With language models, graphics memory matters most: the model has to fit into it. Larger models often give better answers but need more memory and computing power. Add system memory, fast storage and stable cooling and power.
The three routes compared
Existing hardware: low initial cost, limited model size, good for pilots.
New hardware: full control and predictable running costs, but investment, space, power and maintenance.
Hetzner GPU cloud servers: flexible booking, no purchase, ongoing rent; check availability, location and contract terms with the provider beforehand.
Decision aid
Test or pilot only: existing hardware or short rental.
Permanent use with sensitive data: consider your own hardware.
Fluctuating load or quick start: rent a GPU server.
Always: secure access, plan updates, watch cost and utilisation.
Important before buying
First measure in a pilot which model size your tasks really need. This prevents oversizing. Hardware prices and delivery times change, so we deliberately give no figures here.
GPU server AIHetzner GPUhardware for local AIgraphics card language modelsize AI serverOllama hardware
Frequently asked questions
How much graphics memory do I need?
That depends on the model. In a pilot we determine what size your tasks need and size accordingly.
Is a rented GPU server compliant with data protection?
That depends on provider, location and contract. Check data processing agreement and location. We support the check points but do not give legal advice.
Is your own hardware worth it?
With permanent, intensive use often yes, with occasional use rather not. A pilot provides the basis for the decision.