Based on reading https://developer.nvidia.com/blog/record-breaking-nvidia-cuo..., we are honestly pretty similar in terms of technology choices and (verifiable) performance. NVIDIA's found a few new records on VRPTW and pickup-and-delivery problems; but so did we in May and June 2023: https://galgos.inf.puc-rio.br/cvrplib/index.php/en/updates/. We have stopped playing the benchmarketing game, but if PyVRP ran for a few hours I'm sure we could find new records on (some) instances again. And we only need commodity hardware to do that, not beefy GPUs.

The NVIDIA article mentions constant time move evaluations using time warp inside their local search. We use the exact same thing. We used to have a genetic algorithm on top of the local search, like NVIDIA, but recently dropped that in favour of iterated local search, because we found that to be stabler, converge faster, and overall easier to reason about.

I think NVIDIA has a great product, and an even better marketing department.