Every quantum announcement you read this year will lead with a physical qubit count. A thousand. Two thousand. Soon enough, ten thousand. The number is meant to land like a benchmark, the way teraflops once did. It isn't one. A physical qubit on its own is a noisy, fragile thing that holds coherence for microseconds and lies to you about the answer it just computed. What matters — what actually runs a workload — is the logical qubit, the error-corrected abstraction sitting on top of hundreds or thousands of those physical ones. When we sat down to scope Ireland's first sovereign quantum facility for delivery in Q2 2027, the very first decision we made was to ignore the headline number entirely and design backwards from logical qubits.

The number that matters is the one nobody markets

If you ask a physicist how many physical qubits it takes to make one fault-tolerant logical qubit under the surface code, the honest answer is "it depends on your physical error rate, your code distance, and what you're trying to run." The unhelpful but more useful answer is: a lot. Hundreds at the low end, thousands at the high end, and the ratio gets worse the longer your circuit needs to run without a logical error.

So a machine that boasts two thousand physical qubits is not a machine that runs two thousand logical operations. Depending on the architecture and the gate fidelities, it might give you a handful of logical qubits with enough headroom to do real work, or it might give you none at all and serve only as a research substrate. Both are legitimate machines. They are not the same product.

The trap, for a small country building its first machine, is to chase the marketable number and end up with a research substrate dressed in industrial clothing. That is not what Ireland needs. Ireland needs a machine that does something on day one — something a pharma team in Cork or a logistics group in Dublin or a cryptography unit inside a state body can actually queue a job against and get a result that beats what they could do classically, or that they couldn't do classically at all.

Scoping from the workload back

The way we approached it was to start with a short list of workloads we believe will be ready to run by 2027 — not in theory, but in compiled, benchmarked, peer-reviewed form — and ask how many logical qubits, at what code distance, with what T-gate budget, each of them needs.

The list is shorter than people expect. It includes:

• Small-molecule electronic structure problems relevant to drug discovery and catalyst design, in the regime where classical methods get expensive and quantum methods start to look genuinely competitive.
• Certain optimisation problems framed as QAOA or quantum annealing analogues, where the quantum advantage is narrow but real if the problem instance is shaped correctly.
• Specific cryptanalytic and lattice-related primitives that the state has a sovereign interest in being able to study on home soil rather than queueing on someone else's cloud.
• Materials simulation for battery chemistries and superconductor candidates, where the logical-qubit demand is higher but the payoff for Irish industry is direct.

None of these need ten thousand logical qubits. Most of the interesting near-term ones live in the range where a well-designed machine with a respectable physical-to-logical ratio can deliver something useful. That's the design point we're building to.

Why the surface code, and why we're honest about its tax

The surface code is the leading candidate for fault-tolerant quantum error correction for good reasons. It tolerates relatively forgiving physical error rates, it maps cleanly onto a 2D lattice of qubits with nearest-neighbour interactions, and the decoding problem, while hard, is tractable enough that real-time decoders are now a working engineering discipline rather than a research aspiration.

It is also expensive. A logical qubit at code distance d costs roughly 2d² physical qubits, plus ancillas, plus the routing overhead between logical patches when you want them to actually talk to each other. T-gates — the non-Clifford gates you need for any computation that isn't trivially classical — have to be distilled from magic states, and magic state distillation is its own factory of qubits sitting alongside the computation.

I labour the point because the marketing material for our sector tends to skip it. A facility that quotes you "X logical qubits" without telling you the code distance, the assumed physical error rate, the T-gate throughput and the magic state factory footprint is selling you a number, not a machine. Ireland Quantum will publish those numbers. Boring tables. Real engineering. The kind of spec sheet a procurement officer at a hospital or a bank can take to their own technical staff and get a straight answer about what jobs will run.

What "useful quantum advantage" actually means in 2027

The phrase useful quantum advantage has been bent badly out of shape over the last few years. It got conflated with "quantum supremacy" — a narrow academic claim about a contrived sampling problem — and then with vendor demos that solved instances of problems nobody had actually wanted solved. Useful means something specific. It means: a workload that an organisation would pay to run, where the quantum machine returns an answer that is either better, faster, cheaper, or simply impossible classically, on a problem instance large enough to be worth the trouble.

By that bar, quantum computing 2027 is going to look modest. The first useful workloads will be narrow. They will live in chemistry and materials and certain corners of optimisation. They will not replace HPC. They will sit alongside it, hybrid, with the classical machine doing most of the work and handing the quantum co-processor the parts of the calculation it cannot do efficiently itself.

That is the honest scope. A facility built to that scope is one that delivers value on day one and grows. A facility built to chase headlines is one that sits idle while its operators write press releases about milestones.

Sovereignty as a scoping constraint

There is a second axis to the scoping decision that has nothing to do with physics. Ireland's interest in a sovereign machine is not academic. The state, and the firms that operate under Irish law, need somewhere to run sensitive workloads — cryptographic, pharmaceutical, defence-adjacent — without those workloads leaving the jurisdiction.

This shapes the design. It means a facility with on-prem decoders, on-prem control electronics, and a software stack we can audit end-to-end. It means the logical-qubit count has to be sized not for the most demanding workload imaginable, but for the workloads the state and Irish industry will actually queue. Sovereign capacity that sits unused is a vanity. Sovereign capacity that runs at high utilisation, even on a modest logical-qubit count, is infrastructure.

This is why the spec is being written in conversation with the people who will queue jobs on it, not in isolation by the people building it. A machine designed in a room with no users is a machine that gets used by no one.

What we lose by being honest about this

We lose the headline. A facility that opens in Q2 2027 with a credible logical-qubit count and a published code distance will be smaller, on paper, than facilities elsewhere quoting raw physical numbers. Some of those elsewhere facilities will run nothing useful for years. Some will run useful things sooner than expected. Predicting which is which from a press release is impossible, and that's the point — the press release is not the spec sheet.

What we gain is a machine that does work. A machine that the Department of Enterprise can point to and say, here is a job that ran here, here is the result, here is the firm that paid for it. That is the only metric that matters for sovereign infrastructure. Not the qubit count on the launch slide. The job log six months in.

What to do this week

If you run a research group, a pharma team, a logistics function, or a cryptography unit in Ireland and you think you might have a workload that fits the scope I've outlined, the thing to do this week is to write down — in one page — what the problem is, what the current classical cost looks like, and what shape of answer would be useful. Not a quantum-flavoured version of the problem. The actual problem. We are building the early-access list for Ireland Quantum from those one-pagers. The machines will be sized to the workloads, the workloads will not be retrofitted to the machines, and the conversation needs to start now if you want to be in the first cohort that runs jobs in 2027 rather than reading about them.