Jevons Meets Jensen: Cheaper Intelligence and the Physical Clock of Energy

In 1865, William Stanley Jevons noticed that Britain was burning more coal precisely because it had learned to use coal more efficiently. A century and a half later, every generation of accelerators makes intelligence cheaper, and we consume more of it. This closing instalment asks whether cheaper intelligence ultimately increases or decreases the energy intensity of everything around it.

---

This is Part 5 of a five-part AI × Energy series. Energy and Civilization established the 50-year rule for energy transitions. Power Density quantified the spatial constraints. The 84% Gap calculated the capital shortfall. The New Map added geopolitics and supply chains. This synthesis adds the feedback loop.

---

TL;DR

1. Efficiency will not make AI’s energy use disappear. Energy per unit of useful intelligence keeps falling quickly while total electricity consumed by AI keeps rising. Demand for intelligence looks highly elastic, and agents remove the cap that human attention used to impose.

2. AI is also an input into the energy system. Forecasting, grid optimisation, materials search and plasma control can change the behaviour of physical systems far larger than the data centres themselves. The useful comparison is AI electricity consumption against the changes AI causes, not against zero.

3. Nuclear runs on a different clock. It is the right technology for firm, low-carbon load, but new reactors and SMRs arrive mostly after 2030. For the first wave of the build-out, the advantage belongs to reactors that already exist, and to countries such as France that already possess them.

4. Discovery is getting faster than deployment. AI compresses whatever can be represented as information. The energy transition is only partly an information problem. Bits can accelerate atoms. They cannot make them behave like bits.

---

In 1865, William Stanley Jevons noticed something strange about the steam engine. Britain had spent a century making coal dramatically more useful. Yet it was burning more coal, not less.

Greater efficiency had lowered the cost of useful work. That made new applications economical. More factories could use steam engines, more mines could pump water, more machines could replace human or animal labour. Coal consumption rose.

A century and a half later, we may be running the same experiment with intelligence.

Every generation of accelerators makes computation cheaper. Models become more efficient. Inference prices fall. And the response is not simply to perform the same amount of computation using less electricity. We train larger models, generate more tokens, use more test-time compute and put models into places where intelligence was previously too expensive to deploy.

But AI adds another loop. The thing consuming energy may itself improve the energy system. AI can forecast renewable production, optimise grids and industrial processes, search enormous materials spaces, control plasmas and reduce the cost of scientific discovery.

The question is therefore not whether AI consumes energy. It does, and increasingly so. Nor is it whether efficiency will make that consumption disappear. It probably will not. The harder question is whether cheaper intelligence ultimately increases or decreases the energy intensity of everything around it.

That question brings together the previous four parts of this series. Smil gave us the physical constraints: energy transitions are slow, infrastructure-heavy and constrained by power density (Part 1, Part 2). Subran and Zimmer gave us the capital constraint (Part 3). Yergin added geopolitics and supply chains (Part 4). Jevons gives us the feedback loop.

---

The Scale of the Problem

The first mistake in the AI-energy debate is to start with a single query. The second is to start with a futuristic gigawatt data centre. Both can be true and still obscure what is happening.

The IEA estimates that data centres consumed around 415 TWh of electricity in 2024. In its base case, consumption more than doubles to around 945 TWh by 2030, just under 3% of global electricity demand. Accelerated servers, driven mainly by AI, account for almost half of that increase¹.

This is large enough to matter enormously for electricity systems without being large enough to explain global climate change. Data centres account for roughly 0.5% of global fuel-combustion CO₂ emissions today. The problem is the slope, the concentration and the timing.

A 500 MW or 1 GW campus is not a statistical abstraction to the grid node that has to serve it. Electricity demand is concentrating in a handful of places while grid connections, transformers, transmission lines and new generation operate on much slower clocks.

AI is digital at the application layer and stubbornly physical underneath. Models may scale in software time. Power systems do not.

The Misleading Comfort of the 0.3 Wh Query

Epoch AI estimates that a typical GPT-4o text query consumes roughly 0.3 Wh, around one tenth of a widely repeated earlier estimate². For an individual user, that is tiny. Long prompts, large contexts and reasoning workloads can consume substantially more, but the environmental significance of AI is poorly captured by asking whether someone should feel guilty about one more ChatGPT query.

The macro problem is multiplication. If the marginal cost of intelligence falls far enough, we stop using AI only when a human explicitly asks for it. Models begin reading every document, inspecting every image, writing and testing software continuously, simulating designs, monitoring physical systems and interacting with other agents.

The relevant unit then moves from queries per person to something closer to machine cognition per unit of economic activity.

---

Jevons Meets Jensen

Jevons’s original insight was not that efficiency is useless. James Watt’s steam engine genuinely produced more useful work from a unit of coal. His point was that an efficiency gain changes economics. When useful work becomes cheaper, more uses become profitable.

AI looks unusually exposed to this mechanism because demand for intelligence appears highly elastic. Lower inference cost does not simply make an existing chatbot cheaper. It changes what can economically be automated at all.

Better chips reduce the energy required for a computation. Better architectures reduce the amount of computation required for a result. Quantisation and distillation make models smaller. Mixture-of-Experts activates only part of a network. Better inference engines squeeze more tokens out of the same hardware.

And then we spend the savings.

Models get larger. Context windows get longer. Reasoning runs for longer. Applications call models more frequently. Tasks that could not economically justify inference suddenly can.

DeepSeek made the mechanism unusually visible. Its mixture-of-experts architecture, latent-attention techniques and systems optimisation showed that high-level reasoning could be delivered with much less computation than many observers had assumed. The immediate consequence was not fewer GPUs. It was that capable reasoning became cheap enough to use in many more places.

This is also why Jensen Huang’s recurring claim that customers save money by buying more accelerated computing is less paradoxical than it sounds. Our Reading Nvidia series follows the same logic from the supplier’s side, ending with The Return of the Real World, where power, land and capital become part of the compute problem. Better computation can lower the cost of a unit of useful intelligence while increasing the economically rational quantity of intelligence consumed.

But Jevons Is Not a Law

Efficiency improves. Demand expands. But total energy consumption does not have to rise forever.

Rebound depends on demand. Some workloads will become good enough and migrate away from frontier models. Sparse activation, quantisation, distillation, speculative decoding, better memory systems and specialised accelerators can reduce the physical work required for a given result. Electricity and capital are not free.

Human attention may eventually cap interactive AI. Agents weaken that cap. If software can consume intelligence without a person watching every token, the addressable market becomes the number of economically useful machine tasks rather than the number of waking human hours.

For now, the likely asymmetry is simple: energy per unit of useful intelligence keeps falling quickly while total electricity consumed by AI keeps rising. Efficiency is essential to AI economics without necessarily reducing AI’s aggregate energy use.

---

The Other Feedback Loop

If that were the whole story, AI would simply be another rapidly growing electrical load.

But intelligence is an input into other systems. A model can consume electricity while making a cement kiln, transmission line, building, logistics network or power plant more efficient. The electricity consumed by the model may be much less important than the physical system whose behaviour it changes.

The IEA estimates that widespread adoption of AI applications already available today could enable around 1.4 gigatonnes of CO₂ reductions in 2035, several times the projected emissions associated with data centres themselves¹. Getting there would require much wider deployment than we see today, along with better data, infrastructure, skills and regulation.

The useful comparison is therefore not AI electricity consumption against zero. It is AI electricity consumption against the changes AI causes in much larger physical systems.

Grids

Electricity systems are becoming harder to operate as generation becomes more variable and decentralised. Forecasting, congestion management, predictive maintenance and dynamic line ratings are natural machine-learning problems.

DeepMind and Google demonstrated an early version of this by forecasting wind generation 36 hours ahead and increasing the economic value of a 700 MW wind fleet by roughly 20%³. The IEA estimates that AI applications could unlock as much as 175 GW of additional transmission capacity from existing lines under widespread adoption.

AI does not manufacture a transformer, build a transmission line or obtain a permit. It can make scarce physical infrastructure more productive.

Materials

Google DeepMind’s GNoME generated 2.2 million candidate crystal structures and identified around 380,000 as especially stable candidates for experimental synthesis⁴.

That expands the search space dramatically. It also moves the bottleneck.

Berkeley’s A-Lab combined computational databases, machine learning, literature-derived synthesis recipes and robotics. The corrected Nature record reports 36 successful compounds from 57 targets over 17 days and 353 experiments⁵.

Prediction gets faster. Synthesis does not.

Once the search problem accelerates, physical experimentation, characterisation and scale-up become the constraints. AI can tell us where to look much faster. Matter still has to cooperate.

Weather and Fusion

GraphCast can produce ten-day global weather forecasts in under a minute and outperformed ECMWF’s deterministic HRES system on most evaluated variables and lead times⁶. Cheaper forecasting allows larger ensembles, more frequent updates and wider access to high-quality prediction. Grids, agriculture, logistics, insurance and adaptation all depend on decisions made under uncertainty.

Fusion pushes the same question further. Machine learning has already demonstrated impressive control of tokamak plasma configurations, including DeepMind’s work with EPFL⁷. AI can also contribute to materials discovery, simulation and experimental design.

But the history of fusion is, among other things, a history of timelines. Faster simulation or better plasma control does not make engineering, construction, regulation and deployment move at software speed. If AI materially accelerates fusion, the payoff is enormous. The energy argument does not need to assume that it will.

---

Nuclear Runs on a Different Clock

There is an obvious match between AI and nuclear power. Data centres want large quantities of electricity, continuously, with little carbon and limited exposure to intermittency. Nuclear provides exactly that.

The problem is time.

AI infrastructure is being built on a technology clock. A large data centre can go from decision to operation in a few years. Nuclear still runs on an infrastructure clock.

This distinction gets lost in the flurry of agreements between technology companies and nuclear developers. Microsoft, Google, Amazon and Meta show that the technology industry is beginning to secure electricity far beyond its traditional procurement horizon. They do not mean that new nuclear can supply most of the AI capacity being built before 2030.

For the next few years, nuclear’s advantage comes mainly from reactors that already exist.

This is why Microsoft’s agreement around the former Three Mile Island Unit 1 is more immediately relevant than many SMR announcements. Restarting a reactor remains difficult, but much of the physical system is already there. The same logic applies to lifetime extensions and uprates.

New conventional nuclear operates on another timescale. Flamanville is an extreme example, but a useful one: construction started in 2007 and the reactor received authorisation to enter service in 2024. China and Korea have built reactors considerably faster, but recent Western experience makes new conventional nuclear a poor answer to electricity demand appearing over the next four years.

SMRs are meant to change the economics of construction through smaller units, standardised designs and factory manufacturing. The promise depends on repetition. Before the learning curve can accelerate, first units have to be licensed and built, supply chains have to scale, fuel has to be available and factories have to reach sufficient throughput. The IEA expects the first commercial SMRs around 2030, with a more meaningful contribution later.

AI electricity demand is rising now. Much of the nuclear system its largest customers increasingly want will arrive later.

The IEA expects renewables to provide almost half of additional electricity consumed by data centres through 2030. Gas and coal together still provide more than 40% of the increase. Nuclear becomes more important after 2030.

Nuclear may therefore be one of the right technologies arriving too slowly for the first wave of the AI build-out.

A model can diffuse globally in months. A new reactor still has to pass through concrete, steel, regulators, supply chains and time.

The Supply Chain Behind the Reactor

Even the reactor timeline understates the problem. Many advanced reactor concepts require HALEU, enriched above the levels used by conventional light-water reactors. Western commercial supply remains constrained. Then come conversion, enrichment, fuel fabrication, specialised components, qualified suppliers, skilled labour and regulators.

Hyperscalers do improve one part of the equation. They are unusually attractive long-term customers, with strong balance sheets and the ability to sign twenty-year contracts. That can make capital-intensive generation easier to finance.

Money can fund factories and projects. It cannot instantly create industrial learning, licensed designs, trained workers or fuel capacity.

---

How Much Energy Can We Actually Add?

There is a harder objection. The transition already requires replacing an enormous fossil energy system while electricity demand rises from transport, heating, industry and air conditioning. AI adds another fast-growing load.

At the global level, data centres remain a relatively small part of the problem. They consume around 1.5% of world electricity today and could approach 3% by 2030. They account for less than a tenth of global electricity-demand growth over the rest of the decade.

But a data centre does not consume global electricity. It connects to a particular grid.

That is where scarcity appears.

Renewables can be built quickly and are expected to meet nearly half of additional data-centre demand through 2030. Nuclear can provide firm low-carbon power, but much new capacity arrives later. Transmission takes years to permit and construct. Where demand arrives faster than clean generation and grids, the residual supply comes from what is already available. For the next few years, that often includes gas and coal.

This does not mean energy consumption has to fall mechanically as AI expands. Electrification can provide the same services with much less primary energy: an electric motor wastes far less energy than an internal-combustion engine, and a heat pump can deliver several units of heat for one unit of electricity. Renewable generation is still expanding rapidly. Higher energy prices also eliminate uses whose economic value is too low.

But neither can every desirable use of electricity assume that supply will appear on demand.

At gigawatt scale, energy stops being a background input to computing and becomes part of the product constraint.

Scarcity then feeds back into AI. Expensive electricity makes marginal inference less attractive. Workloads move towards abundant power. Training becomes more flexible. Smaller and specialised models take jobs that do not justify frontier inference. Developers become more selective about when additional reasoning is worth paying for.

Jevons pushes demand upwards. The energy system pushes back.

---

France

France is interesting because it already possesses much of what other countries are trying to build.

In 2025, 95.2% of mainland French electricity generation was low-carbon. Nuclear generated 373 TWh, hydro 62.4 TWh, and average direct emissions intensity fell to 19.6 gCO₂e/kWh. France was also a record electricity exporter⁸.

EDF has begun turning that system into an AI infrastructure proposition, initially identifying four industrial sites representing roughly 2 GW of available power and grid access.

The advantage is not simply cheap electricity. It is electricity that is already low-carbon, largely firm, connected to an existing grid and backed by an industrial nuclear system with decades of operating experience.

Grid connections, permitting, cooling, fibre, construction capacity and talent still matter. Low-carbon electricity also has an opportunity cost: it can electrify transport, buildings and industry as well as data centres.

But France illustrates the difference between possessing nuclear power and planning nuclear power. For the AI build-out of the late 2020s, that distinction may matter more than many reactor announcements.

---

Discovery Is Getting Faster Than Deployment

A pattern has emerged.

GNoME can search materials space faster than laboratories can synthesise candidates. AI can identify ways to increase the capacity of an electricity grid faster than new transmission can be permitted and built. Machine learning can improve plasma control faster than a commercial fusion plant can be engineered. Hyperscalers can create demand for nuclear power faster than new reactors can be delivered. AI itself can diffuse faster than the electricity system that powers it can expand.

The bottleneck moves.

AI is extraordinarily good at compressing problems that can be represented as information: search, prediction, optimisation, simulation and design. The energy transition is only partly an information problem. It is also mines, factories, transmission lines, reactors, permitting, financing, construction crews and political consent.

Bits can accelerate atoms. They cannot make them behave like bits.

---

Jevons Meets Smil

The five pieces in this series end with two forces moving in opposite directions.

Intelligence gets cheaper, so we consume more of it. New applications become economical. Compute demand rises and the energy system has to accommodate it.

At the same time, cheaper intelligence makes physical systems easier to predict, optimise, design and automate. Grids can carry more electricity through existing infrastructure. Industrial processes can waste less energy. Materials searches that once took years move into computation.

Both can happen at once.

A ChatGPT query can consume very little energy while global AI creates a major new electricity load. Joules per token can collapse while total compute rises. AI can increase emissions through electricity demand while reducing larger emissions elsewhere.

The balance will not be decided by efficiency alone. It will be decided by how valuable additional computation becomes, how quickly clean electricity can be added, where grids have spare capacity, and how much leverage AI actually creates in the physical economy.

A model can be trained in months. A data centre can be built in a few years. A transmission line can take a decade. A nuclear supply chain can take longer.

Smil tells us why the physical system moves slowly. Subran and Zimmer show why knowing what to build is not enough to finance it. Yergin shows how quickly technology becomes geography and supply chains. Jevons explains why efficiency can increase our appetite rather than satisfy it.

AI brings those constraints together faster.

---

Closing Note

Smil ended Energy and Civilization in Renaissance Florence, with Leonardo and Michelangelo walking through streets still powered overwhelmingly by wood, muscle and draft animals. Intellectual possibility had moved much faster than the energy system beneath it.

Jevons wrote The Coal Question centuries later as Britain experienced the reverse: a new energy system expanded the range of economically possible activity, while greater efficiency made society consume more of the resource powering it.

We may now be watching both processes at once.

Machine intelligence is becoming cheaper extraordinarily quickly. That will almost certainly increase the amount of intelligence we consume. Whether it increases the amount of energy and carbon consumed by the rest of the economy is a different question.

AI will not repeal the constraints of the physical world. It may change how quickly we discover what can be built within them.

The wager is whether that acceleration reaches our solutions faster than it reaches our appetite.

---

This is the fifth and final article in a five-part AI × Energy series on the essential books for understanding energy, technology, and investment:

1. Vaclav Smil, Energy and Civilization: A History: the 50-year rule and why energy transitions cannot be rushed

2. Vaclav Smil, Power Density: why land, not cost, constrains the energy transition

3. Ludovic Subran & Markus Zimmer, The 84% Gap: what climate investment actually requires

4. Daniel Yergin, The New Map: energy geopolitics and the realignment of power

5. The Synthesis: Jevons Meets Jensen (this article): the rebound effect and the nuclear revival in the age of AI compute

---

Sources

• William Stanley Jevons, The Coal Question (1865).
• Tae Kim, The Nvidia Way (2024).
• International Energy Agency, Energy and AI (2025).
• Vaclav Smil, Energy and Civilization (2017) and Power Density (2015).
• Ludovic Subran & Markus Zimmer, Investing in a Changing Climate (2023).
• Daniel Yergin, The New Map (2020).
• Google DeepMind: GNoME, GraphCast, wind forecasting and tokamak plasma control.
• Nature: A-Lab autonomous materials synthesis.
• RTE: French Electricity Review 2025.

Footnotes

1. [Source] IEA (2025). Energy and AI. International Energy Agency. https://www.iea.org/reports/energy-and-ai ↩ ↩²

2. [Source] Epoch AI (2025). How much energy does ChatGPT use? https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use ↩

3. [Source] Google DeepMind (2019). Machine learning can boost the value of wind energy. https://deepmind.google/discover/blog/machine-learning-can-boost-the-value-of-wind-energy/ ↩

4. [Source] Merchant, A. et al. (2023). Scaling deep learning for materials discovery. Nature 624. https://www.nature.com/articles/s41586-023-06735-9 ↩

5. [Source] Szymanski, N. J. et al. (2023). An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624, with subsequent correction. https://www.nature.com/articles/s41586-023-06734-w ↩

6. [Source] Lam, R. et al. (2023). Learning skillful medium-range global weather forecasting. Science 382. https://www.science.org/doi/10.1126/science.adi2336 ↩

7. [Source] Degrave, J. et al. (2022). Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602. https://www.nature.com/articles/s41586-021-04301-9 ↩

8. [Source] RTE (2026). Bilan électrique 2025. Réseau de Transport d’Électricité. ↩