On-prem LLM: how many GPUs, and what will it cost
Enter your scenario to get VRAM requirements, viable GPU configurations, budget scale, and vLLM launch flags you can actually use. VRAM is derived from real formulas, not guesswork.
Results
Viable GPU configurations
vLLM launch flags
What to watch out for here
VRAM = weights + KV cache + 15% runtime overhead; typically within 5-10% of actual. GPU prices are indicative only and are not a quote.
Want us to actually deploy it?
Sizing is the easy part. The time goes into air-gapped deployment, model integration, workflow build and handover. We run Ollama / vLLM + RAG + multi-agent orchestration in production.