Vol. XII · No. 05 · May 2026
Jake Cuth.

Cheap has happened before.
This fast hasn't.

A seven-century stress test of Chris Anderson's claim that AI is the first commodity with infinite demand. Eleven commodities, from Babylonian lamp oil to GPT-4 tokens. The historical record says the pattern is familiar. The speed is not.

Scroll through five interactive charts built from Nordhaus, Fouquet, Epoch AI, IEA, and Goldman Sachs data. Each chart asks whether intelligence's price collapse is unprecedented, and if so, in kind or only in velocity. Toggle commodity lines. Hover for sources. The data answers its own question.


On May 24, 2026, Chris Anderson posted a diagram to X and asked a question that sounded rhetorical: has there ever been a commodity with effectively infinite demand?

The thesis is seductive. LLM inference costs are falling 10-50x per year, yet total token consumption and spend keep rising. The pattern looks like nothing any economist has ever documented. Anderson's diagram shows price falling and demand expanding simultaneously in what he calls "recursive expansion."

The problem is that the thesis, in its strong form, is empirically wrong. Not wrong about elasticity. Intelligence is plainly elastic. Wrong about uniqueness. At least six commodities in the historical record exhibit the same pattern: price collapses of multiple orders of magnitude, demand growth that outruns the price decline, and total spend that rises even as unit prices crater. The economists who documented this pattern have a name for it. They call it the Jevons paradox, and it is 161 years old.

Anderson is right about elasticity. He is wrong about uniqueness. The interesting question is whether intelligence's velocity changes the implications.

Each line traces the real-price decline of one commodity from its "takeoff" year (defined as the first 10x price drop). The y-axis is logarithmic. Toggle commodities on or off. Intelligence's line is the steepest, but the shape is familiar.

Commodities Reference lines
Intelligence's three-year, 1,000x decline is the steepest in the historical record. But every curve on this chart looked "unprecedented" at the same stage.

Each dot is one commodity. The x-axis measures how far the real price fell. The y-axis measures how much total demand grew. Everything above the diagonal consumed more than it saved. The Jevons signature.

Every elastic-demand commodity in the comparison set sits above the 45° line. Jevons is the rule, not the exception.

Three curves normalized to "year zero" of industrial-scale price collapse. Light begins at 1800, compute at 1940, intelligence at 2021. At the same elapsed time, intelligence's slope is roughly 10x steeper than compute's and 100x steeper than light's.

Time window
Intelligence at year 3 has already traversed the distance that took compute 15 years and light 80 years.

Every prior "infinite demand" commodity hit a physical ceiling. For intelligence, the binding constraint is upstream: electricity. The cascade below traces how a 10x drop in token price propagates through the infrastructure stack.

LLM Token Price
1,000x decline in 3 years
Compute Demand
4-5x per year (training compute)
Electricity Demand
945 TWh by 2030 (IEA)
6.7-12% of US electricity by 2028 (LBNL)
1,350 TWh by 2030 (Goldman)
Grid Capacity
$720B US grid spend needed
10+ year lead time for transmission
“The cost of AI will converge to the cost of energy.”
Sam Altman, US Senate, May 2025

GPT-3 quality decline
1,000xin 3 years
GPT-4 quality decline
62xsince Mar 2023
Fastest current rate
40xper year (GPQA)
UK lighting decline
12,000x1300-2000
Data-centre power 2030
945TWh (IEA)
US share by 2028
6.7-12% (LBNL)
Enterprise LLM spend
$8.4Bmid-2025
China tokens/day
180trillion (Feb 2026)

Engine

Every chart is hand-drawn in SVG by a small vanilla JavaScript module. No frameworks, no build step, no analytics beyond Cloudflare's page-view counter. Data is baked into the page as static arrays. Pill filters and hover tooltips redraw locally. Nothing is fetched from an API.

Data sources

Price-of-light series from Fouquet & Pearson (2006) and Nordhaus (1996). Compute from Nordhaus (2007) and Epoch AI. LLM token pricing from Epoch AI's inference cost dataset and a16z's LLMflation analysis (Nov 2024). Solar PV from IRENA via Our World in Data. DNA sequencing from NHGRI. Energy demand from IEA Energy and AI (Apr 2025), LBNL (Dec 2024), and Goldman Sachs Research (Apr 2026). Elasticity estimates from Sorrell (2007) and Saunders (2000). Snapshot vintage: May 25, 2026.

Reading list

Fouquet & Pearson 2006 · Seven Centuries of Energy Services ↗
Epoch AI · LLM inference price trends ↗
a16z · LLMflation report ↗
IEA · Energy and AI (Apr 2025) ↗
Nordhaus 1996 · Do Real Output and Real Wage Measures Capture Reality ↗
Jevons 1865 · The Coal Question (archive.org) ↗

Limitations

Pre-1800 light prices depend on Nordhaus's reconstruction and carry uncertainties of plus or minus 50% at the Babylonian end. LLM token pricing combines list prices across providers; cached, batched, and quantized prices can be 5-10x cheaper. Demand-side data for intelligence are recent and self-reported by OpenAI, Microsoft, OpenRouter, and the China National Data Bureau via Sina Finance. They are company-disclosed, not audited. The "GPT-4 equivalent" quality threshold drifts as benchmarks saturate. All forward-looking power figures (945 TWh, 1,350 TWh, 6.7-12%) are forecasts, not historical data. The Jevons/rebound literature is contested. This lab presents the data and the range, not a verdict on macro rebound.