Intelligence & Machines · Updated July 2026Momentum · accelerating ↗

How close are we to a quantum computer that doesn't break?Can we build a quantum computer that actually works?

or, simply: Can we build a quantum computer that actually works?or, precisely: How close are we to a quantum computer that doesn't break?

Error-corrected qubits have crossed below threshold; the open question is scaling from one logical qubit to the millions a useful machine needs.Today's quantum computers are too error-prone to trust. Fixing that means going from a handful of good qubits to millions.

We are here

One ion hosts an error-corrected spin-cat qubit - A spin-cat encoding inside one atomic ion reduced errors by up to 2.2-fold and extended lifetime by up to 1.5-fold over an unencoded qubit. Next up - IBM targets Starling availability (expected 2029).

01 · Where we stand

State of playWhere fault-tolerant quantum stands right now

The current stage, the honest metric, and the single threshold that gates the next stage. Each threshold is a falsifiable claim with a named next test.How far up the ladder we've climbed, the honest verdict, and the one thing blocking the next step.

The five stagesMaturity ladder

Deployed

Out in the real worldDeployed at scale

Scaling

Making it cheap enough at scaleScaling toward competitive cost

Engineering

Building one that pays for itselfEngineering a system that pays back

Lab demo← HERE

Shown to work in a labDemonstrated in the laboratory

Theoretical

The idea is worked out on paperTheoretical basis established

The honest verdictVerified state

Today's quantum computers are too error-prone to trust. Fixing that means going from a handful of good qubits to millions.Error-corrected qubits have crossed below threshold; the open question is scaling from one logical qubit to the millions a useful machine needs.

7.7e-4Best result so farGoogle Willow, surface-code memory (Jul 2026)

1.0e-12Approximate logical-operation error required for factoring a 2,000-bit number

Blocking the next stepBlocking threshold

Compute while correctingIntegrated Universal Logical Operations Next test: Run a multi-logical-qubit universal circuit through repeated correction with a lower total failure rate than its physical implementation.

The verdict, in five rungsHow far up the ladder↓ next — the thresholds & the gap

01 · The evidence

The thresholds that gate the next stageWhat has to happen next

Each threshold is a falsifiable claim with a named next test; the gap chart shows how far today's metric sits from the goal.Each row is one thing that has to be proven — and how far today's number is from the target.

Below-Threshold Error ScalingBigger codes work better✓ Achieved · Dec 2024

100%

Proven byA 101-physical-qubit distance-7 memory reached 0.143% error per cycle.

Corrected Memory Beats HardwareProtection pays for itself✓ Achieved · Dec 2024

100%

Proven byThe distance-7 logical memory exceeded the best physical-qubit lifetime by 2.4±0.3 times.

Integrated Universal Logical OperationsCompute while correctingEarly

35%

Next testRun a multi-logical-qubit universal circuit through repeated correction with a lower total failure rate than its physical implementation.

Utility-Scale Fault-Tolerant ComputationDo useful work economicallyEarly

Next testIBM's 2029 Starling claim is a visible test: 200 logical qubits and 100 million gates must be independently characterized.

THRESHOLDS - Thresholds for Fault-Tolerant Quantum.

Scale

NOTE - This follows the candidate file's recommended metric and keeps only Google's comparable superconducting surface-code memory sequence. A memory error per correction cycle is not a logical-gate error, and changes in chips, decoders and cycle circuits prevent treating this as a pure hardware trend. The 10^-12 goal is an application estimate per logical operation, not a demonstrated threshold. Neutral-atom, ion, colour-code and post-selected results are kept out of this line because their metrics are not interchangeable.

02 · How we got here

The record behind the verdict

Major events set large; context events set small but never hidden. Everything below the TODAY rule is a schedule, not a result.

1982-19932 events1 shown

Computing Becomes Quantum

Computing Becomes Quantum moved the field from feynman proposes quantum simulation to deutsch defines universal machine. The results narrowed the next question without closing it.

1982

Feynman proposes quantum simulationTheory

Feynman argued that simulating quantum physics efficiently may require a computer governed by quantum mechanics.

Richard Feynman

1985

Deutsch defines universal machine

Deutsch described a universal quantum computer able in principle to simulate any finite physical system.

David Deutsch

1994-20014 events3 shown

Codes Make Reliability Possible

Codes Make Reliability Possible moved the field from shor makes reliability consequential to kitaev introduces topological protection. The results narrowed the next question without closing it.

1994

Shor makes reliability consequentialTheory

Shor gave polynomial-time quantum algorithms for factoring and discrete logarithms, motivating very long reliable computations.

Peter Shor

1995

Quantum error correction appearsTheory

Shor showed that one logical qubit could be encoded across nine physical qubits to correct arbitrary single-qubit errors.

Peter Shor

1996

Threshold theorem establishes scalabilityTheory

The threshold result showed arbitrarily accurate computation is possible if component errors stay below a fixed threshold.

Emanuel Knill, Raymond Laflamme and Wojciech Zurek

1997

Kitaev introduces topological protection

Kitaev proposed fault-tolerant computation using anyons, laying the theoretical foundation for toric and surface codes.

Alexei Kitaev

2002-20152 events1 shown

Correction Enters Hardware

Correction Enters Hardware moved the field from surface-code threshold quantified to trapped ions correct errors. The results narrowed the next question without closing it.

2002

Surface-code threshold quantified

Dennis and colleagues analyzed 2D surface codes and showed protection can improve without limit below a critical error rate.

Eric Dennis, Alexei Kitaev, Andrew Landahl and John Preskill

2004

Trapped ions correct errorsExperiment

Three beryllium-ion qubits encoded, detected and corrected induced spin-flip errors in the first complete QEC sequence.

NIST

2016-20223 events2 shown

The Breakeven Experiments

The Breakeven Experiments moved the field from cat code reaches breakeven to fault-tolerant control beats physical. The results narrowed the next question without closing it.

2016

Cat code reaches breakevenExperiment

A corrected bosonic qubit lasted 320 microseconds, 1.1 times its best physical constituent, without post-selection.

Yale University

2021

Majorana evidence is retractedSetback

A 2018 claim of quantized Majorana conductance was withdrawn for insufficient rigor, setting back topological-qubit evidence.

Delft University of Technology and Microsoft-affiliated researchers

2021

Fault-tolerant control beats physical

A 15-ion Bacon-Shor experiment reported 0.6% preparation-measurement error and 0.3% logical Clifford-gate error.

University of Maryland, Duke University and IonQ

2023-203318 events4 shown

Below Threshold, Not Useful

Below Threshold, Not Useful moved the field from larger surface code barely wins to ibm targets starling availability. The results narrowed the next question without closing it.

2023

Larger surface code barely winsExperiment

Google's distance-5 memory reached 2.914% error per cycle versus 3.028% at distance 3, only a 4% relative reduction.

Google Quantum AI

2023

Correlated event sets error floor

One high-energy event left a distance-25 repetition code at 1.7×10^-6 error per cycle; removing 0.15% of trials cut it tenfold.

Google Quantum AI

2023

Real-time bosonic QEC beats every component

A continuously corrected logical qubit achieved 2.27 +/- 0.07 times the coherence of its best imperfect component.

Yale University and collaborators

2023

Discrete logical qubit passes breakeven

Active correction extended a cavity-encoded logical qubit's lifetime about 16% beyond its best physical constituent.

SUSTech, Fuzhou University, USTC and collaborators

2023

Neutral atoms encode 48 logicals

A 128-atom processor ran sampling circuits with 48 error-detected logical qubits, but full post-selection retained only eight samples.

Harvard University, MIT and QuEra

2024

Low-overhead qLDPC protocol published

Simulations found a 0.7% circuit-level threshold and modelled 12 logical memories for nearly one million cycles using 288 physical qubits; no such processor was built.

IBM researchers

2024

Willow crosses surface thresholdExperiment

A 101-qubit distance-7 memory reached 0.143% error per cycle, improving 2.14-fold for each two-distance increase.

Google Quantum AI

2024

Rare bursts limit scaling

Willow repetition codes encountered correlated bursts about once per hour, producing an observed error floor near 10^-10.

Google Quantum AI

2025

Ocelot tests biased-noise correction

A nine-qubit chip combining five cat qubits and four ancillas showed phase-error detection improving from three- to five-cat repetition codes; overhead savings remain projected.

AWS Center for Quantum Computing and Caltech

2025

Logical magic-state distillation demonstrated

A neutral-atom processor ran 5-to-1 distillation with distance-3 and distance-5 colour codes, producing logical outputs with higher fidelity than their inputs.

Harvard University, MIT, QuEra and collaborators

2025

Neutral atoms repeat correction

A 448-atom architecture reached 0.62% logical error per round at distance 5 versus 1.33% at distance 3 after decoder tuning.

Harvard University, MIT and QuEra

2026

Universal toolkit remains tinyExperiment

Eight trapped-ion physical qubits hosted three error-detected logical qubits for a measurement-free Grover search over eight items.

University of Innsbruck and collaborators

2026

Hardware decoder closes a small latency loop

An FPGA-integrated superconducting test sustained up to 25 rounds with sub-microsecond mean decoding per round and 9.6-microsecond end-to-end feedback after nine rounds.

Riverlane, Rigetti and collaborators

2026

Logical circuits beat selected physical baselines

Peer-reviewed trapped-ion circuits spanning up to 12 logical qubits improved 11- to 800-fold over selected physical-circuit baselines, with post-selection; the ratio was circuit-dependent.

Microsoft and Quantinuum

2026

Neutral atoms demonstrate erasure-aware logic

Metastable ytterbium qubits supported adaptive erasure checks, a [[4,2,2]] encoding and teleportation between code blocks, but could not correct arbitrary Pauli errors.

Princeton University and collaborators

2026

Willow halves the logical error rateExperiment

Google's surface-code memory reached 7.7e-4 logical error per cycle, down from 1.4e-3 in 2024 - correction improving with scale, though still astronomically far from the ~1e-12 a useful algorithm needs.

Google Quantum AI

2026

One ion hosts an error-corrected spin-cat qubitWe are here

A spin-cat encoding inside one atomic ion reduced errors by up to 2.2-fold and extended lifetime by up to 1.5-fold over an unencoded qubit.

MIT, MIT Lincoln Laboratory and collaborators

TODAY · below this line: schedules, not results

2029

IBM targets Starling availabilityDeploymentTarget

IBM targets a client-accessible system with 200 logical qubits and capacity for 100 million gates; this is a vendor roadmap.

IBM

2015-20333 events0 shown

Events outside the declared eras

Events outside the declared eras moved the field from repeated detection reduces failures to darpa utility deadline arrives. The results narrowed the next question without closing it.

2015

Repeated detection reduces failures

A nine-qubit superconducting device cut retrieval failures 8.5-fold after eight error-detection cycles, using post-selection.

Google and UC Santa Barbara

2033

IBM targets billion-gate Blue JayDeploymentTarget

IBM targets 2,000 logical qubits and one billion gates after 2033; no hardware meeting this specification exists today.

IBM

2033

DARPA utility deadline arrivesPolicyTarget

QBI aims to verify whether any architecture can deliver computational value exceeding total system cost by 2033.

DARPA

— end of record · 32 shown, 0 hidden —

32 events · below the TODAY rule = scheduled, not done

03 · The data behind the verdict

Why the meters read the way they do

The learning curves and comparisons that justify each threshold's percentage. Every series is measured, with the source event linked in the timeline above.

The memory works. The application gap is vast.Surface-code memory error is down to 7.7e-4 per cycle; the 1.0e-12 application estimate is 772 million× lower — an unmeasured gap, not a forecast.One dot = one Google surface-code memory result · error probability per correction cycle · log scale · 2023–2026One dot = one comparable Google surface-code memory result · logical memory error per correction cycle · log scale · 2023–2026 · memory error is not gate error; target is an application estimate, not a demonstrated threshold

GOOGLE · AS OF JUL 2026

Announced scale is not achieved capacity.Announced logical-qubit scale spans 200–2,000; every point is a hollow IBM target, not achieved capacity.Hollow mark = announced target · dashed rail = planned year · logical qubits and executable gates · 2029–2033One hollow mark = one IBM-announced target · dashed rail = announced year · labels retain the target unit and value · 2029–2033

IBM · AS OF JUL 2026

The distance-3, distance-5 and distance-7 memories used 17, 49 and 101 physical qubits respectively; the duplicate 2023 points are different code distances, not elapsed-time growth.

NOTE - The distance-3, distance-5 and distance-7 memories used 17, 49 and 101 physical qubits respectively; the duplicate 2023 points are different code distances, not elapsed-time growth.

Suppression strengthened from a marginal 1.039-fold distance-3-to-5 improvement to a 2.14-fold reduction for each two-distance increase on Willow.

NOTE - Suppression strengthened from a marginal 1.039-fold distance-3-to-5 improvement to a 2.14-fold reduction for each two-distance increase on Willow.

Two breakeven landmarks, but not a clean platform series: 2016 used a bosonic cat code, whereas 2024 used a 101-qubit surface code.

NOTE - Two breakeven landmarks, but not a clean platform series: 2016 used a bosonic cat code, whereas 2024 used a 101-qubit surface code.

Distance 33 % error/cycle
Distance 52.9 % error/cycle

COMPARISON - Google's first larger surface code improved logical error by only about 4%, showing how narrow the first below-scaling signal was.

Distance-7 logical memory291 microseconds
Best constituent physical qubit119 microseconds

COMPARISON - Willow's corrected logical memory lasted 2.4 times longer than its best constituent physical qubit.

Demonstrated distance 7101 physical qubits
Projected distance 271457 physical qubits

COMPARISON - Even Google's extrapolated 10^-6 memory would require about fourteen times the physical qubits used by its demonstrated distance-7 memory.

Willow memory error0.00 probability/cycle
Factoring-scale target0.00 probability/operation

COMPARISON - Numerically, Google's memory-cycle error is roughly 1.4 billion times the 10^-12 operation target, though the two metrics differ.

01Bosonic breakeven was more than marginal

Yale's 2023 result is stronger than a simple “16% breakeven” claim: the fully stabilized bosonic logical qubit reported a coherence gain of 2.27±0.07 over every imperfect component involved in correction.

Source · Nature

02qLDPC's overhead claim is modelled

IBM's bivariate-bicycle qLDPC proposal has degree-6 connectivity, a depth-eight syndrome circuit and a simulated 0.7% threshold; its 288-physical/12-logical comparison is a modelled workload at 0.1% physical error, not a processor result.

Source · Nature

03Logical-circuit gains depend on the test

Quantinuum and Microsoft used a 12-physical-qubit code encoding two logical qubits and a 16-qubit tesseract colour code encoding four. The reported 11×-800× range depends on the circuit, baseline and acceptance rule; it is not one universal logical error rate.

Source · Nature

04Why magic-state distillation matters

Logical magic-state distillation matters because Clifford operations alone are not universal: the neutral-atom experiment combined five logical inputs into one improved output and tested both distance-3 and distance-5 colour codes.

Source · Nature

05Ocelot concentrates on phase flips

Ocelot's noise bias is the mechanism: cat qubits suppress bit flips so a repetition code can concentrate on phase flips. The demonstrated device contained five data cat qubits plus four syndrome ancillas.

Source · Caltech

06Decoding must keep pace with syndromes

The June 2026 decoder result addresses the backlog problem: syndrome generation must not outrun classical decoding. Its sub-microsecond mean per round is encouraging, but the measured closed-loop response for nine rounds was 9.6 µs and the test involved only eight qubits.

Source · Nature Communications

07Online tuning improved different things

Google's 2026 controller used error-detection events to tune more than 1,000 analogue control parameters while QEC ran. It improved stability against injected drift by 3.5×, whereas fine-tuning an already calibrated processor reduced logical error by about 20%; these are different claims.

Source · Nature

08Colour-code and surface-code rates differ

The same Google paper also reports a distance-5 colour-code memory at 8.19(14)×10⁻³ error per cycle. Do not merge this with the surface-code series: code family, layout and logical circuit differ.

Source · Nature

09Erasure information is not full correction

Erasure conversion supplies the decoder with the location of certain faults, distinct from an ordinary syndrome. Princeton's [[4,2,2]] experiment exploited that information but did not demonstrate correction of arbitrary unknown Pauli errors.

Source · Nature Physics

04 · What it unlocks

If the remaining tests pass

Downstream capabilities, drawn dashed because they depend on results not yet in.

05 · The players & the money

Who is building it-and what the money saysCapital, institutions and the global race

The teams doing the work, where they are based, and whether the money points to real delivery or only a plan.Company finance, public programmes, institutional leadership and market evidence-kept separate from valuations, forecasts and announced capacity.

The short versionCapital and institutional readout

Fault-tolerant quantum computing is a multi-architecture race backed by venture rounds and national programmes, but roadmaps still outrun demonstrated systems. The strongest evidence is small logical memories and circuits; no company operates a utility-scale universal fault-tolerant computer.

Players

Who is building itCompanies, laboratories and programmes

Google Quantum AI

USA

Superconducting surface and colour codes, Willow processors, decoding and control research.

IBM Quantum

USA

Superconducting hardware, qLDPC research, middleware and a public fault-tolerance roadmap.

Quantinuum

USA / UK

Trapped-ion processors, logical circuits, compilers and quantum cybersecurity software.

Companies, laboratories and programmes working on this problem
Player	Country	What they are doing	Funding	Named investors	Source
Google Quantum AIlab	USA	Superconducting surface and colour codes, Willow processors, decoding and control research. Funded inside Alphabet; no standalone financing round.	Not disclosed	Not disclosed	Source · quantumai.google
IBM Quantumlab	USA	Superconducting hardware, qLDPC research, middleware and a public fault-tolerance roadmap. Roadmap logical-qubit and gate counts are vendor targets, not achieved capacity.	Not disclosed	Not disclosed	Source · ibm.com
Quantinuumcompany	USA / UK	Trapped-ion processors, logical circuits, compilers and quantum cybersecurity software.	$300M equity round · Jan 2024 · $5B pre-money valuation	JPMorgan Chase · Mitsui · Amgen · Honeywell	Source · honeywell.com
PsiQuantumcompany	USA / Australia	Silicon-photonic architecture targeting utility-scale fault tolerance in Brisbane and Chicago. Funding and government-backed sites do not establish that a utility-scale machine has been built.	Series E · $1B · Sep 2025	BlackRock · Temasek · Baillie Gifford · NVentures · Qatar Investment Authority	Source · psiquantum.com
QuEracompany	USA	Neutral-atom logical processors, repeated correction and reconfigurable atom-array architectures.	>$230M financing · Feb 2025	Google Quantum AI · SoftBank Vision Fund 2 · Valor Equity Partners · NVentures	Source · quera.com
Riverlanecompany	UK	Hardware-agnostic quantum error-correction decoding and real-time control stack. Decoder progress is necessary infrastructure, not a complete fault-tolerant computer.	Series C · $75M · Aug 2024	Not disclosed	Source · riverlane.com

06 · Sources

Where every number comes from

17 sources — every figure on this page traces to one.