For quantum computing to reach the point where it is fault-tolerant, scalable, and commercially viable, it’s going to be with the help of key components of advanced computing today, namely AI, open models, and GPU-powered HPC simulation and modeling, all of which conveniently fit in Nvidia’s wheelhouse, according to Sam Stanwyck, director of quantum product for the giant GPU maker.

Those components are going to be important for a broad array of needs that have to come together for quantum systems to flourish, with error correction being a significant challenge, Stanwyck told journalists this week.

“The way to think about this is, AI is becoming the control plane for quantum hardware,” he said. “Qubits are noisy, and the way to manage that noise at the scale we need is with AI models. These models also really need to be open so they can be customized, fine-tuned, and continuously improved on by the quantum community. This is the path to quantum GPU supercomputing, which is a quantum accelerator integrating with the GPU supercomputer, solving valuable problems.”

It has long been understood that quantum computing will be tightly intertwined with classical HPC and AI, and will likely – at least in the early era – be offered as a cloud service. Nvidia has ridden the rapidly expanding AI market for more than a decade, all the way to a market capitalization that has reached more than $4.5 trillion, and sees accelerated HPC as the path to quantum computing. Co-founder and chief technology officer Jensen Huang a year ago said that Nvidia will be a critical accelerator of quantum, adding that “we don’t build quantum computers, and yet we are deeply integrated into the quantum computing industry.”

The company has continued to expand its portfolio of open quantum technologies, including its CUDA-Q programming platform – which includes two libraries, QEC for error correction and Solvers for hybrid algorithms – and cuQuantum for GPU accelerated simulation for CUDA-Q and other frameworks to NVQLink (unveiled late last year) for low-latency interconnectivity between quantum chips and GPUs.

Nvidia has also been growing its family of open source AI models, from Nemotron for agentic AI and Cosmos for physical AI to Gr00t (robotics), BioNeMo (biomedical research), Apollo (AI physics), and Alpamayo (autonomous vehicles).

Now the focus shifts to quantum, with Nvidia this week coming out with the first of several planned open models for the field. The first two of its open quantum models – that will be housed in a family of models called Ising – focus on calibration and decoding, and both address error correction.

“Both of these are targeting the fundamental challenge in quantum computing, which is that qubits are inherently noisy, and that noise is the fundamental bottleneck standing between today's quantum hardware and useful applications,” Stanwyck said. “Today, the very best quantum processors make an error about once in every thousand operations, which is amazing. But to become useful accelerators for scientific and enterprise valuable problems, that number needs to become one in a trillion or even less. The good news is that AI can be the answer for how you manage that noise at scale. It has the potential to enable very rapid progress in closing that gap.”

Error correction is central to scaling fault-tolerant quantum systems. Qubits are extremely fragile and prone to errors. They are highly sensitive to such environmental factors as noise, light, heat, and other qubits, which causes them to decohere, or break apart and lose a specific quantum state. This can lead to errors that can affect the accuracy of a system’s results, which is never good in computing.

Calibration essentially is how users understand the noise in each quantum processor and tune it to achieve the best possible performance, he said. However, no matter how well users can calibrate, they still need to be able to correct the errors in real time using a computer that can do this faster than the errors are being made.

“Both of these are AI-shaped problems, and Ising delivers the world's best performance on both calibration and error correction decoding,” Stanwyck said. “This is the path to quantum GPU supercomputing, which is a quantum accelerator integrating with the GPU supercomputer, solving valuable problems. There’s Ising Calibration, which is a vision language model [VLM] that interprets measurement data coming out of the quantum computer and automatically calibrates and recalibrates the quantum computer based on that data. Then there's Ising Decoding, which is two 3D convolutional neural network models: One optimized for speed, one optimized for accuracy, both for quantum error correction.”

Calibration today is primarily done by human physicists or simple algorithms, which can be inaccurate, take days to accomplish, and doesn’t scale. It difficult to do this with a system with 100 qubits, and commercial quantum systems will need more than a million qubits. Ising Calibration is a 35-billion parameter VLM, which Stanwyck said is 15X smaller than other such systems, that uses AI agents to automate the entire calibration workflow, which cuts the job from days to hours.

“This problem, we need to scale up, and [calibration] scales exponentially poorly,” he said. “If this can be done in an autonomous way, it's a game-changer for scaling quantum hardware. Calibration is one of those problems where human experts are the bottleneck. Calibration is not something you do once and then you're done. It's not a one-time thing. These machines are constantly needing recalibration. Actually, the standard now is calibrating the quantum processor before every single computation. An AI agent running Ising Calibration already does it faster and better, and it gets better as the hardware scales up, not worse.”

With Ising Decoding, what Nvidia actually is offering is pre-decoding. Much of the decoding today is done with PyMatching, an open source Python and C++ library that uses what’s called the Minimum-Weight Perfect Matching algorithm to identify and correct errors. Ising Decoding is designed to work with PyMatching and other decoders and accelerate them. Nvidia’s mode comes in two model variants, one aimed at speed – which is 2.5X faster than alternatives – and the other accuracy, which Nvidia says is three times better. Ising Decoder also needs 10X less data to work.

“When we move to logical qubits, error-corrected qubits, how fast you can do this and how accurately you can do quantum error correction are what determine how useful your quantum processor is, along with its fundamental noise characteristics,” Stanwyck said. “The closer you can get to the quantum computer speed, error correction becomes far more effective and the logical error rate is directly correlated with how much computing you can do with the quantum processor and the problems you can solve.”

Addressing error correction through calibration and decoding was a logical first step, he said.

“Other than building the actual QPU, these are the two most important problems we need to solve today that are bottlenecking hardware capability,” he said. “They're the workloads every quantum team is dealing with right now, and they're both fundamentally AI problems. They're good AI problems. You're processing a very high throughput, you have noisy data at scale and making real-time decisions, which is exactly what AI models are great for.”

That said, the plan going forward is to add other Ising models to address other challenges, including optimizing quantum circuits and programs, bringing in system-level control, and building more optimized algorithms.

Nvidia is position Ising as the top layer for its portfolio of quantum-related offerings. For example, the models can integrate with CUDA-Q for hybrid workloads and NVQLink for real-time control and error correction, and the vendor offers a case where Ising Decoder is being used with the CUDA-Q QEC library running on NVQLink.

Both of the new models already are being used by a range of AI vendors, research institutions, and university labs, and Stanwyck said he expects more to be added to the list in the coming weeks.