Claude as AI Cloud Infrastructure: Optimizing Local Context Overhead

When someone thinks about "running AI on our own infrastructure," the almost automatic reflex is to start calculating how much RAM and GPU to buy. Working on the infrastructure behind guatemalia.com, we reached a different conclusion — and it's worth sharing.

The finding

We run Claude Code directly on our VPS, inside a persistent session. We connect over SSH, work, close the connection, and when we reconnect from a different device the conversation picks up exactly where we left it — full context intact, nothing to rebuild from scratch. That alone is valuable. But what really changed how we think about infrastructure was noticing where the heavy lifting actually happens.

How the workload actually splits

Our server only needs to hold the conversation's context: text, file references, command output. That's compact data — it doesn't require a GPU or large amounts of RAM. The real model processing, the part that consumes serious compute, runs on Anthropic's infrastructure, covered by the subscription we already pay for ($20/month, the Pro plan). In practice, we use Claude's cloud as if it were our own inference cluster, without having to provision it, scale it, or maintain it ourselves.

Why this changes how we plan hardware

The benefit is twofold. First, there's no need to overprovision the server "just in case we use AI more intensively" — that part of the work never lives on our hardware at all. Second, it removes the classic operational friction of AI-assisted development: open a terminal, paste code, run it, review it, repeat. With a persistent session reachable from any client, that cycle gets shorter — there's no need to "reopen context" every time we switch devices or workspaces.

The question that reshapes infrastructure planning

The right question stops being "how much hardware do we need to run AI on our own infrastructure?" and becomes "which part of the work actually needs to live on our server, and which part is already solved by a cloud we're already paying for?" In our case, the answer was clear: almost all of the weight can be delegated. That frees up more local resources for what actually matters — serving our users and customers.