Home » Quantum Computing » Alice & Bob Says It Decoded Cat-Qubit Error Correction on GPUs for the First Time

October 13, 2026

Alice & Bob Says It Decoded Cat-Qubit Error Correction on GPUs for the First Time

A decoder bottleneck has held back fault-tolerant quantum computers for years. Does swapping custom chips for ordinary GPUs actually fix it, or just move the problem?

Alice & Bob says it ran a cat-qubit GPU decoding test for the first time on October 13, 2026. NVIDIA confirmed the test used its new NVQLink architecture, a system built to wire quantum processors directly into banks of graphics chips. Both companies call it a proof of concept. It is not a finished product, and the details are worth reading before the headline.

The quiet bottleneck: decoding

Alice & Bob builds a kind of qubit called a cat qubit. Its physical design suppresses one common error, the bit flip, almost automatically. That leaves the error-correction system free to spend its energy on the error that remains, the phase flip, and the company’s LDPC-cat codes are built around that trade.

Suppressing one error type doesn’t remove the need for a decoder, though. A decoder is a classical computer program sitting next to the quantum chip. It reads a stream of error signals, called syndromes, and works out what went wrong fast enough to fix it before the next step runs. If the decoder falls behind, errors pile up faster than they get corrected, and the whole point of error correction collapses.

This bottleneck shows up across the industry, not just at Alice & Bob. IonQ has tested its own real-time decoder on ordinary laptop-grade chips, and Google has proposed a modular chip design meant to cut the decoding workload by a factor of ten. Everyone chasing a useful fault-tolerant quantum computer runs into the same classical computing wall.

What the demo showed, and why GPUs

The companies connected Quantum Machines’ OPX1000 control electronics to NVIDIA GPUs through Quantum Machines’ Open Acceleration Stack, which moves data between the control hardware and the GPU in microseconds, not milliseconds. That is the plumbing, not the decoder itself. Researchers ran synthetic error data through it, with no live quantum chip involved, and found that decoding got more efficient when the system processed several batches of error syndromes side by side instead of one at a time.

Both companies are upfront that this was not real-time decoding. Juliette Peyronnet, Alice & Bob’s US general manager, said decoding is one of its most demanding classical tasks, and that the next job is making it fast enough to keep pace with the quantum computer. NVIDIA’s Sam Stanwyck framed NVQLink as the piece that can eventually carry that load.

“Showing that our LDPC-cat codes can be decoded on NVIDIA GPUs with existing tools is encouraging for scaling,” Peyronnet said.

The choice of GPUs matters because most quantum decoders built so far run on specialized chips, often FPGAs, which are designed for one narrow job and are expensive to design and slow to improve. GPUs are manufactured by the millions, and they get faster nearly every year because the entire AI industry depends on them. If decoders can ride that improvement curve instead of waiting on bespoke chip redesigns, scaling fault-tolerant quantum computers could get cheaper and faster over time. That is a possibility this test raises, not a cost figure anyone has published.

Why It Matters to You

Nothing here lowers anyone’s bills today, and no company has published a cost comparison between GPU decoding and custom decoder chips. Picture Teresa, an illustrative example and not a real person, a high school chemistry teacher in rural Minnesota who tells her students quantum computers might someday help design better batteries or new medicines. If a student asked how close that is, the honest answer is that researchers are still fighting the classical computer bolted onto the quantum chip, not just the quantum chip itself. What changed this week is the toolset researchers can reach for while they try.

What is still unproven

The test used synthetic, simulated error data, not output from a live quantum chip, so whether this approach works on real hardware is untested. Decoding was not real-time; the companies describe reaching real-time speed as a future goal. No decoding throughput, latency, or accuracy numbers were published. The only measured figure released is the roundtrip latency of the control-to-GPU link itself, described only as microsecond-scale. Every quote in the coverage comes from Alice & Bob or NVIDIA employees, and Quantum Zeitgeist’s own write-up notes that no independent researcher has weighed in. The claim that decoding multiple error syndromes in parallel will keep scaling rests on this one test, not on a published, reproducible benchmark. This is a company announcement, not a peer-reviewed paper, although the underlying LDPC-cat code design was peer-reviewed separately, back in 2024.

Date	Milestone	What It Showed	Status
June 2025	CUDA-Q integrated into Dynamiqs	Linked NVIDIA’s quantum software stack to Alice & Bob’s own QPU simulator	Software integration
March 2026	CUDA-Q decode of Elevator Codes	Decoded simulated syndrome data in 1 hour, 57 minutes for a single run	Simulation, far from real-time
October 13, 2026	LDPC-cat codes decoded via NVQLink	First GPU decoding of LDPC-cat codes; synthetic data; microsecond-scale control-to-GPU latency	Proof of concept, not real-time

Source: Alice & Bob and NVIDIA press materials, reported by The Quantum Insider and HPCwire, Oct. 13, 2026.

The next thing to watch is whether Alice & Bob’s following demo swaps that synthetic data for output from a live chip, and whether the roundtrip time drops from today’s microseconds into something that can outrun the errors a scaled-up quantum processor would produce. That is the real finish line, and nobody has crossed it yet.

Quick answers

What is a cat qubit, and what is a decoder?

A cat qubit is a type of superconducting qubit that Alice & Bob designs so one common error, the bit flip, almost never happens because of the circuit’s own physics. A decoder is the classical program that reads the error signals coming off the chip and figures out what went wrong. If it can’t keep up with how quickly the chip produces new error data, mistakes build up faster than they get fixed, and the computation fails.

Did Alice & Bob run this on an actual quantum computer?

No. The companies say the test used synthetic, simulated error data running through real control electronics and GPUs connected over NVQLink. Whether this approach works with data from a live quantum chip has not been shown yet.

What happens next, and does this make quantum computers cheaper?

Alice & Bob says its next step is optimizing the data pipeline to reach real-time decoding speeds. Until then this stays a proof of concept. No cost figures or timelines have been published, so cheaper or sooner machines are a possibility, not a result.