What every platform has to do
David DiVincenzo's criteria, set out in 2000, remain the simplest checklist: well-defined qubits that can be scaled, a way to initialise them, coherence times long compared with gate times, a universal set of gates, and a way to read the result. Every platform below meets all five in the laboratory. The differences are in how well, at what cost, and how the answer changes when you try to build a machine with many thousands of qubits rather than a few dozen.
| Platform | The qubit | Operating conditions | Gate speed, order of magnitude | Native connectivity | Developers pursuing it |
|---|---|---|---|---|---|
| Superconducting | Transmon circuit on a chip | Dilution refrigerator, tens of millikelvin | Tens of nanoseconds | Nearest neighbour on a fixed lattice | IBM, Google, Rigetti, IQM |
| Trapped ion | Electronic states of a single ion, commonly ytterbium or barium | Ultra-high vacuum, room temperature or modest cryogenic | Microseconds to milliseconds | Any pair within a trap zone | Quantinuum, IonQ, AQT |
| Neutral atom | Rubidium or strontium atom in an optical tweezer | Vacuum chamber, laser cooled | Microseconds | Reconfigurable by moving atoms | QuEra, Pasqal, Atom Computing |
| Photonic | Photon modes in integrated optics | Cryogenic detectors, room-temperature optics | Set by optical switching | Set by circuit layout; measurement-based schemes | PsiQuantum, Xanadu |
| Silicon spin | Electron or nuclear spin in a silicon quantum dot | Cryogenic, around one kelvin | Nanoseconds to microseconds | Nearest neighbour | Intel, Diraq, Quantum Motion |
Superconducting circuits: fast, and hemmed in by wiring
Transmons are the most mature route in terms of engineering headcount and published results, including the surface-code scaling described on sheet 01. Gates are fast, which is what makes real-time error correction plausible, and the devices are made with lithography that borrows from the semiconductor industry.
The costs are coherence and plumbing. Coherence times are measured in tens to hundreds of microseconds, which sounds short until you divide by the gate time and find room for thousands of operations. The deeper constraint is that each qubit needs its own microwave control line running into the refrigerator, and the refrigerator has a finite budget of heat it can remove. Every roadmap from IBM and Google that we have read treats this wiring problem as the one that determines the shape of the machine, which is why modular designs with links between refrigerators feature so heavily.
Trapped ions: slow, precise, all-to-all
Ions confined in electromagnetic traps have the longest coherence times of any platform, seconds and beyond, and the highest reported gate fidelities. Because the ions in a chain share vibrational modes, any two can be entangled directly, which removes the routing overhead that plagues fixed lattices and makes the low-overhead codes on sheet 01 far easier to implement. Quantinuum's designs go further by shuttling ions between zones of a trap.
The trade-off is speed. Gates take microseconds to milliseconds, so a computation that would run in seconds on a superconducting machine might take hours. Scaling beyond a single chain also requires either physically moving ions or linking traps with photons, and both add complexity. The reasonable position, we think, is that ions will reach high-quality logical qubits early and face a harder question about throughput later.
Neutral atoms: the fastest-moving newcomer
Arrays of neutral atoms held in optical tweezers and entangled through Rydberg interactions have gone from curiosity to serious contender in a few years. The attraction is that atoms are identical by nature, arrays of hundreds or thousands are made with the same lasers as arrays of tens, and atoms can be rearranged mid-computation. The 48-logical-qubit demonstration by the Harvard, MIT and QuEra group depended on that reconfigurability.
The uncertainties are about repetition rate and atom loss. Loading and cooling a fresh array takes time, atoms are occasionally lost from tweezers during a run, and the laser systems that drive Rydberg gates are demanding. None of these looks fundamental, but none has yet been shown to be solved at the scale the platform promises.
Photonics and spins: different bets
Photonic approaches, pursued by PsiQuantum and Xanadu, trade the difficulty of making photons interact for the ability to manufacture in existing semiconductor foundries and to operate much of the system without a refrigerator. Most schemes are measurement-based and rely on generating large entangled resource states, which is where the engineering risk sits. Silicon spin qubits, pursued by Intel and others, are tiny and compatible with transistor fabrication but currently lag on qubit count and uniformity. Microsoft's topological programme, based on Majorana zero modes, promises intrinsically protected qubits; its experimental claims have been contested, and a 2018 Nature paper from a collaborating group was retracted in 2021, so we treat it as an open research direction rather than a platform.
How we would choose what to watch
Ask three things of any platform announcement: whether two-qubit fidelity is quoted for a full-scale array or a hand-picked pair, whether the connectivity supports the codes the developer intends to run, and what the plan is for the next factor of ten in qubit count. A platform strong on all three today is not guaranteed to stay so, which is the honest reason this sheet ends without a ranking.