Quantum Error Correction Overtakes Qubit Counts in 2026 Roadmaps

The scoreboard has changed

The 2026 quantum roadmaps show an industry that no longer scores itself mainly on raw qubit counts. The milestones that mattered this year were about error correction: how well hardware can protect and fix information, and how cheaply it can do it. That makes the companies harder to compare. It also gives a better picture of who is actually moving toward a useful machine.

The latest example arrived on October 15. Google Quantum AI said it had run "magic state cultivation" on its Willow processor and cut errors about 40-fold compared with states prepared without the technique. The work was published in Science the same day.2529 Three weeks earlier, IonQ said it had built a real-time error-correction decoder that runs on one off-the-shelf CPU.31 Meanwhile, IBM has not yet shown the processor it scheduled for 2026, and Rigetti has quietly slowed its scaling targets.

Google: from storing qubits to computing with them

Google's result matters because it targets the cost of fault-tolerant computing, not just a headline number. Magic states are special inputs a quantum computer needs to run the full range of operations. Each one can be used only once, so a machine has to keep making them.26 Google cites estimates that up to 95% of a quantum computer's capacity could go to producing them.25 A cheaper way to make good magic states would shrink the size of future machines.

On Willow's 105 superconducting qubits, the team produced logical T-states (one type of magic state) with a reported fidelity of 99.99%.25 The coverage agrees on the main numbers but frames them differently. Social posts quickly linked the result to an earlier "Q-Day," the point at which quantum computers could break Bitcoin's cryptography.23 More careful write-ups stressed that the 99.99% applies to one prepared state, not to the processor's overall accuracy, and that only about 8% of attempts were kept.28 The experiment also moved cultivated states from a distance-three color code to a distance-five surface code, a step called "grafting."24 However, it did not show a measured drop in total computing resources, and Google says larger demonstrations are still needed.24

The right reading sits between those two views. The result builds on Willow's December 2024 "below threshold" demonstration, in which logical error rates fell by about half each time the encoded grid grew from 3x3 to 5x5 to 7x7.21 Critics have pointed out that the earlier result was about storing one encoded qubit without performing logical gates on it.30 The magic-state work begins to close that gap. It shows Google building a key ingredient of universal computation on real hardware, not just preserving a qubit. Google's stated goal is still far off: a useful error-corrected computer framed around 2029, with an eventual machine of roughly one million physical qubits.22 In March 2026 the company also added neutral-atom hardware as a second approach alongside its superconducting chips.22

IonQ: making the classical side keep up

IonQ's September result goes after a different bottleneck. Error correction depends on ordinary computers that read error signals and work out fixes fast enough that the quantum computation doesn't stall. IonQ says its two-part decoder handled simulated circuits with up to 408 logical qubits and more than 31.5 million operations. Under standard noise, it added as little as 0.02% to run time.31 One account says the work used 12 of the 16 cores on an Apple M4 Max. The same account says the slowdown stayed below 12% even at higher error rates.39

The main caveat is in the reporting itself. No physical quantum computer at this scale was used, and the results came from simulated workloads that match IonQ's proposed architecture.39 Trapped-ion systems also have an advantage here: their error-checking cycles take 1 to 5 milliseconds, far slower than superconducting chips, which gives the decoder much more time per round.39 A single CPU keeping up with ions does not mean the same approach would keep up with superconducting chips.

IonQ has also been the most aggressive company in describing what its future machines could do. In April it published the "Walking Cat" blueprint, a design for scaling to 10,000 physical qubits and beyond.34 In September it estimated that a machine of about 20,000 physical qubits could break secp256k1, the elliptic curve used by Bitcoin, in just under 26 days. It placed that capability in about the 2028 timeframe.32 CEO Niccolo de Masi said the company is on track to build a fully fault-tolerant 10,000-physical-qubit system in 2027.32

IonQ's track record should temper those claims. Its original SPAC-era roadmap forecast about 4,000 physical qubits by 2026, and it has not come close to that.35 The nearer test is concrete: one investor brief says the first 256-qubit chips arrived from SkyWater in Q2 2026, with a demonstration targeted for the end of the year.38 Whether that system appears on time will say more about IonQ's credibility than another resource estimate.

IBM: the most detailed plan, now facing its 2026 checkpoint

IBM has published the most specific public plan in the industry, and it has largely delivered on schedule. In November 2025 it shipped the 120-qubit Nighthawk processor and the Loon test chip. It also moved main wafer production to a 300 mm facility in Albany, New York.12 The plan runs from Kookaburra in 2026, the first module meant to store data in qLDPC error-correcting memory with an attached logical processor, to Cockatoo in 2027, which links two modules, and then to Starling.12 IBM says Starling will run 100 million gates on 200 logical qubits by 2029.17 Its qLDPC codes are meant to cut error-correction overhead by up to 90% compared with older methods.17

IBM has also put serious money behind the plan. The company says it will invest more than $10 billion in quantum over five years.13 Separately, a $2 billion Department of Commerce award included $1 billion for IBM to build a dedicated quantum chip foundry called Anderon.20

This is where the coverage disagrees in a way that matters. One industry summary describes Kookaburra as a 4,158-qubit system and says every major vendor reached logical-qubit milestones in the first half of 2026.37 Other profiles describe Kookaburra as a 1,386-qubit multi-chip processor.11 A late-September review found that IBM had not announced Kookaburra's delivery at all.12 The most careful reading is that Kookaburra is still pending. IBM usually shows new processors at its November developer conference, so the delay does not yet count as a slip. If November passes without it, it would be the first real delay in the fault-tolerance plan.12 IBM's other 2026 promise is verified quantum advantage from combined quantum and classical computing. That is a separate claim and would not by itself prove that Starling will arrive on time.18

Rigetti: a quiet reset

Rigetti's story is the clearest case of a roadmap meeting physical limits. In November 2025 it promised a system of more than 150 qubits with 99.7% median two-qubit fidelity by around the end of 2026. It also promised a system of more than 1,000 qubits with 99.8% fidelity by around the end of 2027.1 By August 2026, its 108-qubit Cepheus system was running at about 99.1% median two-qubit fidelity, and the stated target had become about 1,000 qubits at 99.9% fidelity over "roughly a three-year" period.3 Rigetti now aims for about 99.5% fidelity on the 108-qubit system by the end of the year.6

In practice, this is a delay of more than a year for the 1,000-qubit goal, even though management describes the company as on track. The company has been open about the cause. CEO Subodh Kulkarni said coherence times of 25 to 30 microseconds are now the main limit on fidelity. He also said Rigetti is "not quite close" to quantum advantage and is building systems for research applications for now.4 Rigetti's own documents give different figures for its gate-speed goal: the press release says below 50 nanoseconds, while the earnings call says below 40.34 The company plans to publish an updated roadmap later this year.7 Smaller chiplets do perform better: Rigetti reports 99.8% fidelity at 9 qubits and 99.6% at 36.4 That supports its modular approach, but the drop at 108 qubits shows how hard scaling is.

What it adds up to

The industry is splitting into companies with demonstrated error-correction results and companies still raising the quality of their physical qubits. Google has the strongest record of experimental results. IBM has the most detailed engineering plan and the most money behind it. IonQ is making the boldest claims, and most of them still rest on simulations and resource estimates. Rigetti is clearly behind on the error-correction measures that now define progress.

The common thread is cryptography. Breaking secp256k1 appears in IonQ's 2028 projection32 and in the instant reaction to Google's paper.23 Both NIST-standardized signature schemes, ML-DSA and SLH-DSA, resist this kind of attack.32 Whatever any single company delivers, the case for moving to post-quantum cryptography is stronger than the case for betting on a particular vendor. The next tests are IBM's November conference and IonQ's year-end 256-qubit demonstration.