The Classical Side of Quantum: Alice & Bob Decodes Its Error Correction Codes on NVIDIA GPUs

On October 13, the quantum computing company Alice & Bob announced that it had decoded its LDPC-cat error correction codes on NVIDIA GPUs, connected through NVIDIA's NVQLink architecture. The demonstration is a proof of concept: it ran on synthetic data, not a live quantum processor, and it did not achieve real-time decoding. But it is the kind of result that matters in quantum computing precisely because it is unglamorous. The bottleneck between here and a useful fault-tolerant machine is increasingly a classical computing problem, and the industry is starting to treat it like one.

Alice & Bob, based in Paris and Boston, builds its machines around cat qubits, a technology developed by its founders. The company has raised 180 million euros and employs more than 250 people. Its NVIDIA relationship already produced two earlier milestones: the June 2025 integration of CUDA-Q into Dynamiqs, the company's simulation library, and a March 2026 result in which decoding simulated syndrome data for its Elevator Codes took 1 hour and 57 minutes.

What the demonstration actually did

Quantum error correction spreads the information of one protected logical qubit across many noisy physical qubits, then repeatedly checks those physical qubits for signs of disturbance. Those checks produce streams of diagnostic data called syndromes, and a classical program called a decoder reads them to decide which correction to apply.

The Alice & Bob team wanted to answer a narrower question: can measurement data be moved from quantum control electronics to a GPU fast enough, and can LDPC-cat codes be decoded there at all? The test rig used Quantum Machines' OPX1000 controller with the company's Open Acceleration Stack, a framework that connects the controller directly to GPUs with microsecond-scale roundtrip latency, routed through NVIDIA's NVQLink. NVQLink is NVIDIA's open architecture for coupling GPU supercomputing with quantum processors, built for tasks needing fast quantum-classical communication: decoding, calibration, and hybrid applications, across any qubit technology.

The answer was yes. The results showed that decoding becomes more efficient as several syndromes are processed in parallel, which is exactly the scaling property a decoder needs. The team was explicit: this was not real-time decoding. The next step is to optimize the pipeline until GPU decoding can keep pace with the quantum computer.

Why decoding is the bottleneck nobody talks about

Most coverage of error correction focuses on the qubits. The decoder sits offstage, but it has the hardest deadline in the building. Between one round of checks and the next, the decoder must read the syndrome data, compute the most likely errors, and apply corrections. If it falls behind, errors accumulate faster than the system can fix them. For fast qubit technologies like cat qubits, that window is especially tight.

The hardest real-time workload in quantum computing may not run on a quantum computer at all.

That is the understated version of what both companies said. Juliette Peyronnet, Alice & Bob's US general manager, called decoding "one of its most demanding classical tasks" and framed the GPU result as encouraging for scaling, because decoding capacity can now grow with the machine. Sam Stanwyck, NVIDIA's director of quantum product, said errors build up faster than they can be corrected, and it has traditionally been hard to bring state of the art compute to faster qubit modalities like cat qubits.

GPUs fit because decoding thousands of syndromes is a highly parallel problem: one syndrome's diagnosis does not depend on another's. The proof of concept confirmed that decoding grows more efficient as syndromes are processed in parallel, so the decoding layer can scale with future machines rather than becoming a fixed ceiling.

Cat qubits and LDPC-cat codes, briefly

The codes deserve a short explanation, because they explain why Alice & Bob claims its machines need less hardware than competitors'. Cat qubits are designed so that one common error type, the bit flip, is suppressed automatically, with suppression improving as the qubit holds more photons. That leaves the phase flip for active correction, and the company claims this can cut the hardware needed for a useful large-scale machine by up to 200 times versus competing approaches.

LDPC, short for low-density parity-check, is a family of codes borrowed from classical telecommunications, where variants have served in Wi-Fi and 5G for years. Alice & Bob's LDPC-cat codes, introduced in a January 2025 paper, adapt that machinery to the two-dimensional layout of its cat-qubit processors. Lower overhead means fewer physical qubits per protected logical qubit.

A parallel idea from Google: fix the answer after the run

The same day brought a second classical-side result, from Google Quantum AI. In an arXiv preprint, researchers Oles Shtanko, Saswata Roy, Shashwat Kumar and Zlatko Minev proposed virtual error cancellation: a "measure now, mitigate later" approach that uses error records saved during a computation to reduce distortion in the final answer after the computation has finished.

The method uses two decoders. A fast one operates while the machine runs, because some calculation steps depend on earlier measurement outcomes. A second, more careful decoder examines the saved records afterward, free of the deadline. Runs are sorted by whether the two decoders agree. Results from the agreement group get slightly more weight in the final average; results from disagreement groups get negative weights, so their contribution is subtracted. With the right weights, remaining errors cancel on average.

In simulations, the largest end-to-end test reduced the residual logical error bias by more than 1,000-fold compared with ordinary decoding, and by more than 100-fold compared with discarding runs where the decoders disagreed. The tests used surface codes at distances 3, 5 and 7; the distance-7 model had 49 data qubits. The strongest results needed a 10-million-sample learning budget, and the 1,000-fold suppression held when the physical error rate was at or below 2.2 percent, weakening at higher rates.

There is a cost: the weights make the final estimate more sensitive to random variation, so more repetitions are needed for a chosen precision. The study found this needs less additional sampling than existing methods that use error records, approaching roughly half of a single-decoder benchmark under favorable conditions.

What neither result means

Both announcements carry honest limits, and those limits are the story. Alice & Bob's demonstration used synthetic data and was not real-time; what remains is optimizing the data pipeline. Google's method exists only in simulations with simplified noise, where errors were placed on data qubits while the error checks were assumed perfect. Real machines also suffer faults during gates, measurements, and repeated correction rounds, and the researchers name realistic circuit noise as the next test. Neither runs a useful algorithm.

Together they sketch the actual frontier of the field in late 2026. The qubit race gets the headlines; the quiet race is in the classical layer around the qubits. Decoding syndromes in microseconds, estimating noise from saved records, moving measurement data between controllers and accelerators: these are classical computing problems, and the industry is now attacking them with the most powerful classical hardware available.