Renting a dedicated server long-term rewards loyalty badly. Most providers buy hardware, rack it, and recover the capital over 36 to 48 months. After that the box is fully depreciated, still billing at the original monthly rate, and costing you real performance against newer machines that often rent for less. The people who save money on dedicated hardware are the ones who audit what they rent against what the market currently charges, and refresh on a cycle instead of waiting for a failure.
A 2019-vintage Xeon Scalable and a current-generation EPYC are not in the same league. The newer chip typically delivers somewhere between two and three times the memory bandwidth, double the core count at similar clock speeds, and markedly better single-thread performance. If your workload is memory-bound, and databases usually are, the difference shows up directly in query latency. We routinely see MySQL and PostgreSQL workloads drop from tens of milliseconds to single digits purely from a generation jump, before anyone touches the schema.
That performance is money. A server that handles twice the requests per second halves your effective cost per request. If you are vertically scaling because the old box is saturated, you are paying more per unit of work every month the refresh is deferred.
Older nodes draw more power per unit of compute, and in UK datacentres power is a large share of the provider's cost per server. A provider running a hall full of 2018 hardware is spending more on electricity per customer than one running current-generation low-TDP parts. That cost is in your invoice. It is one reason newer generation servers at newer operators often undercut legacy fleets on price for the same or better specs.
Refresh discipline is not universal. Keep the old box when: the application is licensed per-socket and a new server would multiply licensing costs; the workload is latency-sensitive to a specific peering point and the box sits in exactly the right rack; or the application is genuinely idle and the monthly rate is already discounted to legacy levels. Plenty of internal tooling, staging environments and low-traffic services fit that description, and moving them buys nothing.
Once a year, do the following. Pull the last twelve months of bandwidth graphs and check your actual 95th percentile against what you pay for, since an over-provisioned commit is a discount you are handing back. Run a benchmark on your current hardware and compare against published figures for the generation the provider now offers new customers. Ask for a refresh quote: most UK providers will migrate you to a current-generation machine at the new-customer rate if you ask, because churn is more expensive for them than a discount. Finally, check whether your storage is still spinning rust; a move to NVMe alone can retire a class of performance problems.
Do the migration properly. Provision the new server in parallel, sync data over the network while the old box serves traffic, cut over with a DNS TTL lowered in advance, and keep the old server online for a week as a rollback path. On our own fleet this is a routine, rehearsed operation, and it should be with any provider you rent from.
The real reason aging servers stay rented is that nobody owns the decision. The box works, the invoice is auto-paid, and migration is someone's weekend. But deferring a refresh that would pay for itself in four months of billing difference is not stability, it is an unmanaged subscription creep. Put the hardware audit on a calendar the same way you would a domain renewal, and treat every server past its fourth birthday as a candidate for repricing.