Everyone quotes qubit counts. Almost nobody quotes the one that matters.
Key takeaways
- Headline qubit counts are physical qubits, which are noisy. Logical qubits are the error-corrected ones that can do useful work.
- Microsoft and Quantinuum cut an estimate of 300 physical qubits per logical qubit roughly tenfold, reaching four logical from 30 physical.
- Google Willow showed logical error rates falling about 2.14 times for each increase in surface code lattice size, the first hardware evidence that fault tolerance scales as theory predicted.
Around 300 physical qubits per logical qubit. That was the working estimate not long ago, and this year Microsoft and Quantinuum cut it roughly tenfold, reaching four logical qubits from 30 physical.
That ratio, not the headline count, is the number that decides whether a quantum computer can do anything useful.
What a logical qubit actually is
A physical qubit is a real piece of hardware and it is noisy. It decoheres, it picks up errors from its neighbours, and on its own it cannot hold a state long enough to finish a serious calculation.
Error correction spreads one unit of information across many physical qubits so the errors can be detected and undone. What comes out is a logical qubit: slower, far more expensive, and actually usable. The overhead is how many physical qubits each logical one costs, and for years that overhead was the reason the field quoted a million-qubit machine as the price of entry.
What moved in 2026
Google Willow, a 105 physical qubit superconducting chip, showed logical error rates falling by roughly 2.14 times for each increase in surface code lattice size. That is the first hardware-scale evidence that fault tolerance scales the way the theory said it would, rather than being swamped by the errors the correction itself introduces.
Microsoft and Quantinuum ran 14,000 instances of a quantum circuit on logical qubits with no errors, alongside the tenfold overhead reduction. Atom Computing entangled 24 logical qubits built from 112 physical and ran computations across 28. Quantinuum announced 94 logical qubits on its H-series in March. Iceberg Quantum and Diraq are targeting 1,000 logical qubits from 150,000 physical on silicon spin hardware, which is a very different number from the million everyone used to quote.
Why the pattern matters more than any single result
These are unrelated hardware platforms. Superconducting circuits, trapped ions, neutral atoms and silicon spin qubits share almost nothing at the physical layer, and the overhead per logical qubit fell by roughly an order of magnitude on all of them inside a year.
That is the sign of a field where the bottleneck has moved from physics to engineering. One vendor cutting overhead tenfold is a result. Four doing it at once with different hardware suggests the error correction codes themselves got better, and code improvements transfer.
So when the next press release quotes a qubit count, the useful follow-up is which kind, and at what ratio. A 1,000 physical qubit machine and a 1,000 logical qubit machine are separated by roughly two orders of magnitude of hardware, and only one of them can run the algorithms people keep promising.
For a sense of how differently governments are budgeting for this, our explainer on sovereign AI covers the funding side, and the Huawei Ascend roadmap shows the same national-capability logic playing out in classical silicon. If you are sizing real hardware today rather than in 2030, the local LLM VRAM tier list is the more practical read.
