Four Books, One Synthesis: Can We Power the Intelligence Revolution?

Four Events, One Question

In April 2025, the International Energy Agency published Energy and AI, assessing artificial intelligence’s energy footprint. Global datacentre electricity consumption reached 415 terawatt-hours in 2024, roughly 1.5% of world electricity demand. AI workloads accounted for up to 15% of that total, projected to reach 50% by 2030. Total datacentre consumption was on course to more than double. By 2030, datacentres would consume more electricity than all energy-intensive manufacturing combined: aluminium, steel, cement, and chemicals.

The obvious response was higher efficiency. Three months earlier, a Chinese AI lab matched frontier AI performance at a fraction of expected compute cost. DeepSeek’s R1 activated only 37 billion of its 671 billion parameters per query. Nvidia lost $600 billion in a single trading day, the largest single-day market cap loss in Wall Street history. An efficiency miracle, at first glance. Then the paradox. The data told a different story: inference costs had already fallen 1,000x in three years, and cheaper AI had produced more AI, not less. Jevons’ paradox, articulated in 1865, was playing out in silicon.

Capital moved accordingly. Microsoft restarted Three Mile Island Unit 1, a nuclear reactor mothballed since 2019. Amazon acquired a nuclear-powered campus at Susquehanna. Google signed procurement agreements with Kairos Power for small modular reactors. Software companies buying cold nuclear plants. When the world’s most capital-efficient companies do that, it means their demand projections have already outgrown the grid.

Then the supply side. The conflict between Venezuela and Iran disrupted tanker routes and spiked Brent crude. Not an AI story, at first glance. But the underlying logic was identical: the geography of energy supply is heavily contested. Pipelines are leverage. Minerals are weapons. Power grids are strategic assets.

These four developments frame the central question of our era: Can we power the intelligence revolution?

The capital markets have already placed their bets. One trillion dollars in committed hyperscaler spend and a USD 5 trillion Nvidia valuation assume the answer is yes. These are not just bets on AI. They are bets that the energy to run it will be there.

This series attempts an answer.

Four deep dives into foundational books, followed by a synthesis that applies their frameworks to the question none of these authors could fully address: what happens when efficiency gains compound at AI speed?

Why This Series

The Gap in the Discourse

The AI community talks incessantly about scaling laws, parameter counts, and benchmark performance. It talks far less about where the watts come from. Very few practitioners hold a clear mental model of the bottlenecks, the key metrics, or the timelines that will determine whether their models can actually run at scale.

The energy community has developed rigorous frameworks for understanding transitions, constraints, and physical limits. Vaclav Smil has spent forty years quantifying how civilisations harness energy. Daniel Yergin won a Pulitzer Prize chronicling oil’s role in geopolitics. Ludovic Subran runs climate economics research at Europe’s largest insurer. These frameworks offer essential lenses for thinking about AI’s infrastructure challenge. But they were built to explain systems that move slowly. The question at the core of this series is different: how fast can we push the physical boundaries when capital and political will align?

This series aims at bridging this gap. We draw on the best thinking from energy systems analysis to address a core question: can AI realistically be powered at the projected pace? The answer depends on constraints that the AI community sometimes overlooks: how long energy transitions actually take, how much land renewables require, how much capital must flow, and whose political interests must align.

The Thesis in Brief

1. Energy transitions historically take 50 to 75 years because they depend on long-lived infrastructure, not just technological possibility. Coal plants operate for forty years. Cars stay on the road for twelve. Buildings last a century. These asset lives create path dependence: yesterday’s capital decisions constrain tomorrow’s options.

2. Land constraints matter as much as cost curves. Renewable energy is abundant but diffuse. A solar farm delivering the same annual energy as a nuclear plant requires 50 to 100 times the land. When datacentres demand 500 watts per square metre and solar delivers 10. This arithmetic forces choices that cost projections alone don’t explain.

3. Capital must flow at scales comparable to post-war reconstruction. Energy infrastructure has multi-decade payback horizons, regulatory complexity, and political risk that private markets chronically underprice. The bottleneck is not technology. It is the architecture of capital allocation itself.

4. Geopolitics persists regardless of climate ambition. Energy has always been a political instrument. The actors who controlled the previous energy system have consistently fought to shape the terms of the next one. The same dynamic now governs chips and critical minerals.

5. Efficiency may be a trap. In 1865, William Stanley Jevons observed that improvements in coal efficiency increased rather than decreased total consumption. The same pattern appears in compute: GPU efficiency has improved 10x per decade while consumption has grown 1000x. The synthesis confronts the central paradox: AI is simultaneously the largest new load on the grid and potentially the most powerful tool we have for accelerating the transition it complicates.

6. The answer may hinge on nuclear. It is the only low-carbon source with the power density AI infrastructure actually requires, which is why hyperscalers are buying cold reactors, and why it sits at the centre of our synthesis.

Six constraints. Four books that explain the first four, written before AI emerged as a major energy consumer. One synthesis that confronts the last two, and applies all six to the question these frameworks were never designed to answer.

The Reading List

Each of the first four parts covers an authoritative text. The goal: summarise these books so you don’t have to (or give you enough of a taste to want to read them), extract what matters for technology and investment decisions, and connect historical frameworks to contemporary AI realities.

• Part 1: Vaclav Smil, Energy and Civilization (2017) — The 50-Year Rule: Why Transitions Take Generations. Distinguished Professor Emeritus at the University of Manitoba and author of over forty books, Smil reveals why energy transitions take 50 to 75 years, why quality of life seems to saturate somewhere between 75 and 110 gigajoules per capita, and why energy enables but never determines civilisational outcomes.

• Part 2: Vaclav Smil, Power Density (2015) — Watts Per Square Metre: The Land Constraint Nobody Talks About. This companion volume quantifies the most overlooked variable in energy analysis: power density, the rate at which energy flows through a given area. Fossil fuels deliver 1,000 to 10,000 watts per square metre. Solar manages 5 to 10. Wind struggles past 1. Nuclear matches fossil fuels at 2,000 to 6,000.

• Part 3: L. Subran & M. Zimmer, Investing in a Changing Climate (2023) — The 84% Gap: What Climate Investment Actually Requires. Allianz’s chief economist calculated the investment required for the EU to achieve its climate targets and finds an 84% shortfall. Closing the gap demands 480 billion euros annually through 2030, equivalent to building fifteen Channel Tunnels every year.

• Part 4: Daniel Yergin, The New Map (2020) — Five Maps and a Warning: Why Geopolitics Resists Decarbonisation. The Pulitzer Prize-winning historian of oil argues that physical geography, pipeline routes, and installed infrastructure constrain energy transition more than policy ambition can accelerate it.

• Part 5: The Synthesis — William Stanley Jevons & Tae Kim: Jevons Meets Jensen. The final piece applies all frameworks to the core question: can AI realistically be powered at the projected pace?

How to Read This Series

The first four parts are framework providers. Each introduces concepts and models that are meant to reshape how you think about energy: Smil’s 50-year rule, power density arithmetic, Subran’s investment gap, Yergin’s geopolitical maps. You can skim the articles and read only what’s in bold. The key takeaways converge in Part 5, the Synthesis, but we believe there’s value in building the mental scaffolding first.

A Note on Method

Three perspectives dominate discussions of energy and technology:

• In tech circles, techno-solutionism prevails: the assumption that innovation will arrive on schedule to solve whatever problems we prefer not to face today.
• Among critics, a mirror error: dismissing physical constraints as mere obstacles that political will or behavioural change can overcome.
• Among engineers and physicists, determinism: energy and materials dictate outcomes, human choices are downstream.

This series tries a different balance.

We lean toward physical constraints, not because we believe technology cannot surprise us, but because physics provides the only durable foundation. Technology is like an option in this matter: it can deliver extraordinary upside, but it remains just a possibility.

Yet, I am also a humanist. Energy sets the boundaries. It does not determine what we build within them. Florence ran on woodfire and draft animals, like every other city in Renaissance Europe. It still produced Leonardo, Michelangelo, Raphael, and Botticelli in a single generation.

Or as Smil writes in the closing pages of Energy and Civilization:

“During the first decade of the sixteenth century an idler on Florence’s Piazza della Signoria could have passed, in a matter of days, Leonardo da Vinci, Raphael, Michelangelo, and Botticelli, a concatenation of creative talent that is utterly inexplicable by the combustion of wood and the harnessing of draft animals.”

Energy sets the boundaries. What we build within them remains our own affair.