Between June and October 2026, Microsoft defended its topological qubit against a critique published in Nature. Jay Gambetta, director of IBM Research and an IBM Fellow, said that “[w]e are now firmly in the quantum advantage era.” Two companies published estimates of the machine needed to break RSA-2048, one of them in a blog post without a paper.
A reader without a physics degree can test claims like these with seven questions, each answered by a public document, often the paper or post behind the announcement.
Physical or logical, and how many
Infleqtion said on September 24, 2026, that it had entangled 30 logical qubits on Sqale, its neutral-atom computer, using 80 atoms. A physical qubit is one piece of hardware, here a single atom held by a laser. A logical qubit spreads one unit of information across several physical qubits to make errors detectable or correctable.
In a technical post, Chief Technology Officer Pranav Gokhale wrote that the team used the [[8,3,2]] code, “which can detect any error or correct for a lost atom.” The last number in that notation is the code’s distance. A distance-2 code can flag a single error but can correct one only if its position is known, as with a lost atom.
Gokhale wrote that applying loss correction let the team “quadruple the number of good shots,” at “the cost of greater error rates.” Neither the release nor the post said what share of runs the team discarded after its checks, as Quantum Observer reported. He also wrote that Infleqtion planned a full paper “in the coming weeks.” As of November 12, the paper had not appeared.
For any qubit count, a reader should ask three things: physical or logical, which code at what distance, and how many runs the team kept.
Paper, preprint, or press release
On October 22, Iceberg Quantum, a Sydney company that designs fault-tolerant architectures, estimated that 72,000 physical qubits with a 0.1% error rate could factor a 2,048-bit RSA key in an expected 3.7 days. In a February preprint, a paper posted publicly before peer review, Iceberg had estimated about 98,000 physical qubits for a run of up to a month (about 94,000 after a May revision). The October figures appeared in a blog post, with no paper.
In the post, Iceberg reported a mean decoding time of “four microseconds per round of error correction,” against the one-microsecond cycle it assumed for the headline estimate. A decoder is the classical system that works out which errors occurred. Iceberg did not say whether the four microseconds came from a simulation or how a machine would decode in real time.
A reader should note the kind of document behind a claim. Outside researchers can test a preprint’s methods before any journal does. With a blog post alone, a reader has only the company’s numbers.
Useful problem or demonstration
In a review published on September 24, five researchers at Sapienza University of Rome and the University of Tuscia examined 53 experimental studies that used quantum computing in bioinformatics. They found that 36 “did not execute on quantum hardware at all” and that nearly 77% of the evaluated implementations were “restricted to heavily reduced toy datasets.”
Every one of the 40 studies that compared a quantum method with a classical one reported performance “either comparable to or better,” the reviewers wrote. They concluded that “[n]one of the considered articles explicitly demonstrates a quantum advantage over classical approaches.”
For an application claim, a reader should check the size of the problem and the machine that ran it. A team that reports a result on a reduced dataset or a simulated circuit has demonstrated a method on a problem smaller than the one a user needs solved.
Who benefits
In a preprint posted on October 28, researchers at Alice & Bob, a Paris and Boston company that builds superconducting cat qubits, estimated that a machine built on its architecture could factor RSA-2048 in 3.14 days with 22,593 physical qubits, as Quantum Observer reported. Nine of the 10 authors list Alice & Bob as an affiliation. The tenth, André Schrottenloher, is at the University of Rennes and Inria. In its release, the company said its design “needs 40 times fewer physical qubits than Google Quantum AI’s Craig Gidney 2025 estimate.”
Alice & Bob builds the hardware that its estimate favors. In the same release, the company wrote that “[t]he headline figures assume no bit flips during the computation, equivalent to a bit flip time of about 1 century.” A bit flip is an error that turns a 0 into a 1 or the reverse. The authors wrote that an added layer of error correction would relax the requirement to about 1 hour, at a cost of about 45,000 physical cat qubits.
In an October blog post, Alice & Bob reported a bit-flip time of 44 minutes for a single cat qubit. It described the work as “preliminary results on very new techniques” that it still had to replicate in an error-correction experiment. On three of the 12 physical qubits in its Helium 2 processor, it reported 189 to 252 minutes. It wrote that this measurement alone was “not sufficient to claim that all these systems can now be operated as cat qubits and with the same level of control as the previous device.”
A reader who asks who benefits should then find the assumption that favors the claimant, here the bit-flip time, and look for an outside test of it.
Independent checks
Microsoft announced Majorana 1 on February 19, 2025, as “the world’s first Quantum Processing Unit (QPU) powered by a Topological Core,” citing a Nature paper published the same day. Topological qubits are a design meant to protect quantum information from local noise.
On June 24, 2026, Nature published a peer-reviewed critique by Henry Legg of the University of St Andrews and the University of Basel. Legg analyzed transport data that the 2025 paper had not presented. He argued that the claimed readout took place in strongly disordered regimes that appear gapless, lacking the energy gap that topological protection requires. In a reply published alongside it, Microsoft researchers wrote that their radio-frequency measurements “strongly indicate a topological origin.”
“We stand by our results and our roadmap,” Chetan Nayak, Microsoft’s technical fellow and corporate vice president for quantum hardware, told Scientific American. On September 22, Microsoft said it would provide Majorana 2 qubits for “independent testing and evaluation,” with “full access” for the Defense Advanced Research Projects Agency (DARPA), as Quantum Observer reported.
For a contested claim, a reader should ask who outside the company has tested the devices or the data, and where the results appeared. Until DARPA’s evaluators or another laboratory report on Microsoft’s devices, the topological claim depends on Microsoft’s data and its reading of them.
Hardness and classical cost
IBM said in its July 30 release that its processor took about 15 minutes for a task that would “demand infeasible amounts of time with leading classical computing methods.” In their preprint, which has not been peer reviewed, IBM’s authors wrote that they had proved the task’s “asymptotic hardness,” a statement about circuits as they grow. For typical instances, they wrote, the result also depends on two complexity-theoretic conjectures.
Quantum advantage means doing the same task, to the same quality of answer, more cheaply or quickly than the best classical methods can. A reader should therefore ask what the classical computation cost at the size IBM actually ran.
Two weeks later, Hidetaka Manabe and Feng Pan, both of the Singapore University of Technology and Design, and Hanfeng Gu of NVIDIA computed the amplitudes, the numbers that set each output’s probability, for all of IBM’s published output bitstrings. They needed 37.3 minutes on 256 NVIDIA H100 graphics processors, about 159 GPU-hours. They wrote that amplitude evaluation alone “does not constitute a classical simulation of the sampling task itself.”
Under two stated assumptions, they found their estimate of the experiment’s fidelity, a gauge of how closely its output matched the ideal, “numerically compatible with IBM’s fidelity lower bound.” They also projected that, under a convention that weights samples by fidelity, the same 256 processors could produce a sample set matching IBM’s fidelity bound in about 10.6 minutes.
The three became co-authors of the revised preprint. In that September revision, IBM’s team rebuilt its headline experiment around a different circuit. The company released its circuits on a public tracker for others to test, as Quantum Observer reported.
How extraordinary the claim is
Daniel Simon of Amazon Web Services posted a draft to the Cryptology ePrint Archive on August 3, claiming a polynomial-time quantum algorithm for the dihedral coset problem (DCP). In such an algorithm, the running time grows no faster than a fixed power of the problem size.
Combined with published reductions, he wrote, it would yield efficient quantum algorithms for lattice problems including learning with errors. ML-KEM and ML-DSA, formerly CRYSTALS-Kyber and CRYSTALS-Dilithium, depend on a structured variant of that problem, module learning with errors. ML-DSA also depends on a variant of a second lattice problem, module short integer solution. Simon described no attack on either standard.
Simon labeled the paper a “Preliminary Draft” and, in the version available on August 6, marked four proofs as sketches. On August 15, Aparna Gupte of MIT, Seyoon Ragavan of Google and MIT, and Mark Zhandry of Google and Stanford posted “The ePrint:2026/1591 Quantum Algorithm Does Not Solve DCP” and released code for their results in Lean 4, a proof assistant. They wrote that they were “showing directly that the algorithm cannot possibly work.”
On the archive page, Simon wrote that he was evaluating the rebuttal and, on October 7, that he had updated the preprint to address its claims and clarify the proof of Lemma 3. He has revised the draft repeatedly, most recently on November 3, without withdrawing the claim, as Quantum Observer reported.
For a claim that would overturn the hardness assumption behind NIST’s main post-quantum standards, a reader should wait for outside checks before acting. NIST still recommended ML-KEM-768 as the default in its final transition plan on October 27.
A result that answers them
Google Quantum AI researchers answered the first three questions in their December 2024 Willow paper in Nature. They reported a distance-7 code on 101 physical qubits and measured the logical error rate falling by a factor of 2.14 each time the code distance rose by two. They described the result as error correction “below the surface code threshold” in a quantum memory, which stores information without computing with it.
Where the questions stop
The strongest objection to a reader’s checklist is that experts with access to private data have already done the checking. Nayak told Scientific American that DARPA had “moved Microsoft into the final phase of its Quantum Benchmarking Initiative” after independently evaluating the company’s results.
On October 7, DARPA added Atom Computing, Diraq, IBM, and IonQ to that stage, which Microsoft and PsiQuantum had entered in February 2025 through a predecessor program. DARPA published no measurements from its evaluations in that announcement.
Micah Stoutimore, the initiative’s managing director, said on October 7 that the agency increasingly expects someone to build a utility-scale quantum computer, one with computational value greater than its cost, by 2033. If DARPA publishes its evaluators’ measurements, outside researchers will be able to check the six designs in its final stage. If it does not, only DARPA and the companies will have the decisive evidence about those machines.
