For most of the past decade, progress in quantum computing was measured by one number: how many quantum bits, or qubits, the basic building blocks of a quantum computer, a machine could hold. That era is coming to an end, and the hardest problem is no longer simply building more qubits but instead keeping these qubits error-free. Quantum error correction is the discipline of building reliability out of these error-prone qubits, and it has quietly become the field’s central battleground shaping national strategies and investment priorities. Qubits are exquisitely fragile: unlike a classical bit found in ubiquitous laptops, a qubit cannot be inspected and reset without destroying the information it carries. Where the bit in a laptop is a stable switch set firmly to 0 or 1, a qubit is closer to a spinning coin, and the faintest disturbance from heat, vibration, or stray radiation can knock it off course.

Error correction builds reliability out of these unreliable parts. The classical version is intuitive: store each bit three times and take a majority vote, so a single corrupted copy is outvoted. Quantum information forbids this shortcut, since the “no-cloning theorem” prohibits copying an arbitrary unknown state, in other words, one is unable to create an identical copy of an unknown quantum state. The workaround is to spread one unit of information across many noisy physical qubits so the group behaves as a single, far more reliable qubit, known as a logical qubit, much as a large choir stays in tune even when individual singers waver. Information is instead spread across these logical qubits, with operation errors inferred from their relationships.

From Qubit Counts To Logical Counting

The cost of this error-correction is steep, and in quantum computing cost means overhead: the number of physical qubits consumed to produce one usable logical qubit. That ratio runs from two to several thousand by architecture. Driving that ratio down is the central race, and trapped-ion quantum computing firm  Quantinuum leads on encoding rate on commercial hardware, with physical:logical qubit ratios reaching 2:1, while superconducting players such as Oxford Quantum Circuits (OQC), a Firgun Ventures portfolio company, has also reported a 1:1 qubit ratio on their systems.

For years, progress was advertised in physical qubit counts, a figure that means little on its own. Because the conversion ratio between physical and logical qubits varies more than tenfold across designs, comparing raw counts in itself can be misleading. The meaningful unit is now the logical qubit, and even that does not tell the whole story. The reason being, a logical qubit that can merely sit idle is not the same as one that survives a long, deep calculation, so what counts is how many reliable operations a machine performs per second (rQOPS) and how deep it runs before failing.

Two further costs lurk beneath the qubit count. The first is that error-corrected machines run their routine operations almost for free but pay dearly for the handful of special operations that give quantum computing its real power, and producing these on demand is highly resource-hungry.  This requirement known as a magic state factory may occupy as much of a future chip as the logical qubits themselves. The second is timing: errors must be diagnosed and fixed in real time by classical control electronics, and if those decoders cannot keep pace, the machine stalls while it waits. The UK firm Riverlane has built much of its error-correction offering on precisely this layer, with its decoders now deployed alongside partners including Rigetti, Oxford Quantum Circuits, and Oak Ridge National Laboratory, a commercial niche that barely existed three years ago. The same logic extends to error mitigation. Qedma, an Israeli quantum startup, offers their flagship product, QESEM, a software which suppresses errors as a circuit runs and cleans up the rest in post-processing, needing few or no extra qubits, and already reaches users through IBM's Qiskit.

Competing Bets On The Overhead Problem

Three families dominate the quantum error-correction strategic debate, each a distinct wager. The dominant near-term method is the surface code, which arranges about a thousand physical qubits in a grid and uses their redundancy to reconstruct one dependable logical qubit. It is well understood and proven in the laboratory, and it anchors Google’s Willow chip, which in December 2024 crossed a threshold the field had chased for a decade. The engineers enlarged the protective code, and the error rate fell rather than rose, the sign that adding qubits finally helps more than it hurts. The catch is the cost, since a useful machine on this path may need millions of physical qubits.

The second bet, quantum low-density parity-check (qLDPC) codes, promises similar protection with roughly a tenth of the qubits, in exchange for far more intricate wiring, since the qubits that must communicate are no longer tidy neighbours on a grid. IBM’s pivot towards this family was a highly notable strategy shift of 2024, and the Canadian firm Photonic Inc., a Firgun Ventures portfolio company, has gone further, designing its own qLDPC family, the SHYPS codes, which it reported in February 2025 can run any algorithm using up to 20 times fewer physical qubits than the surface code.

The third major bet, bosonic codes, changes the medium entirely, storing information inside the rich internal states of one device, rather like recording a song as a full audio waveform on a single tape instead of spelling out the lyrics letter by letter across a long row of flashcards. Alice & Bob in Paris and Amazon’s AWS have built their architectures here; in September 2025 Alice & Bob kept their cat qubit stable against one of the two common error types (bit flips) in the order of an hour, a substantial improvement from about seven minutes the year before. Other less common quantum error-correction approaches do exist and are focused on making headway. For instance, Microsoft bets on topological qubits, with protection built into the physics itself but this method is unproven at scale and has come under increasing scrutiny, especially following the 2nd generation release of their Majarona chip in June 2026.

A Future Portfolio of Error-Correcting Codes

The plausible future is not one universal code but a portfolio of encodings layered according to the hardware, workload, and noise channels. A code is only attractive if it suits its platform: surface codes generally cater to flat superconducting grids, qLDPC rewards rich connectivity, while bosonic codes reward hardware which suppress one class of errors before software level correction begins. In that sense. error correction is transitioning towards becoming more like designing an industrial system and less like choosing an algorithm.

Google’s Willow showed below threshold exponential error suppression in 2024 and since that period we have seen strengthening progress in this space. Around the same period, Microsoft and Atom Computing reported 24 entangled neutral atom logical qubits, while the following year brought larger demonstrations across leading platforms: Quantinuum’s Helios at roughly 50 logical qubits, and QuEra and collaborators with a simulation of up to 96 active logical qubits in a neutral atom architecture. In two years the technology crossed from theory to working hardware on most of the major platforms. All in all, the winner of the next phase is shaping out to be whoever builds the best marriage of code and machine.