We already had experience with running bare metal infra (hashicorp stack), currently evaluating how to deploy it (currently using GKE). Althrough it should be straight forward since its mostly stateless workload, we it need to be able scale x10 in less than 1h sometimes so we might throw an node autoscaler that provision from cloud.
We recently just started the process of refreshing all the hardware. We looked at cloud again just to cover our bases. It is still cheaper to lease the hardware and pay for colo and a couple of people to manage the hardware, then to pay for the cloud.
[2]: https://engineering.skroutz.gr/blog/skroutz-infrastructure-2...
[3]: https://engineering.skroutz.gr/blog/uncovering-a-24-year-old...
Mostly on prem.
However they are moving workloads to cloud these days.