med-high convictionactive · updated 2026-06-16T00:00:00.000Z
Cost-curve decline → tipping point → non-linear S-curve adoption → forecaster underestimation
A technology's unit cost falls predictably (Wright's Law); once it crosses a price/performance tipping point it beats the incumbent; adoption then goes non-linear along an S-curve; linear-extrapolating forecasters systematically underestimate the speed and scale — but the *timing* of the knee and the eventual ceiling remain poorly forecastable.
The chain
1
A technology's unit cost falls a roughly constant % per doubling of cumulative production (Wright's Law / experience curve).
From 2026-06-16-academic-research-technology-adoption-s-curves: "unit costs decline as a power law of cumulative production"; median learning rate ~21% across 150 technologies; Wright's Law beat five rival laws across 62 technologies (Nagy et al. 2012).
From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: LLMflation — entry-level LLM capability fell ~1,000× in three years ($60→$0.06/M-tokens), an experience-curve dynamic (a16z).
2
Falling cost crosses a price/performance **tipping point** where the new technology beats the incumbent on cost or capability.
From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: RethinkX's framework centers on cost curves crossing a tipping point that opens "a vast new possibility space"; solar reaching <~4¢/kWh and ~$100/kWh lithium packs are the crossings Seba forecast.
(Qualitatively confirmed by the framework and by realized solar/battery parity; the *exact* parity threshold and date for any given technology is what's hard — single mechanism-level sourcing here.)
3
Adoption then goes **non-linear**, tracing an S-curve from ~1-2% penetration through an accelerating phase toward an ~80% ceiling, with steepness set by feedback (learning, scale, network effects).
From 2026-06-16-academic-research-technology-adoption-s-curves: the S-curve is "the dominant stylized fact" of technology usage over time (Geroski 2000).
From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: adoption "is non linear and follows an S-curve … less than 1-2% market penetration … hits a tipping point and accelerates until the product nears about 80%" (RethinkX); steepness set by RMI's four feedback mechanisms.
4
Mainstream forecasters, extrapolating the present linearly and assuming "all else equal," **systematically underestimate** the speed and scale — creating a persistent gap between consensus forecasts and outcomes.
From 2026-06-16-academic-research-technology-adoption-s-curves: IEA "severely underestimated" solar PV/wind growth across scenarios (Lopez 2025); all 26 scenarios under-projected solar (Carrington 2018); models "historically underestimated deployment rates … and overestimated costs" (Way 2022).
From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: forecasters "extrapolate the present into the future in a straight line" and assume cost curves improve "while all else remains equal" (RethinkX).
5
**But** adoption *timing* and the ceiling remain poorly forecastable, so the disruption often arrives later, or in a different segment, than point forecasts claim.
From 2026-06-16-academic-research-technology-adoption-s-curves: Bass `M`/`p`/`q` badly identified and `p` "highly sensitive to the assumed market potential M" (Massiani 2015); learning rates change for ~66% of technologies (Carlino 2025); curves asymmetric (Easingwood 1981).
From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: tony-seba's cost calls were right but his adoption-*timing* calls (TaaS-by-2030, robotaxis, used-ICE-to-zero) missed badly — "right spot, wrong timeframe."
What would falsify this
- Step 1: A tracked technology's cost stops following a power law of cumulative production over a multi-year window (e.g., nuclear-style anti-learning).
- Step 3: A technology crosses cost parity but adoption stays linear / fails to accelerate past the tipping point for years.
- Step 4: Consensus forecasts prove well-calibrated (or *over*-estimate) for a cohort of cost-declining technologies — the underestimation gap disappears.
Contradictions / tensions
- **Outpacing vs. undershooting coexist.** In 2026, solar and grid storage are *beating* mainstream forecasts while EV battery demand is growing *slower* than expected (PHEV shift) and robotaxi/TaaS lags — so Step 4's "underestimation" is segment-specific, not universal (2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves).
- **Aggregate forecastability vs. per-technology instability.** Step 1's reliability (Nagy/Farmer) sits against Step 5's instability (Carlino/Massiani); reconciled by separating cost *direction* (forecastable) from adoption *timing/ceiling* (not).
- **Source incentive.** RethinkX/Seba are advocates; the "underestimation" framing is partly self-serving — independent IEA ex-post analysis (Lopez 2025) is the stronger evidence for Step 4.
Companies
Concepts
Wright's Law (Experience / Learning Curve)Technology Adoption S-CurveCurve Steepness and Adoption VelocityForecastability of Technological ProgressTechnology Convergence
Open questions