Separate layers before sizing
Do not size “an AI VPS.” Separate orchestration (n8n), agent/UI, model weights+KV, and accelerator/external API.
| Architecture | Fits when | Trap |
|---|---|---|
| Hosted-model agent on small Type B | API spend + small always-on host | Credential/egress risk |
| Local small CPU Ollama | RAM/disk for weights+KV | Slow CPU; no tokens/s claims |
| OpenClaw + external LLM | Persistent agent + API keys | Connected-account privilege |
| OpenClaw + local Ollama | Higher RAM hypothesis | Still not an evidenced GPU path |
| Needs VRAM / sustained GPU | Other cluster or external inference | Inventing a Kamatera GPU SKU |
Kamatera hosting is not “local-only privacy” if prompts hit third-party APIs.
Budget RAM and disk—not “7B VPS” labels
Inspect free RAM/swap/disk before model pulls. Worksheet: OS + runtime + download + resident weights + context/KV + concurrency reserve.
Vendor Ollama FAQ floors (about 8 GB for ~3B, 16 GB for ~7B, 32 GB for ~13B, plus disk examples) are attributed hypotheses—revalidate at deploy time. Editorial hypotheses: consider 8 GB+ for one tiny model/one request if resources fit; consider 16 GB+ for a modest quantized 7B-class experiment if weights/KV/OS fit. No tokens/s. Asia 1 TB traffic allotments matter for pulls.
One bounded Ollama reference
Conditions: accepted Setup host; measured RAM/disk.
- Prefer an AiXVPS-style
/opt/aixvps/ollamalayout with an intentionally pinned image after official-doc check—not a blindcurl | sh. - Default bind
127.0.0.1:11434or private network only; authenticated restricted remote access if required; test effective exposure. - Smoke a local API call with a pinned small model tag after license/disk/RAM checks.
- Cleanup named test artifacts only—no indiscriminate volume deletes.
# Illustrative discovery — adjust to your compose layout.
ss -tulpn | grep 11434 || true
curl -sS http://127.0.0.1:11434/api/tags | head
OpenClaw and similar agents
Marketplace/OpenClaw marketing size tables (for example 1 vCPU/2 GB “personal” up through 8 GB+ “local intelligence”) are vendor hypotheses. Before connecting email/calendar/chat: inspect data path, model providers, logs, and a kill/revoke plan. Least privilege; human approval for shell/web actions; secrets outside git; external API spend caps separate from the VPS bill. Hourly billing helps early tests—calendar terminate for abandoned agents.
No evidenced GPU SKU—say so and exit
This research pass did not successfully capture a public Cloud GPU product page. Do not invent GPU prices or SLA. When VRAM or sustained inference is required, use another aixvps cluster with evidenced GPU paths (for example the Vultr AI apps thesis) or external inference—without rankings. AUP: no mining/abuse workarounds.
Observe, stop, resize, split, or leave
| Symptom | Branch | Do not |
|---|---|---|
| OOM / persistent swap | Smaller model, lower context, more RAM, or external API | Ignoring swap as fine |
| Disk full from pulls | Delete named test weights; raise disk; watch Asia traffic | Blind rm -rf on volumes |
| API quota / credential risk | Throttle; rotate keys; disconnect agents | Leaving connected accounts after terminate |
| Need VRAM | GPU-elsewhere / external inference | Inventing Kamatera GPU SKUs |
Link Billing before destroy; link n8n for orchestration without publishing inference ports.