Dispersive microwave readout (+TWPA, Purcell) ro_disp

A resonator coupled off-resonantly (coupling g, detuning Δ ≫ g) acquires a state-dependent shift χ ≈ g²/Δ, so a probe tone carries the state in its phase without exchanging quanta with the qubit — non-demolition by construction. Wallraff and colleagues at Yale established it as circuit QED's readout channel D[591]. Two additions make it usable: an on-chip Purcell filter, which passes the probe while blocking qubit emission into the same line, and a near-quantum-limited parametric amplifier before the HEMT (high-electron-mobility transistor).
Attributes: carrier affinity a = 1.0, fully fabricated. Readout ≈280 ns (10^−6.55 s), non-destructive, mid-circuit capable.

Separation needs a fixed number of signal photons at the amplifier input, so readout time falls with probe power and rises with chain noise. A Josephson travelling-wave amplifier giving 12 dB across 4 GHz near the quantum limit D[592] removes the HEMT's noise penalty, so the amplifier is not the floor. The floor is what hard probing does to the qubit: above a critical photon number the dressed system undergoes measurement-induced transitions out of the computational subspace, and the code inherits leakage — unheralded, sticky, correlated in time, not a Pauli flip. The IQM device shows the split: 99.94% simultaneous assignment fidelity with a 240 ns pulse and shelving into the second excited state, but 99.3% QNDness D[38]. Assignment error is 6×10⁻⁴; the chance the qubit does not survive as measured is an order of magnitude larger — and that is what syndrome extraction consumes. The Purcell filter's linewidth buys speed but, widened, reopens the emission channel it exists to close. What moves the floor: shelving-free discrimination, on-chip isolation, and encodings that turn a failed measurement into a heralded erasure.

The spread is the story: 6×10⁻⁴ on a characterised pair, 5×10⁻³ over 105 qubits, ~10⁻² across a fleet D[38]D[1]C[33]; Zuchongzhi 3.0 reads out at 99.13% D[35]. Readout is the largest single-channel error term at scale, larger than two-qubit gates, and the one whose fleet value has barely moved in two years.

On chip there is no extra process: resonator and Purcell filter are patterned in the qubit's own lithography, the cost being area and frequency planning: each qubit on a feedline needs a distinct resonator frequency in the amplifier's band. Off chip the chain — isolators, circulators, parametric amplifier, HEMT — is the fridge's bulkiest hardware and where a merchant market has appeared: Silent Waves (Grenoble) sells Argo, Carthago and Zephyr, the last launched 2026-03-13, to Rigetti, Alice & Bob, Qilimanjaro, QphoX and CEA C[593]; QuantWare sells a Crescendo TWPA C[594]; QuantumCore is developing one with the Institute for Quantum Computing P[595]. Export exposure is direct: ECCN 3A901, effective 2024-09-06, covers sub-4.5 K electronics and names parametric amplifiers, so a TWPA is controlled in its own right G[301].

Each qubit consumes a resonator, a share of a feedline (eight to ten qubits multiplexed per line), a share of one amplifier's dynamic range, and a digitiser channel with real-time discrimination. The amplifier is the scaling constraint nobody has published a number for: TWPA saturation caps simultaneous tones per device, and each feedline needs its own isolator stack — volume, not silicon. Instrumentation has followed (over 1,000 channels per rack C[526]), but the budget is set downstream: Google's real-time decoder ran at 63 µs latency for d=5 against a 1.1 µs cycle D[1], so readout plus decode defines the loop. At 10⁶ the chain must move into the cold, and no cryogenic digitiser has been shown beyond a handful of qubits.

It serves every superconducting-family architecture: transmons, bosonic cat and GKP encodings, dual-rail erasure qubits, and the same chain in annealers. It requires the transmon's dispersive shift and, for bosonic codes, an ancilla mediating the cavity measurement, and provides the mid-circuit non-destructive measurement every error-correction scheme here assumes — no substitute exists inside the family. Its clock contribution is direct: derived clock = sum of the syndrome round: gate layers + transport + readout + reset is 0.65 µs, of which readout is 0.28 µs — the largest term, against ~0.048 µs per gate layer D[1]D[38] — inside a measured 1.1 µs cycle. No neighbouring empty slot in the technology graph.

Three quantities circulate as "readout fidelity" and are not interchangeable: assignment fidelity, QNDness (the qubit is left in the state reported), and SPAM. The 99.94% headline is simultaneous assignment fidelity with shelving; QNDness on the same device is 99.3% D[38]. IBM's fleet ~10⁻² is undecomposed into amplifier noise, Purcell-limited decay and measurement-induced transitions C[33]. Value conflict: the graph record quotes 280 ns readout, the preprint a 240 ns pulse, so ~280 ns should be read as the full window. No independent laboratory has replicated the 99.94% result.

Not published: a named TWPA or Purcell-filter patent family with assignee and grant year; TWPA yield, unit cost, or the number of multiplexed tones one amplifier sustains before compression, from any supplier; and a fleet-average QNDness from any vendor.

Who.
| Organisation | Role | Country | What exactly they do with this technology | Evidence |
|---|---|---|---|---|
| IQM | developer | Finland | Readout record: 99.94% with shelving, QNDness 99.3% | D[38] |
| Google | developer | US | 99.5% readout across 105 qubits in a 1.1 µs cycle | D[1] |
| IBM | developer | US | Publishes the only at-scale fleet readout error, ~10⁻² | C[33] |
| Silent Waves | supplier | France | Merchant TWPA maker, three shipping products | C[593] |
| QuantumCore | supplier | Canada | TWPA development with IQC Waterloo | P[595] |

Money.

• 2026-05-05 · QuantWare · Series B · $178 M (€152 M) · Intel Capital, In-Q-Tel, ETF Partners · closed C[547]fact
• 2026-04-25 · QuantumCore · listing (CSE: QNCR, 2026-04-14) plus NSERC Alliance grant · $10.7 M raised, $1.7 M grant · NSERC/IQC · closed P[595]
• 2026-01 · D-Wave · M&A (Quantum Circuits) · $550 M · — · announced C[84]

Market & supply chain. The chain, not the chip, is what this node buys: isolators and circulators, HEMTs, the parametric amplifier. The TWPA market is thin — three small merchant suppliers against in-house builds at Google and IBM — so one exit matters more here than anywhere else, and the parts are export-controlled G[301]. G2 and G3 pay for this node; G7 buys the instrumentation.

IP & standards. No TWPA or Purcell-filter patent family with a named assignee and grant year was found; Silent Waves' 2025 Nature Electronics amplifier-isolator is a published design, not a grant C[593]. No litigation identified, and no standards body defines how readout fidelity is reported — hence three numbers quoted interchangeably.

Roadmaps & track record. Silent Waves: three products with named customers (Zephyr promised and launched 2026-03-13) C[593], the most credible record here. QuantumCore: announced 2026-04-25, too recent to judge P[595]. No platform vendor publishes a fleet-average readout target as distinct from best-device fidelity — a notable omission when readout is the largest fleet error term.

Strategic reading. This is a shared chokepoint: four architectures depend on the same amplification and isolation hardware, so concentration risk is systemic in a way qubit-chip choices are not. If fleet readout stays near 10⁻², Λ stalls whatever gates do, since syndrome extraction inherits the measurement error — the case for erasure-style encodings that make a failed measurement heralded. Winners: merchant amplifier suppliers and whichever vendor solves QND rather than assignment. Bargaining power sits with the platform vendors, though export licensing gives regulators a lever.

Confirm or demote within 12–24 months: does any vendor publish a fleet-average QNDness; does fleet readout error fall below 5×10⁻³. Best case by 2029: shelving-free discrimination and on-chip isolation take fleet readout error toward 10⁻³ without slowing the cycle. Worst case: measurement-induced leakage holds QNDness near 99.3% whatever the amplifier does, and readout stays the term that caps Λ. Open questions: how the fleet 1% decomposes; how many tones one TWPA carries before compression.