Quantum Error Correction Race: DARPA Advances Four to Stage C

The verifier now sets the pace

For most of this decade, quantum computing advanced in two ways: company press releases and academic papers. The past few weeks show a third force taking hold, which is the government verifier. On October 7, DARPA moved Atom Computing, Diraq, IBM and IonQ into the final phase of its Quantum Benchmarking Initiative. In Stage C, the agency's independent verification and validation team will check whether each proposed machine can be built as specified and will perform as designed.17 The four companies use four different kinds of qubit: neutral atoms, silicon CMOS spin qubits, superconducting circuits and trapped ions. All four came from the 11 companies selected for Stage B in November 2025.17

This is the most important development of the season, because it changes what counts as proof. The research flow is still heavy. ScienceDaily's quantum section has run stories in recent weeks on a superfluid-helium qubit design said to cut error rates about 100-fold, on non-Abelian anyons shown capable of universal computation using 54 qubits on Quantinuum's H2 processor, and on a Chalmers method that speeds up some operations more than 1,000 times.4 But what the field increasingly needs is checking from outside the company making the claim, and that is exactly what is now being put in place.

Error correction is now a systems problem

The clearest result of 2026 is that error correction is no longer only about building better qubits. It is also about the ordinary computers that have to keep up with them. IonQ announced on September 22 what it calls the industry's first end-to-end real-time quantum error-correction decoder, and it runs on one standard, off-the-shelf CPU.31 The company tested the design on benchmark circuits simulating up to 408 logical qubits across 88 memory blocks and "magic factories", with more than 31.5 million quantum operations in total.335 The CPU was an Apple M4 Max laptop chip. The workload included up to one million T-gates, which are the costly non-Clifford operations that fault-tolerant algorithms rely on.38

Investors responded strongly. IonQ shares rose more than 10% before the market opened and were still up more than 5% in morning trading.2 Some enthusiast coverage called the result a "monumental leap" that solves the decoding bottleneck.7 More careful reporting pointed out that the 408-logical-qubit benchmark ran on a proposed architecture, the "Walking Cat" design published in April, and used simulated syndrome streams rather than live hardware data.38 Other observers added that the result still needs to be reproduced independently and shown on real hardware at scale.5

The careful reading is the right one. The real finding is about economics: if decoding can keep up on commodity processors, the classical equipment around a fault-tolerant machine need not grow out of control. IonQ's own executives describe the goal as cost-to-solution.38 That matters, but it is a result about infrastructure. It is not a working fault-tolerant computer. IBM has made progress on the same problem, reporting decoding times under half a millisecond, fast enough to keep pace with superconducting gates.9

The academic side shows the same theme. Niels Bohr Institute researchers built an FPGA-based controller that tracks changes in qubit relaxation in real time, about 100 times faster than earlier methods. The lead researcher admits, though, that much of the fluctuation they see still has no explanation.8 Chalmers researchers showed that operations on bosonic error-correcting states can be done in one driving cycle instead of thousands, which shortens the window in which errors can build up.3

The logical-qubit scoreboard depends on definitions

Logical qubits have replaced raw qubit counts as the headline number, but the reported figures are less consistent than they look. One widely used tracker lists QuEra at up to 96 logical qubits, Quantinuum's Helios at 48, Atom Computing with Microsoft at 24, and Google's Willow at one verified scaling demonstration.32 That same tracker notes in its table, however, that Quantinuum's and Atom Computing's figures are "detected" errors, not corrected ones, while QuEra's are labeled corrected.32 A separate vendor guide warns that QuEra's 96 is the peak of a single experiment from its 448-atom architecture, not a settled record, because the Nature paper itself describes its deep-circuit results as "dozens" of logical qubits.14

IBM's 70-logical-qubit result with the University of Chicago is the strongest computational claim of the year. The team ran 2,415 logical two-qubit gates and 468 T gates in about 15 minutes, with logical error rates roughly ten times lower than the underlying physical rates. They argue that leading classical simulation methods could not practically reproduce it.369 Here the reporting also splits. IEEE Spectrum, as summarized by one milestone tracker, said none of IBM's three July advantage papers had been peer reviewed, and quoted an outside researcher urging caution on two of them.12 ScienceDaily's archive offers a reminder of how quickly such claims can be overturned: in July, researchers used tensor networks on a laptop to solve a problem once thought to need a quantum computer.4

Google has been quieter. Its Willow chip showed in December 2024 that logical errors fall as the code grows from a 3x3 to a 7x7 grid.13 However, one roadmap analysis found no announcement that Google has completed its third milestone, a long-lived logical qubit.13 In March, Google also started a neutral-atom hardware team in Boulder, a hedge suggesting it does not think superconducting chips alone will get it there.13 Microsoft's Majorana 2 reports qubit lifetimes over 20 seconds6, but Nature has reported continued skepticism among researchers about the company's broader topological claims.10

The roadmaps all point to 2029

The major companies' public roadmaps cluster around the end of the decade, and the clustering is notable. IBM's Starling is meant to run 100 million gates on 200 logical qubits by 2029. It depends on Kookaburra, a qLDPC memory module due this year, which IBM had not announced as delivered by late September.13 Quantinuum says its Apollo fault-tolerant system "remains on schedule for 2029," with the Sol system coming in 2027.13 IonQ opened orders for its 256-qubit Superion platform, with deliveries in 2027, and says it expects lab-setting fault tolerance that same year.12

Money is backing these plans. Quantinuum raised about $1.68 billion in its June Nasdaq IPO and later received a $100 million CHIPS Act award.19 IonQ raised its 2026 revenue guidance to $450–460 million, a figure that now includes the SkyWater foundry it acquired.19 HPCwire's same-day news list also shows a conditional $150 million loan commitment to PsiQuantum.17

The Department of Energy is trying to set its own deadline. Its September 25 roadmap calls for a "scientifically relevant" error-corrected machine by 2028. One close reading of the 58-page report finds that it gives only "illustrative" ranges of 50 to 100+ logical qubits and 10⁴–10⁵ hard operations, with no target logical error rate. The stricter yardstick is the Genesis Q Competition. Its sample agreement requires at least 100 logical qubits on standardized circuits, checked by a DOE verification team, with evaluation expected in September 2028. Taken together with DARPA's Stage C, the direction is clear: in 2028, claiming you met a roadmap will count for less than passing a government test.

The cryptography deadline is already set

Post-quantum cryptography no longer waits on the hardware. Executive Order 14412, signed June 22, and OMB Memorandum M-26-15 require federal high-value and high-impact systems to use PQC for key establishment by December 31, 2030, and for digital signatures by the end of 2031.21 A coming FAR rule is expected to pass similar requirements on to federal contractors.26 The EU roadmap likewise requires high-risk systems to migrate by 2030.22

The threat estimates keep falling. A Google Quantum AI whitepaper from March estimated that breaking Bitcoin's secp256k1 curve would take no more than 1,200 logical qubits and fewer than 500,000 physical qubits.27 A Caltech architecture paper suggested RSA-2048 could fall with about 100,000 qubits. Actual adoption, meanwhile, looks better on paper than in practice. F5 Labs found that 54% of the top million websites support PQC key exchange, but the figure drops to 22% once Cloudflare-hosted domains are removed. Post-quantum certificates are still essentially absent from the public web.22

The takeaway

Read together, the coverage points to one conclusion. The physics is well enough proven that the hard problems are now engineering, verification and migration. The vendors' 2029 dates may or may not hold. Policymakers, though, have chosen not to wait for them: they set 2030 cryptography deadlines and built verification programs that will judge vendor claims in 2028. For an industry used to grading its own results, that outside testing is the milestone to watch.