A personal AI stack on a 1.6GB VPS: operational numbers

#self-hosting#vps#costs

For a few months I have run a personal AI stack on a small VPS: 1 core, 1.6 GB RAM, 40 GB disk, Debian. This note records the measured numbers — what it runs, what it costs, and the operational rules that emerged from running it tight.

What runs on it

  • an agent gateway (the one serving this conversation) over Feishu and WeChat
  • SearXNG as the search backend
  • a Qdrant server for vector memory
  • a Node ingest server for memento
  • a proxy, because the network requires one

At idle the box has ~400 MB free. Everything above runs inside that budget.

The token bill

The LLM behind the agent is DeepSeek, billed per token. To keep the cost visible, a cron job queries the balance API every hour and appends the result to a JSONL ledger (capped at 720 entries). Measured over real weeks of use:

  • 0.033 CNY/hour average burn ≈ 0.8 CNY/day ≈ 24 CNY/month for all LLM usage (chat, cron-generated summaries, vision).
  • The scheduled jobs (nightly summary, hot-topic pushes, balance tracking) cost a few fen per day — below the noise floor.
  • Long sessions with large context show up as discrete spikes in the ledger, which is the point of the ledger: any session’s cost is inspectable afterwards.

Hosting costs a few yuan a month. Token costs about 24 CNY/month. Both are close to zero.

Memory behavior

The constraint that actually bites is RAM, and the failure mode is gradual: long conversations grow the gateway process. A session with 400+ messages took the gateway to ~700 MB resident, leaving ~150 MB free — enough that even spawning a Python subprocess felt risky. Three rules came out of this:

  1. Bound what stays in context. Memory pressure tracks conversation length, so context must be capped and compressed, not accumulated.
  2. Don’t spawn heavy helpers for light queries. A single curl against an API is cheaper than a Python process when you are counting megabytes.
  3. The gateway never runs its own admin CLI. A scheduled job that shelled out to the gateway’s CLI deadlocked on a lock the running gateway holds — the session froze mid-sentence. Config reads, log tails and one-shot scripts go through cron (a separate process) or a file write; never through the gateway process itself.

What the numbers say

The marginal cost of a personal AI stack is near zero — a few fen a day in tokens, a few yuan a month in hosting. The real cost is attention: the restart after a memory spike, the hour spent tracing a deadlock, the surprise that only surfaces because there are 400 MB of headroom and no margin to hide it.

Two practical consequences:

  • Track the ledger. The hourly balance log turns “is this expensive?” into a lookup. It costs nothing to run and has already caught a runaway session.
  • Small machines are honest. On a big box, problems hide in headroom. On a 1.6 GB box every inefficiency surfaces immediately, which makes it a good place to learn what your stack actually does.

What is still open: the memory growth curve of the gateway over very long-lived sessions (days, not hours), and whether the vector DB can share the box with a second model server if I want local embeddings instead of a remote API.