⚛️ Three percent. Say it out loud and it sounds like a rounding error — the kind of number you'd shrug at. So when a Japanese chipmaker announced it had used quantum technology to push a factory's output up "about 3%," the natural reaction was: that's it? All that quantum hype for three measly percent?

Here's the twist. On a semiconductor production line, 3% is not the small part of the story. It's the hard part.

The quiet kind of quantum

Since January 2026, a piece of quantum technology has been clocking in for work every day at a chip plant in southern Japan — not as a demo or a press stunt, but as a tool that helps decide, hour by hour, what the factory does next.

The plant is Lapis Semiconductor's Miyazaki facility, part of the ROHM group, one of Japan's big power-chip makers. The software comes from Quanmatic, a small Tokyo startup whose chief scientist, Waseda University's Professor Nozomu Togawa, works on the math behind quantum optimization. Together they applied quantum annealing to the "front-end" of chip manufacturing. According to the two companies, no one had done that in real production before.

And that phrase, quantum annealing, matters: this is not the quantum computer you've read about in the headlines.

Why shaving 3% off a chip line is brutally hard

To understand the 3%, you have to understand what the front-end of a fab actually is.

A modern chip is built up on a silicon wafer through hundreds of steps: depositing thin films, printing circuit patterns with light, etching them in, cleaning, diffusing, then doing it all again at the next layer. The same machines get used over and over, but the settings change at every stage depending on the product. Lots of wafers move through in different orders, some waiting on a machine, some waiting on each other. Get the sequence slightly wrong and bottlenecks pile up where you can't see them.

[ref src="https://prcdn.freetls.fastly.net/release_image/117406/35/117406-35-cf05760eb318b811e2296fce7ecd0c08-988x250.png?format=jpeg&auto=webp&fit=bounds&width=1200" alt="Diagram showing where the front-end process sits within semiconductor manufacturing" origin="Source: Quanmatic / PR TIMES"]

Working out the best order to run everything (which lot goes to which machine, in what sequence, under which deadline) is a monstrous combinatorial puzzle. There are more possible schedules than anyone could ever check by hand or by ordinary software.

The catch is bigger than it looks. Fabs have been grinding away at this problem for decades. Engineers have built rule-of-thumb algorithms, custom optimizers, elaborate simulations. After all that effort, the realistic ceiling for squeezing out more efficiency tends to be just a few percent. A wafer-systems manager at hard-drive maker Seagate once put it bluntly to Fortune: despite years of advanced scheduling software, the company could only hope to improve efficiency by a handful of percentage points. The easy gains were gone long ago.

So 3% isn't the leftover scraps. It's close to the whole jar — the last bit anyone thought was still in there. And on a line that runs around the clock, a few percent translates into serious money: the same Fortune piece reported that a 7–10% efficiency gain at a large fab could be worth several million dollars a month. Scale that down and you can see why "about 3%" makes executives sit up.

A warm-up at the test bench

ROHM and Quanmatic didn't start with the front-end. They started somewhere simpler.

Their first target was a stage called EDS (Electrical Die Sorting), the step where finished chips on a wafer get their electrical characteristics tested. EDS still has fiddly constraints, especially "setup": the work of reconfiguring equipment and jigs whenever the product type, wafer size, or temperature changes. Setup is dead time, and dead time is waste.

Applying their optimization system there, the companies cut that setup loss by 40%. They rolled it into live production at a ROHM plant in the Philippines and, by mid-2025, called it a first-of-its-kind large-scale industrial use of quantum technology in chip manufacturing. A ROHM engineer told EE Times Japan that wringing the most out of limited resources had always been a core struggle, and that a method which could find near-optimal answers fast from a tangle of constraints was, for the company, a real find.

That EDS win was the proving ground. The question was whether the same idea could survive the jump to the front-end, where the constraints aren't fiddly — they're overwhelming.

What "quantum annealing" actually does

Quick detour, because the word "quantum" gets stretched in two very different directions.

The quantum computers you see in the news — Google's, IBM's — are gate-based machines. They're general-purpose: in principle they can run any quantum algorithm, which is why people dream about them cracking encryption or simulating molecules. The catch is they're maddeningly hard to scale and keep stable, which is why most of that work still lives in the lab.

Quantum annealing is the other branch, and it's a narrower beast. It can't run arbitrary algorithms. It does essentially one thing: find the lowest-energy arrangement of a system, which mathematically is exactly the shape of a hard optimization problem. Translate "what's the best production schedule?" into an energy landscape, and the best schedule is the valley floor. That single trick happens to fit factory scheduling like a glove.

Quanmatic's approach is also hardware-independent. It leans on quantum-classical hybrid algorithms born from annealing research, rather than requiring one specific exotic machine humming in the basement. The point isn't the hardware. The point is the formulation: turning a messy operational decision into a problem this kind of math can chew on.

Onto the messy core

With the EDS recipe in hand, the team built a front-end prototype in 2025, validated it in a pilot at ROHM's Hamamatsu site that April, and switched on full production use at the Miyazaki plant in January 2026. It's still running today.

The mechanism is almost unglamorous. The system continuously recomputes the optimal order for lots and machines, reflecting the factory's real-time state, and recalculates on a set rhythm as conditions shift. That trims the "waiting for people, waiting for parts" dead time that quietly drags a line down. Fewer idle gaps means more finished wafers out the door, and that's where the roughly 3% in production efficiency comes from.

No new machines. No new cleanroom. Just a smarter answer to the question the factory was already asking every hour: what next?

Two races, two finish lines

Step back, and you can see two very different quantum stories unfolding at once.

One is loud. Google's Willow chip crossed a long-awaited error-correction threshold and, in late 2025, claimed a verifiable "quantum advantage" on a benchmark problem; IBM is mapping a path to fault-tolerant machines by the end of the decade; China is pouring resources into its own programs. This race is about raw capability: proving the machines can eventually do things no classical computer can. The genuinely practical payoffs are only just starting to be documented, and most of the action is still about scaling and stabilizing the hardware.

The other story is quiet, and it's the one in Miyazaki. It doesn't promise a general-purpose miracle. It does one narrow thing — optimization — and it does it well enough to run, unsupervised, on a working factory floor and pay for itself. ROHM says it now plans to extend the system to more processes and more sites across the group.

It's a very Japanese kind of breakthrough: not a record on a benchmark, but a few percent earned on the night shift, every night. Whether that counts as "quantum advantage" depends on what you came for. A physicist might say the real prize is still over the horizon. A plant manager might say the prize already showed up in January, and it's measured in wafers.

The headlines will keep going to the lab. But the first place quantum quietly pays its way might just be a factory near you. Where do you think it'll show up first where you live — in a research center, or on a production line?

References