Illustration: a golden chandelier-shaped quantum cryostat glows in a glass shelter at the foot of a huge mountain; a thin lit path climbs toward a summit lost in cloud.

Quantum computing · state of play, Monday 16 November 2026

Quantum computers still can't factor 21

The gates have become good enough for error correction to work, and the estimated cost of breaking real encryption has fallen about a thousandfold in seven years. No machine has yet run Shor's factoring algorithm honestly on a number larger than 15. This page measures the distance.

15
largest number factored by an honest run of Shor's algorithm
In 2001, and again on trapped ions in 2016. Every published 21 skipped the multiplications.
2,405
entangling gates to factor 21
Against 21 gates for the number 15: 115 times more.
99.92%
best two-qubit gate fidelity across a whole machine
98 trapped ions. The best single pair in a lab reached 99.99%.
0.03%
error rate of an error-corrected two-qubit gate
It beat the same gate on bare qubits, with no runs thrown away (preprint, September 2026).
1,000×
fall since 2019 in the estimated qubits needed to break RSA-2048
20 million then; 10,000 to 100,000 in this year's papers. All on paper.
about 100
qubits in today's best machines
"A few hundred at most", by Scott Aaronson's count this February.

The short version

  1. The gates got good. Each extra "nine" of two-qubit fidelity has taken eight to ten years. A whole machine passed 99.9% in 2024, and one pair of ions reached 99.99% in 2025.

  2. Error correction now works at small scale. Bigger codes have beaten smaller ones since 2023, and by a wide margin since December 2024. In September 2026 an error-corrected two-qubit gate beat the bare gate it is built from.

  3. Factoring is stuck at 15. The number 21 needs 115 times more gates. Nobody has run that circuit, and every published "21" skipped the hard part.

  4. Gil Kalai's 2019 predictions are being overtaken, and he has not conceded. He said small error-correcting codes would fail and gate quality would stall. In 2026 he says the codes are "not strong enough so far to refute my theory".

  5. The dates moved in. Scott Aaronson calls an error-corrected run of Shor's algorithm before November 2028 "a live possibility". Company roadmaps and DARPA point at 2029 to 2033. Dated promises of usefulness have mostly been missed before.

1 · Factoring 21

21 is 115 times harder than 15

Illustration: a tiny pocket-watch mechanism on a workbench beside a towering clockwork engine with thousands of gears.

In 2001 a quantum computer factored 15 with a circuit of 21 entangling gates: operations on two qubits at once, where most errors happen. The number 15 is a lucky case: most of the multiplications in Shor's algorithm turn into nothing, or into swapping two wires. The number 21 has no such luck. In August 2025 Craig Gidney of Google wrote out the cheapest circuit he could find for it: 191 two-qubit CNOT gates and 369 three-qubit Toffoli gates on 15 qubits. Counting each Toffoli as six, that is 2,405 entangling gates.

One dot per gate: drag the fidelity to see how many fail

Red dots are gates that failed in one typical run. A run with no red dot gives a clean answer.

98.4%
of runs get through all 21 gates of the circuit for 15
15%
of runs get through all 2,405 gates of the circuit for 21
1.9
failed gates in an average run of the circuit for 21

This page's arithmetic: the chance of a clean run is the fidelity raised to the number of gates. It counts two-qubit gate errors only and treats them as independent. Idle time, measurement and single-qubit errors make it worse; a noisy run can still sometimes give the right answer.

How far away is it?

Why 15 is so cheap: three lucky breaks

Shor's algorithm multiplies a register by a list of constants, each multiplication switched on or off by a control qubit. That is the whole cost. Gidney counts three ways the number 15 escapes it.

  1. Most multiplications are by 1. For 15 the list of constants is 2, 4, 1, 1, 1, 1, 1, 1, so only two of eight do anything. For 21 it is 2, 4, 16, 4, 16, 4, 16, 4, 16, 4. That costs about 4 times more.
  2. The first multiplication is nearly free. The register starts at 1, so the first step only copies a known number in. That halves the work for 15 and saves a tenth for 21: about 1.8 times more.
  3. Multiplying by 4 is a swap. Because 15 is one less than 16, multiplying by 2 or 4 modulo 15 only rotates the bits: two controlled swaps. Modulo 21 the best circuit found by computer search uses 41 Toffoli gates. That costs about 20 times more.

Gidney calls his circuit for 21 more optimised than is realistic: "a more plausible amount of optimization would produce a circuit with 500x the cost of the factoring-15 circuit".

Counted for this page from his circuit file: 15 qubits, 191 CNOT, 369 Toffoli, 10 measurements. A Toffoli counts as six two-qubit gates.

The papers that said they factored 21, and the dog
Illustration: a scruffy dog sits beside a wooden abacus and a chunky old home computer, while a golden quantum cryostat stands under a spotlight behind them.

Photonic and superconducting experiments in 2012, 2019 and 2021 reported factoring 21. Each ran a "compiled" circuit, simplified using facts that only someone who already knows the factors has. Gidney: "they don't do the multiplications, because the multiplications are hard … So I don't count them."

  • Smolin, Smith and Vargo (Nature, 2013) showed that with such shortcuts two qubits can "factor" any product of two primes. The right measure, they wrote, is the length of the period found, not the size of the number.
  • Annealing and variational "records" (143, 56,153 and larger) pick numbers whose factors differ in a few bits. A 2022 preprint claimed a path to RSA-2048 with 372 qubits; a check found the method stops working near 70 bits. A 2024 story that a D-Wave machine had "broken RSA" was about a 22-bit number.
  • In 2025 Peter Gutmann and Stephan Neuhaus matched every such record with a 1981 home computer, an abacus and a dog.

The honest record stays at 15: the 2001 nuclear-magnetic-resonance run, or the 2016 trapped-ion run if the first does not count as quantum. A search for this page found no honest run on 21 or above through November 2026.

2 · The gates today

Each extra nine has taken eight to ten years

Illustration: gloved hands with tweezers steady a row of tiny glowing glass spheres balanced on needle points; one sphere's glow drifts away like smoke.

A two-qubit gate with 99.9% fidelity fails once in a thousand uses. Trapped ions hold the records: 99.3% in 2008, 99.9% in 2016, 99.99% in 2025. Superconducting chips went from 99.4% in 2014 to 99.9% in 2023. Whole machines run years behind the best pair in a lab.

Two-qubit gate fidelity, best published results

Each step up the scale is ten times fewer errors. Tap a point for the source.

  • Trapped ions
  • Superconducting
  • Neutral atoms
  • Filled: average over a whole machine
  • Hollow: best pair in a lab

Around 99%, error correction starts to help. 99.9% is what the code-breaking estimates assume on every gate.

The same data as a table
YearPlatformFidelityScopeSource
Platform by platform, November 2026
PlatformBest pair in a labWhole machineTrade-off
Trapped ions99.99% (Oxford Ionics, owned by IonQ, October 2025 preprint)99.921% over 98 qubits (Quantinuum Helios); 99.914% over 20 qubits in 2024 (company report)Any qubit can reach any other. Each step takes about a thousand times longer than on a chip.
Superconducting99.92% (MIT fluxonium, 2023)99.67% over 105 qubits (Google Willow, 2024); 99.62% over 105 (USTC Zuchongzhi 3.0, 2025)Fast: about a microsecond per error-correction cycle. Each qubit talks only to its neighbours.
Neutral atoms99.85%, or 99.94% when runs that lost an atom are set aside (April 2026 preprint)About 99.5% in runs with hundreds of atoms (Harvard and QuEra)The largest arrays: 6,100 atoms trapped at Caltech. Slow cycles, and atoms get lost.
Spin qubits in siliconAbove 99% on devices from a chip foundry (Diraq and imec, 2025); 99.5% in university devicesA dozen qubits at mostMade like ordinary chips. Still very small.
PhotonicsNo comparable gate figure: the published numbers are for componentsLoss of photons is the main error.
Topological (Microsoft)No gate shown. Nature's editors noted that the 2025 paper is not evidence of the Majorana modes the design needs.Would need far less error correction if it works.
What the largest circuits look like without error correction

IBM's newest chip runs circuits of "more than 7,500 quantum operations", and the next is designed for 10,000. These use error mitigation: many noisy runs are combined into a corrected average.

Mitigation recovers averages, not single answers. At 99.7% per gate, about 3 runs in 10 million get through 5,000 gates untouched. Shor's algorithm needs individual runs to come out right, so it needs real error correction.

3 · How good is good enough

Three nines to start, far more to finish

Illustration: a staircase cut into a cliff where every step is far taller than the last; a small climber with a lantern hauls up the fourth step.

Error correction spreads one "logical" qubit across many physical ones and checks them constantly. It only helps once the physical gates are better than a threshold, about 99% for the most-studied code, the surface code. A code's size is called its distance: distance 3 uses 17 physical qubits, distance 7 about 100. Every recent estimate for breaking encryption assumes 99.9% on every gate of a very large machine.

What Scott Aaronson has said

"Fault-tolerant" means computing on error-corrected qubits from start to finish.

  1. Sep 2024

    How good: "at least on paper", fault tolerance "becomes a net win" once two-qubit gates are "~99.99% reliable". (talk)

  2. Sep 2024

    How fast: from the late 1990s, "the 50% became 90%, became 95%, became 99%, and within the past year, multiple groups have reported 99.9%."

  3. 13 Nov 2025

    When: "a live possibility that we'll have a fault-tolerant quantum computer running Shor's algorithm before the next US presidential election." Factoring 15 that way would count, he added. (post)

  4. 21 Dec 2025

    How good, restated: several platforms now have ">99.9% fidelity two-qubit gates, at or above the theoretical threshold for fault-tolerance." (post)

  5. 29 Apr 2026

    When, for code-breaking: experts he trusts say a machine able to break deployed cryptography "ought to be possible by around 2029." Then: "I'm not a timing guy." (post)

  6. 7 Aug 2026

    Status: "Scalable fault-tolerance is still in the future". (post)

A search of his blog for this page found no statement of how many years each nine takes, and no forecast of the form "another N years at current rates". The eight to ten years per nine comes from the published records in the chart above.

More of his dated statements
  • 3 Feb 2012: offers "US$100,000" for a demonstration "that scalable quantum computing is impossible in the physical world". It is unclaimed. (post)
  • 10 Dec 2024, on Google's Willow: gate fidelity "roughly 99.7% … compared to ~99.5% in 2019". "Please don't ask me to predict how long". Milestones arrive at intervals "measured in years". (post)
  • 20 Nov 2025, a week after the election remark: "a 'live possibility' is short not only of a 'certainty,' but of a 'probability.'" (post)
  • 21 Dec 2025: he now takes more seriously the companies' roadmaps for "2028 or 2029".
  • 18 Feb 2026, in a comment on the estimate of 100,000 qubits for RSA-2048: "State-of-the-art systems right now have ~100 physical qubits, or a few hundred at most." (comment 30)
  • 2 Apr 2026: "I can't give anyone reassurance that (say) a 3- or 4-year timeline from here to CRQC is out of the question". CRQC is a machine that breaks deployed cryptography. (comment)
  • 7 Apr 2026, on factoring: "the actual goal is to factor using Shor's algorithm on a fault-tolerant QC. That has not yet been done."
Illustration: a square grid of small candles whose light gathers into one large steady flame above them; a gust has blown a few out and the large flame holds.

How much hardware a logical qubit costs depends on one measured number, written : how many times the error falls each time the code grows by one step. Google measured in 2024. The code-breaking estimates assume gates ten times better than the threshold, which means near 10.

What one good logical qubit costs

Each dot is a physical qubit holding data. All of them together hold one logical qubit.

27
code distance: the patch is this many qubits on a side
1,457
physical qubits for one logical qubit
2.0 million
physical qubits for 1,400 logical qubits, the size of the 2025 RSA-2048 design

A rule of thumb fitted to Google's 2024 measurement, not a design. At Λ = 2.14 it gives distance 27 and 1,457 qubits for one error in a million, the figure Google quotes. Real designs store idle qubits more densely, which is how the 2025 estimate stays under a million.

The rule of thumb, and what Λ has measured so far

For a surface code of distance , which uses physical qubits, the error per cycle falls as

where is the physical error rate and , about 1%, is the threshold. The constant 0.03 is chosen so that the formula passes through Google's distance-7 result.

ExperimentΛWhat was measured
Google, 20231.04Distance 5 beat distance 3, barely: 2.914% against 3.028% per cycle.
Google Willow, 20242.14Distance 7 on 101 qubits: 0.143% per cycle. The logical qubit outlived the best physical one 2.4 times.
USTC Zuchongzhi 3.2, 20251.40Distance 7 on 107 qubits.
Quantinuum Helios, 2026not comparableA different, small code on ions: 0.0046% memory error per cycle. The company reports a roughly twelvefold cut per step for another code; no paper was found for that.
What a useful machine needs
Machine or taskLogical qubitsOperationsPhysical qubits
"Megaquop" machine (John Preskill's name for the first useful step)about 1001 millionNot specified
IBM Starling, planned for 2029200100 millionIBM says its codes need about a tenth of the surface code's
Break 256-bit elliptic-curve keys (Google and others, March 2026)under 1,200under 90 million ToffoliUnder 500,000
Factor RSA-2048 (Gidney, May 2025)about 1,400about 6.5 billion ToffoliUnder 1 million
Factor 21 (Gidney's circuit, August 2025)15369 ToffoliNo estimate published

Cheaper codes exist on paper. IBM's "gross" code holds 12 logical qubits in 288 physical ones, where the surface code would need about 3,000. It needs long-range wiring between qubits, which IBM first put on a test chip in 2025.

4 · The skeptic

Gil Kalai's predictions, and what the machines did

Illustration: a field of lanterns at night; on one side light rain makes single lanterns flicker at random, on the other a gust of wind darkens a whole patch at once.

Gil Kalai, a mathematician at the Hebrew University of Jerusalem, has argued for twenty years that quantum computers cannot work at scale. His main argument is about noise: a noisy machine, he says, can only produce outputs simple enough for an ordinary computer to imitate, so it can never reach the quality that error correction needs. A second strand says errors on entangled qubits arrive together, in bursts that defeat error correction.

In 2019 he wrote down predictions that "will be tested in the next few years". They were.

Small error-correcting codes will fail

He predicted
Attempts to build distance-5 surface codes "will fail, and the difficulties will be evident already in the baby goals", such as distance 3.
What happened
Google's distance 5 beat distance 3 in 2023. In 2024 distance 7 beat both, with the error halving at each step. A Chinese team repeated distance 7 in 2025.
He says now
"The quantum error correcting codes, they are not strong enough so far to refute my theory" (August 2026). In March 2026 he asked to discuss the 2024 data with Google's team.

Gate quality will stall

He predicted
"The quality of qubits and gates cannot be improved beyond a certain threshold that is quite close to the best currently existing qubits and gates."
What happened
The best pair went from 99.9% to 99.99%. The best whole machine went from 99.4% on 53 qubits to 99.92% on 98.
He says now
Efforts to cut gate errors "will hit a wall" (August 2026).

No quantum computer will beat classical ones at sampling

He predicted
Random-circuit "supremacy" with 50 to 100 qubits will fail.
What happened
Google claimed it on 53 qubits in 2019, and classical computers caught up with that run in 2024. Later runs, on cleaner and larger machines, have not been matched in print.
He says now
His audit of the 2019 data found it "may indeed have been flawed". He grants that the experiment, "if it stands, directly refutes my conjecture".

Errors come correlated

He predicted
"Every pair of gated qubits will be subject to errors with large positive correlation", leading to bursts. He called this secondary: the main argument "does not depend on" it.
What happened
Errors fell exponentially as codes grew, which heavy correlation would prevent. What is left is a floor of rare bursts, about one an hour on Google's chip.
He says now
The correlated-noise conjectures "remain undecided after twenty years" (March 2026).

Bigger codes now beat smaller ones, down to a floor set by rare bursts

Logical error per cycle against code size. The line is a simple model: set how much each step helps, and how often a burst wipes out the whole code.

  • Model
  • Google, 2023
  • Google Willow, 2024
  • Willow, simpler repetition code
  • Quantinuum ions, 2026 (a different code)

Measured points: Google 2023, 3.028% and 2.914% at distances 3 and 5. Willow 2024, 0.143% at distance 7; its distance 3 and 5 points are worked back from the published Λ. Willow's repetition code bottoms out near one error in 10 billion cycles at distance 29. Quantinuum's small code on ions: 0.0046% per cycle (preprint).

So are errors independent enough? For one logical qubit held in memory on about 100 physical qubits, yes. For hundreds of logical qubits computing for days, nobody has measured it. Aaronson's verdict in 2026 is that Kalai is "fighting a losing battle at this point". Kalai calls his own case "a good argument for why quantum computers cannot work, but this is not an ironclad argument".

His argument in a little more detail
  • Noisy machines are weak. His central assertion: what today's noisy devices can robustly produce "can be described by low-degree polynomials", a class of computation far inside what ordinary computers do. The idea comes from his 2014 work with Guy Kindler on noisy systems of photons.
  • Error correction is harder than supremacy. "Achieving quantum supremacy is easier than achieving quantum error-correction." If the first is out of reach, so is the second.
  • The weak link, in his own words. The step from that asymptotic statement to machines of a fixed small size "can be regarded as both a novel and a weak link of our argument".
  • A guess he attached. The effort to control qubits well grows exponentially and fails early: "My guess is that the number will be ≤ 20."
His audit of Google's 2019 experiment

With the statisticians Yosef Rinott and Tomer Shoham, and later Carsten Voelkmann, Kalai has re-analysed the data from Google's 2019 run: one paper in Statistical Science (2022) and preprints in 2022, 2023, 2024 and December 2025, the last using more detailed data that Google supplied.

  • They say the match between Google's predicted and measured fidelity is "statistically implausible", and they question how the chip was calibrated.
  • They point out that the classical running time was off by ten orders of magnitude, which later simulations bear out.
  • No formal reply from Google was found. Aaronson's answer: the 2019 experiment "was long ago superseded anyway".

In March 2026 Kalai wrote that "we would love to have a similar discussion with the relevant Google team about the 2024 distance-3,-5,-7 surface code paper", and that the exponential decay of accuracy, "if confirmed, is overall encouraging".

The replies, and the other skeptics
  • Aram Harrow, 2012, in a written debate with Kalai: if errors were correlated that strongly, ordinary computers would fail the same way, because every memory cell is a small error-correcting code.
  • Scott Aaronson, 2012: the $100,000 offer above. In 2025: "Evidence continues to pile up that we are not living in the universe of Gil Kalai and the other quantum computing skeptics."
  • What Kalai concedes: "I don't take it as axiomatic that scalable quantum computing is impossible" (December 2024), and "My goal is to advance understanding—even if the conclusions eventually go against my own views" (March 2026).
  • Mikhail Dyakonov, a physicist, argues that controlling the continuous parameters of thousands of qubits is hopeless in practice ("When will we have a quantum computer?", 2019). His argument is about engineering, and makes no dated prediction.

5 · Early announcements

Twenty years of early announcements

Illustration: a museum hall with a row of spotlit pedestals, most of them empty with a velvet cloth on the floor beside them, one holding a small modest brass instrument.

The pattern is steady. Hardware milestones, such as more qubits, better gates and working error correction, have mostly arrived when promised. Claims of a speedup or a use have mostly been matched by ordinary computers soon after.

What was claimed, and what happened

  1. 2007Never substantiated

    D-Wave shows a 16-qubit "commercial quantum computer" solving a Sudoku.

    Nothing showed it used quantum effects. Aaronson: "either substantially fake or entirely fake".

  2. 2008 to 2009No speedup

    Hartmut Neven of Google, with D-Wave: image recognition and a car detector "with the quantum adiabatic algorithm".

    The 2008 papers ran on ordinary solvers. The 2009 detector was trained on a D-Wave chip. No speedup was ever shown.

  3. 2015Overtaken

    Google: its D-Wave is 100 million times faster than the standard method on one processor core.

    On problems built to suit it. Other classical methods kept up.

  4. 2018Retracted

    A Microsoft-backed team reports Majorana particles, the basis of its qubit, in Nature.

    Retracted in 2021. Its 2025 chip claim was formally challenged in 2026.

  5. 2019Overtaken

    Google: 200 seconds on its Sycamore chip against 10,000 years on a supercomputer.

    IBM estimated, without running it, that 2.5 days would do. In 2024 it was done classically in 86 seconds.

  6. 2023Overtaken

    IBM: a 127-qubit run shows quantum "utility".

    Simulated within two weeks, once on a laptop.

  7. 2024Kept

    Google: error correction "below threshold", the second milestone on its roadmap.

    Published in Nature and repeated in China in 2025.

  8. Jan 2025Withdrawn

    Jensen Huang of Nvidia: "very useful" quantum computers are 15 to 30 years off.

    Quantum stocks fell about 40% in a day. He took it back two months later.

Illustration: on a night running track a golden cryostat on wheels leads, with a row of server racks on mechanical legs closing the gap behind it.

The 10,000-year task now takes a classical computer 86 seconds

How long each task was said to take an ordinary supercomputer, and how long it has taken since.

  • Classical time claimed at the announcement
  • Classical time achieved since

A claim that a task is beyond ordinary computers is a claim about algorithms nobody has written yet. Three of these four still stand.

Hartmut Neven's adiabatic years, 2007 to 2015
  • 2008: "Image recognition with an adiabatic quantum computer", with D-Wave's founders, recasts image matching as an optimisation problem of the kind D-Wave's machine was built for. It is a recipe, not a measurement.
  • November 2008: "Training a binary classifier with the quantum adiabatic algorithm". The training was done by ordinary heuristic solvers; the paper expects "additional gains" from "the actual quantum processor".
  • December 2009: a Google research post reports a car detector trained on a D-Wave chip, and says it "performs better than those we had trained using classical solvers". No speedup was measured, and the companion paper rests on simulation.
  • 2013 to 2015: the D-Wave Two purchase and the two speed claims in the list above.
  • From 2014 Google built its own gate-based chips under John Martinis. Those produced the 2019 and 2024 results.
  • 2019: "Neven's law", his claim that quantum computing power grows doubly exponentially.

The adiabatic work never showed an advantage. Neven's lab later delivered two of the field's real milestones.

Why "supremacy" became "advantage" and then "utility"

John Preskill coined "quantum supremacy" in 2012 for the first task a quantum machine does that no ordinary computer can. After Google's 2019 paper some researchers objected to the word's overtones, and "quantum advantage" took over. IBM introduced "utility" in 2023 for results that are hard to simulate exactly but claim no proof.

Each change also loosened the claim. "Supremacy" can be refuted by one classical simulation, as in 2024. "Advantage" and "utility" have no agreed test.

6 · Projections

The estimates fell a thousandfold. The machines have about 100 qubits.

Illustration: a shipyard at dawn with the keel and a few ribs of an enormous ship under scaffolding, and faint glowing lines tracing the planned hull.

The estimated machine for breaking 2,048-bit RSA shrank from about a billion qubits in 2012 to 20 million in 2019 and under a million in 2025. Three papers in 2026 put it between 10,000 and 100,000, each with a newer code or a different kind of hardware. None of this is hardware progress: it is better algorithms, on paper.

Qubits needed to break RSA-2048, and qubits built

Every step down the scale is ten times fewer qubits. Tap a point for the source.

  • Estimate, standard code on a chip
  • 2026 estimate, newer codes or other hardware
  • Machine built
  • Company target
The same data as a table
YearWhatQubitsNote

Who says when

Round marks are forecasts, diamonds are company targets, squares are official deadlines.

Kalai's forecast is "never". Jensen Huang's withdrawn one, 2040 to 2055, is off the scale.

Illustration: an old iron padlock on a heavy door with hairline cracks glowing from inside, and beside it a new lock with a lattice pattern being forged.

The sharpest change in 2026 was for the elliptic-curve keys that protect Bitcoin and most web traffic. In March a Google paper with Dan Boneh and Justin Drake put breaking a 256-bit key at under 500,000 qubits and a run of minutes. Google has set itself 2029 to finish replacing such keys. Replacements exist and are standardised, so this application has a known defence.

This year's estimates, side by side
PaperTargetPhysical qubitsRun timeRests on
Gidney, May 2025RSA-2048under 1 millionunder a week99.9% gates, microsecond cycles, a plain grid
Iceberg Quantum, Feb 2026RSA-2048under 100,000about a month, as reportedDenser codes that need long-range links between qubits
Google and others, Mar 2026256-bit elliptic curveunder 500,000minutes99.9% gates on a plain grid. The circuits are withheld; a proof that they exist is published.
A Caltech-led team, Mar 2026256-bit elliptic curveas few as 10,000 atoms"a few days" with 26,000 atoms; RSA-2048 ten to a hundred times longerNeutral atoms that can be moved around, with high-rate codes
IonQ, Sep 2026256-bit elliptic curve19,397 ionsabout 26 days a tryCompany blueprint; slow ion cycles
Alice & Bob, Oct 2026RSA-204822,593 cat qubits3.14 daysCompany paper; a special qubit that suppresses one kind of error

The headline reading, that the job keeps getting ten times easier, hides the catch: each cut assumes hardware nobody has built, such as long-range links, 99.9% gates across 100,000 qubits, or month-long stable runs. Gidney is reported to see no further tenfold cut in his own design without changing its assumptions.

Company roadmaps, and what arrived in 2026
WhoTargetStatus, November 2026
IBMStarling in 2029: 200 logical qubits, 100 million gates. Blue Jay in 2033: 2,000 logical qubits.The 2026 chip that was to combine memory and logic, Kookaburra, was not shown at IBM's November conference. Starling was reconfirmed.
QuantinuumSol in 2027: about 100 logical qubits. Apollo in 2029: full fault tolerance.Helios delivered in 2025. The company says Sol is on track.
IonQ10,000 qubits in 2027; 2 million qubits and 40,000 to 80,000 logical qubits in 2030.No report was found that its 256-qubit step for 2026 has arrived.
GoogleSix milestones ending at a million qubits. No dates.Two reached. The third, a logical qubit good to one error in a million, is not yet claimed.
PsiQuantumAbout a million photonic qubits, in Brisbane and Chicago.Targets around 2027 to 2029.
DARPAJudge whether anyone can build a "utility-scale" machine by 2033.Six companies in the final stage: Microsoft and PsiQuantum from an earlier pilot, and Atom Computing, Diraq, IBM and IonQ since October 2026.

The programme's managing director, Micah Stoutimore, in October 2026: "we increasingly expect that someone will build a utility-scale quantum computer by 2033."

Why a slow machine must be a big machine

Gidney's week for RSA-2048 assumes an error-correction cycle of a millionth of a second, which superconducting chips manage. Ions and neutral atoms take about a thousandth of a second. This page's arithmetic: the same design at that speed would run for about 19 years.

The estimates for atoms and ions get round this by doing far more in parallel or by using denser codes, which is why their run times are days to weeks for the easier elliptic-curve problem.

7 · What it is for

Two proven uses, and a narrow door for data

Illustration: a mountain of papers and glowing data cubes piled against a wall with one keyhole-sized opening; a single bright thread passes through into a calm room with one glowing sphere.

After thirty years, two uses have solid theory behind them: breaking today's public-key cryptography, and simulating quantum physics and chemistry. Aaronson's list in December 2025 is the same two, plus "some modest benefits for optimization, machine learning, and other areas", "eventually".

A quantum computer takes its input through its control lines, one precise pulse at a time. There is no fast way to pour a large dataset in. The strong candidates are problems with a tiny input and an enormous calculation, which describes code-breaking and chemistry and rules out most data-heavy work.

Chemistry: what the fertiliser example needs

The standard example is the molecule at the heart of the enzyme that lets bacteria fix nitrogen at room temperature, where industry uses the hot, high-pressure Haber-Bosch process. Simulating it was estimated in 2021 to need about 4 million physical qubits for under four days.

Two cautions are widely voiced. A simulation would explain how the enzyme works; it would not by itself give a better catalyst. And most of the energy in Haber-Bosch goes into making hydrogen, which no catalyst for the nitrogen step removes. A 2023 study also found the evidence for an exponential quantum advantage in such chemistry unproven.

Why search-type speedups probably do not pay

Many proposed uses in optimisation and machine learning rest on a square-root speedup: a million steps instead of a trillion. Error correction makes each quantum step thousands to millions of times slower than a transistor's. A 2021 analysis from Google concluded that such speedups will not beat ordinary computers on early fault-tolerant machines.

The data-loading problem
  • Aaronson's 2015 essay "Read the fine print": the famous exponential speedups for linear algebra and machine learning assume the data is already loaded into a quantum state and that one number at the end is all you want.
  • In 2018 Ewin Tang, then an undergraduate, showed that for recommendation systems an ordinary algorithm does about as well given the same access to the data.
  • A 2023 survey of quantum memory found that a fault-tolerant version costs so much that it erases the advantage for big-data uses.
The ordinary computer beside it

Every error-correction cycle produces measurements that an ordinary computer must decode before the next step can depend on them. Google's 2024 experiment decoded in real time with an average delay of 63 microseconds; the code-breaking estimates assume 10. IBM, Nvidia and AMD are all building decoder hardware for this. A large quantum computer is also a large classical one.

Where this leaves it

Sources

Read for this page on 16 November 2026. Preprints and company statements are marked in the text; they have not been through peer review. Figures marked as this page's arithmetic are its own.

All sources, by section

Factoring

Gates and error correction

Scott Aaronson

Gil Kalai

Early announcements

Projections

Uses