SkyWater Technology’s foundry line just cranked out a silicon chip whose digits don’t stay put.
The blade's final structure emerges from the controlled conversation between fire and water.
A Japan–U.S. team has built the first spintronic probabilistic bit fabricated inside a standard semiconductor manufacturing process. The device, born from a collaboration between Tohoku University and the National Institute of Standards and Technology, uses SkyWater’s commercial 130-nm CMOS line to print transistors and lower interconnects. Then, in a separate spintronics facility, researchers deposit the superparamagnetic tunnel junctions and upper electrodes. What emerges is not a lab-built anomaly but a voltage-in, voltage-out logic element that produces a fluctuating digital output—and it does so on silicon that any fab can produce.
The paper, published in IEEE Electron Device Letters and submitted on the arXiv preprint server on April 15, 2026, marks the moment probabilistic computing declares its manufacturing maturity. This is not a physics demo. It is a declaration that p-computers will scale on the existing silicon supply chain, not in specialty foundries.
One cartographer favors precise, brittle lines; the other points to the vine's exploratory, probabilistic route-finding.
What Are We Looking At?
A probabilistic computer uses p-bits—bits that stochastically fluctuate between 0 and 1 using intrinsic physical randomness. Conventional binary logic processes deterministic instructions. P-bits harness thermal noise. Spintronics earns its place here because nanoscale magnetic devices generate this randomness naturally; their magnetization wobbles, and that wobble is the computational primitive.
The lineage is clear. In 2021, Tohoku University researchers built what they called the “poor man’s quantum bit,” a nanoscale magnetic tunnel junction whose magnetization direction updated every 8 nanoseconds on average—100 times faster than the prior world record. In 2022, a team from Tohoku, the University of Messina, and UC Santa Barbara presented a scaled-up spintronic probabilistic computer at the International Electron Devices Meeting on December 6. Then, in August 2025, a Nature Electronics paper reported a probabilistic computer using voltage-controlled magnetic tunnel junctions as its entropy source, built on a 130-nm foundry CMOS process and wielding 1,143 p-bits to factor integers using Ising-machine-based invertible logic gates.
The new device cuts the final tether to exotic fab. The team built the CMOS underlay directly at SkyWater, then post-processed the magnetic structures. When the superparamagnetic tunnel junction’s resistance fluctuates, a corresponding digital output voltage fluctuates with it—and is tunable via the input voltage. You put a voltage in, you get a probabilistic voltage out. This is a library-ready logic element.
The Real Disruption Is in the Supply Chain
This story’s center of gravity isn’t a performance leap; it’s that probabilistic computing now speaks the semiconductor industry’s production language. The 130-nm node is mature, low-cost, and operated by a pure-play foundry. Integration here proves that p-computers can follow the same commercial path as any analog mixed-signal block.
By landing in a foundry, spintronic p-bits pivot from a physics experiment to a fabless semiconductor proposition. The vector of disruption runs straight through the AI chip market. Applications that rely on sampling, Bayesian inference, or combinatorial optimization—core workloads for drug discovery, logistics, and edge decision-making—get a native, low-power hardware primitive that deterministic silicon cannot replicate efficiently.
Who Wins, Who Loses
The immediate loser is analog AI compute startups whose differentiation relied on bespoke fabrication. Companies that built their moat around non-CMOS, exotic-material processes now face a direct substitute that scales on the semiconductor industry’s standard infrastructure. A p-computer block built at SkyWater is a block a major IP vendor can license and harden into a standard cell library. The competitive buffer of a custom fab process evaporates the moment a foundry PDK drop lands.
The immediate winner is the edge-inference market. A low-power chip that models uncertainty at the transistor level displaces deterministic accelerators for workloads where inference is about confidence distributions, not point predictions. The same hardware architecture applies to combinatorial optimization problems that currently consume clusters of digital processors.
The signal to watch is intellectual property movement. The 2025 Nature Electronics paper integrating 1,143 p-bits on 130-nm foundry CMOS was the warm-up. Now that the magnetic device itself is demonstrably fabricable on that same node, the industry has line of sight to a production-sourced p-computer. Predictably specific: within 18 months, a fully packaged, foundry-sourced spintronic p-computer chip with at least 10,000 integrated p-bits will run a publicly evaluable optimization benchmark—likely for logistics or molecular docking—on a standard evaluation board. That device will be directly associated with at least one major semiconductor IP partner or fab collaboration.
What Semiconductor Strategists Need to Track Now
The shift is from “if” to “when.” For semiconductor strategists, probabilistic computing should now appear on near-term roadmaps as a licensable compute fabric, not a long-paper project. For AI hardware architects, a CMOS-native stochastic primitive changes the co-design equation radically: sampling-heavy models get a physical target. For investors, the risk is sunk cost in fabless analog AI plays whose process lock-in now looks like a ceiling, not a moat.
The opportunity is straightforward. A computing element that handles uncertainty natively, built at a node that costs little and scales far, opens a path to silicon that does what deterministic logic cannot—without the overhead of cooling, calibration, or quantum error correction.
“Our latest technology made the first step to realize Feynman’s vision with spintronics,” Tohoku’s Shun Kanai said in 2021, referencing the earlier breakthrough the team has now dragged onto a commercial fab floor.
That vision now has a standard-cell footprint. The industry has officially let randomness into the cleanroom—on purpose, and with a commercial foundry’s stamp of approval. The digits are supposed to wobble now.
