Computing limits: why transistor shrinking is no longer enough

Computing limits come in two very different kinds, and mixing them up causes most of the confusion. The first kind are trend lines that engineers noticed and then worked to keep alive. The second kind are hard limits set by the laws of physics, and those sit a very long way below anything running today.

The trend lines are the ones you may have heard of. Moore’s Law tracks how densely transistors, the tiny on-off switches inside a chip, can be packed onto silicon. Dennard scaling described how a patch of chip kept using the same power even as it filled with smaller transistors. Koomey’s Law follows how much computing you get from each unit of energy. Kryder’s Law covers how tightly data can be squeezed onto magnetic disks, and Keck’s Law tracks how much information a single optical fiber can carry.

Every one of them has slowed, flattened or ended. Moore’s Law has slowed, Dennard scaling is over, the energy curve has begun to flag, disk density has slowed to a crawl and fiber capacity is leveling off. But none of these was ever a law of nature. They described how well people were doing at engineering, not what the universe allows.

The story starts with a challenge. In late 1959, the physicist Richard Feynman gave a famous talk arguing there was plenty of room at the bottom, meaning nothing in physics stopped us from building machines and storing information at almost atomic scale. Living cells already did it with DNA, so it was clearly allowed. We just lacked the tools.

He even put up two $1,000 prizes of his own money. One was for a working electric motor smaller than a 0.4-millimeter cube. An engineer named William McLellan claimed it within about a year, building the motor by hand under a microscope with a toothpick and a watchmaker’s lathe. The second prize, for shrinking text by a factor of 25,000 in each direction, took until 1985, when Stanford student Tom Newman wrote out the first page of A Tale of Two Cities with an electron-beam machine.

Meanwhile, Jack Kilby at Texas Instruments showed around the same time that a whole circuit could be carved from one piece of semiconductor. Component counts then doubled so regularly that Gordon Moore drew a line through the trend in 1965. Moore never said shrinking would last forever. He pointed out that there is a sweet spot, around 50 components per chip back then, where each part is cheapest to make. Too many, and manufacturing faults wreck your yields.

For decades, smaller transistors were also faster and used less power, so each new generation was an automatic win. That bonus ended about two decades ago, when chips could no longer get rid of the heat from running every transistor at full speed. Clock speeds stopped rising and the industry switched to putting several processor cores on each chip instead.

Today, progress is mostly bought by scaling up whole systems: more chips wired together, more memory to feed them, and more power and cooling around them. That is the brute-force path behind the $700 billion buildout.

AI shows it best. The amount of computing used to train leading AI models has grown far faster than Moore’s Law could deliver, and very little of that growth came from smaller transistors. Chips still improve, but each generation arrives more slowly and gives back less speed and power saving than the one before.

The catch is that this path has a ceiling too. A grid can carry only so much power, and a region can spare only so much land and water for cooling. Building more cannot go on forever.

For everyday people, the shift explains a few things you may already have noticed. New phones and laptops tend to improve in smaller steps than they did a generation ago. A lot of the impressive new features, especially AI ones, run in huge remote data centers rather than on the device in your hand.

It also means the cost of computing is increasingly tied to the cost of electricity, land and cooling. When companies race to build bigger facilities, those costs show up somewhere, whether in subscription prices or in debates about local power supplies. The explainer does not put a figure on what that means for consumer prices, so treat any confident forecast with care.

Here is the twist: compared with what physics allows, today’s computers are barely getting started. The author, Volkan Gurses, a postdoctoral researcher at MIT who also works with Nvidia, walks through four true limits drawn from quantum mechanics, general relativity and information theory. They cap how fast a machine can compute, how much it can store, how fast it can send data and how little energy a step can waste.

Energy is where we are closest. Chips today still waste far more energy per switch than physics demands, and one idea for closing that gap is reversible logic, where a calculation is arranged so nothing is thrown away and energy can be recovered. A startup called Vaire Computing tested its first chip using this approach in 2025. Data links are also close to their limit, and Nvidia and Broadcom already ship optical connections built right next to the chip so signals leave as light.

The other gaps are enormous. DNA could store data far more densely than any drive, but writing it is still far too slow and expensive. Quantum computers could, in theory, help with raw speed, though today they use vast numbers of atoms for every working unit. The author admits ideas like these have seemed a decade away for a very long time.