LLMs Convert Electricity into Intelligence
The future of AI & power
I believe that intelligence is the ultimate multiplier of human endeavor. Our hunger for intelligence is effectively unlimited. And the more I use AI tools, the more I am convinced that LLMs are a tool for converting electricity into intelligence. A simple corollary follows: electricity demand is on a path to infinity.
There may be obstacles along the way. There are scaling laws to grapple with. As LLMs become more advanced, their compute needs increase quadratically. Electricity generation is costly and requires natural resources. Emissions are an issue, depending on your method of electricity generation (and who you ask). The question is how gracefully politics, markets, and physics can stretch to meet this appetite.
Physics, Scaling Laws, and Entropy
Physics dictates how much energy we need to produce a unit of cognition. The relationship follows power laws: more parameters, bigger datasets, and lower loss all demand more computational muscle. For now, throwing more FLOPs at the problem yields noticeably smarter models. Each new generation comes with a hefty electricity bill.
Supercomputers are getting more energy-efficient, doubling their efficiency every 2.3 years, which is slower than the old 1.6-year pace but still nothing to scoff at. But here's the kicker: while efficiency inches forward, demand for compute power is sprinting ahead. I don't think we've hit the thermodynamic wall yet, but it's not an endless horizon.
The link between energy and cognition isn’t just metaphorical. Claude Shannon, the father of information theory, rooted the definition of information in entropy itself. Landauer’s principle takes it further, quantifying the minimum thermodynamic cost of erasing a bit. Intelligence, at its core, is a physical process. LLMs are not just software running on chips. They are engines of entropy compression, powered by electrons. We are, quite literally, burning joules to generate probabilities over words.
A single GPT-4o exchange currently costs about 0.3 Wh, roughly equivalent to a Google search, and that number keeps dropping as Nvidia’s H-class accelerators boost performance per watt. Yet here’s the catch: aggregate usage is skyrocketing so fast that it outpaces these efficiency gains, a phenomenon reminiscent of Jevons’ Paradox, where cheaper resources just fuel greater consumption.
If that usage were confined to high-value tasks, the ROI would be clear. But increasingly, we're embedding LLMs into trivial workflows: rewriting emails, crafting marketing blurbs, suggesting answers to customer support chats. When tokens are used for the cognitive equivalent of bubble wrap, the watt-to-value ratio shrinks. Not every prompt replaces a $900/hour lawyer. Some just save you the trouble of reading an FAQ. As embedded cognition becomes ambient, energy consumption spreads like fog, harder to trace, but still mounting.
The Rising Value-Density of a Kilowatt-Hour
If one watt-hour can return a paragraph of code, a legal brief citation, or a design mock-up worth dollars, the economic value per kilowatt-hour soars to uncharted territory. A frontier-model token costs fractions of a cent to generate but may replace minutes of expensive labor. I call this the rising value-density of a kilowatt-hour.
That arbitrage sets off two feedback loops:
Capital formation: AI companies are already locking in multi-gigawatt power-purchase agreements (PPAs), which makes new solar, wind, and small-modular nuclear (SMR) projects less risky and more attractive to investors.
Grid Pricing Politics: Big industrial buyers are negotiating sweetheart deals for electricity, often below retail rates. This shifts the cost of grid upgrades onto everyday residential customers, and you can bet that's stirring up trouble. Look at Georgia and Texas, where the pushback is already heating up.
So, how big can the appetite get? Sam Altman has plans for his “Stargate” campuses, generating and slurping up 5 GW. Single-tenant loads will rival entire cities.
But it’s not all smooth sailing. Grid strain is already a headache, with places like Ireland and Northern Virginia slapping moratoriums on new data centers because the power lines can’t keep up. And it’s not just electricity. There might be other bottlenecks lurking. Copper for wiring, water for cooling, even the helium used in some advanced systems, could become scarce.
Integration, Efficiency, and Edge Innovations
You could imagine a future where grid access is the real moat. Whoever controls the electrons controls the cognition. It’s not far-fetched to think that the next great tech giants will vertically integrate their supply chains from solar farms to transformer substations. Altman’s stake in Helion, the fusion startup, is less about science fiction and more about long-term control of the means of cognition. The new aristocracy may be those who can turn sunlight and uranium into tokens with the fewest kilowatt-hours burned per bit.
In this world, cooling will be vital. Shouldn’t we all be investing in air conditioning? I’m interested in a company like Comfort Systems (ticker: FIX), which handles HVAC for some of the biggest data centers in the U.S. As AI keeps growing, I think stocks like FIX should have plenty of runway left. (Nothing here is investment advice).
But don’t count out the efficiency wizards just yet. There are some clever tricks on the horizon, like Mixture-of-Experts, which can slash the number of parameters needed per token to just 10 to 20 percent of current levels. That could mean a lot less energy per output. And let’s not forget about smarter data center designs. Microsoft released a blueprint last year aiming for zero-water evaporation in cooling systems. If these innovations pan out, they could flatten the energy demand curve even as user appetite keeps ballooning.
And yet, all the focus on frontier models misses the quiet shift already underway: inference at the edge. Apple’s GenAI strategy leans heavily on on-device processing. Qualcomm’s chips are optimized for local token generation. Not all cognition requires a data center. Some will come from your phone, your car, your toaster. The more efficient edge gets, the more it redistributes the energy burden across billions of devices, rather than concentrating it in a few hyperscale campuses. Your phone’s battery might become the new bottleneck.
One theory I have long held is that crypto is a store of value for electricity. We can capture sunlight in remote locations (e.g., the Australian Outback) and use that for crypto mining, converting otherwise lost power into something of value. An off-grid solar farm embodies what the Australian Broadcasting Corporation called “energy export without wires.” LLMs extend that logic. Instead of hashing SHA-256 for crypto, we hash the world’s knowledge graph. I can imagine a future in which all surplus joules are poured into either cryptographic scarcity or linguistic intelligence, whichever yields a better return. Picture a solar farm in the Outback mining Bitcoin to power a data center in Singapore, turning remote watts into global smarts.
Countries blessed with abundant renewables, such as Norway’s hydro, Canada’s surplus baseload, and Saudi Arabia’s solar belt, could find themselves exporting cognition the way they once exported aluminum, oil, or steel. You don’t need to ship the electrons. You just rent them to a data center and output a stream of fine-tuned legal documents or design files for export. This shifts the axis of geopolitical leverage. The oil wars of the 20th century may give way to the cognition treaties of the 21st.
Policy, Resources, and the Sustainability Imperative
Energy policy is going to be a roadblock, but I think politicians will eventually get in line. They will realize that their jobs and their countries’ economic futures depend on it. We might see a surge in advanced technologies like SMRs, geothermal plants, and high-voltage DC lines, all spurred by AI’s insatiable demand.
We might even get some creative policy moves, like location-based tariffs or demand response programs to smooth out the peaks. And I think that resource accounting will get more complicated than just tracking CO2. Water, copper, and helium scarcity will all be on the radar. If we don’t get our act together, we will end up outsourcing the problem to places with dirty, coal-heavy grids. The job of “sustainability consultant” will be a critical career path of the future, helping companies to navigate their needs and obligations.
The 20th century treated power as an enabler for steel, fertilizer, and aluminum. I think the 21st century will treat it as the feedstock for cognition itself. Whoever masters the art of turning sunlight and uranium into useful tokens at the lowest marginal joule will shape the economics and geopolitics of the coming decades.
Conclusion: Power as Feedstock for Cognition
For investors, that means underwriting clean power generation is no longer some type of ESG altruism. It is a direct lever on the supply curve of intelligence and, consequently, global growth. For policymakers, grid bottlenecks are no longer a tech-sector niche. They are a national competitiveness choke point. For thinkers and writers, it means that the frontier of ideas is tethered to the physics of power in a way that it has never been before. The (electric) meter is running. The only question is how fast we choose to spin it, and what we mint from the watts we burn.