Juliette Peyronnet's team at Alice & Bob spent March decoding simulated syndrome data, and one run took 1 hour 57 minutes.
The quantum hardware was not the bottleneck.
The classical decoder was.
Alice & Bob builds cat qubits in Paris and Boston, and cat qubits run fast.
Fast qubits mean errors accumulate faster than a decoder can clear them, which is why the company wrote its own LDPC-cat codes to keep the overhead low.
Their pitch is a fault-tolerant machine with up to 200 times lower hardware requirements than competing designs.
None of that lands if the decoder cannot keep pace.
On October 13 the team reported the turning point.
They decoded LDPC-cat codes on NVIDIA GPUs through NVQLink, coupled to Quantum Machines' OPX1000 and Open Acceleration Stack with microsecond-scale roundtrip latency.
The result is a proof of concept on synthetic data, so real-time decoding is still ahead of them.
The useful part is the shape of it: several syndromes decode in parallel, so decoding capacity can grow with the machine instead of being sized for it years in advance.
Sam Stanwyck, Director of Quantum Product at NVIDIA, framed the old problem plainly.
Errors build up faster than they can be corrected.
That is a compute problem before it is a physics problem.
Hourly GPU capacity is how a team experiments with a pipeline like that before committing to a dedicated box.
RTX 3090s at $0.15/hr, storage at $2.49/TB per month, and Buy Now Pay Later: zero upfront, payment after one month.
When a pipeline is too slow, what is the first piece you move onto a GPU?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w42&utm_content=evening
Originally posted on linkedin_personal