Jevons paradox

Original Article ↗ Hacker News Discussion ↗

Jevons Paradox in AI and Nvidia

  • Many comments apply Jevons to AI: cheaper or more efficient training/inference (e.g., DeepSeek R1, synthetic data, RL) could drive more total AI usage and thus more total GPU demand.
  • Some argue the DeepSeek paper is actually bullish for Nvidia: synthetic data pipelines and “thinking models” imply more and better foundation models, hence more GPU usage overall.
  • Others counter that Nvidia’s current margins rely on a few mega-buyers building huge, differentiated datacenters. If AI becomes cheap and commoditized, demand for ultra‑expensive datacenter GPUs and $500B buildouts may shrink even if total AI use rises.

Stock Valuation, Market Dynamics, and Politics

  • Comparisons are made between NVDA and AMZN in the dotcom era. Detractors say the analogy fails because Nvidia already has massive operating income; supporters still see bubble‑like speculation and hope for a “dot‑com‑style” AI crash as a buying opportunity.
  • Several note Jevons applies to resource consumption, not directly to stock prices. Market moves reflect perceptions of future margins and competition, not just volume.
  • A side thread debates a recent high‑profile Nvidia stock sale: some see normal trading on public news; others speculate about political insider knowledge, without evidence.

Efficiency, Constraints, and Demand Curves

  • One line of argument: theory of constraints and finite use cases mean there isn’t an infinite GPU demand curve; at some efficiency level, “good enough” caps spending.
  • Others claim that large labs will always find ways to saturate any available compute (larger, more frequent, or more specialized models), so efficiency gains still increase total consumption.
  • Analogies are drawn to SMT solvers: huge efficiency and price drops didn’t yield infinite or even massive mass‑market demand; adoption is limited by people and workflows, not just cost.

Access to LLMs and Price Sensitivity

  • Several commenters say price does lock out users and organizations:
    • Paid add‑ons for office suites were too expensive for many SMBs.
    • Local SOTA inference often needs 400–768 GB of RAM/VRAM, with hardware costing $15–30k, which is out of reach for most individuals.
  • Lower costs plus local trainability are seen as alleviating:
    • Lack of tuning control,
    • Data ownership/privacy issues,
    • Power waste per useful unit of work.
  • Some remain skeptical, arguing many end users dislike current AI features and that LLMs are “solutions in search of problems.”

Induced Demand, Rebound Effect, and Definitions

  • Multiple comments tie Jevons to induced demand and the rebound effect:
    • Rebound: efficiency → more use, partially offsetting savings.
    • Jevons: efficiency → more than full offset, total resource use rises.
  • Debate centers on whether induced demand is:
    • Just “realized latent demand” along a standard demand curve, or
    • A genuine shift of the demand curve itself.
  • Highway and housing examples illustrate how cheaper travel or lighting can permanently change behavior and urban form.

Energy, Lighting, Transport, and Other Examples

  • Home insulation: cited research suggests initial gas savings erode as occupants raise thermostats, matching Jevons‑like behavior.
  • LEDs: strong disagreement over whether 10x efficiency led to similar or greater increases in total lighting energy:
    • Some point to far more fixtures (accent lights, outdoor, screens) and historical data showing consumption rising >100x as lighting got cheaper.
    • Others doubt a full 10x usage increase and focus on per‑fixture savings and reduced replacement.
  • Transport: commenters discuss EVs and 1950s travel levels; cheaper per‑km driving may increase total kilometers driven, but lifestyle and urban design constraints complicate this.

Scope and Misuse of Jevons

  • Several see Jevons being casually invoked as “cope” to defend high AI and chip valuations, ignoring time lags and competitive dynamics.
  • Others stress that Jevons is empirically uncommon compared to ordinary rebound effects and that its relevance must be analyzed case by case, not assumed.