Five Silicon Qubits Ran on One Microwave Line. The Calibration Math Is the Real Breakthrough.
A RIKEN team drove five silicon spin qubits through a single shared microwave line and kept single-qubit fidelity above 99.99 percent with three running at once, falling to about 99.9 percent with all five. Pairwise calibration made that possible, and the arithmetic behind it explains both why silicon scaling just got easier and why it is still nowhere near done.
99.9 percent sounds like a rounding error away from 99.99 percent. It is not. Not even close. Run a 1,000-gate circuit at 99.9 percent fidelity per gate and only 36.8 percent of runs come out clean, assuming independent errors. Run that same circuit at 99.99 percent and 90.5 percent come out clean. Each extra nine buys a tenfold longer circuit at the same success rate, which is why a result published in Nature Physics on October 15, 2026 deserves a closer read than its headline fidelity suggests.
Researchers at RIKEN's Center for Emergent Matter Science, working with colleagues at National Taiwan University, ETH Zurich, and QuTech at Delft University of Technology, built an array of five silicon spin qubits and drove all five through one shared microwave line. Individually, every qubit cleared 99.99 percent fidelity for a primitive half-turn gate, with the best approaching 99.999 percent. With up to three qubits firing together, fidelity stayed above 99.99 percent. With all five firing together, it landed at about 99.9 percent, a level the paper calls the practical threshold for fault-tolerant operation. RIKEN's own press release frames this as an important step toward larger silicon processors, and on the control problem alone, that framing holds up. Project manager Seigo Tarucha, who leads the silicon quantum computing project under Japan's Moonshot Goal 6 programme, states the long-term goal plainly as a fault-tolerant computer built on silicon spin qubits, with reliable parallel control as an essential step toward that goal rather than as a destination in itself.
Why sharing a line used to break everything
Silicon spin qubits store information in the spin of an electron trapped in a silicon quantum dot, and they are attractive precisely because silicon fabrication is the most mature manufacturing technology on Earth. Scaling them, however, runs into a plumbing problem that has nothing to do with quantum weirdness and everything to do with wires. Each qubit historically needed its own impedance-controlled microwave line to deliver control pulses, and those lines are among the hardest components to cram into a cryostat, as the paper itself notes. Give every qubit its own line and a million-qubit machine needs a million lines, plus the feedthroughs, amplifiers, and heat load that come with them.
Sharing one line among several qubits dodges that ceiling, but it introduced a different failure, one that only shows up when pulses overlap in time and neighbouring qubits start listening to signals that were never meant for them, which is exactly what a working computer forces them to do. When a microwave pulse meant for one qubit leaks into its neighbours, it shifts their energy levels through an effect called the AC Stark shift, which changes the phase of their quantum states and scrambles the timing of their operations. Run qubits one at a time and the problem never appears. Run them together, as any real computer must, and accuracy used to collapse. That collapse is the reason parallel control, not single-qubit fidelity, has been silicon's bottleneck. Control was the bottleneck.
RIKEN's device gives the problem concrete dimensions. Five quantum dots sit in an isotopically purified silicon-germanium quantum well, defined by three overlapping aluminium gate layers, with two charge-sensor dots at the ends for readout and a cobalt micromagnet on top that separates neighbouring qubit frequencies by about 150 megahertz. Idle qubits dephase in about 10.3 microseconds on average, which is short, while driven qubits hold coherence far longer, which is why the team minimised idle time between gates to just 2 nanoseconds and replaced idle operations with driven ones. Shaped Kaiser-window pulses, chosen because their spectrum stays off neighbouring resonances, lifted individual fidelities from about 99.98 percent with conventional rectangular pulses to well above 99.99 percent. Pulse shaping matters here because a rectangular pulse carries spectral sidelobes that spill onto neighbouring resonances, while a Kaiser window suppresses that spill, trading a taller peak drive for a cleaner spectrum and thereby buying back the fidelity that crosstalk would otherwise destroy.
Pairs beat combinations, and the numbers prove it
Here is the genuinely new idea. Instead of calibrating every possible combination of simultaneously driven qubits, first author Yi-Hsien Wu and co-first author Leon Camenzind calibrated pairs. AC Stark phase shifts from different drives turn out to be independent and additive, so measuring how much qubit A drifts when qubit B is driven, and vice versa, predicts the drift when both drive together. No separate procedure for every group is needed.
How much does that save? Run the paper's own rule, N(N-1) pairwise calibrations for N qubits, against exhaustive calibration of every non-empty combination, 2^N minus 1.
| Qubits (N) | Pairwise calibrations N(N-1) | Exhaustive combinations 2^N - 1 |
|---|---|---|
| 5 | 20 | 31 |
| 10 | 90 | 1,023 |
| 100 | 9,900 | About 1.3 x 10^30 |
| 1,000 | 999,000 | About 10^301 |
At five qubits, pairwise calibration is barely a win: 20 measurements instead of 31. At ten qubits it is an 11-fold saving. At one hundred qubits, exhaustive calibration exceeds the number of atoms in a human body by orders of magnitude, while pairwise calibration needs under ten thousand measurements. Quadratic scaling is not a slogan here; it is the difference between a calibration schedule and a fantasy. Fantasy, not schedule.
Quadratic still bites, though, and this is the calculation nobody in the announcement ran. For 1,000 qubits, pairwise calibration means 999,000 measurements. At a purely illustrative one minute each, done serially, that is about 694 days of calibration before the machine computes anything. RIKEN's actual calibration time and parallelism were not disclosed, so 694 days is an illustration of scaling, not a measured schedule. Fortunately, the paper notes that corrections become negligible for sufficiently detuned qubit pairs, which means many of those 999,000 pairs can in principle be skipped. Which pairs, and how the skipping rule holds up in a real thousand-qubit device with fabrication variation, is now the next scaling question, not whether pairs beat combinations.
Wiring tells a parallel story. Group five qubits per shared line and a million-qubit machine needs 200,000 lines instead of 1,000,000, saving 800,000 lines, each of which would otherwise need its own cryostat feedthrough, its own amplifier chain, and its own slice of the heat budget that already strains every cryostat in the field. Group ten per line and it needs 100,000. Those are extrapolations, not RIKEN claims, and grouping size at that scale is unproven, but they show what is actually at stake: line count sets the cryostat and packaging ceiling long before qubit physics does.
What five qubits still cannot do
Now for the number that should temper every celebration. In this same device, two-qubit CZ entangling gates are limited to below 97 percent fidelity, because increasing the exchange coupling into the megahertz range triggers spin-valley mixing. That figure comes from the paper itself, in its discussion of device operation, not from a critic's blog. Single-qubit steering at 99.99 percent alongside entangling gates below 97 percent is a lopsided machine. Steering is not computing. Surface-code error correction typically needs physical fidelity above about 99 percent, with 99.9 percent as the practical target under realistic noise, according to the paper's own introduction. A 97 percent entangling gate fails both bars.
Put plainly: this device can steer five qubits beautifully and entangle them badly. No error correction, no logical qubit, and no useful algorithm has been demonstrated on it, and the authors do not claim otherwise. Anyone who quotes the 99.99 percent figure without the sub-97 percent CZ figure is describing a steering wheel and calling it a car.
One more nuance belongs in the open. RIKEN's abstract rounds five-qubit parallel operation to the 99.9 percent threshold, while table-level values reported from the paper's extended data put four of five qubits above 99.9 percent at longer gate times and the fifth at about 99.8 percent. That gap is small, and it does not overturn the result, but it means the honest sentence is "about 99.9 percent, with one qubit slightly below at longer gate times," not "all five above 99.9 percent." Precision about a tenth of a percent is the whole game in this field, so rounding it away would be a category error.
Limitations
Several limits qualify everything above. First, the circuit-success arithmetic assumes independent, identical gate errors, and the moment you allow errors to correlate across neighbouring qubits, as real microwave crosstalk encourages them to do, the clean exponentials in the table become optimistic ceilings rather than predictions, with leakage making real circuits worse still. Second, the 694-day calibration illustration assumes one minute per calibration and serial execution, neither of which RIKEN disclosed, so treat it as a scaling illustration only. Third, the million-qubit wiring arithmetic extrapolates a grouping size that has been demonstrated only at five, and fabrication variation across a real wafer could break the neat pairwise additivity that makes this scheme work. Fourth, no outside lab has independently reproduced this result; Nature Physics peer review is the only external check so far. Fifth, the remaining five-qubit loss comes from drive-induced decoherence caused by higher microwave power, an effect pairwise calibration cannot cancel, and the authors point to micromagnet redesign as the mitigation, which is a future device, not a demonstrated one, so any roadmap that assumes five-qubit fidelity will simply persist at larger arrays is assuming away the very decoherence mechanism that already caps the current device.
The Strongest Case Against
Consider the strongest case against reading this as a scaling breakthrough at all. Superconducting qubits already run hundreds of qubits with two-qubit gates far above 97 percent, and trapped-ion systems run dozens with two-qubit fidelity above 99.9 percent. Silicon's manufacturing advantage is real, but it is an advantage in making qubits, not in making them compute, and a five-qubit demo with sub-97 percent entangling gates does not close that gap. Worse, the shared-line trick solves a problem that only matters once you have thousands of qubits to wire up, while silicon's actual bottleneck today is entangling gates and readout at even ten qubits. Solving tomorrow's wiring problem before today's entangling problem is progress, but it is progress in the wrong order if it distracts funding and talent from the gate fidelity that actually determines whether error correction ever starts.
That objection has force, and it should change how readers weight this paper, because a field that celebrates control milestones while its entangling gates lag can easily mistake a solved subproblem for a solved system, and funding follows headlines far faster than it follows the full error budget. It does not erase the result. Shared control was a known, named blocker for silicon scaling, and RIKEN removed it for the single-qubit case with a scheme whose cost grows quadratically rather than exponentially. A blocker removed is a blocker removed, even if a larger blocker remains standing behind it. It removed one.
What You Can Do
If you evaluate a silicon quantum pitch, demand three numbers together and refuse to price the pitch on one alone: simultaneous single-qubit fidelity, two-qubit CZ fidelity, and qubits per shared control line. Demand all three. A deck that shows only the first number is hiding the other two, and after RIKEN, you know exactly which questions to ask. If you are a student choosing a specialisation, pairwise calibration and shaped-pulse control are transferable skills that apply to every microwave-driven qubit platform, not just silicon. If you track quantum computing for security policy, keep this result in perspective: five well-steered qubits with sub-97 percent entangling gates pose no near-term threat to RSA encryption, and Shor-scale factoring remains far away on this platform. Watch instead for the next paper from this group: a CZ gate above 99 percent in a shared-line device would matter more than another decimal place on single-qubit fidelity.
The Bottom Line
RIKEN did not build a quantum computer. It removed a specific, quantified obstacle to building one in silicon, an obstacle that had been named in paper after paper as the reason shared control could not be trusted at scale, and the numbers in this article tell you exactly how much that removal is worth, in lines saved, in calibrations avoided, and in circuit depth gained or lost. Pairwise calibration turns an exponential calibration nightmare into a quadratic schedule: 20 measurements at five qubits, 9,900 at one hundred, 999,000 at one thousand, against exhaustive counts that leave the realm of the possible before one hundred. At the same time, dropping from 99.99 percent to 99.9 percent when the fifth qubit joins costs a full order of magnitude of usable circuit depth, and entangling gates below 97 percent mean error correction cannot even begin on this device. Hold both facts at once and the story is clear: silicon's wiring problem just got a credible answer, while its computing problem, the entangling gate, is still waiting for one.