sharpest example is Fernand Braudel, whose long definition of a civilization lists its houses, skills, dialects and cooking without once naming energy.

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 — how much energy a given patch of ground yields or uses over time. 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 in his example the city leaned on farmland and woodland about a hundred times its own area. 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 turns less than half a percent of sunlight into plant matter, so a temperate woodlot could at best supply about 2% of the power density a traditional town burned. 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 new 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 figures are the useful part. On Smil’s numbers, the share of fuel turned into electricity went from under 2% in 1900 to about a quarter by 2000, and was still below a tenth at the end of the Second World War. World output of electricity rose about 11% a year until 1935 and more than 9% a year until the early 1970s; for the rest of the century the rate was about 3.5%. That is the base rate for the century’s defining technology: two generations of fast 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. Smil’s figures for China show energy use per head rising fourfold over the thirty years to 2010; the country passed the United States as the largest user in 2009 and was about 30% ahead by 2015. That compressed 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, Smil puts China’s use per head in 2015, about 95 GJ, at the level France had reached in the early 1970s, with billions more people in India, Southeast Asia and Africa hoping to follow. He calls rising demand a truism, but adds that “none of us can foresee how it will be met”. The desk’s reading: of all the inputs to the AI thesis, demand for power is the one the long record gives least reason to doubt; how it gets met is the open question.

The book closes with a gallery of confident energy forecasts, kept honest by hindsight. Smil cites Glenn Seaborg, chairman of the Atomic Energy Commission, forecasting in 1971 that half of America’s generating capacity would be nuclear by 2000 and that nuclear spacecraft would be carrying crews to Mars. Instead, orders for new reactors in the West all but stopped in the 1980s. The section’s title names the genre: “Electricity’s promise that never ends”.

He applies the same audit to the current favorites. Solar panels without tracking, he notes, run at only 11–15% of their rated output over a year in cloudy places and about 25% in Arizona. His proof that the industry’s growth was policy, not sunlight: “in 2015, cloudy Germany produced nearly three times as much PV electricity as sunny Spain”. 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.