The constraint on computing went back to being electricity
For twenty years the limit on computing was how small you could make a transistor. That stopped being true around 2005, and in 2026 the constraint has moved somewhere much less tractable than a fab: the grid.
A 37 percent jump
Google reported a record 37 percent rise in electricity use driven by its data centres, a figure that landed in July alongside a broader month of very large infrastructure commitments across the industry. Percentages of an already enormous base are the kind of number that stops being an operational detail and becomes a planning problem for somebody else, namely whoever runs the regional grid.
The same month, General Fusion became the first fusion energy company to list on Nasdaq. Those two facts belong in the same paragraph. When the demand curve for electricity bends this hard, capital starts taking long-dated bets on generation that would have looked eccentric a few years earlier.
Power was the original constraint
It is worth remembering that computing started out as an electrical problem. ENIAC drew on the order of 150 kilowatts, and the enduring local legend that switching it on dimmed the lights in west Philadelphia survives precisely because it felt true to anyone who saw the machine. Early computing was rationed by what a building could supply and, just as importantly, by what it could cool.
The transistor and then the integrated circuit made that constraint disappear so thoroughly that a whole generation of engineers never thought about it. For roughly forty years, each new process node delivered more transistors that were also faster and used less power each. Performance arrived as a free gift on a predictable schedule, and the only real question was how quickly the fab could deliver.
2005: the free gift stopped arriving
The mechanism behind that gift had a name. Dennard scaling described how shrinking a transistor let you lower its voltage in proportion, so power density stayed constant as density rose. It broke down in the mid-2000s, when voltages could no longer fall without leakage currents becoming unmanageable. Density kept improving; the free reduction in power per transistor did not.
The visible consequence is one most people noticed without knowing why. Clock speeds stalled around 3 to 4 gigahertz and have barely moved since, and the industry pivoted to multiplying cores instead. That pivot pushed an enormous burden onto software, since a second core only helps if the program can be split across it. The power wall was a hardware limit that quietly became a software problem.
The constraint moved from the chip to the grid
What has changed in 2026 is the scale at which the same limit binds. It is no longer about watts per square millimetre of silicon; it is about megawatts per site, and whether a utility can deliver them within the decade. Interconnection queues, transmission capacity, and siting near existing generation now shape where computing physically happens, in a way that transistor density has not for a long time.
The pattern is familiar to anyone who has watched a constraint migrate rather than disappear. Every era of computing has had one binding limit: tubes and cooling, then memory, then transistor density, then power density on the die. Solving one has never produced an unconstrained system, only a system whose limit sits somewhere new. This one happens to be measured in megawatts and negotiated with utilities, which makes it slower to relieve than anything a fab could fix.
Frequently asked questions
What is Dennard scaling and why did it matter?
Formulated in 1974, it described how shrinking a transistor allowed a proportional reduction in operating voltage, so that power density stayed roughly constant even as transistors got smaller and more numerous. The practical effect was that each new process generation delivered more transistors that were individually faster and less power-hungry. When it broke down in the mid-2000s, density kept improving but the power savings did not, which is why processors gained cores instead of clock speed.
Is this the same thing as Moore's law ending?
No, though the two are frequently conflated. Moore's law is an observation about transistor density, which has continued to improve, if more slowly and expensively. Dennard scaling was about power, and it ended roughly two decades earlier. The distinction matters because it explains why chips kept getting denser while single-thread performance flattened: the transistors kept arriving, but the free power budget to run them all at higher speed did not.
Compare Kirality to the alternatives, see pricing, or browse the AI glossary.