Bottom line. Quantum computing has passed genuine milestones, most importantly error correction that improves as it scales. There is not yet a widely accepted demonstration of a commercially useful problem solved faster than the best classical method. Most confusing headlines mix up three different things: a physical-qubit count, a benchmark designed to be hard for classical computers, and a roadmap target. Five questions sort them out.
Physical qubits are not logical qubits
Physical qubits are noisy. They lose their quantum state quickly, and every operation on them introduces errors. Quantum error correction combines many physical qubits into one logical qubit whose error rate is lower than that of its parts. This works only if the physical error rate is below a threshold. Below it, adding qubits makes the logical qubit better. Above it, adding qubits makes things worse.
In December 2024, Google Quantum AI reported operating below that threshold [1]. On its Willow processor, a distance-7 surface code used 101 physical qubits to hold one logical qubit, with a logical error rate of 0.143% per error-correction cycle. Each step up in code size, from distance 3 to 5 to 7, cut the logical error rate by a factor of about 2.14. The logical qubit also outlived the best physical qubit on the chip by a factor of 2.4.
This is real progress. It is also a measure of the distance remaining. One error-corrected memory took about a hundred physical qubits. Resource estimates for useful algorithms, such as factoring a 2048-bit RSA key, call for around a million noisy physical qubits [2].
A benchmark is not an application
In 2019, Google reported that its 53-qubit Sycamore processor completed a random circuit sampling task in about 200 seconds, and estimated that a leading supercomputer would need 10,000 years [3]. Classical researchers then improved their methods. In 2022, one group solved the same sampling problem with tensor-network techniques on a cluster of graphics processors, in hours rather than millennia [4]. In 2024, Google reported that Willow completed a harder sampling task in under five minutes, against an estimated 10²⁵ years for a classical supercomputer [5].
Random circuit sampling was designed to be hard for classical computers. It was not designed to be useful, and it has no known commercial application. It is a valid test of hardware. It is not evidence that a business problem will run faster.
In October 2025, Google reported a different kind of result. An algorithm it calls Quantum Echoes, which measures a quantity known as an out-of-time-order correlator, ran about 13,000 times faster on Willow than the best classical algorithm it tested on a supercomputer [6]. The result is more meaningful than random sampling in two ways: it can be verified by repeating it on comparable hardware, and it connects to a physical measurement, molecular structure from nuclear magnetic resonance (NMR) data. It is also one group’s report. Comparisons against classical methods tend to tighten as other researchers respond.
A roadmap is a target, not a result
Hardware roadmaps are useful for planning, and they are forecasts. IBM, for example, has announced a target of a fault-tolerant system called Starling by 2029, designed to run 100 million quantum operations on 200 logical qubits [7]. Targets like this should inform when to prepare. They should not be read as delivered capability. The reliable signal is a result that has been published, measured and, ideally, reproduced.
Five questions to ask of any claim
- Physical or logical? Is the qubit count physical or error-corrected? How many physical qubits per logical qubit, and at what error rate?
- Benchmark or application? Was the task chosen because it is hard for classical computers, or because someone needs the answer?
- Against which classical method? Was the comparison made against the best known classical algorithm, tuned with comparable effort, or against a convenient one?
- Measured or projected? Is the result demonstrated, or extrapolated from a roadmap or a resource estimate?
- Reproduced? Has anyone outside the reporting group confirmed it?
| Claim type | What it shows | What it does not show |
|---|---|---|
| Physical-qubit count | Scale of fabrication | Error rates or useful capacity |
| Logical qubit below threshold | Error correction works and improves with size | That useful algorithms can run yet |
| Random circuit sampling | Hardware does something classically hard | Any commercial use |
| Verifiable advantage on a physics task | A checkable result beyond tested classical methods | Advantage on a business problem |
| Roadmap target | A plan and a direction | Delivered capability |
The classical baseline is part of the answer
Question 3 matters more than it seems. In 2018, Ewin Tang showed that a quantum algorithm for recommendation systems, previously thought to offer an exponential speed-up, could be matched by a classical algorithm inspired by it, up to polynomial factors [8]. This approach, called dequantization, was later extended to a broader family of quantum machine-learning proposals [9].
The lesson works in both directions. Quantum-inspired classical methods can deliver some of the promised benefit on today’s computers. And any claimed quantum advantage is only as strong as the classical method it was compared against.
How we apply it
Our quantum work follows from these questions. We prepare systems now for the part of quantum computing that is certain: post-quantum cryptography standards and their migration deadlines. We track hardware milestones by the five questions, not by headlines. Where a problem has a quantum-inspired formulation, we test it against a well-tuned classical baseline before assuming quantum hardware is needed.
Conclusion
Quantum computing is moving from physics experiment toward engineering, and error correction below threshold is the clearest sign. The useful response is neither excitement nor dismissal. Ask of every claim what was measured, against what, and by whom. Then prepare for the parts that are already certain.
References
- Google Quantum AI and collaborators, “Quantum error correction below the surface code threshold,” Nature, vol. 638, pp. 920-926, 2025.
- C. Gidney, “How to factor 2048 bit RSA integers with less than a million noisy qubits,” arXiv:2505.15917, May 2025.
- F. Arute et al., “Quantum supremacy using a programmable superconducting processor,” Nature, vol. 574, pp. 505-510, 2019.
- F. Pan, K. Chen, and P. Zhang, “Solving the sampling problem of the Sycamore quantum circuits,” Physical Review Letters, vol. 129, 090502, 2022.
- H. Neven, “Meet Willow, our state-of-the-art quantum chip,” Google, Dec. 2024.
- Google, “Our Quantum Echoes algorithm is a big step toward real-world applications for quantum computing,” Google, Oct. 2025.
- IBM, “IBM sets the course to build world’s first large-scale, fault-tolerant quantum computer at new IBM Quantum Data Center,” IBM Newsroom, Jun. 2025.
- E. Tang, “A quantum-inspired classical algorithm for recommendation systems,” in Proc. 51st Annu. ACM Symp. Theory of Computing (STOC), 2019.
- N.-H. Chia, A. Gilyén, T. Li, H.-H. Lin, E. Tang, and C. Wang, “Sampling-based sublinear low-rank matrix arithmetic framework for dequantizing quantum machine learning,” in Proc. 52nd Annu. ACM Symp. Theory of Computing (STOC), 2020.