⚡ For years, engineers ran into the same invisible wall. You could make a switch faster, but only by feeding it more power, which meant more heat, which meant the device cooked itself to death. The floor sat stubbornly at the nanosecond: one billionth of a second. A team in Tokyo just went a thousand times faster, to 40 picoseconds, and did it while staying almost cold. Here is how that wall finally fell.
The wall that everyone hit
Every computer is, at heart, a vast crowd of tiny switches flipping between 0 and 1. The faster they flip, the more a chip can do. So for half a century the story of computing has been one of making switches smaller and quicker.
But somewhere in the nanosecond range, the speed gauge stopped moving. A nanosecond is a billionth of a second, already absurdly fast, yet that became the practical floor for the switches inside today's CPUs and GPUs. Push for more speed and the power bill, and the heat, climbed faster than the performance.
That ceiling matters far beyond the lab right now, because the world is trying to run an unprecedented amount of computation. The International Energy Agency estimates that the electricity used by data centers could more than double to around 945 terawatt-hours by 2030, just under 3% of all the electricity humanity generates. A huge slice of that energy never becomes "computing" at all. It leaks out as heat. Anything that lets a chip do the same work while wasting less of it as warmth is, quietly, one of the most valuable things an engineer can build.
Source: JST / The University of Tokyo press release
Why nobody could break it
To understand the wall, picture the most common way chips store a bit magnetically. You have a tiny magnet, and "north-up" means 1 while "north-down" means 0. To write data, you flip the magnet.
The trouble is the flipping. In conventional magnetic memory, the fastest known ways to force that flip in picoseconds essentially work by brute force, slamming the material with a current so intense that it heats up by hundreds of degrees in an instant. The magnet does flip. But a device that endures a furnace blast on every single write does not last. Researchers had demonstrated picosecond flips before; what they could not demonstrate was a picosecond flip that the chip would survive in the long run.
So the field was stuck with an ugly trade-off. You could have speed, or you could have endurance and low power, but the laws of physics seemed to forbid all three at once. Nanosecond switching was the compromise everyone learned to live with.
The detour through an unlikely material
The breakthrough came from refusing to play that game, and from a material most people have never heard of.
A research group led by Professor Satoru Nakatsuji at the University of Tokyo, working with RIKEN, Osaka University and Japan's science agency JST, has spent more than a decade on a class of materials called antiferromagnets. In an ordinary magnet (a ferromagnet), all the atomic spins point the same way, which is what makes a fridge magnet stick. In an antiferromagnet, neighboring spins point in opposite directions and cancel out, so the material shows almost no external magnetism at all. For storing data that sounds useless, until you learn that the internal spin dance of an antiferromagnet is naturally far faster than a ferromagnet's.
The specific material is Mn3Sn, a compound of manganese and tin in which the spins sit on a triangular, honeycomb-like lattice. Nakatsuji's team first made it famous back in 2015, when they discovered it behaves in strange and useful ways despite having no net magnetism. The question that lingered for years was whether you could actually flip its state quickly without melting it with heat, and if so, how the flip really happened. A companion study led by Ryo Shimano at the University of Tokyo, with Nakatsuji among the co-authors, appeared in Nature Materials in December 2025. Using current pulses as short as 140 picoseconds, it managed to watch the flip in real time and resolved two distinct routes: a temperature-assisted one that briefly melts the magnetic order, and a faster "non-thermal" one that does not.
The new work is the payoff: turning that cool, non-thermal route into a working device.
40 picoseconds, and almost no heat
Here is the clever part. The team paired their thin film of Mn3Sn with a layer of the heavy metal tantalum. When a current runs through the tantalum, it gets "sorted" by a quantum effect so that electron spins pile up at the boundary and push on the magnet, a phenomenon called spin-orbit torque. Think of it less like heating a kettle and more like a precise nudge that rotates the spins directly.
Because the push comes from transferring angular momentum rather than from heat, the device barely warms up. That single difference cascades into everything that matters: the team recorded switching with electrical pulses as short as 40 picoseconds, about a thousand times faster than the nanosecond floor. Each bit cost roughly one femtojoule, and the power density came in at a tenth or less of existing picosecond switching devices. And crucially, because it isn't being baked, it endures: more than a hundred billion switching cycles (over 10^11) without degrading.
The result was published online in the journal Science on May 14, 2026. In the framing of the announcement, it is picosecond switching that finally combines a sharp reduction in heating with high endurance.
When light writes directly to memory
Speed and endurance would have been a strong result on their own. But the team added a twist that points at where this is really headed.
Instead of an electrical pulse, they fed the device a pulse of light: a laser at 1.55 micrometers, the same band that carries data through the fiber-optic backbone of the internet. A photoelectric converter turned it into a 60-picosecond burst of current, which flipped the magnet, 250 times in a row without an error. In other words, an optical signal wrote a bit into non-volatile memory more or less directly.
That matters because of a hidden tax inside every data center. Information arrives as light through optical fiber, but computers think in electricity, so there is a constant, energy-hungry conversion happening at the boundary, the "I/O" between optics and electronics. A device that turns light straight into a stored magnetic bit could shrink that tax. The researchers call this "spintronics photoelectric conversion," and it is the part of the work with the clearest line to the data-center energy problem.
A note of honesty is in order, though, and the researchers themselves are careful about it: a faster switch is not the same as a faster computer. A computer is a whole orchestra of components, and speeding up one instrument doesn't speed up the symphony by the same factor. This is a proof of concept at the device level, not a chip you can buy. The road from a lab demonstration to a product is long and littered with failed candidates.
A different bet than Silicon Valley's quantum race
It is tempting to file this under "quantum computing," partly because the official announcement calls it a "quantum switching device." But that label refers to the quantum-mechanical behavior of electron spins doing the flipping, not to qubits or the kind of machine IBM, Google and Intel are racing to build. This is hardware for the ordinary, classical computing we already use, made dramatically more efficient.
That contrast is itself the interesting bit. The American giants are largely betting on a future, exotic kind of computer, the quantum machine, that may or may not arrive on schedule. Japan's spintronics community is making a quieter, more pragmatic bet: take the computers the world already runs on, and attack the thing that's actually choking them right now, which is energy. It is less glamorous than qubits, but the demand is immediate, and Japan has a genuine, decades-deep lead in the underlying physics. The same lab's 2015 discovery of how Mn3Sn behaves is now, eleven years later, turning into a device. That is what a long-term basic-science bet looks like when it pays off.
Whether it scales into real chips is the open question, and a fair one. But for the first time, the nanosecond wall has a thousand-fold hole in it.
In your country, the conversation about AI is usually about what it can do. How often does anyone ask what it costs to keep the lights on, and who should be paying to make that cheaper?
Global Discussion
5 comments