The AI Jevons Paradox: Why Cheaper AI Means More Demand, Not Less
In 1865 William Stanley Jevons observed that as steam engines became more efficient, Britain consumed more coal, not less. Every efficiency gain created new economic uses for coal faster than it eliminated old ones. The paradox has now returned in a form Jevons would recognize instantly.
Satya Nadella said it explicitly. On 27 January 2025, the day markets sold off on DeepSeek R1, Nadella posted: 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. Jensen Huang, asked directly about the paradox in a T. Rowe Price interview, agreed: we made computing so cost effective it created something new. And so a new type of data center emerged. And this new type of data center is called AI factories.
The efficiency gains are unambiguous. Stanford’s AI Index 2025 documents that the cost of querying a model at GPT-3.5-equivalent quality fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024 – a 280-fold reduction in 18 months. Depending on the task, LLM inference prices have fallen anywhere from 9x to 900x per year, with Epoch AI putting the median at roughly 50x per year. Andreessen Horowitz’s LLMflation analysis reaches the same conclusion: token prices are collapsing faster than almost any commodity in economic history.
By classical demand elasticity, this should reduce total spend on compute. It has done the opposite. Epoch AI’s own accounting is stark: global inference capacity is growing about 3.4x per year while token demand is growing about 10x per year. Demand is outrunning supply by a factor of three, at compounding rates. Google reports serving 1.2 billion tokens per second; Meta’s 85,000 employees consumed 60 trillion tokens in a single month. SemiAnalysis notes that some knowledge workers now consume over 100 billion tokens a month personally, with annualized token spend running near 30% of employee compensation for heavy users.
The binding constraint has moved cleanly from silicon to the grid. The IEA’s Energy and AI report forecasts that global data-centre electricity consumption will more than double from about 415 TWh in 2024 to roughly 945 TWh by 2030 – slightly more than Japan’s total electricity consumption today – and rise to about 1,200 TWh by 2035. Accelerated AI servers alone are growing at 30% a year and will account for almost half the net increase. That is more than four times faster than electricity demand growth from all other sectors combined.
The US grid cannot absorb this at pace. Berkeley Lab’s Queued Up 2025 shows roughly 10,300 projects seeking interconnection – 1,400 GW of generation plus 890 GW of storage. Median wait time from interconnection request to commercial operation has doubled to over four years. Only 13% of capacity that requested interconnection between 2000 and 2019 had actually reached commercial operation by end-2024. Bain estimates the world will need 200 GW of AI compute by 2030, roughly 100 GW of it in the US, “on a grid that has seen relatively flat load growth for the past 20 years.”
The strategic response has already begun. Microsoft signed an 835 MW, 20-year PPA with Constellation to restart Three Mile Island. Amazon locked in 1,920 MW of nuclear from Talen through 2042. Google agreed up to 500 MW of SMR capacity from Kairos Power. Meta issued an RFP for 1-4 GW of new US nuclear. These are not ESG gestures. They are hyperscalers buying reactors to sidestep the interconnection queue.
AIMG View. Every efficiency gain in AI has generated more demand than it eliminated. Cheaper tokens are not deflationary for the AI industry; they are a demand accelerant. The bottleneck has shifted from chips to power to permits.
AIMG Bottom Line. Investment theses that model AI compute as a shrinking cost line miss the point. Model it as a growing physical infrastructure line, and locate value in the choke points: grid interconnection, firm power, cooling, transmission and long-dated PPAs.