VPS, GPU cloud, serverless, or sandbox: choosing compute for AI work

Short versionDecision ledger

Rent	When	Example
A VPS	You need a small always-on box with SSH: agents that poll, schedulers, webhooks, small APIs.	DigitalOcean
GPU cloud	You need CUDA hardware for training or steady inference.	RunPod
Serverless	Your load is bursty and you want to pay only while code runs.	Modal
Agent sandbox	The code being executed is untrusted, for example code an LLM wrote.	E2B

Most real setups combine two of these.Not a benchmark

This page is not a benchmark. I have not load-tested these products against each other, and there are no performance numbers here. It is a decision guide based on how each category works and what the official pages say. Category prices below are described in orders of magnitude only; exact rates live on the linked official pricing pages and change often.

The four categories at a glance

Billing-granularity claims in this table come from each vendor’s official pricing or billing docs, checked on the date shown at the top. The rest of the table is my qualitative read of each category, not vendor wording.

Choose by workload

Most people who land on the serverless row want to start without paying, and those free tiers split three ways. Quota tiers, where running out pauses something instead of billing you: Cloudflare Workers for edge requests, Render for a container that spins down when no traffic arrives, Netlify and Vercel for front ends and their functions, and Supabase when the missing piece is a hosted Postgres backend rather than compute. Trials that start without a card: Fly.io, which ends at whichever comes first of a machine-runtime cap or a day count, and Railway, whose trial credit expires into a much smaller monthly allowance. Free plans that want a card first: Koyeb, whose signup defaults to its paid Pro plan until you downgrade, Deno Deploy, where linking a card unlocks the full free limits rather than ending them, and Northflank, which requires a payment method before you can create resources at all. The card mechanics behind that last group are cataloged in credit card traps.

Two adjacent options worth knowing

OpenRouter if what you actually need is model access, not compute. It is a unified API for hundreds of models. A free plan with free models exists (their FAQ says 50 free-model requests per day, or 1000 per day after buying at least $10 of credits). No dollar signup credit is published, only a “small free allowance” with no stated amount, so ignore any post that names one.

Google Cloud if you want a full cloud around your compute. New customers get a $300 / 90 days Welcome credit (their terms say 3 months). A credit card or other payment method is required at signup; Google places a temporary authorization hold of at most $1, not a charge, per their docs. Amount, expiry, eligibility, and the payment-hold nuance are broken down on the dated page: Google Cloud $300 credit.

How to actually decide

Start from the failure mode you cannot accept. If downtime is unacceptable, a VPS or managed serverless beats a spot GPU pod. If a security escape is unacceptable, a purpose-built sandbox beats a hand-rolled container on a VPS. If surprise bills are unacceptable, per-second billing with an idle floor of zero (serverless, sandboxes) beats a monthly box you forget about.

Then look at grants before paying: the credits index tracks what each provider officially gives away, with the date I last checked each number. Comparisons between specific pairs of providers are collected under /compare/, starting with RunPod vs Modal and Contabo vs DigitalOcean.