The Jevons Paradox: Why Cheaper AI Means More AI, Not Less

The Tweet That Called the Bottom

On January 27, 2025, Nvidia lost $589 billion of market value in a single session - a 17% drop, at the time the largest one-day market-cap loss in stock-market history. The trigger was a Chinese AI lab most investors had never heard of two weeks earlier. DeepSeek had released a frontier-class model and published the bill: $5.576 million for the final training run, on de-rated chips, at rental rates.

The logic of the selloff was simple. If world-class AI can be trained for single-digit millions instead of the hundreds of millions frontier runs were widely reported to cost, then the hundreds of billions flowing into chips and data centers are overkill. Efficiency kills demand. Sell the picks and shovels.

That same day, Microsoft CEO Satya Nadella posted a one-line rebuttal citing an argument first published 160 years earlier, in a book about coal:

"Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of."

Fourteen months later, we can score the argument. Nvidia closed today at $177.39, the DeepSeek drawdown long since recovered. The four largest AI spenders have guided 2026 capital expenditure to a combined $625 billion-plus. Google is processing 1.3 quadrillion tokens a month, up roughly 134x in the eighteen months through last October.

We used Barebone to line the DeepSeek episode up against every major efficiency shock since 1865 - coal, bank tellers, bandwidth - to test whether the Jevons paradox actually holds, and, more importantly, where it breaks. Because it does break. And the places it breaks are exactly where the money gets lost.

The Economist Who Worried Britain Would Run Out of Coal

In 1865, the English economist William Stanley Jevons published The Coal Question, a book-length panic about Britain exhausting its coal reserves. Buried inside it was an observation that outlived everything else he wrote.

James Watt's steam engine was dramatically more efficient than the Newcomen engine it replaced - far more useful work per ton of coal burned. The intuitive prediction: better engines, less coal. The actual result: cheaper steam power made engines economic in places they had never made sense before - factories, railways, steamships, mines. British coal output roughly trebled between 1830 and 1850, from about 15 million to 49 million tons a year, and kept compounding to 224 million tons by 1900. Across the nineteenth century, Britain mined roughly two-thirds of the world's coal.

Jevons's conclusion was that treating fuel economy as equivalent to reduced consumption was a confusion of ideas - the very contrary, he argued, was the truth. Efficiency lowers the effective price of whatever the fuel does. And when the price of something useful collapses, people find vastly more uses for it.

That is the entire paradox. It is also not a law of nature - which Jevons never claimed it was. Hold that thought.

The ATM Test

The cleanest modern test began in 1969, when Chemical Bank installed America's first ATM at a branch on Long Island. The machine did the core of a bank teller's job. The forecast wrote itself: tellers were finished.

Economist James Bessen ran the numbers decades later. In 1985, the US had roughly 60,000 ATMs and 485,000 bank tellers. By 2002, ATMs had multiplied to 352,000 - and teller employment had risen to 527,000. By around 2010, with some 400,000 machines installed, teller jobs reached nearly 600,000.

The mechanism was pure Jevons. ATMs cut the number of tellers needed to run a branch from about 20 to 13 between 1988 and 2004. Cheaper branches meant banks opened more of them - urban branch counts rose 43% over the same window - and every new branch hired tellers, who shifted from counting twenties to selling mortgages.

Look at the right-hand side of that chart before filing this under "automation never kills jobs." Teller employment was 350,300 in 2023, and the Bureau of Labor Statistics projects a further 15% decline by 2033. The ATM made branches cheaper. The smartphone made them unnecessary.

We will come back to that kink in the line. It is the most important data point in this article.

The $500 Billion Rehearsal

The 1990s ran the experiment at telecom scale. After the Telecommunications Act of 1996 deregulated the industry, carriers invested more than $500 billion in five years - most of it borrowed - laying fiber-optic cable on the theory that internet traffic would grow without limit.

Here is the uncomfortable part: the demand thesis was right. Traffic kept growing through the crash and never stopped. Streaming, social media, and cloud computing - industries that did not exist when the fiber went into the ground - were eventually built on exactly that glass.

The investors still got carried out. Capacity arrived years ahead of demand: after the bust, an estimated 85 - 95% of the fiber laid in the 1990s sat dark - installed but unused - and one 2002 estimate put the share actually lit below 3%. Global Crossing went bankrupt in 2002 carrying $12.4 billion of debt. WorldCom followed in the largest accounting fraud in US history.

Jevons tells you demand for a cheapening input will expand. It tells you nothing about whether capacity shows up five years early, or whether the people who financed it survive long enough to see utilization catch up.

DeepSeek: Jevons in Real Time

Which brings us back to January 2025. DeepSeek's V3 technical report, published in late December 2024, stated that the model's final training run consumed 2.788 million GPU-hours on Nvidia H800s - the export-compliant chip Nvidia was allowed to sell into China - which at $2-per-hour rental rates works out to $5.576 million. The same paper notes, in plain text, that the figure excludes all prior research, ablation experiments, and infrastructure. It was the cost of the last lap, not the program. Almost nobody read the footnote.

When DeepSeek's R1 reasoning model dropped on January 20, the December paper's numbers landed on every trading desk's screen, and the market priced the headline: frontier AI at a small fraction of the assumed cost. Nvidia fell 17% in a day.

The capex response is now a matter of public record across five earnings seasons - every one of them raised the number. In December 2025, street forecasts had the five biggest hyperscalers (including Oracle) crossing $600 billion of capex in 2026, a 36% jump over 2025. Then the Q4 2025 earnings calls landed in January and February of this year, and the guidance from just the big four blew through that forecast:

Company	2026 capex guidance	Midpoint used
Amazon	~$200B (announced Feb 5, 2026)	$200B
Alphabet	$175 - 185B	$180B
Meta	$115 - 135B	$125B
Microsoft	"at least" $120B	$120B
Combined		$625B+

For scale: that is more than four times what the entire publicly traded US energy sector spends in a year drilling wells, refining oil, and delivering gasoline. Fourteen months of "efficiency kills demand," answered with $625 billion of spending plans - and every earnings call since the panic has revised the number up, not down.

Why Cheaper Intelligence Means More Intelligence

The mechanism shows up most cleanly in the unit economics of inference - what it costs to use a model once it is trained. Every chatbot answer is metered in tokens, the word-fragments models read and write.

Stanford's 2025 AI Index measured the collapse: querying a model at GPT-3.5-level performance cost about $20 per million tokens in November 2022 and about $0.07 per million tokens by October 2024 - a more than 280-fold price decline in eighteen months.

If efficiency killed demand, inference spending should have fallen roughly in line. Instead, usage went vertical. Google disclosed it processed about 9.7 trillion tokens a month in April 2024. By May 2025 - three months after the DeepSeek panic - the figure was 480 trillion. By October 2025, 1.3 quadrillion a month, which Sundar Pichai framed on the Q3 earnings call as more than 20x growth in a year.

Price down two orders of magnitude, volume up two orders of magnitude. That is the coal chart with a 160-year software update.

Episode	Efficiency shock	Demand response
Coal, 1830 - 1900	Watt engine slashes coal per unit of work	UK output trebles by 1850, hits 224Mt by 1900
Bank tellers, 1985 - 2010	ATMs cut tellers per branch from 20 to 13	Branches +43%; teller jobs reach nearly 600K
Inference, 2022 - 2025	GPT-3.5-level cost falls 280x	Google tokens up 134x; 2026 capex guided to $625B+

The economics are not mysterious. When a unit of intelligence costs $20, you ration it for high-value questions. When it costs seven cents, you wire it into every code editor, search box, and support queue you own - and workloads that were absurd at the old price, like agents that burn hundreds of model calls to finish one task, become line items.

The demand was never really demand for tokens. It was demand for cognition, and cognition was supply-constrained by price.

Where the Logic Breaks

Now the part the January 2025 bears deserve to hear said honestly: the Jevons paradox is not a law. It is a special case. Economists who study energy efficiency call demand's response the "rebound effect," and they have measured it for decades. The findings are inconvenient for anyone using "Jevons" as a one-word thesis.

Most efficiency gains do not backfire. Across modern energy uses, measured direct rebound effects run roughly 10 - 20% - efficiency mostly just saves energy. Full backfire, where total consumption rises on net (the true Jevons case), is rare, and researchers reviewing the field have concluded that backfire claims generally do not hold up to scrutiny. The lighting literature puts the rebound from LED and CFL bulbs at roughly 6%. When LEDs made light radically cheaper, nobody lit their lawn like a stadium. Demand for lumens in rich countries was saturated - everyone already had as much light as they wanted.

That is the first condition: Jevons requires unsaturated, price-elastic demand. Coal in 1865 qualified. Lighting in 2015 did not. The open question is which one cognition resembles - and how long its unsaturated phase lasts.

Complements flip into substitutes. The teller chart's kink is the second failure mode. ATMs lifted teller employment for three decades because they complemented the branch system. The smartphone did not make branches cheaper; it made them unnecessary, and teller jobs have fallen roughly 40% from their 2010 area to 350,300. Every Jevons effect has a shelf life, and it ends when the next technology stops complementing the thing you own and starts replacing it. Today's capex assumes AI keeps complementing data-center-scale computing specifically.

Being right about demand does not make the capex safe. That is the fiber lesson. Bandwidth demand grew exactly as the bulls promised - and 85 - 95% of the fiber still sat dark while its financiers went bankrupt. Tokens have their own version of this problem: tokens are not revenue. Reasoning models burn many times more tokens per answer than the models they replaced, and quadrillion-token disclosures include free products. With per-token prices falling 280-fold, revenue grows only if volume outruns deflation. So far it has, by a wide margin. "So far" is the load-bearing phrase in this entire industry.

What This Means

The Jevons paradox is the best one-line answer to "cheaper AI kills AI spending," and fourteen months of rising guidance back it. But the history says it is one-third of a framework, and the investors who stopped at the first third are the ones who bought fiber in 1999. Three questions worth tracking from here:

1. Is demand still elastic? Watch the spread between usage growth and price deflation - token volumes against per-token prices. The day usage growth decelerates toward the rate of price decline, the Jevons phase is ending and AI infrastructure starts trading like a utility. As of the latest disclosures, volume is winning by an order of magnitude.

2. Complement or substitute? Every AI-exposed workflow eventually gets its smartphone moment. Cheap on-device models, or architectures that need radically less centralized compute, would be the kink in this era's teller chart. The signal is not whether AI grows - it is whether the growth still routes through the assets being built today.

3. Is capacity outrunning demand? The fiber builders were not wrong about traffic; they were early, levered, and undifferentiated. Utilization, depreciation schedules, and how the buildout is financed will tell you whether 2026's $625 billion is coal in 1865 - or fiber in 1999.

Jevons promises that the demand shows up. He never promised it shows up on schedule - or that the people who paid for the buildout are the ones who get paid.

---

Data: Barebone | Sources: DeepSeek-V3 Technical Report (December 2024), Stanford HAI AI Index Report 2025, Alphabet Q3 2025 earnings call, hyperscaler Q4 2025 earnings calls (January - February 2026), U.S. Bureau of Labor Statistics Occupational Outlook Handbook, W.S. Jevons, The Coal Question (1865) | Data as of April 2, 2026