The Google’s 105-qubit Willow quantum Processor: A Technical Briefing on Google’s Superconducting Quantum Chip

What is Google’s 105-qubit Willow quantum chip, and why does it matter?

Google’s Willow is a superconducting quantum processor with 105 physical qubits, announced in December 2024 by Google Quantum AI. It marks a clear step beyond the company’s earlier Sycamore device (53–72 qubits). The core advance is not merely more qubits, but the first experimental demonstration that quantum error correction can improve exponentially as the system scales—errors fall as more physical qubits are used to protect logical information. It also set a new record on a random-circuit sampling benchmark and later enabled a verifiable quantum-advantage demonstration with the Quantum Echoes algorithm.

This explanation covers the hardware, the two headline results (error correction and sampling performance), technical specifications, context relative to earlier chips and competing approaches, remaining limitations, and what the progress implies for practical quantum computing.

Hardware Architecture and Key Improvements

Willow uses superconducting transmon qubits arranged in a square lattice (rather than IBM’s heavy-hex layout). Neighboring qubits connect via tunable couplers that can be switched on and off, giving an average connectivity of about 3.47. The chip is fabricated at Google’s dedicated facility in Santa Barbara, California.

Major engineering gains over Sycamore include:

Coherence times: Average T1 (energy relaxation) improved from roughly 20 µs to 68 µs (with some reports approaching 98–100 µs on optimized configurations). Longer coherence means qubits retain quantum information longer before decoherence.
Gate and measurement fidelities: Single-qubit gate errors around 0.035 %, two-qubit (CZ or iSWAP-like) errors in the 0.14–0.33 % range depending on configuration, and measurement errors near 0.7 %. Post-release refinements have pushed single-qubit fidelities to ~99.97 % and entangling gates to ~99.88 %.
Fabrication and design: Larger capacitors, gap engineering to suppress leakage and cosmic-ray effects, optimized participation ratios, reduced flux noise, and better parameter tuning. These changes lower loss and correlated errors.
Cycle speed: Surface-code error-correction cycles run at ~1.1 µs (about 909,000 cycles per second).

Two slightly different Willow chips appear in the literature—one optimized more for error-correction experiments and another for random-circuit sampling—but both share the same 105-qubit generation architecture.

Breakthrough 1: Quantum Error Correction Below the Surface-Code Threshold

Quantum bits are fragile. Environmental noise, imperfect control, and cosmic rays cause errors far more frequently than in classical computers. Quantum error correction encodes one logical qubit into many physical qubits so that errors can be detected and corrected without destroying the quantum information.

Google used the surface code, a leading topological code. They demonstrated distance-3 (3×3), distance-5 (5×5), and distance-7 (7×7) surface-code memories on Willow. Key results published in Nature (December 2024):

Each time the code distance increased by 2 (roughly doubling the number of physical qubits protecting the logical qubit), the logical error rate dropped by a factor of approximately Λ = 2.14.
The distance-7 logical qubit (using 101 of the 105 qubits: 49 data + 48 measure + 4 leakage-removal) achieved a logical error rate of ~0.143 % per cycle.
Critically, this logical qubit lived more than twice as long as its best individual physical qubit (beyond “breakeven”).
A real-time decoder was demonstrated on a related 72-qubit Willow device at distance 5, with average latency of 63 µs.

This is the first clear experimental evidence of exponential error suppression with increasing code size on superconducting hardware—the long-sought “below-threshold” regime. In principle, continuing to scale should drive logical error rates arbitrarily low. Earlier surface-code experiments (including Google’s own 2023 work) had only shown marginal or non-exponential improvement.

Breakthrough 2: Random Circuit Sampling Performance

Willow completed a random-circuit sampling (RCS) task—a standard benchmark for quantum computational advantage—using 103 qubits at depth 40. The experiment finished in under five minutes. Google estimated that the same task would require ~10²⁵ years (10 septillion years) on the most powerful classical supercomputers of the time (e.g., Frontier). Cross-entropy benchmarking fidelity reached 0.1 %.

This continues the lineage begun with Sycamore’s 2019 supremacy claim, but with substantially higher qubit count, better fidelity, and deeper circuits. Classical simulation costs grow exponentially with qubit number and circuit depth, so the gap widens dramatically.

Later Development: Verifiable Quantum Advantage with Quantum Echoes (2025)

In October 2025, Google reported that an improved Willow system ran the Quantum Echoes algorithm (out-of-order time correlators / OTOC) approximately 13,000 times faster than the best classical algorithms on leading supercomputers. The result is verifiable—meaning it can be repeated and checked on other quantum hardware—and has potential relevance to molecular structure learning and materials science. This marks a shift from pure sampling benchmarks toward algorithms with clearer scientific utility.

Technical Specifications at a Glance

Metric	Approximate Value (Willow)	Notes / Comparison to Sycamore
Physical qubits	105	Roughly 2× Sycamore
Average connectivity	3.47	Square lattice + tunable couplers
T1 coherence	68–98 µs (mean)	~5× improvement
Single-qubit gate error	~0.035 %	High simultaneous fidelity
Two-qubit gate error	0.14–0.33 %	Depends on gate type
Measurement error	~0.7 %	—
Surface-code cycle time	1.1 µs	High repetition rate
Error-suppression factor Λ	2.14	Distance 3→5→7
RCS performance	<5 min vs. ~10²⁵ years classical	103 qubits, depth 40

Context, Limitations, and Realistic Outlook

Willow sits in a competitive landscape that includes IBM’s superconducting processors (heavy-hex lattices, different scaling strategy), neutral-atom systems (e.g., QuEra, Pasqal), trapped-ion platforms, and photonic approaches. Superconducting qubits offer fast gates and mature control electronics but require millikelvin cryogenics and still face material and fabrication challenges.

Important caveats remain:

Demonstrations so far focus primarily on quantum memory (preserving logical information) rather than full universal logical gates needed for fault-tolerant algorithms.
Logical error rates (~10⁻³ per cycle) are still many orders of magnitude above the ~10⁻⁶–10⁻¹² levels required for large-scale applications such as factoring or accurate chemistry simulation.
Scaling to thousands or millions of physical qubits will demand further reductions in correlated errors (cosmic rays, residual leakage), better decoding latency, and massive improvements in classical control infrastructure.
Practical, commercially useful quantum computers are widely expected to remain years away—Google leadership has spoken of a multi-year roadmap toward fault-tolerant systems capable of useful workloads.

Nevertheless, Willow closes a critical theoretical and experimental gap that had persisted for nearly three decades since Peter Shor’s foundational work on quantum error correction. By showing that larger codes can be better rather than worse, it validates the surface-code scaling path and strengthens confidence that superconducting platforms can reach the fault-tolerant regime.

In short, Willow is not yet a general-purpose quantum computer. It is, however, the clearest experimental proof to date that the error-correction overhead can be made to work in the right direction—an essential foundation for the larger, more reliable machines that will eventually unlock applications in materials design, drug discovery, optimization, and fundamental physics.

1) Hardware Architecture of Google’s Willow Quantum Processor

Qubit Technology: Frequency-Tunable Superconducting Transmons

Willow is built entirely from aluminum-based frequency-tunable transmon qubits. Transmons are nonlinear superconducting LC resonators that function as artificial atoms. Each qubit incorporates a superconducting quantum interference device (SQUID) loop that acts as a flux-tunable nonlinear inductor.

Operating frequencies typically fall in the 5.9–6.5 GHz range.
Each qubit requires two primary control lines: a microwave XY drive for single-qubit rotations and a baseband Z (flux) bias for frequency tuning.
Qubits are fabricated on a silicon substrate with aluminum capacitor pads and conventional Al-AlOₓ-Al Josephson junctions.
Average inter-qubit spacing is approximately 1.15 mm on related Willow devices.

The design retains the tunability of the earlier Sycamore architecture while incorporating substantial improvements in materials and geometry.

Lattice Layout and Connectivity

The 105 qubits are arranged in a two-dimensional square grid. This layout is deliberately chosen because it maps directly onto the surface code used for quantum error correction—data qubits and measure (ancilla) qubits alternate across the plane.

Average connectivity is 3.47 (typically nearest-neighbor four-way coupling where geometry allows).
The square lattice contrasts with IBM’s heavy-hexagonal topology and prioritizes native compatibility with surface-code stabilizers.
In surface-code experiments, a distance-7 logical qubit uses 49 data qubits, 48 measure qubits, and 4 dedicated leakage-removal qubits, consuming nearly the entire 105-qubit array.

Tunable Couplers

Neighboring qubits are linked by tunable transmon couplers. These are additional transmon circuits embedded in capacitive networks that act as flux-tunable impedances.

When biased to a specific frequency, the coupler completely nulls the interaction between qubits (idle isolation).
Detuning the coupler activates a controllable coupling strength, enabling high-fidelity two-qubit gates (primarily CZ or iSWAP-like interactions).
Tunability is essential for reducing crosstalk, isolating idle qubits during long error-correction cycles, and optimizing gate performance across the full array.

This coupler architecture is a direct evolution of the Sycamore design and remains a defining feature of Google’s superconducting platform.

Fabrication, Packaging, and System Integration

Willow is manufactured in Google Quantum AI’s dedicated fabrication facility in Santa Barbara—one of the few facilities purpose-built for superconducting quantum processors. Key process and design advances include:

Capacitor geometry: Capacitor size increased from ~30 µm to ~150 µm with adjusted topology to improve coherence.
Participation-ratio engineering: Careful control of electric-field participation in lossy materials and interfaces.
Gap engineering: A superconducting gap difference (δΔ/h ≈ 12 GHz) across the Josephson junctions suppresses quasiparticle tunneling caused by ionizing radiation (cosmic rays). This dramatically reduces correlated T1 error bursts.
Noise reduction: Lower current noise in flux-bias lines and improved electromagnetic shielding.
Global optimization: An automated optimizer (sometimes called the “Snake” algorithm) selects operating frequencies and coupler biases across the entire grid to maximize simultaneous gate fidelities.

The qubit chip is bump-bonded to a carrier chip that routes control, readout, and reset circuitry. The complete system operates inside a dilution refrigerator at millikelvin temperatures (approximately 10–20 mK).

Coherence and Performance-Enabling Features

These architectural choices produced a roughly five-fold improvement in coherence:

Mean T1 (energy relaxation time) of 68 µs ± 13 µs on the primary error-correction chip, with some configurations reaching ~98 µs.
T2,CPMG values around 89 µs.

Additional functional features supporting high-fidelity operation include:

Multi-level reset (clearing |1⟩ and higher states).
Dedicated leakage removal (primarily |2⟩ state).
Simultaneous readout of large subsets of the array with measurement errors typically under 1 %.
Fast gate times: single-qubit rotations on the order of 25 ns; two-qubit CZ gates around 42 ns on the 105-qubit device.

Architectural Trade-offs and Design Philosophy

Google prioritizes a system-level co-design approach: every element—qubit geometry, coupler design, control electronics, calibration software, and error-correction circuits—is optimized together rather than in isolation. The square lattice with tunable couplers trades some connectivity density for excellent surface-code compatibility and low crosstalk. Gap engineering and larger capacitors improve resilience to environmental noise at the cost of added fabrication complexity.

Two closely related Willow chips exist: one optimized more heavily for electromagnetic shielding and error correction (slightly lower average T1) and another tuned for maximum coherence and random-circuit sampling. Both share the same fundamental architecture.

Summary of Architectural Significance

Willow’s hardware architecture represents a mature evolution of Google’s superconducting platform. By combining a surface-code-friendly square lattice, highly tunable couplers, refined transmon geometry, and radiation-hardening via gap engineering, the design delivered the first clear experimental demonstration of exponential error suppression with code distance. These physical foundations—rather than qubit count alone—enabled the logical qubit to outlive its best physical constituent and set new performance benchmarks.

The architecture remains firmly in the noisy intermediate-scale quantum (NISQ) to early fault-tolerant transition regime. Further scaling will require continued advances in materials, packaging density, classical control bandwidth, and correlated-error mitigation, but the core design principles validated on Willow provide a clear roadmap for larger systems.

2) Key Improvements

Willow does not simply add more qubits; it raises the quality of every physical resource so that larger codes become better rather than worse. The advances fall into several interlocking categories.

Coherence Time Gains

The most fundamental improvement is longer qubit lifetime.

Average energy-relaxation time (T1) rose from roughly 20 µs on Sycamore-generation devices to 68 µs ± 13 µs on the primary Willow error-correction chip, with some configurations reaching ~98 µs.
Corresponding T2,CPMG values reached approximately 89 µs.
This represents an approximately five-fold increase in the time qubits can retain quantum information before decoherence.

Longer coherence directly improves every subsequent operation: single-qubit gates, two-qubit gates, idling during measurement, and the overall fidelity of multi-cycle error-correction sequences.

Fabrication and Materials Engineering

Willow was produced in Google’s purpose-built Santa Barbara fabrication facility. Specific process and design changes include:

Capacitor size increased from ~30 µm to ~150 µm with revised topology, reducing dielectric loss.
Participation-ratio engineering that minimizes electric-field exposure to lossy interfaces and materials.
Superconducting gap engineering across the Josephson junctions (gap difference δΔ/h ≈ 12 GHz). This suppresses quasiparticle tunneling triggered by ionizing radiation (cosmic rays), sharply reducing correlated T1 error bursts.
Lower noise in flux-bias lines and improved electromagnetic shielding.
Optimized circuit parameters selected by global algorithms (e.g., “Snake” optimizer) that balance frequency collisions, coupler biases, and simultaneous gate fidelities across the full array.

These changes address both intrinsic materials loss and extrinsic environmental noise—two dominant error sources that previously limited scaling.

Gate and Measurement Performance

Higher coherence and refined control translated into lower error rates under simultaneous operation:

Operation	Typical Mean Error (Willow)	Notes
Single-qubit gates	0.035 %	Simultaneous randomized benchmarking
Two-qubit gates (CZ)	0.33 %	Error-correction optimized chip
Two-qubit gates (iSWAP-like)	0.14 %	Sampling-optimized chip
Measurement	0.67–0.77 %	Terminal or repetitive

Post-release refinements have pushed single-qubit fidelities to ~99.97 % and entangling-gate fidelities to ~99.88 % across the full 105-qubit array. Gate durations remain fast (single-qubit ~25 ns, CZ ~42 ns), preserving high cycle rates.

Reset, Leakage Removal, and Idle Performance

Willow incorporates multi-level reset that clears the |1⟩ state and higher levels, plus dedicated leakage-removal protocols focused on the |2⟩ state. Dedicated leakage-removal qubits sit at the boundaries of the surface-code lattice. These features prevent population from accumulating in non-computational states during long error-correction runs—an essential requirement for sustained below-threshold operation.

Error-Correction Scaling (The Defining Advance)

The most consequential improvement is the first experimental demonstration of exponential error suppression with increasing surface-code distance.

Logical error rate falls by a factor Λ ≈ 2.14 each time the code distance increases by 2 (3×3 → 5×5 → 7×7 lattices).
The distance-7 logical qubit (101 physical qubits) achieves a logical error of ~0.143 % per cycle and lives more than twice as long as its best constituent physical qubit (beyond breakeven).
Real-time decoding was demonstrated on a related 72-qubit Willow device at distance 5 with average latency of 63 µs.

Earlier surface-code experiments, including Google’s own 2023 work, had shown only marginal improvement or remained above threshold. Willow is the first superconducting processor to operate clearly below the surface-code threshold at multiple distances.

System-Level and Algorithmic Performance

Random-circuit sampling with 103 qubits at depth 40 completed in under five minutes—an estimated 10²⁵-year classical equivalent on contemporary supercomputers.
Later refinements enabled the Quantum Echoes (OTOC) algorithm, delivering the first verifiable quantum advantage on hardware (~13,000× faster than classical counterparts).

These results reflect not only better individual components but also tighter co-design among hardware, calibration software (including machine-learning and reinforcement-learning optimizers), and error-correction circuits.

Architectural Continuity with Targeted Evolution

Willow retains the square-lattice layout and tunable-coupler architecture of Sycamore while systematically raising every performance metric. The design philosophy prioritizes system-level quality over raw qubit count: every element—from Josephson-junction gap profile to simultaneous multi-qubit calibration—is engineered so that scaling the code distance improves rather than degrades logical performance.

Remaining Challenges and Context

Despite these gains, logical error rates remain orders of magnitude above the levels required for large-scale fault-tolerant algorithms. Correlated errors, residual leakage, and classical control overhead still constrain further scaling. Nonetheless, the improvements realized in Willow convert a long-standing theoretical hope—exponential error suppression with code size—into experimental reality on superconducting hardware.

In summary, Willow’s key advances are a multi-fold coherence boost, refined fabrication that attacks both materials loss and radiation-induced errors, higher simultaneous gate and measurement fidelities, robust leakage management, and the first clear below-threshold surface-code scaling. Together they mark a decisive step from proof-of-principle devices toward architectures capable of sustained fault-tolerant operation.

3) Breakthrough 1: Quantum Error Correction Below the Surface-Code Threshold

The core achievement of Google’s Willow processor is the first clear experimental demonstration that a superconducting quantum system can operate below the surface-code threshold. In this regime, increasing the number of physical qubits that protect a logical qubit reduces the logical error rate exponentially rather than increasing it. This result, published in Nature in December 2024, resolved a nearly 30-year challenge in the field and marks a decisive transition from proof-of-principle experiments to scalable error-corrected architectures.

Why the Surface-Code Threshold Matters

Quantum error correction encodes one logical qubit into many physical qubits so that errors can be detected and corrected without destroying the quantum information. The surface code is the leading candidate for near-term superconducting hardware because of its high error threshold (theoretically ~1 % physical error rate under realistic noise models) and its natural fit to a two-dimensional nearest-neighbor lattice.

Theory predicts that when the physical error rate p lies below a critical threshold *p*thr, the logical error rate ε scales exponentially with code distance d:

$\varepsilon_d \propto \left( \frac{p}{p_{\rm thr}} \right)^{(d+1)/2}$

The suppression factor is commonly expressed as

$\Lambda = \frac{\varepsilon_d}{\varepsilon_{d+2}} \approx \frac{p_{\rm thr}}{p}$

A measured Λ > 1 is the experimental signature that the system has crossed below threshold. Prior superconducting experiments, including Google’s own 2023 surface-code work, had remained above or only marginally near threshold, so larger codes either failed to improve or improved only weakly.

Experimental Realization on Willow

Google demonstrated below-threshold performance on two Willow processors:

A 72-qubit device used for a distance-5 surface code with integrated real-time decoding.
A 105-qubit device used for a distance-7 surface code (the primary scaling result).

The distance-7 logical qubit employed 101 of the 105 physical qubits: 49 data qubits arranged in a 7×7 grid, 48 measure (ancilla) qubits that extract stabilizers, and 4 dedicated leakage-removal qubits at the boundaries. The code operated in the ZXXZ variant of the surface code.

Key quantitative results with a neural-network decoder:

Code Distance	Approximate Physical Qubits	Logical Error Rate per Cycle (εd)	Suppression Factor Λ
3	~9–17	Higher	—
5	~25–49	Intermediate	—
7	101	(1.43 ± 0.03) × 10⁻³ (0.143 %)	2.14 ± 0.02

Each increase of code distance by two reduced the logical error rate by a factor of approximately 2.14. An alternative decoder (ensembled matching synthesis) yielded a slightly lower but still clearly above-unity Λ of 2.04 ± 0.02.

Beyond Breakeven: Logical Lifetime Exceeds Physical Lifetime

The distance-7 logical qubit achieved a lifetime of roughly 291 µs. This exceeded the lifetime of its best individual physical qubit (approximately 119 µs) by a factor of 2.4 ± 0.3. In other words, the encoded qubit outlasted every physical component from which it was constructed—the practical definition of “beyond breakeven.” Earlier surface-code demonstrations had not crossed this threshold.

Real-Time Decoding

On the 72-qubit Willow device, Google integrated a real-time decoder for the distance-5 surface code. The error-correction cycle time was 1.1 µs. The decoder maintained an average latency of 63 µs while processing up to one million cycles continuously, preserving below-threshold performance. Real-time decoding is essential for any useful computation: classical feedback must keep pace with the quantum cycle rate so that errors can be corrected before they accumulate beyond the code’s correction capability.

Probing the Limits: Repetition Codes and Correlated Errors

To stress-test the hardware, the team ran repetition codes (which protect against only one type of error) up to distance 29. Logical performance eventually saturated, limited by rare correlated error events that occurred approximately once every hour—or once every 3 × 10⁹ cycles. These residual bursts are far less frequent than on earlier devices (previously occurring on second timescales) and recover more quickly (~1 ms), thanks in part to gap engineering that suppresses quasiparticle tunneling. Their origin is not yet fully understood and remains an active research focus.

Technical Enablers

Several hardware and software advances made the result possible:

Improved coherence (T1 ≈ 68 µs mean on the primary chip).
Lower simultaneous gate and measurement errors.
Dedicated leakage removal and multi-level reset.
Advanced decoding (neural networks and matching algorithms) that extract more information from the syndrome data.
Precise frequency optimization and crosstalk mitigation across the full array.

The square-lattice layout with tunable couplers maps naturally onto the surface-code geometry, minimizing the overhead of routing stabilizers.

Significance and Remaining Gap

This experiment provides the first unambiguous experimental confirmation of the surface-code threshold theorem on superconducting hardware. It converts a long-standing theoretical prediction into laboratory reality and validates the scaling path that underpins Google’s fault-tolerant roadmap.

At the same time, the absolute logical error rate (~1.4 × 10⁻³ per cycle) remains many orders of magnitude above the levels required for large-scale algorithms (typically 10⁻⁶ to 10⁻¹² or better). Extrapolations that assume the measured Λ continues unchanged indicate that code distances of order 70–80 would be needed to reach practical error rates—corresponding to thousands of physical qubits per logical qubit. Further reductions in physical error rates, better handling of correlated noise, and higher decoder accuracy will be required to close this gap.

In short, Willow’s below-threshold surface-code demonstration is the pivotal hardware milestone of 2024. It proves that error correction can improve exponentially with scale on a real device, establishing a solid foundation for the larger, more reliable logical qubits needed for useful quantum computation.

4) Breakthrough 2: Random Circuit Sampling Performance

The second major achievement of Google’s Willow processor is a dramatic new record on the random circuit sampling (RCS) benchmark. In under five minutes, Willow generated samples from a highly complex quantum circuit that Google estimated would require approximately 10²⁵ years—10 septillion years—on the most powerful classical supercomputers available at the time. This result, announced alongside the error-correction work in December 2024, extends the lineage of quantum computational advantage claims that began with the 2019 Sycamore experiment and demonstrates the practical impact of Willow’s improved qubit quality and scale.

What Random Circuit Sampling Measures

RCS is a deliberately constructed task designed to be as hard as possible for classical computers while remaining straightforward for a quantum processor. The quantum device executes a sequence of random single- and two-qubit gates on a large number of qubits, then samples the resulting probability distribution by measuring the qubits in the computational basis. The output is a set of bitstrings whose statistical properties (most commonly quantified by cross-entropy benchmarking, or XEB, fidelity) can be compared against classical simulations for smaller instances.

Because the Hilbert-space dimension grows exponentially with qubit number and circuit depth, exact classical simulation rapidly becomes intractable. RCS therefore serves as a system-level stress test of coherence, gate fidelity, connectivity, and calibration quality under simultaneous operation of nearly the entire chip. It is not a practical application, but it remains the classically hardest benchmark routinely performed on current quantum hardware.

Willow’s RCS Result in Detail

Google performed the experiment on a Willow device optimized for sampling performance (Chip 2 in the official specification sheet). Key parameters:

Qubits used: 103 (two of the 105 physical qubits were inoperable or unused)
Circuit depth: 40 cycles
XEB fidelity: approximately 0.1 %
Runtime on Willow: under five minutes
Estimated classical runtime: ~10²⁵ years on the leading supercomputers of late 2024 (e.g., Frontier-class systems)

The device achieved a circuit repetition rate of roughly 63,000 shots per second. Supporting hardware metrics on this chip included:

Mean T1 ≈ 98 µs ± 32 µs
Simultaneous single-qubit gate error ≈ 0.036 %
Simultaneous two-qubit (iSWAP-like) gate error ≈ 0.14 %
Terminal measurement error ≈ 0.67 %

These figures reflect the coherence and fidelity gains realized in the Willow generation and explain why deeper, wider circuits remained feasible.

Comparison with Prior Benchmarks

Generation	Approximate Qubits	Depth / Cycles	Approximate XEB Fidelity	Classical Estimate	Quantum Runtime
Sycamore (2019)	53	20	~0.2 %	~10,000 years	~200 seconds
Intermediate (2024 updates)	~67	32	~0.15 %	Significantly higher	—
Willow (2024)	103	40	0.1 %	~10²⁵ years	< 5 minutes

Willow roughly doubled the circuit volume (qubits × depth) relative to earlier high-fidelity demonstrations while still producing a detectable XEB signal. The exponential growth in classical simulation cost with both qubit count and depth accounts for the jump from thousands of years to septillions of years.

Google’s classical-cost estimates considered a range of scenarios—from idealized unlimited-memory tensor-network contraction to more realistic memory-constrained, GPU-parallel implementations. Even under optimistic assumptions about classical resources and algorithmic efficiency, the gap remained vast.

Technical Context and Verification

Because full classical simulation of the 103-qubit, depth-40 circuit is intractable, fidelity was estimated via a “patched” method: a small number of two-qubit gates are removed to create classically simulable sub-circuits whose XEB can be measured exactly and then extrapolated. The observed fidelity tracked predictions from a digital error model that incorporates measured single-qubit, two-qubit, and readout errors, lending confidence that the full-circuit performance is consistent with device characterization.

The result builds on continuous improvements in frequency optimization, crosstalk mitigation, and simultaneous gate calibration—capabilities that also underpinned the surface-code breakthrough. Higher coherence and lower error rates allowed the team to push both width and depth while retaining a usable signal above the noise floor.

Significance and Caveats

The RCS record reconfirms that superconducting processors can access computational regimes far beyond the practical reach of classical supercomputers for this specific task. Combined with the below-threshold error-correction result, it shows that Willow improves both the quality of individual operations and the overall system scale at which those operations remain coherent.

Important limitations remain:

RCS is a synthetic benchmark with no known practical application.
The 0.1 % fidelity means that only a small fraction of the sampled bitstrings are free of error; the quantum advantage is statistical rather than deterministic.
Classical simulation algorithms continue to improve, and future hardware or algorithmic advances could narrow (though not eliminate) the reported gap for circuits of this size.
The result does not yet constitute a useful, verifiable quantum advantage for a scientific or industrial problem—that milestone arrived later with the Quantum Echoes algorithm in 2025.

Nevertheless, within the established framework of RCS benchmarking, Willow set a new standard. It demonstrated that the same hardware advances enabling exponential error suppression also deliver orders-of-magnitude larger separations from classical simulation, reinforcing confidence in the scalability of the superconducting platform.

In summary, Willow’s random-circuit-sampling performance is the clearest system-level expression of its hardware quality: 103 qubits at depth 40, executed in minutes rather than cosmic timescales. Together with the surface-code threshold crossing, it positions the Willow generation as a pivotal step from noisy intermediate-scale devices toward the fault-tolerant machines required for practical quantum computation.

5) Verifiable Quantum Advantage with Quantum Echoes (2025)

The core topic is Google’s October 2025 claim that Willow ran a scientifically meaningful algorithm—Quantum Echoes—faster than the best known classical methods, with a result that can in principle be checked on another quantum processor. Unlike random circuit sampling, this task measures a physical observable rather than producing an uncheckable sample distribution.

What Changed After the 2024 Benchmarks

Willow’s December 2024 results established two hardware facts: surface-code error rates can fall as codes grow, and RCS at 103 qubits and depth 40 is classically intractable under current simulation methods. Those results still left an important gap. RCS samples are hard to verify independently, and the task itself has no direct scientific use.

On 22 October 2025, Google reported in Nature (“Observation of constructive interference at the edge of quantum ergodicity”) that an improved Willow system had executed the Quantum Echoes algorithm—an implementation of out-of-time-order correlators (OTOCs)—in a regime the company described as beyond classical simulation. Google’s estimate: about 13,000× faster than the best classical algorithms on a Frontier-class supercomputer. One reported comparison put a ~2.1-hour Willow run against ~3.2 years on Frontier.⁠

How Quantum Echoes Works

An OTOC measures how a local disturbance spreads through a many-body quantum system—the quantum analogue of the butterfly effect.

The protocol has four conceptual steps:

Evolve the system forward with a circuit $U$ .
Apply a small local perturbation $B$ (a “butterfly” operator on one or a few qubits).
Evolve backward with $U^\dagger$ .
Measure a local observable $M$ .

If there is no perturbation, forward and reverse evolution cancel and the system returns to its initial product state. With the perturbation, the mismatch produces an “echo.” Quantum interference can amplify that echo, making the measurement sensitive to fine details of the dynamics.

Google used a more demanding second-order OTOC ( $\mathrm{OTOC}^{(2)}$ ), which repeats the scramble–perturb–unscramble sequence. That extra loop increases classical simulation cost while remaining an expectation value—a number that another quantum computer of similar quality can, in principle, reproduce.⁠

Typical experimental scale:

About 103 qubits participated in forward and reverse random-circuit evolution in the main Willow runs.
A widely cited beyond-classical instance used 65 qubits at 23 cycles.
The project collected on the order of one trillion measurements, enabled by millions of shots in tens of seconds.⁠

Why Google Calls It “Verifiable”

RCS advantage rests on a classical cost estimate: if no known classical algorithm can reproduce the sample statistics, the quantum device is said to win. That claim can shrink when better classical methods appear.

Quantum Echoes outputs an expectation value. In Google’s framing:

The same circuit can be rerun on Willow.
Another quantum processor with comparable qubit count, fidelity, and speed should obtain a statistically consistent answer.
The quantity is also related to physical systems (scrambling, chaos, molecular correlations), not only to a synthetic sampling distribution.

No second processor of Willow’s combined scale and fidelity had independently reproduced the full beyond-classical instance at announcement. Verification was therefore a claimed property of the task, not a completed cross-lab replication.⁠

Hardware Requirements

Time-reversal is unforgiving. Gate errors, decoherence, and calibration drift destroy the echo instead of amplifying it. Google attributed the result to post-release Willow improvements:

Metric (reported for improved Willow)	Approximate value
Single-qubit gate fidelity	99.97%
Entangling-gate fidelity	99.88%
Readout fidelity	99.5%
Gate times	tens–hundreds of ns
Measurement throughput	millions of shots in tens of seconds

Those figures are system-scale simultaneous-operation numbers, not isolated best-qubit lab results. The algorithm needs both precision and volume: many gates and many shots to extract a weak interference signal from noise.

Classical Cost and Red-Teaming

Google said it spent roughly 10 person-years attacking the result classically and implemented about nine simulation algorithms. The published advantage is therefore against the best methods the team and collaborators could field in 2025, not against a proof that no better classical algorithm exists. Reviewers noted the same caveat that followed the 2019 supremacy paper: classical algorithms improve.⁠

Scientific Use Case: Toward Molecular Structure

OTOCs are used in studies of quantum chaos, black-hole information scrambling, and many-body thermalization. Google and collaborators also explored a nearer-term application: combining Quantum Echoes with nuclear magnetic resonance (NMR) data as a “molecular ruler.”

In a proof-of-principle with University of California, Berkeley researchers, the method was applied to two organic molecules to estimate distances between hydrogen atoms. That demonstration was not itself a beyond-classical chemistry calculation. It was an early attempt to connect the OTOC observable to laboratory molecular structure—the same family of information NMR and MRI extract, but potentially over longer distances if the method matures.⁠

How This Differs from RCS

Feature	Random circuit sampling (2024)	Quantum Echoes (2025)
Output	Sampled bitstrings	Expectation value (OTOC)
Classical hardness	Exponential simulation of the distribution	Simulation of interfering time-reversed dynamics
Independent check	Hard; relies on classical cost models	Designed to be repeatable on another quantum device
Scientific content	Benchmark only	Scrambling, chaos, possible NMR/Hamiltonian learning
Reported gap	Minutes vs. ~ $10^{25}$ years	~13,000× vs. Frontier-class methods

RCS still probes raw circuit volume. Quantum Echoes asks whether that volume can be spent on an observable that physicists already care about.

Limits and Open Questions

The advantage is algorithm- and hardware-specific. A better classical OTOC simulator would shrink the gap.
Cross-platform verification of the largest instances had not been completed by other groups at announcement.
The NMR/molecular-ruler work remains early; it does not yet replace conventional spectroscopy or deliver a commercial chemistry workflow.
The experiment still uses noisy physical qubits, not fault-tolerant logical qubits. It sits between NISQ benchmarking and useful fault-tolerant computation.
Google’s broader timeline language after the result still pointed to years, not months, before broadly useful machines.

Why It Matters in the Willow Story

The 2024 papers showed that Willow’s errors can be suppressed by larger codes and that its raw circuits outrun classical simulation. Quantum Echoes used those same hardware strengths—speed, simultaneous fidelity, and measurement throughput—on a task whose answer is a physical number rather than a pile of random bits.

That is the real shift. Willow moved from “this chip can do something classically hard” to “this chip can estimate a checkable many-body observable faster than known classical methods.” Whether that observable becomes a routine tool for chemistry, materials, or fundamental physics depends on independent replication, better classical competitors, and the next steps toward logical gates—not only on the 13,000× headline.

6) Context, Limitations, and Realistic Outlook

Where Willow Sits in the Field

Willow is Google Quantum AI’s superconducting flagship: 105 transmon qubits, square lattice, tunable couplers, fabricated in Santa Barbara. By mid-2026 it remained the platform behind three public milestones:

2019 lineage continued: beyond-classical random circuit sampling.
December 2024: first clear below-threshold surface-code memory on superconducting hardware.
October 2025: Quantum Echoes (OTOC) presented as a verifiable, beyond-classical observable.

Google’s own roadmap is a six-milestone sequence toward a large error-corrected machine. Willow covers early milestones (beyond-classical computation and a quantum-error-correction prototype / below-threshold memory). The next published targets are a long-lived logical qubit and logical gates between logical qubits. Those had not appeared as completed peer-reviewed demonstrations by mid-2026. The long-range destination is on the order of one million physical qubits bundled into far fewer stable logical qubits; Sundar Pichai has publicly spoken of a useful error-corrected computer around 2029. That is a goal, not a delivered system.⁠

Competing approaches matter because they attack different bottlenecks:

IBM (superconducting, heavy-hex): larger raw qubit counts, modular scaling, a 2029 target around hundreds of logical qubits rather than a million physical qubits.
Neutral atoms (QuEra, Atom Computing, and Google’s own Boulder effort added in March 2026): better physical-to-logical conversion in some high-rate codes; Google frames atoms as complementary for qubit count, superconductors as complementary for circuit depth and speed.
Trapped ions (Quantinuum and others): higher two-qubit fidelities, slower gates, different scaling economics.

No platform had a utility-scale, fault-tolerant machine in 2026. Willow’s distinctive claim is experimental proof that surface-code errors can fall as the code grows on a fast superconducting chip—not that Google already owns the end-state architecture.

What Willow Is Not

These distinctions are easy to blur in headlines and worth stating directly.

It is not a universal fault-tolerant computer. The headline error-correction result is a quantum memory: one logical qubit stored and protected. It is not a demonstrated factory of logical two-qubit gates, T gates, or multi-logical algorithms.
It is not close to breaking cryptography. A distance-7 logical qubit used ~101 physical qubits at ~0.143% error per cycle. Useful factoring estimates typically need thousands of logical qubits at error rates nearer $10^{-10}$ to $10^{-12}$ . That is many orders of magnitude in error rate and many orders of magnitude in qubit count.⁠
RCS and Quantum Echoes are not production applications. RCS is a synthetic sampling benchmark. Quantum Echoes measures a physically interesting correlator and has a plausible NMR / Hamiltonian-learning path; the molecular-ruler work was proof-of-principle, not a replacement for laboratory spectroscopy.
“Verifiable” in 2025 meant the task can be repeated on similar hardware. It did not mean an independent lab had already reproduced the full beyond-classical Echoes instance.

Concrete Technical Limits

Error rates versus algorithm needs.
Google’s own Willow paper noted that reaching a $10^{-6}$ logical error rate at then-current physical performance would require about a distance-27 code and ~1,457 physical qubits per logical qubit. Practical algorithms often want still lower errors. The demonstrated $\Lambda \approx 2.14$ is the right direction; it is not yet the required magnitude.

Memory versus computation.
Protecting an idle logical state is necessary and was the missing experimental proof. Computation requires:

logical single- and two-qubit gates below threshold
magic-state distillation or another non-Clifford resource
real-time decoding that keeps up as many logical qubits run at once
hours-to-days of stable operation, not million-cycle memory experiments

Correlated errors.
Repetition codes out to distance 29 showed a floor set by rare bursts about once per hour ( $\sim 3 \times 10^9$ cycles). Gap engineering reduced cosmic-ray–driven quasiparticle tunneling, but later work found residual correlated phase / frequency-shift bursts lasting ~1 ms. Codes that assume independent errors handle these poorly. Until that floor is understood and suppressed, large-distance scaling is not automatic.⁠

Classical competition.
RCS advantages have historically narrowed when simulation methods improve. Quantum Echoes was, as of late 2025–2026 commentary, the Willow-era result least quickly “de-quantized,” but it still rests on a classical-cost estimate, error mitigation, and extrapolation beyond the sizes where exact checks are possible. Advantage claims should be treated as provisional until they survive independent algorithms and independent hardware.

Engineering overhead.
A million-physical-qubit superconducting machine implies cryogenics, wiring, calibration, and classical decode bandwidth far beyond a 105-qubit chip in a single fridge. Drift already forces heavy calibration; 2026 reinforcement-learning control work on Willow improved logical stability under injected drift, which is progress on operations, not a substitute for more qubits.⁠

A Realistic Scaling Picture

A useful way to read the next decade is as stacked gaps, not a single remaining invention:

Gap	Approximate status after Willow	What still has to happen
Below-threshold memory	Demonstrated at d = 5 and d = 7	Hold $\Lambda > 1$ at much larger d
Logical lifetime	~2.4× best physical qubit	Hours of stable logical memory
Logical gates	Not the 2024 headline result	Below-threshold logical two-qubit gates
Logical error rate	$\sim 10^{-3}$ per cycle	$10^{-6}$ to $10^{-12}$ depending on algorithm
Qubit count	105 physical	$10^4$ – $10^6$ physical for useful logical counts
Verification	Echoes designed to be repeatable	Independent hardware replication
Applications	NMR-style proofs of principle	End-to-end chemistry / materials workflows

If $\Lambda$ Λ stays near 2 and correlated bursts do not worsen, error rates can be pushed down by growing the code. The cost is steep: each extra two in distance roughly doubles the protection but consumes far more physical qubits, couplers, and control lines. That is why Google’s million-qubit language and IBM’s “fewer physical qubits, more logical qubits by 2029” language can both be sincere and still describe machines that do not exist yet.

Public timelines in 2026 clustered around late decade for a first scientifically useful fault-tolerant device (Google ~2029 language; U.S. DOE discussion of an ambitious 2028 scientific machine). Those dates are program targets under optimistic engineering, not guarantees. Missing a date would not erase Willow; it would mean the overhead and correlated-noise problems took longer than the press cycle.⁠

How to Read the Result Without Over- or Under-Selling It

Fair reading: Willow is the first superconducting processor to show the textbook signature of the surface-code threshold on real hardware, then reuse that hardware quality on an observable (Quantum Echoes) that is closer to physics than RCS. That combination is a genuine phase change in evidence quality.

Over-selling: treating 105 qubits, a 13,000× Echoes headline, or a 2029 slide as equivalent to drug design, materials discovery, or broken public-key crypto.

Under-selling: treating the result as “just another noisy chip.” For thirty years the field could not show that bigger surface codes help on superconducting devices. Willow showed they can.

The honest outlook is linear-sounding and still hard: keep lowering physical error, keep killing correlated bursts, demonstrate logical gates, grow both code distance and logical-qubit count, and put those logical qubits on problems whose answers can be checked. Willow made that path experimentally plausible. It did not finish the path.