How far can computing go? Fundamental physical limits versus the scale of data centers

The article, published by MIT Technology Review in its “Explains” series, argues that today's computing progress comes mainly from building more (more chips, bigger data centers), while the other kind of progress, making chips better, is running out. Note: the text received is truncated at the end (the last sentence is cut off), so this summary covers only what appears in the downloaded content.

The text opens with Richard Feynman's 1959 lecture, “There's Plenty of Room at the Bottom,” in which he argued that no physical law prevents writing the Encyclopaedia Britannica on the head of a pin or arranging matter atom by atom. Feynman offered two $1,000 prizes. The first, for an electric motor smaller than a 0.4 mm cube, was soon won by William McLellan, who built it by hand with a toothpick and a watchmaker's lathe under a microscope. The second, for shrinking a page of text 25,000 times on each side, went unclaimed for a quarter century until Tom Newman, a Stanford student, won it in 1985 using electron-beam lithography, transcribing the first page of “A Tale of Two Cities.”

The article then traces the descent: Jack Kilby, at Texas Instruments, combined a transistor, a capacitor and three resistors on a single piece of germanium in 1958, and Gordon Moore, in 1965, extrapolated the component-doubling trend that would become Moore's law. Moore, the text stresses, was talking about the number of components that minimizes cost per component (50 in 1965), not unlimited shrinkage. For four decades this was compounded by Dennard scaling (formulated in 1974): smaller transistors switched faster and consumed proportionally less power. Around 2004 that ended: chips could no longer dissipate the heat, clock speeds stalled just above three gigahertz and Intel canceled its promised four-gigahertz processor. The industry moved to multiple cores and, later, to scaling the entire system.

According to the article, six of the largest hyperscalers are on track to spend more than $700 billion this year on compute infrastructure, roughly double the 2025 record. In AI, training compute has doubled roughly every six months since 2010, far faster than Moore's law, and little of that progress is due to miniaturization. The laws of Moore, Dennard, Koomey (computations per joule), Kryder (magnetic storage density) and Keck (optical fiber capacity) are slowing, have ended or are flattening. But the text insists these are observations about human-engineered systems, not fundamental physical laws.

The real physical limits lie far below anything in any data center. The article first measures how far we have come: in almost eight decades, transistor density has grown by 10 orders of magnitude (today more than 100 million per square millimeter), computing speed by 12 orders, and energy per operation has fallen by 17 orders, from about 100 joules in a vacuum-tube machine to one femtojoule per logic-gate switch.

Four limits mark the “bottom,” derived from three fundamental theories:

1. Computing speed (quantum mechanics). Seth Lloyd, in a 2000 Nature paper, imagined the “ultimate laptop”: one kilogram of matter in one liter. By E = mc², it contains about 9×10^16 joules, and spending it all on computation would yield about 5×10^50 operations per second, some 35 orders of magnitude more than a current leading-edge chip. It would be a plasma at about a billion degrees, more like the universe in the first minutes after the Big Bang than a laptop. It is the largest of the gaps and the least bridgeable: the article compares it to the difference between a nanosecond and 200 million times the age of the universe.

2. Energy per bit erasure (thermodynamics). Rolf Landauer showed in 1961 that erasing a bit releases at least about 2.9×10^-21 joules at room temperature. A current chip spends several hundred thousand times more per switch, operating less than six orders of magnitude above the Landauer minimum. Conventional logic can close only part of that gap, because a reliable switch needs a margin against thermal fluctuations that keeps it between a hundred and a few thousand times above Landauer. The alternative route is reversible computing (Charles Bennett, 1973), which discards no information and pays no erasure toll. According to the article, there are already commercial attempts: the startup Vaire Computing tested its first chip in 2025, with resonant circuits that recover part of the switching energy instead of dissipating it as heat.

3. Memory density (general relativity). Jacob Bekenstein, studying black-hole thermodynamics in 1981, determined that the information in a bounded region is limited by its size and energy. For a region like Lloyd's laptop, that would be about 10^42 bits, nearly 30 orders of magnitude more than a device of that size stores today. The most concrete example of exploiting three dimensions at the molecular scale is DNA: Feynman conservatively estimated 125 atoms per bit, and DNA stores it in about 30. What keeps it out of data centers is not density but cost and write speed, still many orders of magnitude from practical.

4. Communication (information theory). Alexander Holevo set in 1973 the limit on the information a stream of quanta can carry. A one-watt channel at the Holevo limit could deliver on the order of 10^17 bits per second; today's optical links deliver roughly a millionth of that at the same power. It is about six orders of magnitude from the limit (a factor of a million) and is the limit the industry will likely reach first. The text notes that Nvidia and Broadcom already sell optical interconnects integrated alongside the silicon, so the signal leaves the package already as photons, and that quantum computers are beginning to use single-photon interconnects: both worlds would converge in the Holevo regime.

As for quantum computing, the article notes it could approach the speed limit if it managed to turn every degree of freedom in a piece of matter into a reliable computing element, but today the opposite is true: trillions of atoms are spent per qubit, and it takes tens to thousands of fragile qubits to form one reliable enough.

The text acknowledges the obvious objection: a gap of 35 orders of magnitude may be a mirage, like sailing toward the horizon. It concedes that reversible logic, quantum computing and DNA storage have been “ten years away” for 40 years, and that it will probably never be practical to build a computer that uses all the energy in a piece of matter. Even so, it argues that these gaps are “invitations.” There is a ceiling, set by what can be built (manufacturable devices, available energy, capital, land and cooling water), and a floor, set by what a joule, a second and a gram of matter can compute, store or transmit. Over the past century, progress came along both paths at once; now it comes mostly from building more, and the ceiling is coming into view: a power grid can carry only so many gigawatts and a region can give up only so much land and water. The article breaks off here, at the start of its final sentence.

Core idea: approaching physical limits pays off twice, as technological development and as fundamental science, because a prototype system that nears a limit forces stagnant paradigms to give way and at the same time serves as an experiment on the physical laws themselves.