The Jevons Paradox: Flawed Consensus View On Efficiency

Mornings With Markman - January 27th, 2026

Jan 27, 2026

When DeepSeek released its R1 model in late January 2025, claiming performance comparable to OpenAI’s frontier models at a fraction of the training cost, tech stocks cratered. Nvidia alone lost $600 billion in market value in a single day. The market’s logic seemed straightforward: more efficient AI means less demand for chips, less demand for data centers, less demand for power.

Satya Nadella posted his response within hours.

“Jevons paradox strikes again!” the Microsoft CEO wrote on X, linking to a Wikipedia page explaining the concept. “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.”

The reference was to William Stanley Jevons, a Victorian economist who observed in 1865 that more efficient steam engines did not reduce coal consumption. They increased it. Because steam power became cheaper per unit of work, industries adopted it across every sector of the economy. British coal consumption tripled by 1900.

Nadella was not offering a history lesson. He was explaining why the infrastructure buildout would continue.

The Math of Abundance

The pattern Jevons identified repeats across technologies. When James Watt improved steam engine efficiency, factories did not burn the same amount of coal to do the same work. They burned more coal to do vastly more work. When vehicles became more fuel-efficient, people drove farther. When computing became cheaper per transistor, the world did not use the same amount of compute. It built smartphones, cloud services, and AI models that consume orders of magnitude more.

Wei Wang, a computer architect at ByteDance, wrote in July 2025 that Jevons Paradox is playing out in real time as breakthroughs in GPU efficiency make compute cheaper and more capable, driving explosive new demand for computational resources. Efficiency gains, Wang argued, are necessary but not sufficient for sustainability. The rebound effect can exceed 100%, meaning efficiency improvements result in faster resource consumption.

The data supports this. Meta CEO Mark Zuckerberg raised 2025 AI spending to $60 to $65 billion just days after DeepSeek’s release, declaring that scaling up infrastructure remains a long-term advantage. Microsoft disclosed a $13 billion AI revenue run rate on its FY25 Q2 call, up 175% year over year. Efficiency did not slow investment. It accelerated it.

$600 Billion in Hyperscaler Capex

The numbers have become difficult to comprehend.

CreditSights projects combined capital expenditure for the five largest hyperscalers (Amazon, Microsoft, Alphabet, Meta, Oracle) will reach approximately $602 billion in 2026, a 36% increase from 2025. Roughly 75% of that spend, around $450 billion, is directly tied to AI infrastructure. Goldman Sachs projects total hyperscaler capex from 2025 through 2027 will reach $1.15 trillion, more than double the $477 billion spent from 2022 through 2024.

Individual company guidance tells the same story. Alphabet revised its 2025 capex upward three times, reaching $91 to $93 billion versus $52.5 billion in 2024. Microsoft spent $34.9 billion in capital expenditures in a single quarter, a 74% year-over-year increase. Amazon raised 2025 capex guidance to $125 billion, a 61% increase.

This spending is not slowing. Capital intensity, measured as capex as a percentage of revenue, has reached levels previously unthinkable for technology companies: 57% at Oracle, 45% at Microsoft. These ratios resemble industrial or utility companies, not software firms.

The Grid Bottleneck

The true constraint is not model efficiency. It is physical infrastructure.

Data Center Knowledge reported in January 2026 that approximately 70% of the U.S. power grid was built between the 1950s and 1970s and is approaching the end of its life cycle. Electricity demand is rising faster than the grid was designed to handle. Power availability has already emerged as a limiting factor for data center developers, with one analysis finding that constraints were extending construction timelines by 24 to 72 months.

PJM Interconnection, the largest U.S. grid operator serving over 65 million people across 13 states, projects it will be six gigawatts short of reliability requirements in 2027. Joe Bowring, president of independent market monitor Monitoring Analytics, told CNBC he has never seen the grid under such projected strain.

The bottleneck is not generation capacity alone. It is transmission and distribution infrastructure. Interconnection queues have swelled to historic levels, with gigawatts of ready-to-build generation projects waiting years for grid connection studies and upgrades. While new solar farms and battery storage can be constructed in months, the transmission lines needed to deliver their power often require a decade or more of planning, permitting, and construction.

Long-Term Power Contracts Signal the Shift

Hyperscalers are responding by locking down dedicated power supply.

In September 2024, Constellation Energy announced a 20-year power purchase agreement with Microsoft to restart Three Mile Island Unit 1, now renamed the Crane Clean Energy Center. The reactor will supply 835 megawatts to Microsoft data centers beginning in 2028. In June 2025, Constellation signed another 20-year agreement with Meta for the entire 1.1 gigawatt output of its Clinton Clean Energy Center nuclear reactor in Illinois, starting June 2027.

Constellation’s stock rallied 57.9% in 2025. Its acquisition of Calpine closed on January 7, 2026, creating a footprint encompassing 21 nuclear reactors and over 50 natural gas plants. Management projected the addition would increase adjusted earnings per share by 20% in 2026.

Amazon secured a 1.92 gigawatt power purchase agreement from the Susquehanna nuclear plant and invested $500 million in small modular reactor development. The pattern is clear: hyperscalers are moving from electricity consumers to infrastructure investors.

The Investment Thesis

The goal of efficiency in AI is not to use less power. It is to do exponentially more work with the power available. Every gain in chip efficiency or cooling performance lowers the cost per unit of intelligence. As cost drops, new industries find it viable to integrate AI into their workflows. Applications that were once too expensive become essential utilities.

Gartner projects data center electricity demand will grow 16% in 2025 and double by 2030, rising from 448 terawatt hours to 980 terawatt hours globally. BloombergNEF expects U.S. data center power demand could reach 106 gigawatts by 2035. The IEA’s central scenario projects data center electricity consumption reaching 945 terawatt hours globally by 2030, with AI’s share rising from roughly 5 to 15% today to potentially 35 to 50%.

The infrastructure layer is where value is captured. Eaton reported Q3 2025 data center orders up approximately 70% year over year, with revenue up approximately 40%. Vertiv revenue jumped 29% year over year to $2.6 billion, with organic orders up 60% and backlog reaching $9.5 billion. These companies provide the electrical equipment and thermal management that AI growth requires regardless of which models win.

The narrative of green software solving energy constraints misses the physics. Efficient algorithms cannot bypass the need for copper, transformers, and baseload power. When the market fully prices the reality that efficiency is a catalyst for demand rather than a substitute for it, the infrastructure assets will reprice accordingly.

Share this with someone who thinks AI efficiency will reduce energy demand.

Satya Nadella X post: “Jevons paradox strikes again!” January 27, 2025 — Fortune https://fortune.com/2025/01/27/microsoft-ceo-satya-nadella-deepseek-optimism-jevons-paradox/
William Stanley Jevons, The Coal Question, 1865: efficiency leads to increased consumption — Wikipedia https://en.wikipedia.org/wiki/Jevons_paradox
Wei Wang (ByteDance): “Jevons Paradox in AI data centers,” rebound effect exceeding 100% — SIGARCH, July 14, 2025 https://www.sigarch.org/the-jevons-paradox-why-efficiency-alone-wont-solve-our-data-center-carbon-challenge/
Meta Zuckerberg raised 2025 AI spending to $60-65B after DeepSeek — AI Proem (Substack), February 2025