The honest answer is that it depends on the shape of your load over time, not on which technology is newer. Here is how we actually talk it through with customers.
Look at the shape of your load
Plot your CPU and memory use over a representative month. If the line is roughly flat, you are paying an elasticity premium for elasticity you never use, and dedicated hardware will be cheaper for the same performance. If the line has a 5× daily peak, fixed capacity means buying the peak and idling through the trough.
Dedicated wins on sustained throughput
Databases, video transcoding, CI runners and anything disk-bound benefit from not sharing. There is no noisy neighbour, no steal time, and NVMe attached directly to the board behaves very differently from network storage under sustained write load.
Cloud wins on variance and on lifespan
Short-lived environments, unpredictable growth, staging that only exists during working hours, and anything you might want to delete next week. Paying by the hour for something that lives for six hours is unambiguously correct.
Count the operational overhead honestly
A dedicated machine is a machine. Someone has to patch it, monitor its disks and decide what happens when a PSU dies. If your team has no appetite for that, the cheaper line item is not the cheaper option.
Failure behaviour differs, not reliability
A cloud instance can be rebuilt elsewhere in minutes but shares a fate domain with its host. A dedicated machine fails less often and takes longer to replace when it does. Neither is more reliable in the abstract — decide which failure mode your architecture already handles.
The answer is often both
Dedicated hardware for the database and the steady-state workers, cloud instances for the burst tier and the environments that come and go. Running them on the same network in the same facility means you get that split without paying for traffic between the halves.
Runs the edge and the maintenance calendar in HOU1. Writes the postmortems, including the ones that are not flattering.