NVIDIA is stretching its cuOpt optimization solver to 100-million-variable scale, signaling a push to make large-scale decision optimization a GPU workload.
An NVIDIA developer blog post describes bringing mPDLP, an extension of PDLP-style methods, into cuOpt to handle decision-optimization problems beyond 100 million variables. The coverage is mostly NVIDIA's own channels, alongside a separate post on how its chips are made.
Solving huge logistics and scheduling problems on GPUs widens GPU demand beyond AI model training and inference. That said, this is a vendor-published performance claim; independent verification and customer deployments have not been confirmed, so real-world impact remains open.
Watch for reproducible public benchmarks, comparisons with established solvers, and concrete customer deployments.