The cheap server that needs an expensive engineer.
When someone sells you a €3,000 hardware local SLM, the real annual cost starts elsewhere. Here is the full list of lines that get forgotten at budgeting time.
Hardware is the visible line. The other five decide whether the project pays off.
- 01
Hardware (GPU + server + storage)
The visible number. A server with a GPU able to run a useful SLM costs €2,500 to €15,000. Amortized over 3-5 years.
- 02
MLOps and deployment
Configure runtime, weights, quantization, inference server, observability. It is a technical role, not a tutorial fix.
- 03
Dedicated or subcontracted staff
An MLOps engineer costs €55,000 to €80,000 gross per year in Spain. Outsourcing runs €60 to €120 per hour. Without this role the system degrades.
- 04
Model updates
New models ship every few months. Migrating weights, revalidating prompts and adjusting to a new version is recurring work, not a one-off.
- 05
Security and compliance
Network isolation, encrypted backups, logging, audit, patches, hardening. What you already do for any corporate server, now also here.
- 06
Continuity and replacement
When the MLOps person leaves or hardware breaks. A documented plan must exist before it happens.
The total annual cost of running a local SLM in production usually lands at 3-5 times the initial hardware price, depending on in-house vs subcontracted operations. That factor rarely appears in the vendor budget you receive.
Local SLM yes, when the real saving beats the hidden lines
In SVP the system is designed to run in dedicated cloud or the client's local server. The decision is not dogmatic: it comes from the real total-cost math and your operation's privacy requirements.