A Chip Device Promises 1000x Speedup Without Overheating
Scientists have unveiled a new device based on the antiferromagnet Mn3Sn, which allows processors to operate a thousand times faster without generating additional heat. The technology could drastically reduce data center energy consumption, with a chip prototype expected by 2030.
A quiet revolution in Tokyo: Why 40 picoseconds rewrite the rules for NVIDIA and Intel
[The Gist]: What's Really Happening
When a paper by Professor Tomo Nakatsuji's group from the University of Tokyo and RIKEN was published in Science on May 14, 2026, the world of AI chips should have paused. But it didn't. And that's a shame. The researchers demonstrated a switching element based on the antiferromagnet Mn3Sn that performs a binary state switch in 40 picoseconds. That's 1000 times faster than current silicon-based AI accelerators. And it generates a minuscule amount of heat.
Why is this important not in theory but in practice? Because today, large data centers spend up to 40% of their total electricity consumption not on computation but on cooling. The laws of physics dictate: the faster you push electrons through wires, the more they heat up. This is called Joule heating. And as long as we use the electron's charge for computation, we can't cheat thermodynamics. But the Tokyo team found a loophole: they use spin, not charge.
Spintronics is not a new concept. But Nakatsuji's team is the first to prove that speeds of 40 picoseconds can be achieved with a temperature rise of only 8 Kelvin. Previous attempts at ultra-fast memory resulted in heating of hundreds of Kelvin, instantly killing any commercial prospects. The physicists at the University of Tokyo fundamentally changed the mechanism: instead of thermal switching (where the state changes due to heat), they use spin-orbit torque. Electrons transfer angular momentum to each other rather than simply colliding with the conductor walls.
The third and most important nuance: the device is non-volatile. That means it retains its '0' or '1' state after power is removed. Today's DRAM must refresh the charge of each cell thousands of times per second, or data disappears. This consumes a huge amount of energy just to 'stay alive.' The new switch is a step toward computers that consume no energy in standby mode.
Timeline and Context
Actually, this story didn't start in May 2026. Back in February 2026, a Chinese group from Tsinghua University led by Cheng Sun published a paper in Nature where they achieved full switching of chiral antiferromagnetic order (Mn3Sn) without an external magnetic field. The Chinese showed it was fundamentally possible. The Japanese, in Science, went further: they not only switched it but did so in record time.
The key technological maneuver by the Japanese was using a tantalum (Ta) interlayer. In their Mn3Sn/Ta configuration, a spin-polarized current is generated that switches the magnetic moment. But there's a catch, quietly discussed in the corridors: deterministic switching still requires a small external magnetic field. Without it, you can't guarantee which state the cell will switch to. This is a fundamental problem that the Tsinghua group claims to have already solved in their work. Competition between Tokyo and Beijing in this field will only intensify.
In May 2025 (exactly one year before the Science publication), the same group had already published preliminary results in Nature. And in late May 2026, a preprint by Xiaokang Li and colleagues appeared on arXiv, showing an alternative path: switching Mn3Sn using a thermal pulse and a 0.1 mT field. That's nearly 100 times smaller than previously needed. So the industry is moving toward eliminating external magnets from multiple directions.
Who Wins and Who Loses
The first and biggest loser is, surprisingly, NVIDIA itself. Why? Because their business is built on selling ultra-expensive, high-power GPUs. The new technology promises to reduce computing energy consumption by orders of magnitude. If data centers can achieve the same performance at 10% of the energy cost, they'll buy fewer chips. Moreover, the von Neumann architecture, where memory and processor are separate, will become obsolete. Non-volatile ultra-fast memory can be integrated directly into the compute die. NVIDIA, whose empire is built on HBM memory and discrete GPUs, will find itself in Intel's position from ten years ago.
Japan wins. This isn't just a 'technological breakthrough.' Japan is deliberately betting on post-silicon electronics. Companies like Kioxia (formerly Toshiba Memory) and Sony have extensive experience in manufacturing spintronic devices—they've been producing MRAM (Magnetoresistive RAM) for niche applications for years. Now they have the scientific foundation for mass production. Expect Japan's Ministry of Economy, Trade and Industry (METI) to launch a national project to commercialize Mn3Sn memory as early as 2026.
TSMC wins. Because the Taiwanese have the most experience in heterogeneous integration. The new material can't simply be 'plugged into' existing FinFET processes. New methods for thin-film deposition and etching must be developed. TSMC already began experimenting with antiferromagnetic materials in their R&D lines in 2025. Samsung and Intel are lagging behind. Production of such chips will start 2-3 years later, and by then TSMC will have captured the market.
As for data center cooling equipment suppliers (Vertiv, Schneider Electric, Stulz), this is an existential threat. If new processors barely heat up, the multi-billion-dollar market for liquid and immersion cooling systems could collapse faster than anyone expects.
What the Media Isn't Saying
The main non-obvious insight I heard from an acquaintance engineer at RIKEN concerns not Mn3Sn itself, but tantalum (Ta). Tantalum is a conflict mineral. 60% of the world's tantalum reserves are in the Democratic Republic of Congo, and its mining finances armed conflicts. Moreover, tantalum is already used in huge quantities in capacitors for smartphones and laptops. If the technology goes mainstream, demand for tantalum will skyrocket. The price of tantalum (currently around $200-300 per kg) could increase 5-10 times. This will create a bottleneck that all scientific journals are silent about.
The second point concerns material inhomogeneity. In the lab, scientists grow perfect Mn3Sn crystals on silicon substrates. But on a 300 mm diameter wafer (standard for semiconductors), it's impossible to ensure perfect crystal lattice uniformity across the entire area. Defects in just a few atoms will cause some cells to switch in 40 picoseconds, others in 100, or not at all. The chip yield could be as low as 10-20%, making them economically unviable even with fantastic performance. The Chinese group at Tsinghua claims to have found a way around this problem, but their work is also still lab-scale.
The third and most cynical point. The Tokyo researchers used a 60-picosecond optical pulse for switching, operating in the standard telecom C-band. This means data can be transmitted over existing fiber optic networks and directly 'imprinted' into memory without conversion to electricity. This kills not only cooling but also network cards, routers, and switches. The next 10 years will be a time when photonics and spintronics merge. Cisco, Arista, and Broadcom could lose their bread and butter because their network chips become unnecessary.
Forecast: Next 30 Days and 90 Days
Next 30 days.
Don't expect any news from NVIDIA or AMD—they'll stay silent to avoid panic among investors. But Intel might make an unexpected move: they have a huge patent portfolio in spintronics dating back to the 2010s (Marvel Technology). Within 30 days, Intel could announce the creation of a joint lab with the University of Tokyo. Also, expect stocks of tantalum mining companies (Global Advanced Metals, AMG Minerals) to rise 5-10% on speculative interest.
Next 90 days.
In three months, perhaps the most important event will occur: IEEE (Institute of Electrical and Electronics Engineers) will launch a working group to standardize testing of antiferromagnetic devices. Currently, no one knows how to measure '1000x speedup' on an industrial scale. Without standards, there's no certification; without certification, no sales. If the working group is formed, it's a green light for investors. If not, the technology will remain in labs for another 5 years.
Also, within 90 days, at least two major reports from analyst firms Gartner and IDC will be released, where they will include 'ultra-fast spintronic memory' in their semiconductor industry roadmaps for the first time. This will legitimize the technology for Fortune 500 companies. Until then, everyone will say 'it's interesting but not practical.'
And a final insight. Don't watch Tokyo; watch Beijing. The Chinese from Tsinghua University have already solved the 'external magnetic field' problem. If they demonstrate their solution working at room temperature on an 8x8 cell array in the next 90 days (which won't be difficult), the scales will tip toward China. In that case, given the 'Chinese pace,' their chip prototype could appear not in 2030, as the Japanese predict, but as early as 2028. The arms race in Silicon Valley is over. The race in Post-Silicon Valley has begun, and only two players are involved: Japan and China. The US and Europe missed this moment.
— Editorial Team
No comments yet.