How much VPS power for an AI agent: the crucial distinction between API or local model

Anyone deploying an AI agent on a VPS encounters a question straight away that precedes any hardware choice: does the agent call an external API (OpenAI, Claude, Gemini) or does it run the model locally? A guide published on 29 July on Contabo's blog attempts to put concrete numbers behind this distinction, with a sizing table for those needing to dimension their server before ordering, not after.

API-driven agents: lean configuration

When an agent merely orchestrates calls to an external provider, the VPS runs only a coordination process: a PostgreSQL or SQLite database, a web dashboard, perhaps a Docker container or two for sandboxes. The heavy lifting—actual inference—stays on the API provider's GPU cluster, not the customer's server.

For Paperclip, a multi-agent orchestration platform, official documentation specifies a bare minimum of 1 vCPU and 1–2 GB RAM, but the realistic threshold for production use with roughly 20 concurrent agents, PostgreSQL database and active dashboard climbs to 4 vCPU and 8 GB RAM, with 80 GB NVMe storage. Hermes Agent, designed for a single agent, has an even lower minimum: 1 vCPU and 1 GB RAM if all inference is routed through external APIs. The recommended offering for this tier is Cloud VPS 4, which according to Contabo's official price list provides 4 vCPU, 8 GB RAM and 100 GB SSD at €5.50 per month including VAT (price for the first 24 months).

Browser automation and RAG raise the bar

Agents that autonomously browse the web—lead research, scraping, form filling via Playwright or Selenium—change the picture: each headless Chromium instance consumes 1–2 GB RAM by itself. The practical minimum for a stack with browser automation therefore rises to 16 GB, the recommended product being Cloud VPS 6.

The same applies to RAG pipelines with vector databases (Qdrant, ChromaDB, Weaviate): a mid-sized collection, between 100,000 and one million vectors, needs 2–4 GB additional RAM and 50–100 GB extra storage for index and embeddings. A complete RAG stack—agents plus vector database plus embedding pipeline—runs comfortably on 16 GB RAM and 200 GB NVMe.

Local inference: you'll need a GPU

Those choosing to run the model locally—typically via Ollama, vLLM or llama.cpp—do so for compliance or data privacy reasons, or because API token costs become significant at scale. In this case the VPS becomes the inference server and required RAM jumps from a handful of gigabytes to 16–64 GB depending on model size.

Model size (Q4 quantisation)	VRAM required	Reference GPU
7B	8–12 GB	RTX 5080 (16 GB) or higher
13B	16–20 GB	RTX 5080 tight, RTX 5090 (32 GB) more comfortable
34B	24–32 GB	RTX 5090 (32 GB)
70B+	48+ GB	RTX PRO 6000 Blackwell (96 GB) or multi-GPU

For a 7B model via Ollama alongside agents, the guide recommends 8 vCPU, 32 GB RAM and 200 GB NVMe, with optional GPU; for 13B and larger you need 8+ vCPU, 64+ GB RAM, 500+ GB NVMe and a dedicated GPU with 16 to 48 GB VRAM, typically on Cloud VDS or GPU VPS.

One detail for those working with GDPR-bound European clients: hosting an API-driven agent on a Contabo VPS in the EU Lauterbourg hub keeps the orchestration layer, agent state and customer data on German infrastructure, even though model calls go to Anthropic or OpenAI—a data-residency position requiring neither local inference nor a 32 GB server.

Paperclip and Hermes Agent are 1-Click applications available on Contabo VPS and VDS; Cloud VPS, Cloud VDS and GPU VPS are Prime Software products.