← OrreryThe Ecliptic is an Orrery publication

Cover Story · The Reality Gap desk

A Logical Qubit Is Not a Unit

One 98-ion machine was credited this year with 94, 48, 2 and 1 logical qubits, and all four counts are true. A register of nine published demonstrations shows what each count was bought with, what the US Department of Energy's new 100-logical-qubit prize implies in error and hardware, and the column no count reports: time.

94 → 1
logical qubits credited to one machine, Quantinuum's 98-ion Helios, across three 2026 papers with no upgrade between them; the error per correction round fell to 5 × 10⁻⁶ as the count fell
~10⁻⁷
logical error per qubit per operation implied by the DOE's illustrative first-generation target of 100 logical qubits and 10⁵ hard operations, about 50 times below the best published memory result (The Reality Gap estimate)
18,100 vs 217,700
physical qubits to hold that target in logical memory at the best published suppression factors, Λ = 12 (trapped ion) and Λ = 2.14 (superconducting); excludes magic-state factories and routing (The Reality Gap estimates)
~180,000×
how much slower per correction round the qubit-efficient route runs: about 0.2 s on Helios against 1.1 µs on Google's 2024 surface-code result

The register

The count and the quality move in opposite directions

110100Helios, iceberg (Feb)1.04IBM / U. Chicago (Jul)1.39Helios, concatenated (Feb)1.67Microsoft / Atom (2024)2.0Infleqtion Sqale (Sep)2.67Helios, Helix d=6 (Sep)10 (data qubits)Helios, colour code d=7 (Oct)73Google, surface code d=7101Physical qubits per logical qubit, log scalepost-selected (runs discarded)repeated correction
Physical qubits per logical qubit for eight of the nine register entries, on a log scale (the Harvard-led entry, which reports several configurations, is omitted). Teal results kept runs by post-selection; navy results ran repeated correction without discarding runs. The Helix ratio counts data qubits only. Ratios are The Reality Gap estimates from the qubit counts each source states; the IBM ratio uses the July release's instance. Source: The Reality Gap TAC-0033. Not illustrative.

Nine demonstrations published between 2024 and October 2026 claim between about 1.04 and 101 physical qubits per logical qubit. Every claim below about 3 physical qubits per logical qubit kept its results by discarding runs in which an error was detected.

  • Same machine, four counts: at 94 logical qubits about three runs in four were discarded; at 48, each correction cycle kept about 62% of runs; at 2 (Helix, distance 6) none were discarded and the error was 4.6 × 10⁻⁵ per qubit per cycle; at 1 (colour code, distance 7) it was 5 × 10⁻⁶ per round.
  • Detection has a horizon: Helios keeps 94.5% of runs per detection cycle at distance 2 and 97.2% at distance 4, so fewer than 1% survive past about 80 and 160 cycles (The Reality Gap estimate).
  • IBM's 70-qubit circuit kept 2,051 of about 3.5 million shots, about 0.059%. The release, the revised arXiv abstract and secondary coverage describe the experiment differently, and the report leaves them unreconciled.
  • The best result is narrow: the distance-7 point rests on 3 failures in 29,000 runs, with Λ = 12 (+3/−2), Clifford circuits only and offline decoding. It is a preprint.

The prize

The DOE fixes the count and negotiates the quality

101001k10k100k1M10MHelios today: 98 ionsPrototype: 10 logical, 10³ operations, ε ≤ 10⁻⁴Λ = 12 route370 (d=5)Λ = 2.14 route4,490 (d=15)First generation: 100 logical, 10⁵ operations, ε ≤ 10⁻⁷Λ = 12 route18,100 (d=11)Λ = 2.14 route217,700 (d=33)Next generation: 1,000 logical, 10⁷ operations, ε ≤ 10⁻¹⁰Λ = 12 route433,000 (d=17)Λ = 2.14 route5.2M (d=51)Physical qubits for logical memory, log scaleHatched: machines not yet built (The Reality Gap estimates)
Physical qubits needed to hold the logical qubits of each DOE illustrative tier at the error budget the tier implies, on a log scale, for the two best-characterised published suppression factors, with the code distance required. Hatched bars denote machines that do not yet exist. All values are The Reality Gap estimates for logical memory only, assume Λ stays constant beyond measured distances, and exclude magic-state factories, routing and spares. Source: The Reality Gap TAC-0033, from DOE webinar slides and published results.

The Quantum Genesis Q Competition, opened on 17 September 2026, plans up to $215 million. Its main pool is shared among companies that show at least 100 logical qubits by about autumn 2028; the error rate, round count and definition of a hard operation are agreed with each applicant separately.

  • The budget depends on the document: the slides' table pairs 100 logical qubits with 10⁵ hard operations (ε ≤ 10⁻⁷); the press release's "hundreds of millions" of operations implies about 10⁻¹⁰.
  • Detection-only machines are excluded: on the register that removes the 94-, 70-, 30- and 24-qubit results as they stand.
  • Even split, width bonuses: a $100 million Phase 2 pool is shared evenly, with two $50 million bonus pools at 150 and 200 logical qubits. Going from 100 to 200 qubits costs about 18,100 more physical qubits and can earn both bonuses; going from 10⁵ to 10⁶ operations costs about 7,200 more and earns nothing (The Reality Gap estimate).
  • Roadmaps fall short of the table: company targets for 2027–2028 that state an error rate sit at 10⁻⁵ to 10⁻⁶, enough for 10³ to 10⁴ operations at 100 or more logical qubits.

The clock

Add time and the ranking inverts

0.01 s1 s1 min1 h1 day1 yrFirst generation: 10⁵ operationsSuperconducting, 1.1 µs/roundabout 0.1 sTrapped ion, 0.2 s/round5.6 hoursNext generation: 10⁷ operationsSuperconducting, 1.1 µs/round11 sTrapped ion, 0.2 s/round23 daysMinimum run time at one correction round per operation, log scaleTrapped ion needs ~12× fewer qubits; superconducting runs ~180,000× faster per round
Minimum run time for the operation counts of the DOE's first- and next-generation illustrative tiers, on a log scale, multiplying operations by the published round time of each platform. These are lower-bound estimates by The Reality Gap, not measured runs: real logical operations span more than one round, and parallelism is ignored. Round times from published Helios and Google results as reported in TAC-0033.

A Helios correction round takes about 0.2 s; Google's takes 1.1 µs. Neither the logical-qubit counts nor the DOE's table carries a rate, and on a time basis the twelvefold qubit advantage of the trapped-ion route reverses.

  • Run-time floors: at one round per operation, 10⁵ operations take at least 5.6 hours at 0.2 s per round and about 0.1 s at 1.1 µs; 10⁷ take at least 23 days against 11 s (The Reality Gap estimates).
  • Qubit-seconds per round at the first-generation budget: about 36 on the Λ = 12 route against about 0.0024 on the Λ = 2.14 route, a ratio near 15,000 (The Reality Gap estimate).
  • An independent benchmark agrees: Sandia's QUOPS reportedly placed Helios ahead on circuit size (1,504 against 216) and far behind on rate (303 operations per second against about 20 million).
  • Base case (The Reality Gap's view): a 100-logical-qubit machine under repeated correction is shown by 2028 and supports 10³ to 10⁴ hard operations, short of the 10⁵ in the DOE's illustration.
The Ecliptic, Issue No. 019, cover analysis, drawn from The Reality Gap TAC-0033 (November 14, 2026). Qubit counts, error rates, acceptance rates and round times are as published by the sources TAC-0033 cites, several of them preprints reported through secondary analyses; error budgets, physical-qubit counts, survival horizons and run-time floors are The Reality Gap estimates. Interpretation is The Reality Gap's.
THE ECLIPTICan Orrery publication
Content is produced by AI systems with human editorial oversight.