NVIDIA Ising: First Open-Source AI Models for Quantum Computing

The most underappreciated number in the Ising announcement is the gap between where quantum hardware stands and where it needs to be. Today’s best quantum processors produce an error roughly once every 1,000 operations. Useful quantum applications — drug discovery, cryptography, complex optimization — require error rates of approximately one in one trillion. That is a six-order-of-magnitude gap, and closing it through hardware improvements alone would require decades of iterative engineering on qubit coherence, gate fidelity, and environmental isolation.

This is where Ising’s approach becomes significant. Rather than waiting for perfect qubits, NVIDIA is betting that AI-driven error correction can bridge the gap computationally. The Ising Decoding models, despite being remarkably small (912K and 1.79M parameters), achieve 2.5x speed and 3x accuracy improvements over pyMatching, the existing standard. The projected latency of 2.33 microseconds per round — potentially dropping to 0.11 microseconds with 13 GB300 GPUs running FP8 — approaches the real-time requirements for continuous error correction during quantum computation. If these numbers hold in production, AI-powered decoding could effectively buy the quantum industry years of runway while hardware catches up.

In January 2025, Jensen Huang told the world that useful quantum computers were ‘decades away,’ sending quantum stocks tumbling. Eighteen months later, NVIDIA is launching purpose-built AI models for quantum computing and backing PsiQuantum at a $7 billion valuation. This whiplash demands scrutiny.

The timeline reveals a deliberate strategic arc rather than a genuine change of mind. While Huang was publicly dampening quantum expectations in early 2025, NVIDIA was already collaborating with Google on quantum processor design using its Eos supercomputer. By June 2025, Huang had reversed course, calling quantum an ‘inflection point.’ By September, NVIDIA was writing billion-dollar checks. The January 2025 comments may have served a dual purpose: resetting inflated market expectations while giving NVIDIA time to assemble its quantum platform (CUDA-Q, NVQLink, and now Ising) before competitors understood the playbook. The result is that NVIDIA enters the quantum AI space with a comprehensive stack — hardware interconnects, software frameworks, and now open-source models — while rivals are still reacting to the market shift Huang himself triggered.

Releasing Ising under permissive licenses on GitHub and Hugging Face appears generous, but it follows NVIDIA’s proven playbook: give away the software to make your hardware indispensable. CUDA made NVIDIA GPUs the default for AI training. Now CUDA-Q, NIM microservices, and Ising aim to make NVIDIA GPUs the default for quantum computing infrastructure.

The adoption signals are already validating this strategy. IonQ, IQM, Atom Computing, and EeroQ are all integrating Ising models, which run optimally on NVIDIA hardware. Cornell and Sandia are deploying the decoding models in research environments where today’s tooling choices become tomorrow’s institutional standards. The projected performance of Ising Decoding at 0.11 microseconds latency specifically requires 13 GB300 GPUs — NVIDIA’s own hardware. Every quantum lab that adopts Ising is implicitly adopting the NVIDIA compute stack. With the quantum computing market projected to exceed $11 billion by 2030, NVIDIA is not selling quantum computers; it is selling the AI infrastructure that every quantum computer will need to function, regardless of which company builds the qubits.

While most coverage focused on Ising’s error correction capabilities, the EeroQ and Conductor Quantum demonstration of autonomous quantum computing labs may be the more immediately transformative application. Their proof of concept showed AI models not just correcting quantum errors after the fact, but autonomously operating the calibration and tuning processes that currently require teams of specialized physicists.

The calibration bottleneck is a practical constraint that rarely makes headlines but fundamentally limits quantum computing’s scalability. Every time a quantum processor needs recalibration — which happens frequently due to environmental drift — highly trained researchers must manually interpret complex measurement data and adjust parameters. Ising Calibration, a 35B-parameter vision language model, automates this by directly interpreting calibration imagery and prescribing adjustments, compressing days of work into hours. This matters because it changes the economics of quantum computing operations. A quantum data center that requires a fraction of the human oversight per processor can scale to hundreds or thousands of machines in ways that are currently impossible. Dr. Brandon Severin of Conductor Quantum frames this as AI moving from assistant to driver of scientific discovery — and in the narrow domain of quantum lab operations, Ising makes that transition concrete rather than aspirational.