Artificial intelligence is colliding with a hard physical limit: the energy and heat of today’s silicon-based chips. As models grow from billions to trillions of parameters, the bottleneck is no longer clever algorithms but the materials that shuttle bits and spins around the hardware. A new class of magnetic effects and compounds, from exotic “p-wave” states to rare-earth-free alloys, is emerging as a credible way to break that ceiling and power the next wave of AI systems.
Instead of pushing electrons through ever-smaller transistors, researchers are learning to compute with magnetism itself, routing information through spin waves, spiral orders, and altermagnetic patterns that waste far less energy. If these lab breakthroughs can be engineered into real devices, the result could be AI accelerators that are not just marginally better but orders of magnitude more efficient than the GPUs that dominate data centers today.
The race to reinvent AI hardware around magnetism
I see a clear pivot under way in AI hardware research, away from squeezing more performance out of conventional CMOS and toward rethinking the basic carriers of information. Magnetic materials are at the center of that shift because they promise to store and process data with minimal charge motion, which is where most of the heat and power loss occur in current chips. The stakes are obvious in hyperscale data centers, where training a single frontier model can consume megawatt-hours of electricity and push cooling systems to their limits.
In that context, the idea that a new magnetic state could cut energy use by several orders of magnitude is not a curiosity, it is a potential platform change. Work on p-wave magnetism, altermagnetism, spin waves, and magnetic whirls all points in the same direction: information can be encoded in the orientation and collective motion of spins rather than in the flow of electrons. If those concepts can be translated into manufacturable devices, the next generation of AI accelerators may look less like shrunken GPUs and more like hybrid spintronic processors built around tailored magnetic lattices.
P-wave magnetism: a new state with radical efficiency potential
The most striking development is the experimental confirmation of a magnetic state known as p-wave magnetism, which had been theorized but never observed in a real material. Researchers working with nickel iodide, or NiI2, have shown that a carefully tuned crystal can host a spiral arrangement of spins that behaves in a fundamentally different way from conventional ferromagnets and antiferromagnets. In this configuration, the magnetization varies in space in a pattern that allows information to be encoded in subtle phase relationships rather than simple up or down orientations.
One of the scientists involved, identified as Jun, has argued that these p-wave magnets could save “five orders of magnitude” of energy compared with today’s electronics, a claim that underscores how far this approach could move the needle if it scales beyond the lab. That estimate is grounded in the way p-wave states suppress dissipative currents and instead rely on collective spin behavior, as described in work on p-wave magnets that highlights the role of Jun and the emphasis on energy savings by Which. For AI, that kind of efficiency gain would not just trim power bills, it would enable architectures that keep far more computation on chip without thermal throttling.
Inside the spiral: how NiI₂ rewrites magnetic order
To understand why NiI2 is attracting so much attention, it helps to look at the geometry of its spins. In this material, the nickel atoms sit on a triangular lattice, and the spins do not simply align or alternate; they twist into a spiral pattern that winds across the crystal. This spiral magnetic order, visualized as light blue arrows on the triangular network of Ni atoms, creates a texture that supports p-wave magnetism and allows for rich spin dynamics that are impossible in simpler lattices.
That spiral configuration is not just a visual curiosity, it is the mechanism that enables new ways to route and manipulate spin information. By adjusting temperature, fields, or strain, engineers can in principle tune the spiral and thereby control how spin waves propagate through the material, a prerequisite for practical devices. The detailed description of this spiral magnetic order in NiI2 shows how the triangular lattice and the emergent p-wave magnetism are tightly linked, giving AI hardware designers a concrete blueprint for materials that can host complex spin textures at useful temperatures.
From lab curiosity to memory workhorse
The obvious question is whether p-wave magnetism is a fragile phenomenon that only appears in exotic lab conditions or a robust state that can underpin real devices. Experimental work by MIT physicists has started to answer that by demonstrating p-wave magnetism in NiI2 under practical room-temperature conditions, not just in cryogenic setups. That matters because AI accelerators need to operate in dense racks and edge devices, where elaborate cooling is expensive or impossible.
Researchers at MIT have framed this discovery as “a new spin on memory,” arguing that p-wave magnets could form the basis of nonvolatile storage and logic elements that combine the speed of SRAM with the persistence of flash. In their experiments, they have shown that the p-wave state can be switched and read in ways that are compatible with spintronic architectures, opening a path to memory arrays that compute as they store. The work on MIT p-wave magnetism makes clear that this is not just a theoretical curiosity but a candidate for integrated spintronic memory that could sit alongside or even replace parts of today’s DRAM and MRAM in AI accelerators.
Altermagnetism and the third pillar of spintronics
While p-wave states are reshaping how physicists think about magnetic order, another concept, altermagnetism, is expanding the taxonomy of magnetism itself. Altermagnets exhibit a newly validated magnetic state that behaves differently from both ferromagnets and antiferromagnets, effectively forming a third fundamental form of magnetism. In these materials, spins are arranged in patterns that cancel net magnetization yet still produce strong spin polarization in momentum space, a combination that can be exploited for ultra-fast, low-loss electronics.Researchers have argued that this altermagnetic state could make electronics up to a thousand times faster by enabling spintronic devices that switch without the inertia and heating associated with charge-based currents. One report describes how a New Type of Magnetism Discovered That Could Make Electronics 1000x Faster is rooted in this altermagnetic behavior, explaining What Altermagnetism is and why it matters for device speed. For AI workloads that depend on rapid matrix multiplications and memory access, integrating altermagnetic layers into interconnects or logic blocks could slash latency and energy per operation in ways that conventional scaling cannot match.
Spin waves: computing with ripples instead of currents
Beyond static magnetic states, a parallel line of work is focused on using spin waves as carriers of information. Instead of moving electrons, these devices launch ripples in the collective spin orientation of a material, a mode of transport that can be far less dissipative. Spin-wave technology offers a way to build logic and signal-processing networks where data flows as waves, with interference patterns performing computation in analog fashion.
One group has demonstrated a large spin waveguide network built from yttrium iron garnet, or YIG, a material known for having the lowest attenuation currently measured for spin waves. By carefully patterning YIG structures, they created the largest spin waveguide network to date and showed that it can route signals with minimal loss, a crucial step toward practical spin-wave circuits. The report on this YIG spin waveguide network underscores how Jul experiments with YIG are turning spin waves from a physics demonstration into a platform that could underpin AI accelerators focused on pattern recognition and signal filtering.
Spin-wave AI devices and tenfold efficiency gains
As these spin-wave concepts mature, they are already being translated into AI-specific devices that promise dramatic efficiency gains. Spin-wave technology can implement neural network operations by encoding signals in wave amplitudes and phases, then letting interference perform weighted sums. Because the waves propagate without moving charge, the energy cost per operation can be far lower than in CMOS-based MAC units, especially for analog inference tasks at the edge.
One analysis notes that spin-wave technology offers a complementary path to the AI hardware ecosystem, potentially improving energy efficiency tenfold for certain workloads. It highlights how Spin and New magnetic tech can make AI more efficient by routing information through magnonic circuits instead of resistive interconnects. The description of Spin waves as a new magnetic tech that makes AI more efficient, especially in the Aug context of analog accelerators, shows why chip designers are exploring hybrid architectures that pair digital control with magnonic compute blocks.
Ion-controlled interference: steering spin waves for on-chip AI
To move from proof-of-concept waveguides to full AI processors, engineers need fine-grained control over how spin waves interfere. That is where ion control comes in. By modulating the local magnetic properties of a material with ions, devices can dynamically adjust the phase and amplitude of spin waves, effectively programming interference patterns that implement different neural network weights or logic functions.Researchers have reported the Development of a High Performance AI Device Utilizing Ion Controlled Spin Wave Interference in Magnetic Materials, describing how ion gating can tune spin-wave paths in a way that is compatible with integrated fabrication. In that work, the team shows that an Development of a High Performance AI Device Utilizing Ion Controlled Spin Wave Interference in Magnetic Materials can realize reconfigurable magnonic circuits that execute AI tasks with high throughput and low power. For AI hardware, this suggests a future where weights are not stored in static SRAM arrays but in the interference patterns of spin waves, updated by tiny ion-induced adjustments rather than large charge movements.
Silicon-compatible magnetic whirls and on-chip integration
Even the most elegant magnetic phenomena will struggle to matter for AI if they cannot be integrated with existing silicon processes. That is why work on silicon-compatible magnetic whirls is so important. Using advanced fabrication and characterization methods, researchers have created free-standing magnetic layers that host a robust family of magnetic whirls, topological structures that can act as information carriers. These whirls can be moved, created, and annihilated with modest stimuli, making them candidates for dense, low-energy memory and logic.
Crucially, the method used to generate these whirls is compatible with silicon platforms, which means they can be layered onto or alongside conventional CMOS. A report on how this method revealed that the free-standing layers are able to host a robust family of magnetic whirls, and that the work has been published in Nature Materials, underscores the Potentially transformative impact of such structures. The description of silicon-compatible magnetic whirls shows how Feb experiments are closing the gap between exotic spin textures and manufacturable AI chips that can store and manipulate data in topological states.
AI-designed magnets and the rare-earth problem
While new magnetic states promise performance gains, the materials themselves raise another challenge: supply chains. Many high-performance magnets rely on rare earth elements that are expensive and geopolitically concentrated. To avoid building the next generation of AI hardware on fragile resource foundations, researchers and startups are using AI to design alternative compounds that deliver strong magnetic performance without rare earths.
One example is Materials Nexus, a company that used AI to design, synthesize, and test a rare-earth-free magnet material dubbed MagNex in roughly three months, a timeline that would have been unthinkable with traditional trial-and-error methods. Reporting on how the company, called Materials Nexus, designed, synthesised and tested MagNex illustrates how generative models and high-throughput simulations can compress the materials discovery cycle. In a separate interview, Mining Weekly speaks with Robert Forest, the chief technical officer and co-founder of Materials Nexu, about how artificial intelligence was used to eliminate the need for rare earths in magnet design, a conversation captured in a Mining Weekly interview that shows how Jun discussions around AI-driven materials discovery are already reshaping industrial magnet supply chains.
Altermagnetic materials tailored for the AI era
Altermagnetism is not just a theoretical curiosity; specific materials with altermagnetic properties are being identified as promising candidates for AI hardware. These compounds can address key drawbacks of today’s magnetic random access memory, which relies on ferromagnetic layers that can introduce stray fields and scaling challenges. By contrast, altermagnets can offer strong spin polarization without net magnetization, reducing interference and enabling denser integration.
Researchers have highlighted how Materials with altermagnetic properties could address key drawbacks of today’s magnetic random access memory, and how a newly validated magnetic state forms a third fundamental form of magnetism that is particularly well suited to spintronic devices. One report frames this as scientists identifying a promising new magnetic material for the AI era, explaining how a newly validated magnetic state could underpin faster, more efficient memory and logic. For AI accelerators, integrating such altermagnetic layers into caches and interconnects could reduce latency and power while avoiding some of the scaling headaches of conventional MRAM.
From physics breakthroughs to the next AI wave
Across p-wave magnetism, altermagnetism, spin waves, magnetic whirls, and AI-designed alloys, a common pattern emerges: the frontier of AI performance is shifting from software tricks to the quantum behavior of materials. Each of these advances tackles a different bottleneck, from memory density and interconnect speed to energy efficiency and resource security. Together, they sketch a plausible path to AI hardware that is not just incrementally better but qualitatively different, with computation encoded in spins, waves, and topological textures rather than in hot, resistive currents.
The remaining challenge is engineering, not physics. Device designers must translate Jun’s vision of p-wave magnets saving five orders of magnitude of energy, the promise that New Type of Magnetism Discovered That Could Make Electronics Faster, and the tenfold efficiency gains suggested by groundbreaking magnetic technology into manufacturable chips that can survive the brutal economics of data centers and consumer devices. If they succeed, the next wave of AI will not be powered by ever-larger GPU farms but by quiet, cool, magnetically driven processors that make today’s hardware look as dated as vacuum tubes.
More from MorningOverview