Everyone is screaming about the grid.
Open any publication right now, and you will find hand-wringing over data centers draining local substations, burning fossil fuels, and plunging cities into darkness just so a chatbot can write a poem about a toaster. The narrative treats energy as a fixed pie. Take a slice for machine learning, and grandmother loses her heating.
This panic is completely backwards.
I have spent the last five years watching corporations misallocate capital because they treat power constraints as a permanent roadblock instead of an aggressive pricing mechanism. The crisis is not that computation requires too much energy. The crisis is that energy markets are ancient, sluggish monopolies terrified of dynamic demand.
Let us dismantle the lazy consensus.
The Grid Fallacy
The mainstream argument goes like this: artificial intelligence demands astronomical compute. Compute requires electricity. Therefore, modern algorithms are stealing watts from residential neighborhoods and accelerating climate collapse.
This logic treats the electrical grid like a static pool of water. It assumes supply cannot scale because transmission lines are expensive and permitting takes a decade.
That view ignores market realities.
Compute is mobile. Heavy industry is not. If a steel mill wants to move, it requires billions in fixed infrastructure, rail links, and specialized labor pools. If a training cluster needs a gigawatt, you can pack the servers into shipping containers and drop them next to an unexploited geothermal vent in Iceland, a flare gas wellhead in Texas, or a stranded hydro plant in rural Washington.
Companies do not build massive data centers in dense urban cores where electricity is scarce. They build them where power is stranded, cheap, and wasted.
The Economics of Megawatts
Let us look at how power is actually priced.
Utilities operate on regulated, cost-plus models designed for the 20th century. They build capacity based on decade-old demand forecasts, pass the cost to ratepayers, and guarantee a profit to shareholders. When a massive load arrives, the system groans because it cannot pivot.
Advanced computing changes the math entirely.
Tech giants are not waiting in line for local utilities to upgrade transformers. They are writing direct checks for dedicated generation. They are funding next-generation nuclear small modular reactors, geothermal exploratory drilling, and dedicated solar-plus-storage arrays.
I have seen firms commit hundreds of millions of dollars to merchant power purchase agreements that subsidize brand-new clean energy infrastructure. Without this surge in demand, those capital projects would never break ground. The influx of tech capital is funding the green transition at a scale governments abandoned years ago.
When you hear that a single training run consumes as much electricity as a small town, ask yourself a better question: how many clean energy projects did that training run subsidize into existence?
The Flexibility Paradox
Critics love to paint machine learning workloads as insatiable energy hogs that run flat-out, 24 hours a day, regardless of grid conditions.
This is technically false.
Inference—running models to answer user queries—requires low latency and steady power. But training—the massive, multi-week computational grinds that create the models in the first place—is remarkably elastic.
Training runs can be paused. They can be throttled. They can be migrated across continents in seconds depending on where electricity is cheapest and cleanest at that exact hour.
Imagine a scenario where a cluster in Ohio automatically ramps down its power consumption by eighty percent during a summer heatwave, selling its contracted power back to the local grid at a massive profit, and resumes heavy compute at three in the morning when wind generation peaks and demand flatlines.
That is not a strain on the grid. That is a shock absorber.
Modern algorithms act as dynamic load balancers for intermittent renewable energy. They provide a predictable, massive baseload buyer for solar and wind farms that would otherwise have to curtail production when the wind blows too hard or the sun shines too bright.
The Real Bottleneck
If energy is not the ultimate limit, what is?
It is not electrons. It is copper, transformers, and regulatory inertia.
Try buying a high-voltage step-down transformer today. You will find lead times stretching past three years. The manufacturing base for heavy electrical equipment withered away during decades of low industrial growth.
This is where the political narrative fails hardest. Politicians love to blame tech companies for consuming power, while their own zoning laws, environmental review boards, and protectionist tariffs block the construction of new substations, transmission corridors, and manufacturing plants.
The restriction is entirely man-made.
When a region blocks a data center because of grid fears, they are usually protecting inefficient local monopolies from competition while starving their own tax base of the revenue needed to modernize infrastructure.
What Actually Happens Next
The tech industry will not stop scaling because of a lack of power. Instead, the relationship between computation and energy will invert.
Energy companies will become technology companies, and technology companies will become energy producers. We are already watching hyperscalers buy equity stakes in nuclear startups and secure exclusive rights to geothermal fields.
The companies that win the next decade will not be the ones with the cleverest algorithms. They will be the ones that master vertical integration of energy generation.
Stop worrying about whether your laptop is draining the local dam. Start worrying about why our energy grids are regulated like municipal water systems in an era that demands the velocity of software.
The electrons are there. The capital is waiting. Only the bureaucracy stands in the way.
