Vaclav Smil

In the [YOU] on AI Field Guide

The cycle documents the collapse of the imagination-to-artifact ratio with the enthusiasm of a builder who has watched barriers fall. The developer in Trivandrum who achieves a twenty-fold productivity multiplier in a week. The solo builder who ships a revenue-generating product in a year of solitary AI-augmented work. The hundred-dollar-per-month subscription that delivers leverage previously available only to well-funded teams. Smil's framework does not deny these achievements. It demands their physical ledger. That hundred-dollar subscription is currently priced below the marginal cost of inference, subsidized by billions of dollars of investor capital. When pricing reflects physical cost, the democratization claim encounters an economic barrier that is downstream of a thermodynamic one.

The cycle's five-stage model of technological transition—threshold, exhilaration, resistance, adaptation, expansion—maps cleanly onto what Smil has documented as the S-curve of technology adoption, with one addition that the cycle's framework does not make explicit: the curve always bends, and the bend is determined not by the technology's capabilities but by the physical infrastructure required to support them. The steep middle section of the AI S-curve is the period of maximum enthusiasm and minimum constraint. The deceleration will come not because people stop wanting AI but because the grid, the semiconductor supply chain, and the data center construction pipeline have timelines that software iteration does not. The institutional response the cycle calls for must be built during the steep section, or it will not be built at all.

The cycle's most morally significant claim—that AI democratizes capability across geography and income—is also the one most vulnerable to Smil's quantitative method. The developer in Lagos requires, at minimum, reliable electricity, sufficient bandwidth, an adequate device, and the financial capacity to pay for the service. Sub-Saharan Africa's average per-capita electricity consumption is roughly one-fiftieth of the United States'. The average Nigerian fixed broadband speed is roughly one-tenth of the American average. The hundred-dollar subscription represents 50 to 70 percent of the median Nigerian monthly wage. The infrastructure required to close these gaps takes decades to build. Declaring the democratization accomplished because the software is available is like declaring hunger solved because the recipe exists.

Origin

Born in 1943 in Pilsen, Czechoslovakia, Smil trained as a natural scientist and emigrated to Canada in 1969 following the Soviet suppression of the Prague Spring. He spent his career at the University of Manitoba, where he became Distinguished Professor Emeritus, building a body of work that Bill Gates has described as indispensable and that defies easy disciplinary classification: energy studies, materials science, environmental analysis, food systems, population dynamics, and technology history, all organized around the conviction that quantitative analysis is the minimum condition for honest thinking about the world.

Smil's approach to technology assessment is grounded in historical pattern recognition across hundreds of transitions. The printing press. The steam engine. Electrification. Nuclear power. The internet. In every case, the advocates of the transformation underestimated the time required, overestimated the displacement of incumbents, and failed to account for the inertia of physical infrastructure. His 2023 book Invention and Innovation: A Brief History of Hype and Failure documented this pattern with twenty detailed case studies. The AI revolution had not yet arrived at its moment of maximum hype when the book was written. By 2026 it had.

Key Ideas

The Weight of Weightless Computation. Every token has a cost that is not visible in the interface: electricity from a power plant, heat dissipated through cooling systems, water evaporated through cooling towers, silicon refined from sand, copper drawn into wire. Training a frontier model consumes the annual electricity of thousands of American homes. Inference, multiplied across hundreds of millions of users, is transforming data centers into one of the largest industrial energy demands of the twenty-first century. The AI revolution is an energy revolution, and energy systems change at the speed of construction, not the speed of code.

The Fifty-Gigawatt Problem and Infrastructure Inertia. By Smil's February 2026 estimate, the United States needs to add approximately 50 gigawatts of new electrical capacity by 2030 to support AI growth—the equivalent of fifty cities of one million inhabitants. A natural gas plant takes three to five years. A nuclear plant takes a decade. A major transmission upgrade takes seven to ten years. A semiconductor fabrication facility takes four years and costs twenty billion dollars. The infrastructure inertia is governed by physics, not ambition. Software capability doubles in months. The physical foundation it requires does not.

The Jevons Paradox of Intelligence. Efficiency improvements do not reduce total consumption when demand is elastic. The Berkeley study the cycle cites documented that AI-augmented workers did not work less—they took on more tasks, expanded into adjacent domains, filled every available moment with additional production. A twenty-fold productivity multiplier absorbed into expanded output produces, approximately, twenty times the computational demand. The Jevons Paradox of intelligence predicts that the democratization of cognitive capability will produce an expansion of aggregate computational demand, not a reduction—and the physical infrastructure must scale to meet that demand or ration access.

The S-Curve Always Bends. ChatGPT's adoption curve was the steepest in the history of consumer technology. The curve will decelerate. Not because people stop wanting AI, but because the physical world imposes constraints that exponential extrapolation ignores. The bend is determined by the binding constraint in the infrastructure chain—grid capacity, semiconductor supply, data center construction, water availability—and the binding constraint is the longest timeline in the chain. Planning for the bend is not pessimism. It is the only intellectual posture that has ever produced planning adequate to the scale of the challenge.