NVIDIA Unveils Ising, Open-Source AI Models Built to Accelerate Quantum Computing

NVIDIA has spent two decades teaching the world that the hard part of computing is not the chip — it is the software stack wrapped around it. On April 14, the company applied that same playbook to quantum computing, announcing Ising, what it calls the world's first family of open-source AI models purpose-built to make qubits usable. The pitch is blunt: the most fragile, error-prone machines in computing need an AI control plane, and NVIDIA wants to own it.

Named after the landmark Ising model in statistical physics, the new family targets the two unglamorous problems standing between today's noisy quantum processors and anything resembling a useful computer — calibration and error correction. NVIDIA says its decoding models run quantum error correction up to 2.5x faster and 3x more accurate than the traditional open-source standard, while needing roughly ten times less training data.

"AI is essential to making quantum computing practical," said Jensen Huang, NVIDIA's founder and CEO. "With Ising, AI becomes the control plane — the operating system of quantum machines — transforming fragile qubits to scalable and reliable quantum-GPU systems."

Why Error Correction Is the Whole Game

Quantum computing's central problem is that qubits are exquisitely sensitive. They decohere, they drift, and they make mistakes constantly. Sam Stanwyck, NVIDIA's director of quantum product, framed the gap in stark terms during the company's briefing: "Today, the very best quantum processors make an error about once in every 1,000 operations, which is amazing, but to become useful accelerators for scientific and enterprise problems, that number needs to become one in a trillion or even less."

Closing that gap of nine orders of magnitude is the job of quantum error correction, which spreads a single logical qubit across many physical ones and continuously measures for errors. The catch is that the correction has to happen in real time, faster than new errors accumulate — a brutal latency requirement that has made the decoder, the classical software interpreting those measurements, a notorious bottleneck.

This is where Ising plants its flag. The Ising Decoding family consists of two variants of a 3D convolutional neural network — one optimized for speed, the other for accuracy — that perform real-time pre-decoding for surface-code error correction. NVIDIA benchmarks them against pyMatching, the open-source decoder most quantum research groups currently rely on, and reports the 2.5x speedup and 3x accuracy gains there. The models are small by large-language-model standards — on the order of one to two million parameters — because they have to run inside an unforgiving real-time loop, not in a chatbot.

The second pillar, Ising Calibration, is a vision-language model that reads measurements coming off a quantum processor and reacts to them. NVIDIA says it lets AI agents automate the continuous tuning that quantum machines require, collapsing a process that takes days into one that takes hours. Calibration is the kind of painstaking, expert-driven chore that scales terribly as qubit counts climb, so automating it is arguably as consequential as the decoding speedups.

Open Models, Closed Stack

The "open" framing matters, and NVIDIA leaned on it hard. The Ising models ship pre-trained on GitHub, Hugging Face, and build.nvidia.com, bundled with a "cookbook" of quantum workflows, training data, and NIM microservices so developers can fine-tune for their own hardware. Crucially, the models can run locally, which lets quantum groups keep proprietary device data in-house — a real concern in a field where calibration data effectively encodes the secrets of your chip.

But openness at the model layer sits inside a thoroughly NVIDIA-owned stack. Ising complements CUDA-Q, NVIDIA's hybrid quantum-classical software platform, and integrates with NVQLink, its QPU-GPU interconnect built for exactly the real-time control and error-correction traffic that decoding demands. The strategy rhymes with everything NVIDIA has done in AI: give away the models, anchor the ecosystem, and sell the silicon and interconnect underneath. Ising now joins an open-model portfolio that already spans Nemotron for agents, Cosmos for physical AI, and BioNeMo for biology — quantum is simply the newest vertical to get the treatment.

The early adopter list reads like a census of serious quantum players. Ising Calibration is already in use at Atom Computing, Academia Sinica, Fermilab, Harvard's engineering school, IonQ, IQM Quantum Computers, Q-CTRL, Infleqtion, and Lawrence Berkeley National Laboratory's Advanced Quantum Testbed, among others. Ising Decoding is being deployed at Cornell, Sandia National Laboratories, SEEQC, UC San Diego, UC Santa Barbara, the University of Chicago, USC, and Yonsei University. That breadth — national labs, hardware startups, and universities across multiple qubit modalities — suggests NVIDIA is positioning Ising as neutral infrastructure rather than a bet on any one quantum architecture.

What to Watch

NVIDIA cites analyst firm Resonance projecting the quantum computing market to surpass $11 billion by 2030, a figure explicitly tethered to progress on error correction and scalability. By embedding itself in that critical path now, NVIDIA is making sure that whichever qubit technology wins — superconducting, trapped-ion, neutral-atom, or otherwise — the surrounding control plane runs on its GPUs.

The numbers to scrutinize next are independent ones. NVIDIA's 2.5x and 3x figures are its own benchmarks against pyMatching, and the quantum community is rigorous about reproducibility; expect adopters like Sandia, Fermilab, and the academic deployers to publish their own decoding results on real hardware over the coming months. The harder question is whether AI-accelerated decoding can keep pace as logical qubit counts grow, since the real-time latency budget only gets tighter at scale. If Ising's decoders hold up under that pressure on physical machines — not just in benchmarks — NVIDIA will have done to quantum error correction what it did to deep learning: turned an academic bottleneck into a product line. Watch the peer-reviewed numbers, and watch how quickly the rest of the stack consolidates around CUDA-Q and NVQLink.