When Intelligence Becomes Cheap: Jevons, AGI and the Quantum Wildcard

Update, July 12, 2026: This essay has been expanded and folded into our full five-part research piece, including six charts, Japan and South Korea, the tax question and AGI timelines: The AI Bubble Can Burst Without AI Failing. The version below remains as originally published.

The Paradox of Cheap Intelligence

Every generation of AI becomes more efficient.

Models get smaller. Chips get faster. Caching improves. Routers send easy questions to cheaper systems. Quantization reduces the number of bits required for each calculation. A task that consumed a large model yesterday can run on a specialized model tomorrow.

The intuitive conclusion is that AI will therefore use less energy.

The evidence points in the opposite direction.

The International Energy Agency says energy use per AI task has recently fallen by at least an order of magnitude per year. Yet electricity demand from data centers rose 17 percent in 2025, while demand from AI-focused facilities rose 50 percent. The IEA expects global data center electricity use to more than double to around 945 terawatt-hours by 2030.

How can every task become cheaper while the total bill keeps rising?

Because we do not keep the number or the nature of the tasks constant.

When intelligence becomes cheaper, we use much more of it. We also invent new uses that were previously uneconomic: reasoning models, video generation, persistent agents, automated research and eventually robots that are continuously perceiving, planning and acting.

This is the Jevons effect applied to intelligence.

What the Jevons Effect Actually Means

In the nineteenth century, economist William Stanley Jevons observed that improvements in the efficiency of coal use did not reduce Britain's coal consumption. More efficient steam engines made coal economically useful in more places, so total demand increased.

The effect is often described too casually. Efficiency does not always increase total consumption. The outcome depends on how strongly demand responds to lower prices.

AI appears unusually sensitive to that response.

When an inference workload becomes ten times cheaper, companies do not merely run the same workload and pocket the savings. They move from occasional use to continuous use. A chatbot becomes an agent. One agent becomes a team. A monthly analysis becomes live monitoring. A human-triggered task becomes a background process that runs across every customer, transaction, machine or software repository.

Total energy can be understood with a simple relationship:

Simple text generation may become almost negligible. But an agent that reads thousands of documents, calls tools, writes and tests code, consults other agents and retries failures is not one prompt. It is a small digital operation.

The economic unit is moving from the token to the completed outcome.

A Fermi Estimate: The Researcher and the Agent

A Fermi estimate is not a measurement. It is a transparent order-of-magnitude calculation built from explicit assumptions. Its purpose is to show which variables matter and whether an idea is roughly plausible.

So let us compare the operational energy surrounding one human researcher with one continuously running AI research agent.

This is deliberately not a claim that an agent is equal to a scientist. It is a way to compare the physical support systems behind one hour of possible cognitive work.

Scenario A: one human researcher

Assume approximately 1,700 working hours per year. For the work-related energy envelope, use broad illustrative ranges:

• office or laboratory operations: 4,000 to 8,000 kWh per year;
• commuting and professional travel: 2,000 to 4,000 kWh per year;
• total operational envelope: 6,000 to 12,000 kWh per year.

That produces roughly 3.5 to 7.1 kWh per working hour.

This is not the researcher's metabolism. It is the building, equipment, transport and institutional environment allocated to the work. A wet laboratory can be far higher; a remote theorist can be much lower.

Scenario B: one persistent AI agent

Assume an agent generates or processes the equivalent of 20 tokens per second continuously. Over a year:

20 tokens x 60 x 60 x 24 x 365 = approximately 630 million tokens

Now assume a wide energy range of 1 to 10 Wh per 1,000 tokens, including a rough allowance for server and data center overhead. Actual values vary enormously by model, hardware, batching, context length and workload.

That produces:

• light case: approximately 630 kWh per year;
• heavy case: approximately 6,300 kWh per year;
• per operating hour: approximately 0.07 to 0.72 kWh.

Under these assumptions, the agent uses roughly five to one hundred times less operational energy per available hour than the support envelope around the researcher.

That sounds like an energy-saving story.

Now apply Jevons.

One hundred million persistent agents would consume roughly 63 TWh per year in the light case and 630 TWh in the heavy case. The first figure is manageable inside the projected growth of data centers. The second is already a large fraction of the IEA's entire 945 TWh global data center projection for 2030.

One hundred million agents is not a forecast. It is a stress test. It shows how quickly a cheap unit becomes a system-level constraint when it is replicated across workers, companies, devices and processes.

The Fermi estimate therefore produces two conclusions at once:

1. Digital cognitive work can be dramatically more energy-efficient per operating hour.
2. That efficiency can unlock enough additional demand to increase total electricity consumption.

Both are true.

Four Caveats the Calculation Needs

First, a current agent is not a scientist. It may produce useful code or analysis while lacking research taste, real-world understanding or the ability to select the right problem.

Second, training is excluded. Frontier training runs consume substantial energy, although that cost is spread across all later use of the model.

Third, embodied energy is excluded. Chips, data centers, laboratories, buildings and vehicles all require energy and materials to manufacture.

Fourth, an agent does not automatically remove the human energy envelope. If researchers keep commuting to the same laboratories while adding agents, the new electricity use comes on top of the old system.

The estimate is useful because it exposes the variables. It is dangerous only when presented as precision.

AI, AGI and ASI Are Not the Same Thing

The energy discussion becomes more important as capability expands. But the labels are often used carelessly.

There is no universally accepted AGI threshold, so it is better to treat the terms as regions on a capability map rather than release names.

Google DeepMind's Levels of AGI framework separates performance, generality and autonomy. That is useful because a system can be superhuman at mathematics yet poor at household planning, or highly autonomous while still making frequent mistakes.

AGI will probably not arrive as a universally agreed date. It will arrive as an argument.

Companies will automate large parts of work while researchers continue debating whether the systems qualify. The economy may cross the practical threshold before the vocabulary catches up.

Where We Actually Are in 2026

Current systems have crossed remarkable narrow thresholds. They write and debug code, use computers, pass difficult examinations, discover algorithms and operate as agents for hours.

They remain unreliable.

METR's 2026 measurements estimate a roughly 12-hour 50-percent task horizon for the public frontier on its suite of self-contained software, machine-learning and cybersecurity tasks. That does not mean an AI can replace a professional for twelve hours. METR explicitly warns that its tasks are unusually well-specified, low-context and automatically verifiable, and that estimates above sixteen hours are currently unreliable.

This is the jagged frontier: astonishing competence next to basic failure.

Today we have powerful AI and increasingly autonomous agents. Whether that deserves an early AGI label is less important than the operational fact that the length and complexity of tasks they can attempt keeps growing.

Why AGI Would Intensify the Jevons Effect

Current AI demand is limited by capability. Many valuable tasks cannot yet be trusted to an agent.

AGI would remove a large part of that limit.

If a system can reliably handle most cognitive work, the number of economically viable AI tasks expands from millions of workflows to a significant fraction of all desk work. Organizations would not buy one AGI query. They would provision persistent capacity.

Every employee could have a team of agents. Every software repository could be continuously maintained. Every scientific hypothesis could be assigned to parallel research processes. Every machine could have a planning layer. Every organization could simulate decisions before making them.

The demand curve would move from human-triggered use toward continuous machine-speed activity.

AGI could make each task more efficient while multiplying the number of tasks by orders of magnitude.

That is Jevons at civilizational scale.

ASI Could Move the Ceiling

ASI adds a more radical possibility: intelligence improving the systems that produce intelligence.

Elements of this loop already exist. Google DeepMind's AlphaEvolve has discovered improvements used in computing and AI training. Anthropic reports that AI writes a substantial share of its code and is increasingly used in AI research. But full recursive self-improvement - an autonomous system designing, training and validating a superior successor - has not been publicly demonstrated.

If it does emerge, an ASI would logically work on its own constraints:

• more efficient algorithms;
• better chip architectures and memory systems;
• improved cooling and power electronics;
• new battery, nuclear, geothermal or fusion technologies;
• automated factories and laboratories;
• better ways to verify research and coordinate agents.

The energy ceiling would no longer be fixed. Intelligence could help raise it.

But intelligence does not remove physics.

A better reactor design still needs materials, permits, construction and testing. A drug still needs biological validation. A chip still needs a fabrication plant. A power grid still contains transformers that take time to manufacture.

ASI could make invention move at the speed of compute while deployment remains constrained by the speed of matter.

That mismatch may become the defining bottleneck.

The Quantum Wildcard

Quantum computers are often inserted into this story as a magical final accelerator.

They are not faster GPUs.

Modern AI training is dominated by huge volumes of relatively simple linear algebra over classical data. GPUs are exceptionally good at that. A quantum computer does not automatically make transformer training faster, and loading classical datasets into quantum states can erase theoretical speed advantages.

Research on quantum machine learning remains active, but Nature Machine Intelligence identifies classical-to-quantum data representation as a central bottleneck. Quantum advantage is more plausible when the data is naturally quantum, as in chemistry, materials and quantum systems.

The realistic relationship is therefore not:

quantum computer replaces the AI data center

It is:

classical AI system calls a quantum processor for a narrow problem that fits quantum computation

What Quantum Could Add

Materials and chemistry

Quantum computers may eventually simulate molecules and quantum materials that are prohibitively difficult for classical systems. Better catalysts, batteries, superconductors, medicines and semiconductor materials could indirectly improve the entire AI infrastructure stack.

Optimization

Some scheduling, portfolio, routing and industrial optimization problems may benefit from quantum or hybrid quantum-classical methods. The advantage will be problem-specific, not universal.

Cryptography

A sufficiently capable fault-tolerant quantum computer could threaten widely used public-key cryptography. This is the most immediate practical consequence because organizations must migrate before such a machine exists. NIST finalized its first post-quantum cryptography standards in 2024 and continues expanding implementation guidance in 2026.

AI for quantum

The arrow also runs in the other direction. AI is already useful for quantum-device design, calibration, error decoding and control. Classical AI may be one of the technologies that makes useful quantum computing possible.

Where Quantum Stands

The field is progressing, but headline qubit counts are difficult to compare.

Google's Willow program demonstrated that logical error rates can fall as the error-correcting code is scaled, an essential milestone toward useful fault-tolerant computing. In March 2026, Quantinuum reported up to 94 error-detected logical qubits and 48 error-corrected qubits from 98 physical qubits using a different architecture and code. These are important experiments, but they are not general-purpose replacements for classical supercomputers.

IBM's current roadmap targets a 200-logical-qubit, 100-million-gate fault-tolerant system called Starling for 2029. That is a company target, not a guaranteed date.

Quantum is best understood as a specialized coprocessor in a future heterogeneous computing system: CPUs for control, GPUs and other accelerators for AI, and quantum processors for the narrow classes of problems where they provide a verified advantage.

The Feedback Loop

Now the pieces connect.

Cheaper classical AI increases total demand through the Jevons effect.

Persistent agents make digital cognitive work cheaper per available hour, as the Fermi estimate illustrates, but their replication pushes total electricity demand upward.

AGI expands the number of tasks that can be automated.

ASI could improve algorithms, chips and energy systems, raising the physical ceiling.

Quantum computing could help with selected materials, chemistry and optimization problems that improve the infrastructure underneath AI.

Those improvements make intelligence cheaper again.

The result is not one technology replacing another. It is a compounding system:

AI improves compute and energy. Better compute and energy enable more AI. More AI creates more demand for improvement.

This loop can produce extraordinary abundance. It can also create extraordinary pressure on grids, supply chains, capital markets and governance.

The Real Limit May Be Matter, Not Intelligence

The common mistake is to treat energy efficiency as an endpoint.

It is more likely to be an invitation.

If intelligence becomes one hundred times cheaper, society will not ask the same number of questions for one percent of the cost. It will build systems that were previously absurd: permanent research teams, continuous software development, personalized education at population scale, automated laboratories and machine intelligence embedded in the physical world.

The Fermi estimate shows why that can be efficient at the unit level.

Jevons shows why the total can still explode.

AGI and ASI expand the demand side by making more cognitive tasks possible. Quantum computing may expand the supply side indirectly by helping solve narrow scientific and engineering problems.

None of this guarantees AGI, ASI, practical quantum advantage or unlimited growth. Every trend can bend into an S-curve. Verification may become the bottleneck. Energy projects may arrive too slowly. Quantum roadmaps may slip. Intelligence may prove harder to generalize than current benchmarks suggest.

But if the capability curve continues, the central economic question will not be whether intelligence is expensive.

It will be what happens when intelligence is cheap enough to be everywhere.

Sources

• IEA: Key Questions on Energy and AI, 2026
• IEA: Energy and AI
• METR: Task-Completion Time Horizons of Frontier AI Models
• METR: Frontier Risk Report, February-March 2026
• Stanford HAI: Artificial Intelligence Glossary
• Google DeepMind: Levels of AGI
• Anthropic Institute: When AI Builds Itself
• Google DeepMind: AlphaEvolve
• Google Quantum AI: Quantum error-correction milestone
• Quantinuum: Skinny Logic
• IBM: Quantum Roadmap to 2030
• Nature Machine Intelligence: Seeking a quantum advantage for machine learning
• Nature Communications: Artificial intelligence for quantum computing
• NIST: Post-Quantum Cryptography

The Fermi estimate in this essay is an original illustrative calculation. Its assumptions are shown explicitly and should not be interpreted as a measured comparison between human and AI research productivity.