GPUs are entering quantum error correction through calibration. The real-time loop is still unproven.

Verified

On October 14, 2026, Riverlane published results from running its Local Clustering Decoder (LCD) on a single NVIDIA H100 GPU to recalibrate decoder weights. A decoder is the classical software that infers which errors hit the qubits. Its weights are the error probabilities it assumes. On a surface code, one training epoch took 2.8 seconds on the H100 against about three minutes on one CPU core, 65x faster. Source: From Minutes to Seconds: Faster Decoder Recalibration for Quantum Error Correction, Riverlane.
Riverlane reports the H100 decoded about 4.2 million surface-code syndromes per second against roughly 59,000 on one CPU core, a 71x gain, and 155 million per second on a repetition code, a 121x gain. Against 32 CPU cores, the H100 was still 3x faster on the surface code and 2x faster on the repetition code. Source: same.
Riverlane says the calibration runs "between runs of an experiment rather than during them." It reports the learned weights cut the logical error rate by "anywhere from several percent to as much as a factor of two" against the starting prior. Source: same.
In the same post, Riverlane wrote that future fault-tolerant machines "will require a heterogeneous classical computing stack, with specialised systems handling different error-correction tasks," and called calibration "a concrete, well-defined workload for the GPUs they are already placing beside quantum hardware." Riverlane sells a separate real-time QEC system, Deltaflow. Source: same, and Riverlane and NVIDIA explore faster calibration for quantum error correction, Riverlane blog.
On October 13, 2026, Alice & Bob SAS reported decoding its LDPC-cat codes on NVIDIA GPUs over NVIDIA NVQLink, using Quantum Machines' OPX1000 controller and Open Acceleration Stack, which the release says provides "microsecond-scale roundtrip latency." The test used synthetic data. The release states: "Though a proof of concept and not constituting real-time decoding, the results show how decoding can scale by becoming more efficient as several syndromes are decoded in parallel." Source: Alice & Bob press release via Public.
In the same release, Sam Stanwyck, Director, Quantum Product at NVIDIA Corporation (NASDAQ: NVDA), said: "Alice and Bob's work shows how NVQLink can in fact accelerate the real-time decoding needed to turn cat qubits into fault tolerant quantum-GPU supercomputers." Source: same.
On August 25, 2026, IonQ, Inc. (NYSE: IONQ) researchers posted a paper describing an end-to-end real-time decoding stack for a trapped-ion fault-tolerant architecture that "runs on a single CPU." It was benchmarked on workloads of up to 408 logical qubits and up to one million T gates. Assuming a 1 to 5 ms error-correction cycle, decoding delay stretched the computation by less than 0.3% at a two-qubit gate error rate of 10⁻⁴, and less than 12% at 5×10⁻⁴. Source: Real-time decoder for a MegaQuOp quantum computer using a single CPU, arXiv:2608.25027.
Quantum Computing Report says the CPU was an off-the-shelf Apple M4 Max using 12 cores. Source: Quantum Computing Report.
In September 2026, IQM Quantum Computers adopted NVIDIA CUDA-Q Logical for its Halocene QEC product line, with NVQLink for GPU decoding, and Anyon Computing released an open-source control plane built on NVQLink. Sources: Quantum Computing Report on IQM and Quantum Computing Report on Anyon.

Beneath the label

The pipeline flagged three GPU-and-decoding items in one week, and the coverage read them as one story: GPUs are coming to quantum error correction (QEC). Read the primary releases side by side and the story splits in two.

The Riverlane result is real and specific, and it is offline. Calibration is batch work. You take a pile of recorded syndromes, the measurement results the decoder reads, and decode them many times under different candidate weights. Every shot is independent. That is exactly the shape of problem a GPU is built for, and the 71x number reflects it.

The Alice & Bob result is about the other job: decoding live, while the machine runs. Here the release is careful. It used synthetic data, it tested whether data could move from controller to GPU, and it says in plain words that it was not real-time decoding. The NVIDIA quote in the same release says the work shows NVQLink "can in fact accelerate the real-time decoding." Those two sentences do not describe the same result. The company's own description is the narrower one.

IonQ's paper is the counterweight. For trapped ions, with cycle times of 1 to 5 ms, the full real-time stack fits on one off-the-shelf CPU with small delay. No GPU required.

The read

The verified facts, restated. GPUs have a demonstrated, measured win in decoder calibration, which runs between experiments. GPU real-time decoding over NVQLink has been shown on synthetic data and is described by the company that ran it as not yet real-time. One trapped-ion vendor shows real-time decoding at scale on a CPU.

The read: NVIDIA's foothold in QEC is coming through the batch layer first, meaning calibration, simulation, and compilation, not the live correction loop. That is a strong position anyway. It puts GPUs in the quantum rack for a workload that recurs every time the hardware drifts. But it is a different claim from "GPUs decode quantum computers in real time," and the releases this month support the first claim, not the second.

What this would mean if it holds:

First, decoder hardware demand will split by qubit type. Hypothesis: slow-cycle platforms such as trapped ions can run real-time decoding on general CPUs, as IonQ's paper argues. Fast-cycle platforms, which NVIDIA's Stanwyck himself grouped with cat qubits as "faster qubit modalities," need either dedicated real-time hardware or a GPU path that hits microsecond budgets on live data. The second has not been shown in these releases.

Second, the GPU's first job is to keep the decoder accurate. Riverlane's own framing is a split stack: GPUs for learning and recalibration, dedicated systems for the live loop. Reading: that makes GPUs a support system for QEC, budgeted like test equipment, before it makes them the critical path.

Third, watch the qualifiers in vendor claims. The gap between Alice & Bob's "not constituting real-time decoding" and NVIDIA's "can in fact accelerate the real-time decoding" is the kind of gap that disappears in secondary coverage. The primary text is the one to quote.

Watch

A live-data, real-time GPU decoding result with a latency number. Alice & Bob says it is now optimizing the pipeline for real-time performance. A published end-to-end latency on real qubit data, measured against the code's cycle time, is the single result that would move GPUs into the live loop.
Calibration in production. Whether a hardware vendor runs Riverlane's GPU calibration before every experiment, as Riverlane proposes, rather than in a demo.
IonQ's CPU decoder on hardware. The arXiv result is benchmarked on compiled workloads under an assumed cycle time. A run against real device data would confirm that slow-cycle platforms can skip GPU decoding.
NVQLink deployments that name the workload. IQM and Anyon have adopted NVQLink. The useful disclosure is which job the GPU does in each: calibration, simulation, or live decoding.

If item 1 lands with a latency inside a superconducting or cat-qubit cycle within 12 months, GPUs are in the live loop and this read is too cautious. If it does not, the GPU's role in QEC stays where Riverlane put it: fast, valuable, and offline.