Tech AI News - 테카이

Why does power consumption keep rising as efficiency improves — AI and the Jevons paradox

· 5 min read

This post was translated from the Korean original by AI.한국어 원문 읽기 →

AI inference costs are falling by nearly half every quarter, yet data-center power demand keeps accelerating instead. Efficiency isn't reducing consumption — it's increasing it. The "Jevons paradox," first observed more than 150 years ago in the coal era, is repeating in almost the same shape.


Background — why this topic now

The data-center power crunch has already been covered often. The solution usually offered is: "as models and chips get more efficient, the power problem will ease." And inference costs really are falling noticeably. Yet over the same period, power-demand forecasts haven't come down — if anything, they've been revised upward. If efficiency gains aren't the solution to the power problem but are instead driving demand even higher, we need to revisit the premise itself. The framework that explains this paradox is the "Jevons paradox," proposed in 1865 by British economist William Stanley Jevons.

Key data & current state

  • Rate of AI inference cost decline: the cost to hit a given performance bar has been falling by roughly 47% per quarter since 2023 (Epoch AI estimate, reported by Let's Data Science)
  • Data-center power-demand forecast: global data-center electricity consumption is projected to more than double, from about 415 terawatt-hours (TWh, roughly 1.5% of global electricity) in 2024 to about 945 TWh by 2030 (IEA, "Energy and AI" report, April 2025)
  • Big tech's actual numbers: in its 2026 environmental report, Google said its 2025 power demand rose 37% year over year — about 3.5 times its 2019 level — and that its total greenhouse gas emissions rose 18% (Axios, June 2026)

Interpretation: the fact that per-unit cost falls by nearly half every quarter while overall consumption forecasts get revised upward is a signal that falling unit prices aren't holding demand down. In other words, "how much is being used" is growing faster than "how fast the unit price is falling."

Deep dive

1) What is the Jevons paradox — from coal to AI

In his 1865 book The Coal Question, Jevons pointed out that James Watt's improvements to the steam engine sharply cut the coal needed for the same amount of work, yet Britain's total coal consumption didn't fall — it rose. As efficiency improves, the cost of using a resource drops, and that lower cost opens up new uses that were previously too expensive, ultimately increasing total consumption. SIGARCH Computer Architecture Today, a publication in the computer-architecture research community, applies this same structure directly to the data-center power problem, noting that as chip and cooling efficiency improve, that very efficiency frees up the headroom to run that many more GPUs, so total power demand doesn't fall.

2) Why the same structure repeats in AI inference

When the per-token price falls, the amount of inference you can run on the same budget rises. On top of that, the recent spread of reasoning models and agentic use means each query consumes far more tokens internally. In other words, two trends overlap: cost per token goes down, but tokens consumed per query goes up. The International Energy Agency (IEA) estimates inference already accounts for 80–90% of AI compute, and that by 2030 inference will account for roughly 75% of AI power demand. That means the structure driving power demand is shifting from a single training run to billions of accumulated everyday inference requests. A paper by researchers at Georgia Tech and UC Hastings College of the Law (arXiv, 2025) likewise argues that the AI environmental debate has focused so heavily on data centers' direct energy and water use that this kind of indirect rebound effect from efficiency gains hasn't been properly addressed.

3) Why does the total keep rising even as efficiency rises

Google's case illustrates this well. Google has said it sharply cut the energy consumption per query of its Gemini models, but its 2026 environmental report shows the opposite direction at the aggregate level. Overall power demand rose 37% year over year, and total emissions — including the production of the AI hardware it deploys — rose 18%, the largest increase on record. Sopra Steria points to cases like this to warn that looking only at per-unit efficiency gains and concluding that total environmental impact has fallen is a mistake. Analysis: efficiency gains lower "the power it takes to do the same amount of work," but they also lower the threshold for "how much work can be done," which creates new demand. Which of the two effects dominates determines the outcome we actually observe.

Implications for Korea

Korea isn't exempt from this same structure. As Naver, Kakao, and cloud providers keep investing in data centers, there's growing concern that delays connecting to the greater Seoul area's power grid and the pace of new generation capacity can't keep up with the pace of AI infrastructure expansion. Bringing in more efficient domestic or overseas AI chips shouldn't be expected to reduce power demand on its own — it's worth planning grid expansion and renewable-energy procurement on the premise that the efficiency gain itself is likely to pull in even more services and usage. Enterprises, too, should keep in mind that treating AI adoption purely as a power- or cost-saving measure can backfire into the paradox of total costs rising from increased usage.

Outlook & variables to watch

  • Positive variables: expanded investment on the supply side — low-power inference-specific chips, expanded renewable and gas generation, and small modular reactors (SMRs)
  • Risk variables: the spread of reasoning and agentic usage increasing compute per query, offsetting the effect of falling unit costs
  • Checkpoints: the length of utilities' new grid-connection queues, and the trend in big tech's annual environmental reports in total power and total emissions (not per-unit efficiency figures) — total numbers are the ones to watch

Conclusion

  • AI efficiency gains really are happening fast, but by themselves they aren't a solution that reduces total power consumption.
  • The Jevons paradox, observed with coal back in 1865, is showing up in the falling cost of AI inference in almost the same shape.
  • To understand the power problem, look at total consumption and total emissions, not per-unit efficiency figures.

References

  • #ai
  • #artificial-intelligence
  • #jevons-paradox
  • #data-centers
  • #power-consumption
  • #ai-efficiency
  • #rebound-effect
  • #carbon-emissions