A team of researchers at the University of Tokyo has demonstrated a switching device that could make processors run up to 1,000 times faster without generating additional waste heat. Published in the journal Science in May 2026, the study describes a non-volatile switching element built from two materials: manganese-tin (Mn3Sn) and tantalum. The implications, if the technology can be commercialized, range from laptop battery life to the global energy footprint of AI data centers.
The Problem: Heat Is the Enemy of Speed
To understand why this matters, you need to understand the central constraint of modern computing: heat. Every time a transistor switches state from 0 to 1, or 1 to 0, it generates heat. The faster you switch, the more heat you generate. This is not a design flaw. It is physics.
For decades, chip designers have pushed switching speeds higher by shrinking transistors and packing more of them onto silicon. But as speeds have increased, thermal management has become one of the dominant costs in terms of both engineering complexity and energy consumption. The fans spinning inside your laptop, the elaborate liquid cooling loops in gaming PCs, the massive chilled-water systems underneath hyperscale data centers: all of it exists to deal with the heat produced by billions of transistors switching at gigahertz speeds.
At the data center scale, this is not a minor inconvenience. A large AI data center today draws power comparable to that of tens of thousands of homes, and a substantial fraction of that electricity goes directly to cooling, not computing. As AI workloads grow, the thermal wall is becoming an existential challenge for the industry. The question researchers have been asking for years is: Can you switch faster without generating proportionally more heat? The Tokyo team’s answer, at least at the laboratory scale, appears to be yes.
The Device: Spintronics Meets Antiferromagnetism
The device is built on a field called spintronics, which stores and processes information using the magnetic orientation of electrons rather than their electrical charge. Conventional memory and logic devices store a 0 or 1 by moving electrons around, which requires energy and generates heat. Spintronic devices instead flip the spin direction of electrons between two stable magnetic states using far less energy and far less heat.
The Tokyo team used a specific class of magnetic material called an antiferromagnet. In conventional ferromagnets such as iron, cobalt, and nickel used in traditional magnetic storage, the electron spins all point in the same direction, creating a net magnetic field. In antiferromagnets, adjacent spins point in opposite directions and cancel each other out, producing no net external magnetic field. This makes them magnetically invisible from the outside, which has historically made them difficult to read and write.
The specific material used is manganese-tin (Mn3Sn), a chiral antiferromagnet whose internal magnetic state can be switched between two stable configurations using electrical pulses. Those two configurations represent 0 and 1. Crucially, the state persists after power is removed, making it non-volatile, unlike DRAM, which loses its data the moment power is cut.
The device is constructed as a thin-film stack: a layer of Mn3Sn on top of a layer of tantalum, a heavy non-magnetic metal, deposited on a silicon substrate. When a short electrical pulse passes through the tantalum layer, a quantum mechanical effect called spin-orbit torque transfers angular momentum to the Mn3Sn layer, flipping its magnetic state. No large current required. No significant heat generated.
The Numbers: 40 Picoseconds
The headline result is switching speed. The device flipped its magnetic state in as little as 40 picoseconds, that is 40 trillionths of a second. Conventional non-volatile switching devices for data processing operate in the nanosecond range, roughly 1,000 picoseconds or more. DDR5 DRAM, currently the fastest widely-deployed memory technology, switches in approximately 208 picoseconds. The Tokyo device at 40 picoseconds is five times faster than DDR5 DRAM, and it retains its state when power is cut, which DRAM cannot do.
The team also demonstrated switching using 60-picosecond photocurrent pulses, generated by combining a telecom-band laser with a photoelectric converter. This means the device can be written using optical signals converted into electrical ones, opening a path toward optical computing architectures in which data moves as light rather than electricity, further reducing energy requirements.
On durability, the results are equally striking. The device survived billions of test cycles without degradation, a critical threshold for any memory technology that needs to function reliably over years of use. Previous ultrafast switching approaches have often involved temperature spikes of several hundred Kelvin during operation, which destroy materials over time. The Mn3Sn device, operating through spin-orbit torque rather than thermal excitation, appears to sidestep this problem entirely.
Why Non-Thermal Is the Key
One of the most important questions the researchers addressed was whether the switching mechanism was genuinely non-thermal, meaning driven by the angular momentum of the electrical current rather than by the heat it generates. Earlier work at Tokyo’s lab, published in Nature Materials in late 2025, had visualized the switching mechanism in real time using ultrafast optical pulses. The conclusion was unambiguous: the switching was driven by the current’s angular momentum via spin-orbit torque, not by heat.
This is a crucial finding. It means that as you switch faster using shorter pulses with less total energy, you do not necessarily generate more heat. The thermal coupling that plagues silicon transistors is fundamentally decoupled from the switching mechanism in Mn3Sn. That is what makes the 1,000x speed claim meaningful rather than merely aspirational.
The Data Center Angle
The implications for data centers are potentially enormous. A large AI data center today might draw power equivalent to 80,000 homes. Estimates suggest that if this technology were deployed at scale, the same facility could theoretically operate on the power equivalent of around 800 homes, a 100-fold reduction. This figure assumes the technology scales and that manufacturing challenges are solved, but even a fraction of that improvement would be transformative for an industry that is currently one of the fastest-growing consumers of electricity on the planet.
The timing is significant. The AI compute boom has driven data center power demand to levels straining electrical grids across the United States and Europe. The industry has responded with liquid cooling, more efficient rack design, and nuclear power agreements. These are solutions to a symptom. A switching architecture that generates less heat at the point of computation would attack the cause.
There is also a memory angle that is often overlooked. One of the most energy-intensive operations in modern computing is not arithmetic but moving data between processors and memory. DRAM is fast but volatile and power-hungry. Non-volatile flash storage is dense but slow. A non-volatile memory technology that switches at sub-DRAM speeds and generates minimal heat could collapse the memory hierarchy, dramatically reducing the energy cost of data access. For AI inference workloads, where models are loaded into memory and queried millions of times per second, this would be particularly valuable.
The Road to Commercialization: 2030 at the Earliest
The researchers are appropriately cautious about timelines. The current device is a laboratory demonstration, a proof of concept that the underlying physics works and that the material can survive extended operation. The team has set a target of a practical prototype chip by 2030. Commercialization at scale would come after that.
The path from here to there involves several non-trivial challenges. Mn3Sn films need to be deposited with sufficient uniformity and quality to function reliably at the wafer scale. The device architecture needs to be integrated with CMOS logic in a way that is compatible with existing semiconductor manufacturing flows. Reading the magnetic state, not just writing it, needs to be demonstrated at comparable speeds and energy levels. And the entire system needs to operate at the density required for modern memory arrays, where billions of cells must coexist on a single chip without interfering with one another.
None of these are insurmountable. Antiferromagnetic materials have been studied for decades, and there is a growing body of engineering knowledge around their deposition and patterning. The team is actively engaging with industry partners to begin the transition from laboratory demonstration to manufacturable design.
Bottom Line
The University of Tokyo’s antiferromagnet switching device is a genuine scientific breakthrough, not a press release dressed up as research. The paper is published in Science, the results have been independently covered by Tom’s Hardware, TechRadar, Live Science, and Phys.org, and the underlying physics has been built through multiple peer-reviewed studies over the past two years.
What it is not is a product. The 2030 prototype target is realistic. Commercial deployment at scale is likely a decade away. A data center built in 2026 will operate under today’s thermal physics for its entire useful life. The gap between laboratory demonstration and production silicon is measured in years and billions of dollars of engineering investment.
But foundational breakthroughs have a way of compressing timelines when the industry has a compelling reason to move fast. And with AI compute demand driving data center energy consumption toward levels that are politically and economically unsustainable, the industry has never had a more compelling reason. If Mn3Sn can deliver even a fraction of what the Tokyo team has demonstrated, the race to scale it will be well-funded, well-staffed, and very fast.
Source: Tsai, H., et al. “Picosecond ultralow-power switching device based on an antiferromagnet.” Science, 392(6799), 761-765 (2026). Additional reporting via Live Science, Tom’s Hardware, TechRadar, Phys.org, and Blocks & Files.