Two camps, both risking serious money, both acting rationally and on good information, say the exact opposite at the same moment, and both sound entirely certain. In The Beginning of the End for Big AI<\/a>, Nicolas Colin argues, and he speaks for a lot of other operators, that now is the time to walk away from big artificial intelligence. The central labs burn billions, charge prices for compute that almost no customer can reconcile, and the smart move, he says, is to bring a good-enough model onto your own machines and keep the data, the weights and the control in house. In the same stretch of weeks, as Colin traces elsewhere<\/a>, Alphabet raises fresh equity for the first time since 2005, about 85 billion dollars, the largest raise of its kind on record, and a space company lists at 1.75 trillion. The filings call it infrastructure, to be built at almost any price.<\/p>\n\n\n\n Both camps are right about their numbers. But only one of them can be right about its bet.<\/p>\n\n\n\n And that is where the error starts. The apparent certainty in this market is not insight. It is a bet that follows inevitably from where the person already stands.<\/p>\n\n\n\n The argument filling every conference and every investor letter has a simple shape. Does compute stay central, or does it move to the edge. On one side the large cloud companies with their frontier models: agentic systems that run around the clock without human pace become the largest class of compute there is, and the value pools where that load is processed, in the big data center. On the other side many small providers and models running on local hardware: the moment such a model solves ninety-five percent of the problem at a fraction of the cost, the work moves to the edge, and the expensively planned central capacity sits half empty.<\/p>\n\n\n\n I take both sides seriously. The central bet is backed by the sheer demand for compute, the decentral one by cost, by data protection, and by the wish not to depend on a handful of providers. You can build a closed, intelligent case for either. Central structures carry one quiet extra advantage that neither camp says out loud: they are easier to steer, to control and, for states, easier to regulate. I will pick that thread up further down.<\/p>\n\n\n\n Because both sides are arguing about the same thing without noticing. They argue about where intelligence is produced. Not whether.<\/p>\n\n\n\n Nobody talks about the one quantity that makes the whole fight small: the amount. For the first time in history, human cognition can be manufactured. Not represented, not simulated, but produced, on demand, around the clock, in any place with power and a network. A question that once cost an analyst a full day is answered in seconds, in parallel, as often as you like.<\/p>\n\n\n\n The human brain is a remarkable machine with a set of inconvenient limits. It sleeps a third of the day. It ages. It sits in exactly one place. It takes twenty years to train and cannot be copied. Manufactured cognition knows none of these limits. It runs through the night, it is available in Tokyo and Toronto at the same time, and the next unit is one compute job away, not one generation.<\/p>\n\n\n\n And because it gets cheaper, it is not used less, it is used more. A mid-sized company that used to pay for a thorough review of its contracts once a year will run it every night. Every drop in price opens a new layer of uses that nobody asked for before. The AI Corner does the math in Y<\/a>ou Are Overpaying for Intelligence<\/a>: a new frontier-class model at roughly a tenth of the previous cost per completed task. So the models themselves turn cheap and interchangeable, the chips spread out, and the place of production swings back and forth, as it has three times already in the short history of computing, from the mainframe to the personal computer, into the cloud, and now a step back toward the edge.<\/p>\n\n\n\n Neither of those is the decisive point. The place of production is open. Only one condition is not negotiable: energy.<\/p>\n\n\n\n A look back is worth it here, because the pattern is old. Human progress has always been the opening of a new form of energy for a new capability. The steam engine opened production. The combustion engine opened mobility. Kerosene opened the sky. Each rung made more energy usable, and with each rung a society’s energy consumption rose. How much energy a society turns over was never a side effect of where it stood. It was the measure of it.<\/p>\n\n\n\n And each of these rungs did more than open a capability. It redistributed power. Whoever held the coal and the rail or shipping line held the factory and the trade. Whoever controlled the oil set a century of geopolitics. The form of energy never decides only what becomes possible, it always decides who owns it. That is the part people miss at the start of each rung and learn the expensive way at the end.<\/p>\n\n\n\n Thinking has already taken the first step on this ladder. With computers and digitization we began, some thirty years ago, to power the first, mechanical rung of thinking: to calculate, to sort, to combine along fixed rules. That was machine thinking, but a narrow, rigid kind. It followed rules that humans had written first.<\/p>\n\n\n\n What comes now is the next wave on the same ladder. No longer only calculation along rules, but something that looks like reasoning, like language, like connecting things nobody had connected before. Electricity no longer powers only the machine that counts. It powers the machine that concludes. The old rung could sum a spreadsheet in milliseconds but could not say which line in it looks suspicious. The new one says it. This is not bigger software. This is the next rung on which we turn energy into capability, and this time the capability is thinking itself.<\/p>\n\n\n\n If that holds, then winning takes more than access to the right models and the right chips. It takes, above all, a robust and cheaply scaling energy infrastructure that delivers power to where machine intelligence is needed, in hard physical plant and a meter on the wall.<\/p>\n\n\n\n Because power is the one thing that grows one to one with manufactured cognition. A better model saves compute steps, a worse price shifts demand, the place moves. The kilowatt-hour stays. It is physical, it is scarce, and it cannot be produced by a clever argument.<\/p>\n\n\n\n How serious this is shows up on the electricity bill, and not the data center’s, but the public’s. In the largest US power market, the wholesale price has risen by about three quarters within a year, from roughly 78 to 137 dollars per megawatt-hour, and the price for firm capacity, the guarantee of delivering on peak days, jumped more than 800 percent in a single auction, with data centers as the main driver, as Tom’s Hardware<\/a> reports, citing the grid regulator. A meaningful share of that lands on ordinary households’ bills. Peter Perri describes in Regulate This<\/a> how one company puts a hundred thousand graphics processors on the grid in nineteen days, on mobile turbines under an exemption, while a regular power plant takes four years to build. And one large investor, as Perri traces elsewhere<\/a>, sets up a dedicated vehicle just to buy land next to power plants, a 430-megawatt campus, a large lab as anchor tenant.<\/p>\n\n\n\n While everyone stares at the models, the binding constraint is the turbine and the permit. The scarce factor is the kilowatt-hour, not the model.<\/p>\n\n\n\n Here the earlier thread returns. Energy has a property the model does not: it can be concentrated, owned and regulated. A power plant stands in one place, a grid has an operator, a permit has a sender. Once thinking is energy-driven, the question of who owns the energy takes on new weight. Whoever holds the source holds access to the thinking of the next rung.<\/p>\n\n\n\n But the control has a gap, and it sits in the model itself. A model can only be governed if you run and sell it as a service, centrally, in one place, under one hand. Open models with freely available weights slip that grip. They can be copied across any border and run in any number of places, switched off by no one. So whoever wants to control the thinking cannot pin it to the model, the model is nothing without chips and power. They have to pin it to the one thing that cannot be copied, to the energy. That is exactly why the reach for control moves from the model to the chip<\/a> and to the power plant.<\/p>\n\n\n\n You can already see it. When a facility’s power for its data centers was challenged in the US, the government stepped in and declared the electricity supply for that compute a matter of national defense, as Perri describes in Regulate This<\/a>. Once the power for thinking becomes a defense question, control over the source is long since a matter of state.<\/p>\n\n\n\nThe fight over where<\/h2>\n\n\n\n
The how much<\/h2>\n\n\n\n
The energy ladder<\/h2>\n\n\n\n
Power infrastructure as the backbone of sovereignty and prosperity<\/h2>\n\n\n\n
Energy stays power<\/h2>\n\n\n\n