Bosonic cavity qubits — cat and GKP on the map bar: Cat & GKP cavities

The architecture in the report (§8.3)

8.3.2 Cat & GKP cavities — Bosonic cavity qubits — cat and GKP (cat)

The idea. The qubit is a microwave resonator mode, the Bosonic cavity mode (cavity), whose dominant error is photon loss. In the Cat-code encoding (biased noise) two-photon dissipation holds two coherent states apart: bit-flips fall exponentially with mean photon number n̄, phase-flips rise linearly, so a line of d cats, not a d × d patch, corrects the rest; lattice density is the price. The bet: photons buy the distance a lattice buys with qubits, if the bias survives the entangling gate. The alternate GKP grid encoding shares the hardware, not the bet: it corrects small displacements and passes up an analogue syndrome.

What it needs. The encoding requires two-photon dissipation on a mode, and Bosonic concatenation a biased or GKP inner qubit: the architecture rests on the phase-flip rate during a two-qubit gate. Every measured entangling gate is ancilla-mediated (g_bos), through a transmon that does not preserve the bias; the Bias-preserving cat–cat CNOT that would replace it is an empty slot with one simulated proposal (Jul 2026) [S][81]. Six layers offer one technology each. Lithographed modes replace 3D machined cavities, and the 3D branch conflicts with the lattice: centimetre-scale bodies cannot tile a dense 2D array (open [82]).

Where it stands. Seven machines, six devices (Table 8.3): Alice & Bob Boson 4, Helium 2 ("Galvanic Cat"), Helium (on-premise system) and the planned Graphene (2030) [R][828], [873] — the undated Lithium (48 cats) and Beryllium (250) chips stay unlisted until Alice & Bob dates them; AWS Ocelot; Nord Quantique single-mode GKP and Tesseract. The largest, Helium (18 cats), deployed June 2026 without logical data [C][87], [88]. Bit-flip times reach 430 s (Boson 4) and a preliminary 44-min mean (Helium 2) [C][77]; only Ocelot ran an outer code: repetition at 1.75% (d=3, n̄=1) and 1.65% (d=5, n̄=1.5) per 2.8 µs cycle, distance and photon number moving together, with CX phase-flip 9.6(4)×10⁻² against bit-flip 3.5(4)×10⁻³ [D][80]. Helium 2's 5.8×10⁻² is a single-qubit Z gate, not a two-qubit error, and Ocelot's 9.6×10⁻² is a phase-flip rate per error-correction cycle, not a gate error, so Table 8.3 shows no two-qubit error for the architecture [C][77]. Nord Quantique single-mode GKP: 8.1×10⁻³ logical error per round, SPAM below 10⁻³ at survival 0.24 × 0.39 [D][82].

Exceptions and borrowings. No off-slot cell, no better-fitting architecture; the shared absences, real-time decoder and interconnect, mark an early stage. The Nord Quantique pair: its decoder cell is an Atlas gap, ∅G-auton (stabilisation without classical feed-forward), so layer 8 is drawn too narrowly; its code cell reads Bosonic concatenation though neither runs an outer code, like Boson 4, a single cat with zero logical qubits on its roadmap [C][92]. GKP joins by hardware, not by the bias bet: for its idea the architecture is drawn too widely. Borrowers: Aurora (continuous-variable photonics) takes the GKP and code technologies; the dual-rail architecture's Aqumen Seeker, SUSTech biased-erasure cavity qubit and D-Wave gate-model dual-rail system take cavity and gates.

The next test. Two expectations to March 2028 (§5.2). First, Helium's roadmap rung, one logical qubit at 10⁻² [R][92]: phase-flip ≤ 10⁻² per cycle at n̄ ≈ 10 confirms; Helium still without logical data at end-2027 demotes. Second, a measured two-cat gate holding bit-flip time within 10× of idle; if 2028 opens with ancilla-mediated gates only, the 126,133- and 758-cat estimates, which assume 10⁻³ phase-flip per cycle, lose their premise [S][95], [96]. GKP owes a published GKP–GKP gate.

What the builders say. AWS chose cat qubits because they suppress bit flips in the hardware itself, as its Ocelot chip showed [D][827], and estimates that the approach cuts the cost of error correction by up to 90% [C][827]. Alice & Bob plans Graphene, 100 logical qubits for early industrial quantum advantage, by 2030 [R][828]; it reported bit-flip times above one hour, four times what that machine needs, before peer review [C][77], and built its on-premise Helium system to encode a first logical qubit in as few as 18 cat qubits [C][87], [88]. Nord Quantique builds the correction into each qubit, one cavity per logical qubit [C][829], and expects its first machines with more than 100 logical qubits by 2029 [R][830].

Machines of this architecture (register: 7, of them 6 devices, 6 gate-capable). Alice & Bob — Alice & Bob Graphene (planned, 2,000), Alice & Bob Helium (on-premise system) (deployed, 18), Alice & Bob Helium 2 ("Galvanic Cat") (demonstrated, 12), Alice & Bob Boson 4 (demonstrated, 1); Nord Quantique — Nord Quantique Tesseract (demonstrated, 2), Nord Quantique single-mode GKP (demonstrated, 1); AWS Center for Quantum Computing — AWS Ocelot (demonstrated, 9).

Cells outside the architecture's slots (2). Nord Quantique single-mode GKP: L8 ∅G-auton (Atlas gap — no classical decoder: autonomous, feed-forward-free stabilisation with analog closest-vect); Nord Quantique Tesseract: L8 ∅G-auton (Atlas gap — no classical decoder: autonomous stabilisation with heralded discard; no real-time decoder).