Reading the AI Era Through Jevons' Paradox: Why Demand Rises and What Talent Will Be Needed

As generative AI advances quickly, the market often worries that efficiency from AI will take away demand and jobs from existing industries.

But if you look at the history of technology and the basics of economics, the opposite is more likely to happen. A useful concept for understanding this change is “Jevons’ Paradox” from economics.

In this piece, I use this paradox as the main thread to look — based on facts — at the reversal of demand happening across AI infrastructure (hardware), labor (software engineering and white-collar work), and energy.

What Is Jevons’ Paradox

Jevons’ Paradox is a phenomenon described by the 19th-century economist William Stanley Jevons. At the time, when steam engines became more fuel-efficient, people expected that coal consumption would go down. But in reality, efficiency made the cost of using a steam engine drop sharply, so steam engines spread across every industry, and total coal consumption actually went up a lot.

“When the efficiency of using a resource improves, the consumption per use goes down, but because the cost of use drops, total demand instead expands exponentially.”

This same principle is now being repeated in the AI era, on both the hardware and labor sides.

1. The Paradox in the Memory Industry: KV Cache Compression and HBM Demand

A clear hardware example is the market reaction to “TurboQuant,” published by Google Research in March 2026.

TurboQuant: Redefining AI efficiency with extreme compression was introduced as a technique that compresses the KV cache (key-value cache) of an LLM by at least about 6x, and speeds up attention processing on the H100 by up to 8x. Right after this announcement, the market quickly judged that “the memory capacity needed for AI drops to one-sixth, so this is a headwind for memory companies,” and the share prices of major memory-related names like Samsung Electronics, SK hynix, and Micron Technology (MU) fell sharply for a short time.

But this is exactly the kind of moment where Jevons’ Paradox happens.

flowchart TD
    A["KV cache compression<br/>(TurboQuant, etc.)"] --> B["Memory needed per use goes down"]
    B --> C1["[The conventional view]<br/>Memory demand declines"]
    B --> C2["[What Jevons' Paradox shows]<br/>Long-context processing cost drops sharply"]
    C2 --> D["Always-on reference of longer context<br/>More concurrent connections<br/>Multi-agents running 24 hours"]
    D --> E["Total memory (HBM/DRAM) usage<br/>across the data center actually surges"]
    classDef old fill:#fee2e2,stroke:#dc2626,stroke-width:2px;
    classDef new fill:#dcfce7,stroke:#16a34a,stroke-width:2px;
    class C1 old;
    class C2,D,E new;

In fact, after passing through this temporary shock, the market again recognized that total memory demand across AI data centers would be pushed even higher. The share prices of major companies like Micron (MU) have surged more than 3x from the bottom of the past correction.

100

over 320

shock low
after the newsnow

* A relative index showing how major memory names like Micron (MU) rose more than 3x from the bottom of the correction.

Even if the amount needed per use goes down, the marginal cost of processing drops, so companies start running AI at a scale that was not possible before because of cost. So the total market does not shrink — it grows larger.

2. The Paradox of Inference Cost: Lower Inference Cost and the Explosion of Data Center Power Demand

The same structure shows up not only in memory, but also in “the drop of AI inference cost itself.”

Model architecture optimization and quantization keep improving, so the generation cost and power needed per token keep going down year by year. At first glance, this looks like it should lead to lower power use in data centers, but the reality is the opposite.

Because cost dropped, the number of AI calls around the world (API calls and autonomous local execution) has grown to an astronomical number. AI that used to run only when a human typed a prompt has now evolved into “loops where agents think and verify autonomously in the background, 24 hours a day.” Lower unit cost leads to an explosion in usage, and as a result the total power demand of the whole data center is pushed even higher. This is exactly Jevons’ Paradox on the energy side.

3. The Reversal of “Abandoned Technology” Caused by Lower Cost

This phenomenon — “as cost drops, demand spreads into areas that were once considered low value and abandoned” — also shows up as a “redefinition of use.”

For example, Google’s decision to adopt a 300MW/30GWh “iron-air battery,” one of the largest in the world, from the US startup Form Energy at its new AI data center in Minnesota is very symbolic.

Google: data centre powers up with 100-hour batteries (Energy-Storage.News) reports on this.

Aspect	Features of the iron-air battery
How it works	A very low-tech approach that uses the chemical reaction of iron “rusting (oxidizing)” in air.
Drawbacks	Very heavy and large. Energy efficiency (50–70%) is worse than lithium-ion (over 90%).
Benefits	Manufacturing cost is a fraction. It can store energy over a very long cycle of 100 hours (about 4 days) or more, and it never catches fire.

This was a low-tech that had long been abandoned as “too heavy to use” for mobile or EV applications. But the moment it connected with the huge demand for stationary storage (BTM: Behind-the-Meter) at AI data centers — “to smooth power safely on-site over a long period” — its value flipped 180 degrees into the best solution.

Just as “trivial areas (niche problems) that could not be automated because of low ROI” in labor are being rescued by AI, in the hardware and infrastructure space too, low-tech that had been abandoned as “low efficiency” is starting to be re-evaluated on the new main battlefield of the AI era.

4. The Paradox in Labor: Automation of Hands-On Work and Rising Demand for “Architects”

This explosion of demand from efficiency does not stay with physical resources like infrastructure. It spreads in exactly the same shape into the structure of our “labor (jobs).”

As the computer scientist Andrew Ng points out, in an age where GitHub Copilot has passed 20 million users and 84% of engineers use AI daily, the cost of coding (the work of giving things shape) has dropped sharply (related post).

Because of this, there is a temporary distortion in the market — job postings for junior roles with a lot of routine work have dropped (down 40% compared to 2022). But the demand for the job itself has not been lost: on a certain platform, the total number of engineering job postings for new graduates is over 400,000 per week. Because the barrier to coding dropped, “long-tail demand” — building systems for trivial tasks and niche problems that were once abandoned as “not worth the cost (ROI)” — is being dug up all at once around the world.

That said, the human role required here is changing a lot. If I compare it to architecture, the following shift in roles is happening.

• Carpenter (hands-on work of building as instructed): This is the area where AI coding tools provide support with very high accuracy and speed. The share of simple work — just writing the code you are told to, or just making boilerplate text and documents — will go down.
• Architect (the role of drawing the whole blueprint): This is the area where humans are needed. The value of orchestrating the whole system — weighing architecture, security, maintainability, redundancy, and the cost-performance that prevents shadow IT — jumps up instead, precisely because you now hold a tool with 10x horsepower (AI).

A senior engineer who has steadily experienced everything from customer interaction to design to programming, and who has an overview of the whole structure, can now focus on this “architect” role, and their importance is rising even more.

This structure — “AI assists the hands-on work, humans design the whole structure” — applies the same way to all white-collar jobs: defining risk in legal, building strategy in marketing, making management decisions in finance.

For example, in marketing, you leave “mass-producing banners and making routine reports” to AI, and concentrate on “building strategy based on the customer’s essential insight (architect).”

In finance, you automate “daily journal entries and data aggregation,” and focus on “simulating management decisions based on future forecasts (architect)” — a division of roles like this.

Precisely because the burden of the work itself goes down, the ability to think for yourself about the story of “what should we solve in the first place,” and to build it with AI as a tool, becomes important.

Conclusion: What You Need to Survive the AI Era, Read Through Jevons’ Paradox

The essence of the change happening now is not “the disappearance of jobs or industries,” but “a large-scale shift of roles that comes with efficiency.”

As Jevons’ Paradox shows, when the marginal cost of some work approaches zero, we no longer need to compete on “speed or volume of output.”

What will be needed in white-collar work from now on is not a person who competes with AI on speed or volume of output.

It is a person who listens thoroughly to the customer’s words, draws out from the dialogue “what is truly needed” — something even the customer has not noticed — and builds it into a single story called “the service the customer is most happy with.”

AI is not the goal. It is only a tool for giving shape to that story.

The human, as the lead, takes a clear sense of purpose and guides the tool called AI, and raises satisfaction through contact with the customer that only a human can do.

And as Jevons’ Paradox shows, because the marginal cost has dropped, we can now reach even the niche areas that did not pay off before and were left untouched.

That is why, for a person with this sense of purpose, an era full of opportunity is beginning — one where you can give shape to ideas you once gave up on, and take on problems you could not tackle before, one after another.