Three modalities can plausibly carry useful quantum computing into the next decade: superconducting transmons, trapped ions, and photonic networks. We picked superconducting for Ireland Quantum 100 after working through the engineering, the supply chain, the software stack, and the timeline to first useful climate workload. This is the long version of why — written for engineers and technical decision-makers who want the trade-offs without the marketing layer.
What we were actually optimising for
Hardware choice is downstream of mission. Ours is climate-science compute on Irish soil inside the next twelve months: variational chemistry for carbon-capture sorbents, photovoltaic candidate screening, battery cathode discovery, and grid-scale optimisation. That biases the answer in three concrete ways.
First, we need gate speed. Variational quantum eigensolvers and QAOA-style circuits run thousands of shots per parameter update, and chemistry workloads layer outer optimisation loops on top. A platform with sub-microsecond two-qubit gates lets you finish a meaningful sweep in a working day. A platform with millisecond gates does not.
Second, we need a credible supply chain for cryogenics, control electronics, and fabrication. A sovereign machine that depends on a single boutique vendor for a single irreplaceable component is not sovereign.
Third, we need a software stack engineers can actually use. The chemistry teams we want as first-cohort users already write in Qiskit, PennyLane and Cirq. Forcing them onto an exotic toolchain pushes first useful result out by a year.
With those three constraints fixed, the comparison becomes much narrower than the open question of "which qubit modality wins."
The superconducting transmon case
A superconducting qubit is a non-linear LC oscillator built from aluminium on silicon, with a Josephson junction providing the anharmonicity that lets you address the |0⟩↔|1⟩ transition without leaking into |2⟩. The transmon variant — capacitively shunted, charge-noise insensitive — is the workhorse of the field for good reason. It is lithographically defined, so once you have a fabrication recipe you can iterate on chip layout in weeks rather than years.
Operating point sits below 15 mK in a dilution refrigerator, because k_B T at the qubit frequency (typically 4–6 GHz) has to be far smaller than the level splitting. Single-qubit gates run in tens of nanoseconds via shaped microwave pulses. Two-qubit gates — cross-resonance, CZ via tunable couplers, or parametric — land in the 40–300 ns range depending on architecture. Coherence times in the 100–300 µs range are now routine on well-engineered devices, which gives you a healthy ratio of coherence to gate duration.
The topology we are building is heavy-hex. It trades raw connectivity for reduced spectator errors during two-qubit gates and is well-matched to the surface-code variants that the field is converging on for fault tolerance. Heavy-hex is not the densest layout, but it is forgiving in fabrication and it routes cleanly to the surface-code primitives we will need when we move past NISQ workloads.
Trapped ions: what they do better, and why we still didn't pick them
A trapped ion qubit — typically Yb+, Ba+ or Ca+ — stores quantum information in two long-lived electronic or hyperfine states of a single atom held in vacuum by RF Paul traps. It is genuinely the cleanest qubit physics on offer. Coherence times are measured in seconds, not microseconds. State preparation and measurement fidelities are excellent. Gates are mediated by laser-driven phonon modes shared across the ion chain, which means every ion in the chain can in principle interact with every other — full all-to-all connectivity inside a register.
For some algorithms — anything with deep all-to-all entanglement structure — that connectivity is worth a lot. SWAP overhead on a heavy-hex superconducting lattice is real and it eats your circuit budget.
So why didn't we pick it?
- Gate speed. Two-qubit gates on ions sit in the tens of microseconds to low milliseconds. For a chemistry VQE running tens of millions of circuit executions per problem, that is the difference between a result this quarter and a result next year.
- Scaling beyond one chain. A single ion chain saturates somewhere in the tens of qubits before mode crowding and laser-addressing complexity bite. Beyond that you need photonic interconnects between chains or shuttling architectures (QCCD). Both are credible research directions. Neither is something I want on the critical path of a twelve-month delivery.
- Laser systems. Trapped-ion stacks are laser-bound. Stable, narrow-linewidth lasers at multiple wavelengths, with precise beam delivery to individual ions, are an entire engineering discipline of their own. You take on that discipline whether you want to or not.
None of this is an argument that trapped ions are wrong. It is an argument that they are wrong for the workload and timeline we committed to.
Photonic qubits: the long bet we admire from across the room
A photonic qubit encodes information in degrees of freedom of light — polarisation, time-bin, dual-rail path, or continuous-variable squeezed states. The appeal is obvious. Photons travel at room temperature. They route naturally through fibre. They are the only modality that interconnects to networking without a transduction step, which makes them the natural substrate for any future quantum internet.
Two architectures dominate the conversation: measurement-based / fusion-based linear-optical computing using single photons and detector-driven gates, and continuous-variable approaches using squeezed light and Gottesman-Kitaev-Preskill encoding.
The hard problems are well known. Deterministic two-qubit gates between single photons are not native — they require ancilla photons, heralded entanglement generation, and high-efficiency number-resolving detectors. Loss is the dominant error and it compounds non-locally. The path to fault tolerance demands either heroic detector efficiencies or aggressive multiplexing of probabilistic gates, and the resource overhead in current proposals is substantial.
Photonic systems will likely be the right answer for distributed quantum computing and for some specialised sampling problems. They are not, today, the right answer for a single-site machine running variational chemistry with a working-day turnaround.
The software stack decision
This part gets less attention than it deserves. Quantum hardware comparison usually focuses on coherence times and gate fidelities. The thing that determines whether real users get real work done is whether their existing code runs.
The superconducting ecosystem maps cleanly onto OpenQASM 3 as the common intermediate representation. Qiskit, Cirq, PennyLane, and the Braket SDK all compile to gate sets that are native or near-native on transmon hardware: parameterised single-qubit rotations plus a two-qubit entangler (CZ, CNOT, or ECR depending on the device). A chemistry researcher who has written a UCCSD ansatz in PennyLane can target our backend through a transpiler pass, not a rewrite.
Trapped-ion stacks have their own SDKs and their gate sets (XX or Mølmer-Sørensen interactions) transpile sensibly, so this is not a fatal point against them. Photonic stacks are further from the gate-model lingua franca, particularly the continuous-variable ones, which require thinking in terms of modes and Gaussian operations rather than qubits and circuits. For users coming from the standard quantum-chemistry literature, that is a steeper ramp.
If you want to dig further into the workload side of this, we go deeper on the chemistry pipeline in the writeup of climate workloads on Ireland Quantum 100.
Sovereignty, supply chain, and the boring engineering
A 100-physical-qubit superconducting machine has a known bill of materials. Dilution refrigerator from one of three or four serious vendors. Microwave control electronics from a small but real ecosystem. Cryogenic amplifiers, attenuators, isolators, circulators — all sourceable. Chip fabrication has multiple credible foundries. Wirebonding, packaging, magnetic shielding, vibration isolation — these are engineering problems with engineering answers.
That matters because "sovereign quantum compute" has to mean something operational. It cannot mean a single black box from a single vendor. The superconducting ecosystem is the only one of the three where you can assemble a machine from components with at least two viable suppliers in every critical category, and where the in-country skills base — RF engineering, cryogenics, FPGA control — already exists in Ireland and the wider EU.
Trapped-ion supply chains are thinner, particularly on the integrated-photonics side that the next generation of ion traps depends on. Photonic stacks consolidate around very few vendors of high-efficiency photon-number-resolving detectors and squeezed-light sources. Both are perfectly valid bets for a research group. They are harder bets for a piece of national infrastructure.
You can read more about the broader programme on the Ireland Quantum 100 overview.
Where to start this week
If you are a researcher, climate-tech founder, or engineering team weighing which modality to learn against: install Qiskit or PennyLane, pick a small VQE on H₂ or LiH from the standard tutorials, and run it on a public superconducting backend. Then run the same circuit on a public trapped-ion backend. Time both. Look at the transpiler output. Look