Five Errors per Million: The Quantum Record That Rewrites the Qubit Budget

Start with the number: five errors per million operations. On October 19, Quantinuum announced that its commercial Helios quantum computer held an error-corrected logical qubit at a fidelity of 99.9995 percent, the best quantum error-correction result ever published and the company’s second record in two months (press release), following September’s Helix-code record on the same 98-qubit trapped-ion hardware, this time using triangular color codes to cut the error rate another tenfold.

A 22-page preprint lets you do better than repeat the press release: Quantinuum published the full distance-scaling data in “Quantum error correction well below the color code threshold,” making the central claim checkable arithmetic. So we checked it, caveats and all, plus one clock-speed calculation that reframes the race.

First, the raw data, because this is where the story actually lives, in three rows the press release summarized and the preprint documents:

Code distance	Physical qubits	Logical error per round
d = 3	13	5.5 × 10−4
d = 5	37	3.8 × 10−5
d = 7	73	5.0 × 10−6

Each step up in code distance divides the remaining errors by 12. That suppression factor, Λ = 12, is the paper’s most important number. Google’s Willow chip measured Λ ≈ 2.14 on its surface code (Nature, December 2024), falling from 0.65 percent per cycle at distance 3 to 0.143 percent at distance 7, which means every distance step buys Quantinuum roughly six times more error reduction than Google, a gap that becomes a chasm when compounded over several steps.

Begin with the vendor’s boldest claim, which we verified from the published suppression factors: Quantinuum says its approach could support 100 logical qubits on about 30,000 physical qubits, versus roughly 900,000 for the superconducting implementation it analyzed (HPCwire). For a triangular color code, the data-qubit count is (3d²+1)/4, which reproduces all three published points exactly. Target a logical error of 10⁻⁹ per cycle, enough for billion-operation fault-tolerant workloads. Quantinuum’s curve reaches it at distance 15, costing 338 physical qubits per logical qubit, or about 33,800 for 100 logical qubits against the vendor’s figure of 30,000: our independent re-derivation lands within 13 percent of their number, and the claim survives.

The same exercise on Google’s Willow parameters gives distance 45 at 4,049 physical qubits per logical qubit, about 405,000 for 100. Quantinuum’s own figure was 900,000, assuming a color-code implementation on superconducting hardware plus overhead our simple estimate omits. Same order of magnitude. Honest conclusion: the 30-fold qubit-count advantage is real arithmetic, and it comes almost entirely from Λ = 12 compounding against Λ = 2.14.

One more check the press release left on the table. It claims the result is 100 times better than Google’s best published number. At distance 7, Google measured 1.43 × 10⁻³ per cycle against Quantinuum’s 5.0 × 10⁻⁶ per round, a ratio of about 290 to 1. Press releases rarely undersell themselves.

Then comes the first caveat: everything above concerns quantum memory, while computation is gates, and Quantinuum’s measured logical gate error is 7.0 × 10⁻⁵ per two-qubit Clifford gate at distance 5, ten times worse than the memory number. John Preskill’s “megaquop” milestone, one million reliable operations, needs per-gate error near 10⁻⁶. With the fitted gate suppression factor of 9 per step, distance 7 reaches only 7.8 × 10⁻⁶; distance 9 crosses the line at 8.7 × 10⁻⁷ for 122 physical qubits per logical qubit. The catch: that two-block benchmark needs 244 physical qubits and Helios has 98, so Quantinuum cannot run the experiment that would verify its own extrapolation. Worse, the gate suppression factor is fitted from just two distances, with error bars of +8/−4. The memory record is rock solid; the gate story is a forecast wearing a lab coat. One is measured. The other is modeled.

Now the second caveat, the one that reframes the race: trapped-ion gates are slow, with a two-qubit ion gate taking roughly 100 to 200 microseconds against tens of nanoseconds on a superconducting chip. That is a thousandfold gap in clock speed, which no error-correction code closes, since an algorithm’s gate count is fixed by mathematics. Take the standard reference estimate for breaking RSA-2048, about 20 million noisy qubits running for 8 hours on superconducting hardware (Gidney and Ekerå, 2021), and run the same circuit a thousand times slower: 8 hours becomes roughly 8,000 hours, or eleven months per key. That is the trade.

Treat that as a Fermi estimate, good to an order of magnitude either way. But the shape is honest: Quantinuum has won the qubit-efficiency war while losing the clock-speed war. Ion traps need far fewer qubits and far more time, which favors quantum chemistry and punishes anything time-bound. Fault tolerance is a race in logical operations per second per dollar, and on that metric the contest is wide open.

Finally, the money, because a $15 billion valuation demands it. Quantinuum went public in June at $60 a share, raising $1.68 billion (deal details). Against 2026 revenue guidance of $28 to $32 million, that is roughly a 500-times price-to-sales multiple, which prices in something close to a decade of flawless execution. The balance sheet’s real story is the war chest: $2.1 billion in cash against an adjusted EBITDA burn of about $68 million a quarter, roughly 7.7 years of runway at the current pace (Q2 results; 10-Q). Revenue is lumpy: first-half 2026 revenue of $13.2 million was down from $21.2 million a year earlier on contract timing. Markets are not buying revenue; they are buying the physics, plus seven years of funded attempts to turn physics into product. Cash is the moat.

The Strongest Case Against This Article

The superconducting rebuttal, at full strength: Google’s Willow runs gates a thousand times faster, IBM’s Starling roadmap targets 200 logical qubits by 2029 on the faster modality, and real algorithms are time-bound. A machine needing eleven months to break one RSA key is strategically a different animal from one needing eight hours. Every record here covers one or two logical qubits of memory; useful computation needs thousands doing gates, and the only published gate data stops at distance 5 with truck-sized error bars. No peer review yet, the 900,000-qubit baseline is the vendor’s own choice, and this is the second record in two months, which is either genuine acceleration or a well-managed drumbeat. None of that erases Λ = 12, though it does make the coronation premature, because qubit efficiency is only one axis of a multi-axis race and the axis Quantinuum trails on is the one measured in seconds.

What This Does Not Prove

Here are the honest bounds. We have not read the preprint itself; its numbers come through the press release and technical reporting, and the headline 5 × 10⁻⁶ carries a one-sigma band of 3 to 9 × 10⁻⁶. Our 30,000-qubit audit verifies the vendor’s arithmetic from the vendor’s parameters, not the underlying experiment. The clock-speed-tax math is a Fermi estimate, not a simulation, movable an order of magnitude either way. The 500-times multiple uses IPO pricing, not current market cap. And Helios is a 98-qubit machine: everything about 100 logical qubits, Apollo in 2029, and “less than a million physical qubits” is extrapolation, not demonstration.

What You Can Do

Four moves, depending on who you are. If you run security infrastructure: the clock-speed tax buys you time, but not complacency. An ion-trap machine needing months per RSA key still validates harvest-now-decrypt-later; keep migrating to NIST post-quantum standards. This result shortens no deadline, but the physics risk is real and compounding. If you invest: a 500-times multiple prices in Apollo executing on schedule, so the $2.1 billion cash pile is the downside protection. Compare against IonQ and Rigetti multiples first. If you are choosing what to study: quantum error-correction code design is now the scarcest high-leverage skill in the field. Λ is the metric that moved, sixfold in a single paper. And if you just follow the field: watch for the peer-reviewed preprint, a distance-7 logical-gate demo (needs 146+ qubits, so watch Helios upgrades or Apollo hardware), and whether Λ holds at larger distances, where threshold behavior has a habit of breaking.

The Bottom Line

Five errors per million is a genuine record, and the arithmetic behind it checks out: a suppression factor of 12 per distance step compounds into a real 30-fold qubit-count advantage over superconducting approaches, verified from the published numbers. Memory, however, is not computation, and the gate record needs one more distance step on hardware that does not exist yet, the thousandfold clock-speed gap means fewer qubits at the cost of far more time, and a 500-times sales multiple means the market has already priced the physics as solved. Physics is ahead of schedule; everything else, gates, machines, revenue, is still a forecast. Watch the gates.

Related

• Quantum’s Error-Correction Bill Just Fell 97%: 3,000 Physical Qubits Instead of 100,000
• Quantinuum Raised $1.68 Billion at a 410× Revenue Multiple. Two Days Later, It Fell Below Its IPO Price.
• A Quantum Computer Just Cracked a 15-Bit Key. Bitcoin Uses 256. The Math Between Them Is Changing Faster Than Anyone Expected.