Your Worst Qubit Decides What Your Computer Can Do

Frontier Signal · Deep Dive · Quantum & Investing · 8 min read · Wednesday, 7 October 2026

Take two qubits on the same chip. Same design, same wafer, same fabrication run, same cooldown. One holds its quantum state for 77 microseconds. The other holds it for 139.

Nobody could reliably say why. Every superconducting quantum company lives with this: you build a chip, you measure it, and the qubits come back uneven for reasons buried in the variables of fabrication. Teams have chased what they can see down a microscope.

A collaboration out of Fermilab's Superconducting Quantum Materials and Systems Center set out to settle it. It did something unusual: it kept the answers hidden.

22 QUBITS · 7 CHIPS · 10 CHARACTERISATION TECHNIQUES
Qubit performance was withheld from the characterisation team until measurements had concluded. Every finding below is a correlation across 22 devices, not a before-and-after build.

The study is "Identifying materials-level sources of performance variation in superconducting transmon qubits", in Applied Physics Reviews (volume 13, issue 2, article 021416). It has been public as a preprint, arXiv:2503.14424, since March 2025; Fermilab publicised it on 1 October 2026.

We say so because almost every account this past week has implied the work is days old. It is not. The science is unchanged, and the industry has spent the interval scaling the processes it examines.

The collaboration: Fermilab, Rigetti Computing, NIST Boulder, Ames National Laboratory, Iowa State University, Northwestern University and the United Kingdom's National Physical Laboratory, twenty-two authors. Seven chips, from the Pritzker Nanofabrication Facility and Rigetti's Fab-1, mostly bare niobium, at least one pairing niobium with tantalum.

One disclosure. The publisher's page refuses automated access, so the text used here is the preprint.

What a transmon is. A small circuit of superconducting metal, usually niobium, patterned on silicon and cooled to a hair above absolute zero. A current then sloshes back and forth in a pattern that holds quantum information.

What coherence is. The pattern does not last. T1 measures the time before the qubit's energy leaks away; the paper leans on the quality factor, how well the circuit holds energy rather than losing it.

Where the energy goes. Into imperfections you cannot see: the few-nanometre skin of oxide that forms on niobium the moment it meets air, and the shape of the trench etched beside it.

What the study did. Measure twenty-two qubits. Separately, have other teams examine the same devices with ten techniques, from optical and electron microscopy to mass spectrometry. Correlate only afterwards.

Know which qubit is the good one and you will find reasons it is good. Characterisation produces enormous data about any sample, and a human looking for an explanation finds a plausible one. That is why clinical trials are blinded.

Withholding the results turned plausible stories into testable correlations, and it is why the negative finding carries weight: a blinded study reporting that the defects everyone polices do not predict performance is far stronger than an unblinded one.

Surface oxide thickness. For bare niobium qubits, a thicker oxide at the top surface trends with a lower quality factor. A few nanometres of material, forming spontaneously, invisible to the inspection steps a fab runs.

Etch sidewall angle. Where the metal meets the substrate, the wall can be steep or taper. Increases in that angle, a smoother decline to the substrate, trend with a reduction in quality factor. Controlling one edge is an etch-recipe change, not a new technology.

Trench depth. Deeper trenches trend with better quality factors, saturating above roughly 70 nanometres. A target rather than an instruction to keep going.

What does not matter. Scratches, stray particles and the other macroscopic defects inspection catches showed no clear statistical correlation. Teams have been fixing the problems they could see.

The spread, in real numbers. One bare-niobium chip ranged from 77 to 139 microseconds. A niobium-and-tantalum chip held 184 microseconds consistently. Uniformity is a materials and process property, and it is achievable.

Every logical-qubit announcement rests on an assumption rarely examined. Error correction spreads one logical qubit across many physical ones, and how many you need depends on how good the physical ones are. Not the best. The typical ones, and in practice the worst. We walked through that arithmetic in the year quantum computers learned to fix their own mistakes.

Andrew Bestwick, Rigetti's chief technology officer and a co-author, framed it plainly: as processors scale, performance is set not by the best qubits but by the worst, because one bad qubit can drag down any computation that depends on it.

So a paper about oxide thickness matters more than most qubit-count records. A thousand qubits with a wide spread is not a thousand usable qubits. Narrowing it is a manufacturing problem.

Rigetti Computing (NASDAQ: RGTI) is the cleanest listed read, a roadmap story not an earnings one. It is a named co-author and one of the two fabrication sites, so the findings sit inside the company that would apply them. Better uniformity means higher usable qubit counts on future systems, not revenue next quarter. Treating it as a trading catalyst misreads it.

IBM (NYSE: IBM) and Alphabet (NASDAQ: GOOGL) run the other large superconducting programmes and are too diversified for a materials result to register financially. Both own their fabrication and can act quickly.

Process tools and metrology are where the money would eventually go: etch geometry, surface passivation and measuring a few nanometres of oxide. None of those firms discloses revenue tied to quantum fabrication, and quantum volumes are trivial next to logic and memory, so it is the destination of the engineering, not exposure today. The adjacent read is manufacturability itself: a startup that made working qubits on a commercial 300 millimetre line.

National laboratories are the quiet winners. Fermilab, NIST, Ames and NPL supplied capabilities no single company has. Superconducting qubits now need instruments that live in public institutions.

An industry read-through, not a claim that any of these companies uses these findings. Quantum foundries and fabrication services are most affected: if oxide and etch geometry are the controlling variables, holding them to specification across a wafer becomes the product. Cryogenic and control-electronics suppliers are unaffected, worth naming because much quantum investment went into equipment around the chip while the constraint sits on it. Materials and surface-treatment specialists are the least glamorous and most implicated, and mostly private.

Nothing in this paper makes a single qubit better. It tells people where to look.

Every finding is a correlation across twenty-two devices, a small sample for claims this specific, as the authors say. A trend across existing chips is not the same as building two batches that differ only in sidewall angle. Until somebody does that, these are strong hypotheses, not validated specifications.

The scope is narrow: niobium transmons, with one tantalum-paired chip. It says nothing reliable about fluxonium, other superconducting designs, or the competing modalities, trapped ions, neutral atoms, spins in silicon, or a machine built around light.

And the trends are trends. The 70 nanometre saturation point is useful, but none of this arrives as a number a fab can dial in. Turning it into a recipe is months of work inside each company.

If you rebuild to the oxide, sidewall and trench targets this study points at, does your quality factor actually move?

That is the difference between a correlation study and a process improvement, and the thing to ask any vendor citing this work. An excellent map is not a destination.

It measured twenty-two transmon qubits on seven chips from two facilities, then had separate teams examine those devices with ten techniques without being told how each qubit had performed. The correlation came afterwards.

Three nanometre-scale features track qubit quality: thicker top-surface niobium oxide is worse, a larger etch sidewall angle is worse, and a deeper trench is better to around 70 nanometres. Visible defects showed no clear relationship.

Because error correction needs many physical qubits that are all decently good, and a chip's usefulness is set by its worst qubits rather than its best.

No. It appeared in Applied Physics Reviews earlier this year and has been public as a preprint since March 2025. Fermilab publicised it on 1 October 2026.

Not yet. These are correlations across twenty-two devices, not a demonstrated improvement. The confirming experiment has still to be run.

The journal page refuses automated access; the arXiv preprint is the text used.

Frontier Signal explains frontier technology in plain English. Research findings reported as correlations should not be read as validated process specifications. This is general information, not investment advice.