What Nvidia did
For thirty years, the unit of computing was the chip. Nvidia’s quiet structural achievement of the last three years was to move that unit up a level: the rack became the computer. Seventy-two accelerators, wired together by a switch fabric fast enough that software treats them as one enormous processor, cooled by liquid because air stopped working. That change was a performance decision. Its consequences were an energy decision.
Figure 1.1 — The rack stopped being a shelf and became a furnace
Read the figure as a slope, not as five numbers. A pre-AI enterprise rack drew five to ten kilowatts, which is a large domestic oven. The Hopper generation reached roughly forty. Blackwell’s NVL72 landed at one hundred and twenty to one hundred and forty, and reporting on the 2026 Vera Rubin generation puts a fully configured rack between one hundred and ninety and two hundred and thirty kilowatts depending on how it is specified.1 The next rack architecture is quoted at six hundred, with megawatt-class densities named beyond it. Two things break at that slope, and both matter later in this book.
The first is cooling. Air carries roughly three thousand times less heat per unit volume than water, which is why liquid cooling stopped being an exotic option and became the only option somewhere between forty and one hundred and twenty kilowatts per rack.2 An enormous amount of the world’s existing data centre floorspace cannot host this hardware at all without being rebuilt. The second is that essentially all of that power leaves the building as heat. This is not an engineering failure; it is thermodynamics doing its job. A data centre is best understood as a machine that converts electricity into two products: tokens, and warm air. The tokens are the ones you sell.
What Google did
If Nvidia raised the numerator, Google spent fifteen years shrinking the denominator. The industry measures this with PUE, power usage effectiveness, the ratio of total facility power to the power that actually reaches the computers. A PUE of 2.0, common in the 2000s, means every watt of computing carried a second watt of chillers, pumps, transformers and lighting. Google drove its fleet toward roughly 1.1, and in doing so taught the entire industry how: hot-aisle containment, custom power delivery, machine-learning control of the cooling plant, and siting facilities where the power and the climate are favourable rather than where the real estate is convenient. Google also did something less discussed. It designed its own accelerator, the TPU, and it designed it at rack scale before rack scale was fashionable. That matters to us in Chapter 12, because a company that designs its own silicon can also decide to design silicon that survives orbit and it has. But notice what fifteen years of efficiency work implies for the future. If your PUE is already 1.1, perfecting your cooling entirely buys you nine percent. The overhead has been squeezed out. What remains is the load itself, and the load is growing at the slope in Figure 1.1. Efficiency is no longer where the answer is.
The wall
So the industry turned to the grid, and found a queue. The specifics are worth carrying with you, because the whole orbital argument rests on them being true and durable: Projects reaching commercial operation in the PJM territory in 2025 had spent an average of about eight years in the interconnection queue.3 New high-capacity connections in the dense hubs, Northern Virginia, Phoenix, Dallas are quoted at four to seven years, and EPRI’s 2026 analysis warns that a data centre relying only on the grid could face up to ten years to energisation in some regions.4 Large power transformers, which every new substation needs, run two to four years of lead time.5 EPRI’s same analysis puts US data centres at nine to seventeen percent of national electricity consumption by 2030.4 Notice the shape of this problem. It is not that electricity is expensive. In much of the United States, industrial power is cheap by world standards. It is that electricity is not available at a specific place on a specific date, and no amount of money compresses a transformer factory’s order book or a transmission permitting process into the eighteen months an AI roadmap runs on. This is the sentence to keep: The binding constraint on artificial intelligence is no longer silicon, and it is no longer the price of power. It is the queue for power.
An engineer reading that should immediately ask the right follow-up question, which is not “how do we build more grid?”, many capable people are on that but rather: is there anywhere that a gigawatt of primary energy is already sitting, unqueued, unpermitted, and unowned? There is exactly one such place, and we have been able to reach it since 1957.