The Quantum Scoreboard Changed: Now It Counts Errors, Not Qubits

On a clear night, the data came back beautiful. Forty-eight logical qubits holding a line through operation after operation, error after error caught and returned to the ledger. Deep time has a way of settling arguments, and the argument being settled right now is whether quantum computers are real machines or a permanent promise. Watching the numbers arrive through this summer, I have been struck less by any single record than by something quieter: the field has changed the instrument it measures itself with. Astronomers know how disorienting that is — the night sky does not move, but when you switch your telescope from a camera to a spectrograph, suddenly a pinprick of light becomes a story about temperature and age and distance. The stars were always there; the instrument revealed what mattered about them.

For a decade the headline metric was simple and satisfying: physical qubits, and more of them. Fifty, then a hundred, then a thousand in some designs. It was the kind of number you could put on a slide and feel the room nod. But a physical qubit is not a working bit the way a shareholder counts shares; it is a raw material that leaks. Left to itself, it drifts into noise within microseconds — the technical word is decoherence, which is a polite way of saying that the system is forgetting itself. To the engineers who actually build these machines, the difference between owning a qubit and trusting a qubit is the entire subject. What the industry learned to do, painfully, is to take many noisy physical qubits and knit them into one logical qubit that can actually be relied on. The transition happened gradually enough that nobody noticed when the true scoreboard changed. It changed to this: how few errors survive.

When the scoreboard changed

Let me think about what that shift means, because it is the whole story in one move. A logical qubit is an act of engineering, not a piece of hardware. You encode it across several physical qubits and run error correction around it, so that a single failing atom cannot destroy the computation. The cost is brutal — dozens of physical qubits to protect one logical bit — but the payoff is the only thing that matters: the error rate drops by orders of magnitude. Two groups demonstrated the point this summer, and the numbers are worth reading slowly, the way you read a spectrum when you are not sure yet what it is telling you.

The first group works with ion traps, which are exactly what they sound like: individual atoms held in electric fields, each atom a qubit, each one given a near-silent environment in which to think. Quantinuum’s machine in this family is called Helios, built around 98 physical qubits. On top of that hardware they demonstrated 48 logical qubits with an average two-qubit gate fidelity of 99.921 percent. A gate at 99.9 percent sounds like a rounding error until you remember what is being asked of it: a computation that touches a million gates multiplies those tiny imperfections into catastrophe. Fidelity is compounded, and compounding punishes the small defect. Then, in August, the same group reported logic fidelity closing in on five nines — 99.999 percent territory. I should say plainly that I am not auditing their instrument stack from here; the claim is theirs, cross-checked by two independent channels in the reporting I have read. But even a skeptic has to admit the direction is striking, and direction matters more than any single reading.

The other route to the same summit

The second group travels a different road. IBM builds superconducting circuits — electric currents on chips, cold enough to behave quantum mechanically — and has bet its roadmap on a target few would have risked a few years ago. Working with the University of Chicago, on July 30 IBM reported 70 encoded logical qubits completing 2,415 logical two-qubit operations and 468 logical T gates. The T gate is worth pausing over. It is the hardest gate to protect because it resists the usual error-correction tricks, which means it is the best test of whether the correction is actually working or merely theatrical. Every computation that matters eventually needs T gates, so 468 of them executed successfully is not a cosmetic achievement; it is evidence that the machinery performs under its least friendly workload.

IBM calls this set of operations a foundational criterion for quantum advantage — the threshold where a fault-tolerant machine begins to look like a machine rather than a laboratory. Their stated target is 2029: a system called Starling that would support 200 logical qubits and one hundred million quantum operations. I have been in this field long enough to know that roadmaps slip; I am more interested in the fact that the roadmap now exists in that form at all. Five years ago the honest statement was “we do not know how to get there.” Today the honest statement is “we know how, and we are arguing about the schedule.” That is progress of a specific, measurable kind — the kind that shows up in a ledger rather than in a press release.

Suppression, not just possession

The word I keep circling back to is suppression. Possession of qubits is inert; suppression of error is progress. A single data point from the industry’s daily reporting in late August makes the point sharply. Researchers working with Google’s Willow processor optimized the timing of measurements and reduced logical error rates by up to 40 percent. Let me slow that down, because it is easy to skim past the most interesting sentence of the summer. No new hardware. No bigger machine. The change was in the ordering of when and how you look at the system — which measurements happen first, how the system is interrogated mid-computation — and it bought a 40 percent reduction in logical errors. That is a reminder that we are early enough in this discipline that engineers are still finding errors inside their own error correction, and it is a good sign, not a bad one. The remaining waste is a design problem, and design problems have a way of being solved. Or rather, they have a way of being solved when the incentives line up — and this year, the incentives are lining up faster than the roadmaps.

This is the moment to be honest about the difference between a demonstration and a product. Every number above is a laboratory result measured under controlled conditions, not a service you can bill at the end of the month. Quantinuum plans to offer the Helios system through Oracle Cloud Infrastructure, which is a serious commercial step — the cloud is where enterprise customers will first meet these machines — but availability is not the same as usefulness. IBM’s own schedule concedes the distance: a foundational criterion now, a fault-tolerant machine by 2029. The long view says the same thing it always says about deep technology: the timeline is longer than the headlines and shorter than the skeptics claim. Both statements are true at once, and the discipline is to hold both.

The institutions are reading the same sky

It is worth stepping back from the chips to the institutions, because they are drawing the same conclusion from the same data. On August 25 the United States Treasury announced a quantum readiness working group, established under Executive Order 14412, charged with coordinating the federal financial system’s migration to post-quantum cryptography. Translate that into plain language: the people responsible for moving trillions of dollars are no longer treating quantum computers as a curiosity for physics seminars. They are treating them as a dated threat to encryption and as a schedule to plan around. A working group is the bureaucratic form of the long view — nothing glamorous, but the ledger gets kept. The same instinct that drives the physicists — how do we make the error rate go down — reappears at the Treasury as: how do we make the vulnerability go away before the machines arrive. Encryption that takes today’s computers centuries to break might fall to a fault-tolerant machine in hours; the migration to post-quantum cryptography is, in effect, the financial system buying insurance against a date it cannot see but can now estimate.

None of this proves the machines will meet the 2029 schedule. It proves something more modest and, in its way, more durable: the people who build the machines and the people who depend on them now share a working assumption that fault tolerance is arriving. When a government coordinates a cryptography migration on that assumption, the assumption has begun to harden into infrastructure planning. Institutions rarely move early; they move when the evidence is adequate. Their movement this summer is itself a data point, and I would weigh it alongside the gate fidelities.

What the long view actually settles

Let me close with the perspective I try to keep when the hype cycle swings. I have watched technology forecasts for long enough to trust only one habit: check what the measurements say, not what the announcement says. By that habit, this summer’s data is genuinely good. A 40 percent error reduction from timing alone, a system approaching five nines, seventy logical qubits executing hundreds of protected gates — these are not press-release adjectives. They are numbers that would have been dismissed as fiction five years ago. Picture the ledger concretely: 2,415 operations, each one checked and corrected, the errors marked in red and then struck through, until the struck-through lines outnumber the surviving ones by orders of magnitude. That image, more than any single record, is what changed this year.

But measured wonder has a discipline to it. The discipline says: record the 99.921 percent, and also record that no one has yet run a commercially valuable computation on these machines. Record the Treasury working group, and also record that a working group is a beginning, not an outcome. Evidence keeps the awe honest; it does not forbid the awe. I still find it remarkable that a machine can hold a single logical bit stable long enough to be useful, and that the path to useful machines now looks like an engineering problem rather than a physics miracle. The poetry and the spreadsheet are not competitors here; they are the same act at different magnifications.

Deep time has a way of settling arguments, and the argument over whether quantum computing is real has, I think, been settled — not by a single record but by a steadily improving ledger of suppressed errors. What remains is the ordinary, unglamorous work of making the error rate fall further, operation by operation, gate by gate. That is the long view, and it is the only honest one. The scoreboard no longer counts qubits. It counts how many mistakes survive the night.