Quantum computing has a noise problem. Every qubit in existence is imperfect — affected by thermal fluctuations, electromagnetic interference, cosmic rays, and the fundamental fragility of quantum states. Without error correction, computations on today’s quantum processors become unreliable after just a few hundred operations. That’s nowhere near enough for the problems quantum computers are supposed to solve.

Quantum error correction (QEC) is the set of techniques designed to fix this. And after years of incremental progress, 2025 and early 2026 have delivered some genuinely significant results. Let me walk through what’s happened, what it means, and where the hard problems remain.

The Core Challenge

A quick recap for context. Classical computers store information as bits (0 or 1) and correct errors through redundancy — storing the same bit multiple times and taking a majority vote. Quantum computers can’t do this directly because of the no-cloning theorem: you cannot copy an unknown quantum state.

QEC works differently. It encodes a single “logical qubit” across multiple “physical qubits” in a way that allows errors to be detected and corrected without measuring (and thereby destroying) the quantum information. The most studied approach is the surface code, which arranges physical qubits in a 2D grid and uses auxiliary “syndrome” qubits to detect errors on their neighbours.

The overhead is substantial. Current estimates suggest you need somewhere between 1,000 and 10,000 physical qubits per logical qubit, depending on the error rates of the physical qubits and the precision required. This is why a quantum computer with a million physical qubits might only give you a few hundred logical qubits — and a few hundred logical qubits is roughly what you need for commercially interesting problems like breaking RSA encryption or simulating complex molecules.

Google’s Below-Threshold Demonstration

The biggest milestone came from Google’s quantum AI team. In late 2025, they published results from their Willow processor demonstrating that increasing the size of a surface code actually reduced the logical error rate — a property called “below threshold” operation.

Why is this important? Because it’s the fundamental requirement for QEC to work. If adding more physical qubits to your error correction scheme makes things worse instead of better (because each additional qubit introduces more noise than it corrects), then QEC is a dead end. Google’s demonstration that a distance-5 surface code outperformed a distance-3 code, which outperformed uncorrected qubits, was the first convincing evidence that we’re past this critical threshold on real hardware.

The numbers: Google achieved a logical error rate of approximately 0.14% per error correction cycle with their distance-7 surface code — roughly a 2.1x improvement over the distance-5 code and a 4x improvement over the distance-3 code. The scaling trend matched theoretical predictions. This was a genuine breakthrough, even though the absolute error rate is still far too high for practical computation.

IBM’s Roadmap and the Heron Processor

IBM has taken a somewhat different approach. Rather than focusing purely on surface codes, they’ve been developing a strategy that combines error correction with error mitigation — techniques that reduce the impact of errors without fully correcting them.

Their Heron processor, released in late 2024 and refined through 2025, achieves two-qubit gate error rates below 0.1% — among the best in the superconducting qubit world. IBM’s thesis is that if you make the physical qubits good enough, you can get useful results with lighter-weight error correction schemes that require fewer physical qubits per logical qubit.

In January 2026, IBM demonstrated a 12-logical-qubit system using a combination of their best Heron hardware and a hybrid error correction/mitigation protocol. The system ran a 200-gate-depth circuit with measured fidelity above 90%. That’s not enough for breaking encryption, but it’s getting into the range where certain quantum chemistry simulations become meaningful.

The philosophical difference between Google and IBM matters. Google is betting on brute-force scaling of surface codes: build enough physical qubits and the error correction math works in your favour. IBM is betting that smarter error handling techniques can reduce the overhead, making useful quantum computing possible sooner with fewer qubits. Both approaches have merit. Neither has won yet.

Microsoft’s Topological Qubit Claim

Microsoft has been the most controversial player in QEC. For over a decade, they’ve pursued topological qubits — a fundamentally different type of qubit that would be inherently resistant to many forms of error, dramatically reducing the error correction overhead.

In early 2025, Microsoft published a paper in Nature claiming they had demonstrated a topological qubit using Majorana zero modes in a semiconductor-superconductor heterostructure. The claim was bold, given that previous attempts (including Microsoft’s own retracted 2018 paper) had failed to convincingly demonstrate topological protection.

The reception has been cautious. Several independent groups have questioned aspects of the experimental data, and as of February 2026, no independent replication has been published. If the result holds, it’s potentially transformative — topological qubits could reduce the physical-to-logical qubit ratio from thousands-to-one to tens-to-one. If it doesn’t hold, Microsoft will have lost another half-decade on an approach that may not be physically realisable.

My read: the evidence is suggestive but not conclusive. I’d put the probability of topological qubits being commercially viable within the next decade at around 25%. The potential payoff is enormous, but so is the uncertainty.

Quantinuum’s Trapped Ion Progress

While superconducting qubits get most of the attention, Quantinuum’s trapped ion approach continues to deliver the highest-fidelity gates in the industry. Their H2 processor achieves two-qubit gate fidelities above 99.8% — significantly better than any superconducting system.

In late 2025, Quantinuum demonstrated real-time quantum error correction on 12 logical qubits encoded using a colour code (an alternative to the surface code that’s naturally suited to trapped ion architectures). The system ran 14 rounds of error correction with logical error rates of approximately 0.03% per round — the lowest achieved by any platform.

The catch with trapped ions is speed. Gate operations are roughly 1,000 times slower than superconducting qubits. So while each operation is more accurate, the total computational throughput is lower. Quantinuum’s bet is that the dramatically lower error rates more than compensate for the speed penalty by reducing the total number of operations needed.

Where We Actually Stand

Let me be blunt about where the field is.

We’ve proven that QEC works in principle on real hardware. That’s the Google result, and it’s a big deal. Two years ago, this hadn’t been demonstrated.

We don’t yet have enough logical qubits to do anything commercially useful. The largest demonstrations are 12-15 logical qubits. Commercially interesting problems require hundreds to thousands.

The physical qubit overhead remains massive. Even with the best current hardware, you’re looking at 1,000+ physical qubits per logical qubit for the precision needed for real problems.

The timeline to useful fault-tolerant quantum computing is 5-10 years. That’s not marketing optimism — it’s based on hardware roadmaps from Google, IBM, and Quantinuum, combined with realistic scaling projections. The most aggressive estimate puts a 100-logical-qubit machine at 2028-2029. The more conservative estimate says 2031-2033.

What to Watch in 2026

Three things will tell us a lot about the trajectory of QEC this year:

Google’s next surface code demonstration. They’ve announced plans for a distance-11 or higher surface code on their next-generation processor. If the error rate continues to scale as theory predicts, the path to useful QEC becomes much clearer.

Independent replication of Microsoft’s topological qubit. If another lab confirms the result, the entire competitive landscape shifts. If not, the surface code and colour code approaches remain the most credible paths.

Quantinuum’s scale-up. They’ve promised a 56-physical-qubit system in 2026 that should support more logical qubits with higher code distances. Trapped ions at that scale with maintained fidelity would be a strong data point for the technology.

The Honest Summary

QEC went from theoretical to experimental in 2025. It hasn’t yet gone from experimental to practical. The milestones are real and encouraging, but anyone telling you quantum advantage from error-corrected machines is imminent is ahead of the evidence. We’re in the hard middle phase — the principles work, the engineering doesn’t yet scale, and the timeline depends on which approach wins the race between error rates and qubit counts.

It’s worth paying attention to. It’s not yet worth restructuring your technology strategy around. Give it three to five years.