🎲 The world's AI boom has an electricity problem you have probably heard about. The surprising part is the fix one lab is chasing: instead of fighting the randomness inside a chip, it wants to burn randomness as fuel. A team in Japan and the United States has just built the first working version of a "probabilistic bit" using the ordinary factory processes that make everyday computer chips. No superconductors. No giant refrigerator. Just a tiny magnet that cannot make up its mind.

A bit that keeps flipping a coin

The bit in a normal computer is parked at either 0 or 1. The qubit in a quantum computer lives in a superposition of both. A p-bit does something else again: it flips between 0 and 1 on its own, many times a second, like a coin tossed over and over.

The trick is that you can load the coin. Nudge an input voltage and you set how often it lands on 1 versus 0. Wire a lot of these biased coins together, let them settle, and the whole network drifts toward the answers to a certain class of hard problems. The sweet spot is combinatorial optimization (think delivery routes, shift scheduling, the classic traveling-salesman puzzle), probabilistic inference, and the kind of sampling that machine learning leans on. According to Tohoku University, on those problems a machine like this can explore huge numbers of candidate states several orders of magnitude faster than general-purpose hardware such as GPUs.

A cousin of the quantum computer, not a copy

Both probabilistic and quantum computers get pitched as "new-concept" machines beyond the everyday digital computer, and the names rhyme on purpose: "p-bit" was coined as a deliberate play on "qubit." But Tohoku is careful to say the two are fundamentally different. A qubit holds a genuine superposition of 0 and 1 and can be entangled with other qubits. A p-bit does neither. It is a classical device that simply fluctuates very fast.

That difference is also the practical selling point. Today's leading quantum machines have to be chilled to near absolute zero inside dilution refrigerators. Spintronics p-bits run at room temperature and can be made on the same silicon lines that already turn out memory and logic chips. The basic p-bit cell even resembles the one-transistor-plus-magnetic-junction structure used in commercial MRAM memory, with a single change: the magnet is built unstable on purpose so that it wobbles.

The idea has a long pedigree. Tohoku points back to a 1981 lecture by Richard Feynman, better remembered for launching quantum computing, in which he also floated the notion of using probabilistic processes to compute efficiently. The modern p-bit was formalized in 2017 by Kerem Camsari, Supriyo Datta and colleagues. And in 2019, a Tohoku-Purdue team that included some of the same Tohoku researchers behind the new result wired up an 8-bit probabilistic computer that could factor small numbers, connecting individual spin devices to a microcontroller with cables.

Printing the wobble onto silicon

Those cables were the bottleneck. To matter for real workloads, a probabilistic computer needs not a handful of bits but thousands, eventually millions. That means leaving the hand-wired benchtop behind and printing the spin devices and their control circuitry together on one chip.

That is what the new work demonstrates. The team, led by Shunsuke Fukami at Tohoku's Research Institute of Electrical Communication and William Borders at the US National Institute of Standards and Technology (NIST), combined Japanese and American fabrication steps. The transistors and lower wiring were made on SkyWater Technology's 130-nanometer CMOS process in the US; the deliberately fluctuating spin element and the upper wiring were added at Tohoku's Nano-Spin facility.

Then they checked that the resulting circuit behaved the way a p-bit should. Two things mattered: the output voltage genuinely fluctuates over time, and its time-average can be dialed up or down by the input voltage. Both held. According to the team, this is the first reported demonstration of a spintronics p-bit built on a single substrate using a semiconductor integration process. The result appeared in IEEE Electron Device Letters in late May 2026.

The promise, and where it actually stands

Here is where restraint matters. It is tempting to file this under "the answer to AI's energy crisis," and the energy framing is real enough. The International Energy Agency estimates that data-center electricity ran to about 415 terawatt-hours in 2024, roughly 1.5 percent of world demand, and could approach 1,050 TWh in 2026; electricity used specifically by AI-focused data centers jumped about 50 percent in 2025. If data centers were a country, the IEA notes, they would rank among the world's largest electricity consumers. Hardware that chews through optimization and sampling problems while sipping power is exactly what that picture invites.

But what exists today is a single p-bit, demonstrated at the device level. It is proof that the manufacturing path exists, not a product. Tohoku puts it plainly: the achievement brings scaling toward a million bits within range, and the next job is to develop the device and circuit technology to actually get there. That is years of work, not a ship date.

Even so, there is something quietly radical here. For seventy years, computing has been a campaign to stamp out uncertainty, to make every bit hold its value perfectly. This is a bet on the opposite instinct: that the next gains might come not from suppressing the wobble in our machines, but from learning to aim it. How is the energy cost of AI talked about where you live — as a problem to engineer around, or as a reason to rethink how we compute in the first place?

References