French quantum computing company Alice & Bob has unveiled plans for "Carbon," a quantum computer that would build 200 logical qubits from 3,600 physical cat qubits as early as 2030, according to a new roadmap announced on November 18, 2026. A logical qubit is a mechanism that spreads quantum information across multiple physical qubits and protects it through error correction. Under the previous roadmap published on December 4, 2024, the company planned to build 100 logical qubits by 2030 with "Graphene." The new plan revises the error-correction scheme and the configuration that links multiple cryogenic cooling units, raising the logical qubit target to 200.
Physical Qubits per Logical Qubit Cut From About 67 to 18
In Alice & Bob's latest roadmap, "Beryllium" is positioned as the 2028 development target ahead of Carbon. Beryllium is planned to build 6 logical qubits from 400 physical qubits, enabling the sequence of logical operations that quantum algorithms require. Carbon, by contrast, is designed to increase the number of logical qubits by linking multiple chips of the same 400-qubit scale and combining repetition codes with two-dimensional LDPC codes.
| Planned machine | Physical qubits | Logical qubits | Physical qubits per logical qubit | Error-correcting code | Target logical error rate |
|---|---|---|---|---|---|
| Beryllium (2028) | 400 | 6 | About 67 | Repetition code | |
| Carbon (2030) | 3,600 | 200 | 18 | 2D LDPC code and repetition code |
Dividing physical qubits by logical qubits gives about 67 for Beryllium and 18 for Carbon. Carbon sharply reduces the number of physical qubits per logical qubit while also setting a stricter logical error rate target. However, because the two machines differ in error-correcting codes, hardware configuration, and target logical error rates, this ratio reflects differences in design rather than an improvement in error-correction efficiency measured under identical conditions.
Carbon is planned to use nine "ab 400-x" modules, each with 400 physical qubits. Because there is a limit to the scale of quantum chip a single cryogenic cooling unit can house, the company will first connect two cooling units with coaxial cables, then expand to three and nine. Separate quantum processors will be connected using "Bell pairs," pairs of qubits that share quantum entanglement. The design is meant to make multiple modules operate as one quantum computer, including calibration of individual chips, compilation that converts quantum circuits into a form that can run on real hardware, and decoding that estimates errors from measurement results.
Cat Qubits' Error Resistance and Remaining Challenges
The "cat qubits" Alice & Bob develops use quantum states formed in a resonator of a superconducting circuit. Their defining feature is that hardware itself can suppress "bit flips," in which the states corresponding to qubit 0 and 1 swap. "Phase flips," in which the relative phase of a superposition of 0 and 1 is inverted, remain, so separate error correction is still required.
A peer-reviewed paper in Nature (DOI: 10.1038/s41586-024-07294-3) published in 2024 by Réglade, Bocquet, Gautier and colleagues reported that for a single physical cat qubit implemented in a superconducting circuit, the bit-flip time exceeded 10 seconds, while the phase-flip time exceeded 490 nanoseconds. The research team also succeeded in generating and visualizing quantum superposition states, showing that the phase could be controlled while the suppression of bit flips was maintained.
The repetition codes and two-dimensional LDPC codes Alice & Bob uses are mainly intended to detect and correct the phase-flip errors that remain in cat qubits. In a repetition code, the information of one logical qubit is spread across multiple physical qubits for protection. In two-dimensional LDPC (low-density parity-check) codes, by contrast, the number of qubits involved in each check is kept small while multiple logical qubits can be stored in a single code block. The conditions under which this approach can reduce the number of physical qubits are also examined in research on two-dimensional codes published in 2025.
Can 2D LDPC Codes Reduce the Qubits Needed for Quantum Memory?
In a paper in Nature Communications (DOI: 10.1038/s41467-025-56298-8) in 2025, Ruiz, Guillaud, Leverrier, Mirrahimi, Vuillot and colleagues compared two-dimensional codes and repetition codes through numerical simulation. The authors include Ruiz and Guillaud of Alice & Bob. The study used a model in which errors occur at each operation in a quantum circuit, reflecting noise calculated for each cat qubit operation. It assumed a ratio of one-photon loss rate to two-photon stabilization rate of , an average photon number of 11 in the resonator, and a phase-flip probability of about 0.1% per operation of a single physical qubit.
A key metric for evaluating an error-correcting code's performance is "code distance." It represents the minimum number of physical qubit errors needed to cause an error in a logical qubit without being detected by the code. The research team calculated the probability of logical errors from phase flips using the Monte Carlo method, extrapolating from the calculated results for code distance 22. They also estimated the error probability arising from bit flips with a separate formula and added it to the phase-flip error probability.
Under the condition of holding 100 logical qubits in error-corrected quantum memory, the number of physical cat qubits required, including auxiliary qubits for checks, was estimated at 758 for the two-dimensional code and 2,100 for the repetition code. The total logical error probability per logical qubit per correction cycle was predicted to be for the two-dimensional code and for the repetition code. Although the error probabilities differ, both designs fall below the shared target of , and the calculations show the two-dimensional code can substantially reduce the number of physical qubits required.
However, the two-dimensional code is not always advantageous. In this numerical model, the two-dimensional code achieved the target with fewer physical qubits when the phase-flip probability was sufficiently below the error-correction threshold. For the codes compared, the two-qubit operations (CNOTs) needed for one error check number two for the repetition code and four for the two-dimensional code. The authors believe this difference in circuit depth is a factor behind the two-dimensional code's lower phase-flip threshold.
In addition, while a two-dimensional code can store multiple logical qubits in one block, operating on each one individually requires additional computation circuitry. The research team proposed a configuration that combines a memory layer of 758 physical cat qubits with a separate computation layer. This layer would contain pathways that mediate operations between logical qubits, and circuits that generate the special quantum states needed for universal quantum computation.
Error-Correction Experiments on Three Prototypes Fall Short of Threshold
In an announcement on October 20, 2026, Alice & Bob reported that it had built and operated three prototypes as the first phase of France's quantum computer development program "PROQCIMA." They comprise two prototypes with 6-qubit chips and one with a 12-qubit chip. According to the company, it succeeded in running CNOT gates while maintaining the cat qubits' bit-flip suppression, reading out information about errors from auxiliary qubits, and then resetting them for the next check. It also said it performed a non-Clifford gate, one of the operations needed for universal quantum computation, on physical cat qubits.
For error correction, the company implemented a repetition code of distance 3 using three data-holding qubits and ran multiple correction cycles. However, the logical error rate measured across the tested qubit groups did not reach a level below the error-correction threshold. At the November 18 roadmap announcement, the company explained that these experiments brought it closer to the threshold, but the October experiment announcement does not publish the specific threshold value used for the judgment or measurement results for different code distances.
Alice & Bob also developed a "digital twin" that reproduces the quantum processor's behavior using measurements taken on real hardware. Simulations with this model predict that getting below the error-correction threshold requires consistently long quantum-state coherence across the entire chip.
Scaling Up Error Correction and Linking Cooling Units Are the Next Challenges
In quantum error correction, if the physical error rate falls below a certain threshold, increasing the code distance reduces the logical error rate. To move from the stage where Alice & Bob ran multiple correction cycles of a distance-3 repetition code to larger codes, it will need to confirm that this relationship holds on real hardware. It will be important to align physical noise and operating conditions, measure the logical error rate per correction cycle for each code distance, and compare them along with statistical errors.
The PROQCIMA experimental results do not list logical error rates, measurement errors, or numbers of trials for each prototype. As for two-dimensional LDPC codes, the aforementioned Nature Communications paper shows performance from numerical simulation, but within the scope of public materials, no independent real-hardware replication using the same code can be confirmed.
In Carbon's plan, technology to connect multiple cooling units will also be needed in parallel with verifying this error-correction performance. As the connection scale expands to two units and then three, key metrics will be how often Bell pairs can be generated between units and the "fidelity," which expresses how close the generated quantum states are to the ideal. If generating Bell pairs takes time, errors arising from qubits waiting in the meantime must also be considered.
Furthermore, the logical error rate including quantum operations across cooling units, and the number of physical qubits required for error-correction auxiliary qubits, inter-unit connections, and computation processing, will also be subject to evaluation. Once the breakdown of these is clarified, the design target of "18 physical qubits per logical qubit" that Carbon sets can be concretely verified in terms of real-machine performance.
In its November 18 roadmap announcement, Alice & Bob cites materials simulation using the Ising model and Hubbard model as uses for Carbon. To assess advantages over classical computers on these problems, it will be necessary to compare not only the scale of problems that can be handled but also the time needed to obtain results of the same accuracy. What scale of computation can be completed while continuing error correction across cooling units will be a basis for judging the practical meaning of the 200-logical-qubit goal.
