# IonQ, ORNL Use Generative AI To Cut Quantum Optimization Costs > Source: > Published: 2026-09-17 19:26:25+00:00 TL;DR — Key Takeaways - Researchers from ORNL, IonQ, Nvidia and UT developed DQAOA-GPT, which uses generative AI to create quantum circuits directly instead of repeatedly tuning them. - In a 100-variable benchmark, DQAOA-GPT held runtime near 28 seconds as subproblems grew, while conventional DQAOA rose from about 34 seconds to more than 11 minutes. - The experiment used Nvidia’s CUDA-Q platform and cuQuantum software to simulate quantum circuits on a single H200 GPU. Researchers at Oak Ridge National Laboratory, IonQ, Nvidia and the University of Tennessee, Knoxville have developed a generative AI method that cuts the computational cost of tuning circuits for a distributed quantum optimization technique. The work is being presented this week at [IEEE Quantum Week](https://qce.quantum.ieee.org/2026/) in Toronto, where it received a third-place Best Paper award in the Quantum–GenAI Co-Design & Co-Discovery track. ## Training AI To Generate Quantum Circuits The framework, called DQAOA-GPT, combines the distributed quantum approximate optimization algorithm, or DQAOA, with a generative AI model trained on high-quality quantum circuits. It then uses that model to generate new circuits directly instead of repeatedly tuning each one. DQAOA works by breaking a large problem into smaller subproblems that can be solved separately and folded back into the overall solution. As those subproblems grow, however, standard DQAOA requires more circuit evaluations and classical optimization steps, driving up the computational cost. To build DQAOA-GPT, the researchers used ADAPT-QAOA, an adaptive method that builds circuits iteratively, to generate high-quality reference circuits that taught the model to map each subproblem to a suitable circuit. In a benchmark problem with 100 variables, the researchers tested what happened as the smaller pieces handled by DQAOA grew in size. IonQ said the generative approach held runtime nearly flat at about 28 seconds, while the conventional method rose from about 34 seconds to more than 11 minutes. The company said the quality of the AI-generated solutions also roughly doubled as the subproblems grew. Here, solution quality refers to relative accuracy against the best-known results from earlier work, rather than a direct comparison with a classical solver. ## GenAI Changes the Cost Curve The significance of this research is in that very relationship between size, quality and cost. Larger subproblems can capture more of the interactions in the original problem and produce better solutions, but they also make the conventional tuning process much more expensive. The results suggest generative AI could allow researchers to work with larger pieces of a problem without the same increase in runtime. If that result holds at larger scales, it could make hybrid quantum and high performance computing approaches more practical for complex optimization problems in areas such as materials design, logistics and networks. Nvidia was represented among the research team, and the experiments relied on its hardware and software. The experiments ran on ORNL’s Defiant2 system using a single Nvidia H200 GPU, with Nvidia’s CUDA-Q platform and cuQuantum software used to simulate the quantum circuits. That gave the researchers the same computing environment for testing both the conventional and generative approaches, making the runtime comparison more controlled. Another important limitation: no quantum processor was used. Because every circuit was simulated, the work does not demonstrate quantum advantage or establish how DQAOA-GPT performs on quantum hardware. The paper describes the study as benchmark-scale validation. The researchers say the framework could eventually be distributed across multiple GPUs and computing nodes because its subproblems can be handled independently. ORNL researchers said they are extending the framework to real-world applications and larger HPC systems. Access the research paper at [this link](https://arxiv.org/abs/2607.20225).