Quantum Error Correction Leaps Cut Qubits Needed to Crack RSA

Why early April mattered

Round-ups of the week of April 10, 2026 put AI and quantum computing at the top. The quantum story that mattered most was not a new chip with a bigger qubit count. Two research groups showed that a machine able to break today's public-key encryption could be much smaller than the field had assumed. Error correction was the main reason, and AI tools helped get there.

The two results came out within days of each other. A group of Caltech physicists published a design for an encryption-breaking quantum computer that needs only tens of thousands of qubits, and announced a company, Oratomic, to build it5. In the same window, Google researchers released a version of Shor's algorithm that they say is ten times more efficient than the best earlier method5. Neither group has hardware that can break encryption today. Still, the coverage broadly agreed that capable machines may now be years away rather than decades5.

The effect was quick. Cloudflare, which protects a large share of web traffic, said it was moving its deadline for quantum readiness up to 2029. One of its researchers called the papers a real shock3. Google and Microsoft have since set 2029 as the deadline to finish their own post-quantum cryptography migrations22.

The breakthrough was in error correction, not qubit counts

Oratomic's advance is an error-correction result. Qubits are fragile, so information is spread redundantly across many physical qubits to form one more reliable "logical" qubit. In neutral-atom machines, that has historically meant 100 to 1,000 atoms per encoded qubit3.

The Caltech team adapted quantum low-density parity-check (qLDPC) codes to neutral-atom hardware. These codes link qubits that sit far apart in the array, and in return they fit far more logical qubits into an array of a given size5. The overhead figures differ slightly between outlets. TIME reported that the scheme needs about three atoms per qubit, roughly a hundredfold cut in particle count3. Quanta described a code that builds one logical qubit from four atoms and tolerates 20 to 24 catastrophic errors. An earlier strong qLDPC code needed 12 physical qubits per logical one and handled up to 12 such errors5. Both accounts make the same point: the overhead dropped by orders of magnitude.

The resulting estimates are striking. The team's simulations suggest that 10,000 atoms could break common RSA in about a century, while 100,000 atoms could do it in roughly three months5. Elliptic-curve cryptography looks weaker: about three years with 10,000 atoms, or a few days with 26,0005. Google's separate paper estimated that most cryptocurrencies could be broken in minutes by a machine with fewer than 500,000 qubits5. Princeton's Jeff Thompson called Google's tenfold cut in space-time cost hugely significant5.

AI's role, and the pushback

The AI angle is what pushed this story into general news. Robert Huang, an Oratomic co-author, used OpenEvolve, an open-source tool that runs large language models such as Gemini and Claude through an evolutionary search. Before that, he said, the team's key algorithms performed about 1,000 times worse than they needed to3. Quanta reported that the model also found an efficient decoder, the routine that identifies errors and works out how to fix them5. Co-author John Preskill said he was surprised by how far the qubit count fell, but stressed that humans still drove the research3.

The skepticism is real and should be weighed. The paper had not been peer-reviewed when it was covered. Thompson, who also runs the rival neutral-atom startup Logiqal, said many of its assumptions are untested and that it is easy to shrink a machine on paper by assuming better qubits3. He also called the team's assumptions about operation speed aggressive. The design expects the full error-correction cycle, including replacing stray atoms, to run once every millisecond5.

There was also a new kind of secrecy. Google used a zero-knowledge proof to show that its method works without publishing how it works5. Oratomic briefed U.S. officials before publishing and chose carefully what methodology to release3. In my view, this caution about publishing is the clearest sign of how seriously the researchers themselves take the timeline.

How the companies positioned themselves

The papers landed in a crowded season. On March 24, Google announced a neutral-atom hardware team in Boulder led by physicist Adam Kaufman22. Hartmut Neven framed it as a split: superconducting chips scale more easily in circuit depth, while neutral atoms scale more easily in qubit count22. This put Google in the same modality as Oratomic. That is notable because Huang had worked at Google Quantum AI before co-founding the startup3. Google said it had been studying atomic approaches for years3.

The contrast with IBM is the clearest strategic split in the field. IBM's roadmap stays fully superconducting and relies on qLDPC codes to close the efficiency gap, while Google is hedging across two modalities.

Other April news fit the same pattern. NVIDIA released its Ising family of open-source models for AI-based processor calibration and error-correction decoding, claiming decoding up to 2.5 times faster and 3 times more accurate than traditional methods6. Rigetti deployed Cepheus-1, a 108-qubit system built from twelve 9-qubit chiplets, described as the largest modular quantum system so far21. Discover's April 11 overview chose stability and efficient error correction as the defining themes, citing Quantinuum's 98-qubit Helios and the 10,000-atom Shor's algorithm result4.

The logical-qubit leaderboard is crowded, and messy

Error correction is clearly the field's main focus. How to score it is less clear, and the tallies do not agree.

QuEra is widely credited with the largest count: 96 logical qubits from 448 neutral atoms, using a high-rate [[16,6,4]] qLDPC code at about 4.7 physical qubits per logical one, published in Nature in January 20261115. Yet one tracker cites a 48-logical-qubit QuEra result instead20. Quantinuum's Helios is listed with 48 fully error-corrected logical qubits by some outlets and 50 by others112220. Some trackers note that several headline counts rely on error-detection codes, which discard bad runs instead of fixing them, and mark those numbers separately15.

Some claims should be discounted outright. One post said Google ran 12 logical qubits at once in March 202614. Trackers that follow Google closely list only a single verified logical qubit on Willow and report no announcement that Google has reached its next milestone, a long-lived logical qubit1522. Another report listed IBM's Kookaburra at 4,158 qubits in its 2025–2026 lineup13, but as of late September, IBM had not announced delivering it22. A June press release from AIX Global Innovations claimed fault tolerance on rented IBM Heron hardware, with about one physical qubit per logical qubit1. It is a company announcement backed by a self-published report, and no independent reporting has confirmed it.

The baseline is still Google's Willow. Its distance-7 code reached a logical error rate of 0.143% per cycle, and each step up in code size cut errors by a factor of about 2.1411. That is real exponential suppression. It is still orders of magnitude worse than the roughly one-in-a-million error rates that large algorithms are thought to need6.

What has happened since

Several events after April support the view that error correction is moving from physics experiments to engineering. In June, Microsoft and Quantinuum reported error correction during an actual computation, not just while storing data, across up to 12 logical qubits. Errors fell between 11-fold and 800-fold compared with unprotected runs11. In July, IBM and the University of Chicago encoded 70 logical qubits to solve a problem that classical computers cannot practically handle. Logical error rates were 10 times lower than physical ones19. IBM's research director Jay Gambetta said the field is now in the quantum advantage era19. In September, Infleqtion reached 30 entangled logical qubits from 80 physical ones20.

The roadmaps are converging on the same years. IBM targets Starling, with 200 logical qubits, in 202920. Quantinuum plans Sol for 2027 and the fault-tolerant Apollo for 202922. The Atom Computing and Microsoft system Magne aims for 50 logical qubits by late 202627. Analysts caution that quantum roadmaps have often slipped28.

The takeaway

I read April 2026 as the point when the main question in quantum computing changed from "how many qubits?" to "how few?" Better codes, decoders found with AI help, and tighter algorithms cut resource estimates faster than hardware grew. The hardware still has to catch up, and the boldest assumptions remain unproven. But security teams were right to act on the estimates without waiting for the machines.