A University of Tokyo-led team has shown a magnetic switching device that works in 40 picoseconds with greatly reduced heat, but this is not a finished 1,000x faster computer.
The most interesting part of the new Japanese chip research is not the headline claim that computers are suddenly about to become 1,000 times faster. They are not. The real story is more practical, and for AI infrastructure, more important: researchers have demonstrated a way to flip a binary magnetic state at picosecond speed without leaning on the heat-heavy behavior that makes today's faster chips so difficult to scale.
According to the University of Tokyo's Institute for Solid State Physics, the work was published online in Science on May 15, 2026, under the title Picosecond ultralow-power switching device based on an antiferromagnet. The team includes Hanshen Tsai, Takuya Matsuda, Kouta Kondou and Satoru Nakatsuji, with researchers connected to the University of Tokyo, RIKEN, Osaka University and other institutions. The material at the center is Mn3Sn, a manganese-tin antiferromagnet that can hold and switch a binary magnetic state.
The device was switched using a 40-picosecond electrical pulse. A picosecond is one trillionth of a second, so this sits in a very different speed class from the nanosecond-scale switching usually discussed around conventional CPU and GPU operations. That is where the 1,000x language comes from. It is a comparison between nanosecond and picosecond switching, not proof that a server processor can now run 1,000 times faster in a data center rack.
This distinction matters because AI hardware is already full of optimistic language. Faster switching is valuable, but a computer is not only a switch. It is memory, logic, interconnects, packaging, software, thermal design and manufacturing yield all working together. A single device that flips quickly does not automatically become an Nvidia-class accelerator, a replacement for HBM, or a drop-in CPU upgrade.
Still, the result deserves attention. The University of Tokyo release says existing CPU and GPU technologies face sharp increases in energy use as operating speed rises, making sub-nanosecond operation difficult. Earlier routes toward picosecond switching have faced durability problems because heating can reach hundreds of degrees. The new antiferromagnetic approach uses spin-orbit torque based on angular momentum transfer, which the researchers say allows fast operation with greatly reduced heat generation.
That is the useful point for founders, chip investors and data center operators. The fight is no longer only about peak performance. It is about performance per watt. AI demand has made electricity a strategic bottleneck, not a back-office utility bill. The International Energy Agency has projected that global data center electricity consumption could more than double to around 945 terawatt-hours by 2030 in its base case, driven in large part by AI workloads. Anything that reduces wasted heat while preserving speed earns a serious look.
The team also showed switching using a 60-picosecond photocurrent pulse generated by combining a telecommunications-wavelength laser with a photoelectric conversion element. That is an early demonstration of an opto-spintronics path: light is converted to an electrical signal and then connected directly to nonvolatile magnetic writing. If this can be developed further, it could matter for systems where data movement is as expensive as computation itself.
Why Nonvolatile Switching Matters
Today's AI systems burn energy not only when they calculate, but also when they move and refresh data. Nonvolatile devices hold information without constant power. That is why magnetic memory and related spintronic ideas keep returning in conversations about future compute. If a switch can be fast, low-power and stable, it could help shrink the gap between memory and logic, or at least reduce the penalty of shuttling data across a chip.
The durability claim is another reason this work stands out. The University of Tokyo release says the device combines picosecond switching with high endurance because it does not depend on the severe temperature rise seen in earlier approaches. That does not settle commercialization, but it answers one of the first questions engineers ask about any new memory or switching device: does the mechanism look stable enough to be worth scaling?
The harder questions now move from physics to engineering. Can Mn3Sn-based devices be fabricated reliably at chip scale? Can they be integrated with CMOS processes without adding too much cost or complexity? Can the switching behavior remain consistent across millions or billions of devices on a wafer? Can this help logic, memory, or only a narrower class of storage and signal-conversion tasks?
Those questions are not footnotes. They are the difference between a Science paper and a product roadmap. The release itself frames picosecond switching as still in the research and development stage, which is the right level of caution. The market should treat this as a serious materials and device breakthrough, not as an imminent upgrade cycle.
For AI companies, the lesson is straightforward. The power wall will not be solved by one invention. It will be attacked from every layer: better GPUs, custom accelerators, advanced packaging, optical interconnects, memory closer to compute and new switching devices that waste less energy as heat. The University of Tokyo-led work belongs in that stack of possibilities. If it scales, it could become part of the answer to a very expensive problem. If it does not, it still points in the right direction: the next era of compute will be judged as much by what it does not waste as by how fast it runs.
Also read: AI hacking is turning DeFi security into a balance sheet risk • Polymarket Is Turning Insider Trading Into an AI Compliance Test • The UK proved it can walk away from Palantir and save millions
