S-Curve Deceleration

In the AI Story

The S-curve's steep section generates the most confident predictions and the largest forecast errors. Nuclear power in the 1960s-1970s was projected to provide 50%+ of U.S. electricity by 2000; the actual figure was ~20%, and it has since declined to ~10%. The projections extrapolated the steep construction rates of the 1960s without accounting for cost overruns, safety incidents (Three Mile Island 1979), public opposition, and regulatory tightening that slowed and eventually halted new plant construction in most Western countries. Supersonic passenger aviation—the Concorde, Boeing 2707 project—was expected to replace subsonic flight by the 1980s; instead, the Concorde was retired in 2003 and no commercial supersonic flight exists today. The projections assumed demand would sustain premium pricing and that operational costs (fuel, maintenance, noise constraints) would decline with scale; both assumptions failed. The pattern: the steep section looks permanent from inside it, and it never is.

AI adoption data through 2025-2026 shows characteristic steep-section growth. ChatGPT: 100 million users in two months. Claude Code: $2.5 billion run-rate revenue within months of December 2025 threshold. AI-assisted code: rising from negligible to 40%+ of new commits in two years. Enterprise AI adoption: surveys showing 60-70% of Fortune 500 companies deploying generative AI in some capacity by mid-2025. The numbers are real and verified. The question is what happens next. Do they continue at current rates (implying near-total AI saturation of knowledge work within 2-3 years)? Do they decelerate gradually (following a classic logistic curve to 70-80% saturation over 5-10 years)? Do they decelerate sharply (encountering a binding constraint that shifts the curve into an earlier plateau)? The data through early 2026 cannot distinguish these scenarios because all three produce similar growth rates during the steep section. Only the approach to the bend reveals which curve the data trace.

The constraints that bend AI's S-curve are enumerated across the preceding chapters of this volume: energy availability (fifty-gigawatt requirement, grid construction timelines), semiconductor supply (four-year fab construction, EUV bottleneck, geographic concentration), water resources (millions of gallons per facility, competing demands in scarce watersheds), financial sustainability (inference costs currently below marginal costs, subsidized by investor capital), and institutional adaptation (retraining, regulation, educational reform). Each constraint can be addressed through investment and planning; none can be addressed instantaneously. The timeline for addressing the full set of constraints—building the physical infrastructure, securing the resources, developing the institutional frameworks—is measured in years to decades. If AI adoption continues at current rates, it will encounter these constraints before the infrastructure to support unconstrained adoption is complete. The curve bends when it hits the constraint, not when planning completes.

Smil's prescriptive claim, developed across Growth and Numbers Don't Lie, is that recognizing the curve will bend is not pessimism but realism—and that realism enables better planning. Organizations and policymakers who assume exponential growth will continue build strategies that become obsolete when growth decelerates, producing stranded assets (overcapacity built for demand that never materializes), resource misallocations (capital committed to expansion when consolidation would serve better), and workforce disruptions (training programs for jobs that saturate before graduates enter the market). Organizations that plan for the bend—maintaining flexibility, avoiding overextension, building capacity in increments that match observable demand rather than extrapolated projections—navigate the deceleration with less dislocation. The historical evidence favoring this approach is overwhelming: the firms and countries that survived nuclear power's deceleration, fiber-optic overcapacity in 2000-2002, and cleantech's stumble in 2008-2012 were those that planned for the bend or got lucky; those that extended on the assumption that growth would continue suffered bankruptcies, stranded capital, and lost decades.

Origin

The S-curve adoption model originates in population biology (Verhulst's 1838 logistic equation) and entered technology studies through sociologist Everett Rogers's Diffusion of Innovations (1962). Smil employs the S-curve across his work on energy transitions, material consumption, and technology adoption—not as theory but as empirical observation. His contribution is systematic quantification: measuring adoption timelines across hundreds of technologies, calculating time-to-saturation, and demonstrating that virtually no technology escapes the pattern. His Growth (2019) documents the universality with almost tedious thoroughness, precisely because tedious thoroughness is what defeats the recurring human tendency to believe this time is different.

The application to AI appears in Smil's 2025 Pictet essay and 2026 Bankinter webinar, where he notes that AI growth rates are spectacular but that 'spectacular' and 'sustainable' are different adjectives. The Vaclav Smil—On AI volume's Chapter 8 extends the framework by specifying the physical constraints that will bend AI's curve: energy, water, semiconductors, institutional capacity. The chapter's contribution is not prediction—Smil does not forecast when the bend occurs or at what adoption level—but preparation: identifying the variables that will determine the bend so that planning can address them rather than being surprised when extrapolation fails.

Key Ideas

Universal pattern. Every technology adoption follows an S-curve—slow start, rapid middle, eventual deceleration; AI's current position in the steep section does not exempt it from the pattern that has held for every predecessor.

Extrapolation fallacy. The steep section generates forecasts assuming current rates continue indefinitely, producing absurd endpoints (total AI saturation in two years) that reveal the forecast's structural flaw; the curve always bends before the absurdity is reached.

Constraint revelation. The factors that decelerate growth are often invisible during the steep section—appearing only as growth approaches their limits—making early constraint identification essential for avoiding late-stage surprises.

Planning for the bend. Organizations that assume exponential growth will continue build for a future that does not arrive; those that build flexibility and plan for eventual deceleration navigate transitions with less stranded capital and fewer workforce dislocations.

Realism enables adaptation. Recognizing the curve will bend is not defeatism but the prerequisite for building institutions, allocating capital, and training workforces at scales matched to plausible rather than fantastical demand trajectories.