The 100x Problem
Why AI's Energy Impact Has No Historical Precedent
This isn’t just a story about how much energy AI uses (it’s all over the news and LinkedIn… constantly these days)…. it’s about how an entire economic revolution is concentrated into a single, innermost loop: converting electricity into intelligence.
At the India AI Impact Summit earlier this year, Demis Hassabis, the CEO of Google DeepMind and a Nobel laureate, made a striking claim: artificial general intelligence will deliver “the equivalent of 10x the Industrial Revolution, happening 10x faster.”
Some people heard that as AI hype. I heard it as an energy statement. And once you think about it as an energy statement, the implications are stark.
Let me explain.
The Industrial Revolution Was Broad
The Industrial Revolution was arguably the most transformative economic event in human history. It reshaped every sector of the global economy. But here is the thing: it did so across an extraordinarily wide surface area.
Steam engines powered factories, railways, ships, and mining operations, mechanization transformed textiles, agriculture, steel production, chemical processing, and construction. The revolution ran on coal, then oil, then natural gas, then electricity. And it took 150 years.
It was a thousand simultaneous transformations spreading across the entire economy over the better part of two centuries. Energy demand grew enormously, but it diffused: across fuels, across sectors, across continents, across generations.
AI Is Not Broad. AI Is Narrow.
Not narrow in its economic impact. Hassabis may well be right that AI’s economic impact rivals or exceeds the Industrial Revolution. The narrowness is in its physical pathway.
Every dollar of economic value that AI creates flows through a single bottleneck: electricity, to a data center, to a GPU. There is no coal equivalent. There is no steam equivalent. There is no distributed mechanical pathway. Unlike every prior technological revolution, AI is structurally incapable of diffusing its energy demand across multiple fuels, sectors, or delivery mechanisms.
Consider the comparison:
The Industrial Revolution: dozens of sectors, multiple fuels (coal, oil, gas, electricity), diffused over 150 years. Energy demand grew massively but spread across the entire economy.
Electrification: many sectors (residential, commercial, industrial), one primary fuel, diffused over 50+ years. Transformative, but the load was distributed across millions of end uses.
The Internet: one sector (telecom/data), one fuel (electricity), but relatively low energy intensity per unit of economic value. The internet’s total electricity footprint, even at scale, remained a modest share of total load.
AI: one sector (compute), one fuel (electricity), extreme energy intensity per unit of economic value, concentrated in specific geographic locations, arriving fast.
Each successive technology revolution has been more concentrated in its energy pathway than the last.
The 100x Problem
Now go back to the Hassabis claim. 10x the Industrial Revolution. 10x the speed.
If the Industrial Revolution’s total energy transformation played out across dozens of fuel types, dozens of sectors, and 150 years, and AI delivers a comparable or greater economic transformation through a single energy vector in a single decade, the concentration factor is not 10x. It is orders of magnitude higher.
That is the 100x Problem. Not 100x more energy in absolute terms, though the numbers are staggering. The real problem is that a 100x economic revolution is structurally incapable of diffusing away from electricity conversion. Electricity is the only pathway. And it is a pathway that was built for a world where demand was broad, distributed, and gradual.
The grid has handled large demand growth before. In other cases, demand grew across millions of endpoints, over decades, with natural geographic and temporal distribution. And a lot of that demand is price sensitive.
AI demand does not work this way: it does not respond to price signals the way any previous category of demand has responded. In my previous article introducing the Compute Heat Rate™1 (https://computeheatrate.substack.com/p/the-compute-heat-rate), I described a framework that quantifies this last point: the price tolerance of AI workloads ranges from roughly $250/MWh on the low end to well over $3,000/MWh on the high end, compared to traditional industrial loads that curtail at $50 to $100/MWh. Even the cheapest AI workloads can tolerate electricity prices 3 to 6x higher than what shuts down a steel mill.
It is not just that AI uses a lot of electricity. It is that AI’s electricity demand cannot diffuse, cannot substitute to another fuel, and will not curtail at prices that could be catastrophic for other consumers.
What now?
Hassabis may be right. He may be half right. He may be wrong about the magnitude but right about the speed, or right about the magnitude but wrong about the timeline. It does not matter as much as you would think. Because even at a fraction of his claim, the concentration of economic impact through a single energy pathway creates grid dynamics that have no precedent.
The question is not whether AI will use a lot of electricity. Everyone knows that. The question is whether the grid, the markets, and the millions of existing electricity consumers are prepared for a new demand class that concentrates the energy footprint of an entire economic revolution into a single, innermost loop.
More to come.
Hans Royal is the originator of the Compute Heat Rate™ (CHR) framework. All views are his own and do not represent those of any employer or affiliated organization.
Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance
and Its Implications for Wholesale Market Repricing (February 28, 2026).
Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318