Help me choose AI hosting

Hosting means renting a computer in a data centre. You need it for an agent that answers messages day and night, a website or app that other people use, a model you want to run yourself, or a heavy job such as generating thousands of images.

You do not need it to try a model: most model providers have a free or pay-as-you-go playground, and a small open model can run on a recent laptop with Ollama. Start there if you are still exploring, and rent hosting when you know what should run all the time.

If your app already runs somewhere (a web host, a Vercel or Netlify project, a company server), adding AI usually means calling a model API from that app — the API is the new cost, not the hosting.

Model: a large file of learned numbers (weights) plus the software to run it. Open-weight models such as Llama or Qwen can be downloaded and run by anyone; closed models are only available through their provider's API.

API: a web address your code sends a question to and gets an answer from. You pay the API provider per use (usually per million tokens, roughly per 750,000 words) and do not need a GPU yourself.

Agent: a program that takes a task, asks a model what to do, calls tools (search, e-mail, a database) and repeats until the task is done. Hermes Agent, OpenClaw and n8n are examples. The agent itself needs little computing power; the model does the heavy thinking.

Inference: running a model to get an answer. Training and fine-tuning: changing a model's weights with your own examples, which needs far more GPU time than inference.

VPS (virtual private server): a slice of a server with its own operating system that you manage over SSH. GPU instance: the same, with one or more graphics cards for AI work. Serverless: you upload code or a model and pay only while a request runs; the provider starts and stops the machines.

A small VPS with 1–4 CPU cores and 2–8 GB of memory is enough for most agents, chatbots and automation tools that call an API; the agent-plan table below starts at 2 GB. It is billed per month and runs around the clock.

A GPU is usually needed when the model runs on your own hardware. Small quantised models can also run on a CPU with llama.cpp or Ollama, but they generate text much more slowly; we have not measured how much. On a GPU, what matters most is its memory (VRAM): the model's weights, plus working memory for the conversation, must fit. A model that does not fit will not run on the GPU, or runs far slower when part of it is moved to the CPU.

Rough guide: a model of 7–8 billion parameters needs a 24 GB GPU at 16-bit precision, or much less with 4-bit quantisation. A 70-billion model needs about 140 GB at 16-bit (two 80 GB GPUs) and still more than 40 GB at 4-bit. The VRAM tool calculates this for a specific model and context length.

For occasional GPU work, serverless GPU billing or an hourly instance you stop afterwards is usually cheaper than a monthly GPU server.

With an API, cost follows usage: a quiet month is cheap, a busy month is expensive. With your own GPU, cost follows time: the same hourly price whether it answers one question or a million.

Running your own model makes sense when you need a specific open model, want prompts and answers to stay on hardware you rent, need steady high volume, or want to fine-tune. It is rarely cheaper for light or irregular use.

You can mix both: an agent on a small VPS that calls an API today can switch to your own model later without changing hosting provider for the agent.

On a self-managed VPS or GPU instance you are responsible for operating-system updates, firewall rules, SSL certificates, monitoring and backups. Guides and one-click templates help, but problems are yours to fix.

Managed agent platforms and serverless services run the software for you: you configure the agent or upload the code, and the provider keeps it running. The trade-off is less control and, often, usage-based pricing that is harder to predict.

The advisor lists, for each option, what the provider does and what is left to you, so you can compare effort as well as price.

Monthly plan: a fixed price for a fixed server. Watch for introductory prices that renew higher and for long-term prepayments.

Hourly or per-second instance: you pay from start to stop, including idle time. A stopped instance may still cost money for its disk; a deleted one does not.

Serverless: you pay only while requests or jobs run, plus a minimum per request or container start. Cheap when idle, more expensive per hour when busy.

Per-token API: you pay per million input and output tokens. Output tokens usually cost more than input tokens.

Spot, preemptible or interruptible: spare capacity at a discount that the provider can take back with little warning. Fine for batch jobs that can restart; not for an always-on service.

Reserved or committed: a lower hourly rate in return for paying for months or years in advance. Only worth it when you know the load.

Hosting cost is what the provider charges for the server or GPU. Project cost is everything you need to run the project: hosting, API usage, domains, other software subscriptions and the hours you spend maintaining it.

Prices on AIHostCompare come from official provider pages on the date shown next to them, without VAT or sales tax. They are listed in US dollars, or in euros where the provider lists euros; you can show them in EUR or USD with the currency control at the top of the page. A converted price is marked “≈”: it uses the ECB reference rate of the date shown, and your bank's rate and fees will differ. Always check the price on the provider's own page before you order.

A price change a provider has announced applies from 00:00 UTC on its date, so the advice and the calculators use the price of the billing month you choose. When a provider switches to a price that changes all the time, there is no fixed price for those days: the advice does not suggest that price, and the cost calculator shows no monthly total for a month that includes them, only the known part before the change, rather than a guess.

The region of a server decides where data is stored and processed. Most providers let you choose a region per server; some serverless and API services choose for you.

When you call a model API, your prompts go to that API's provider. Read its data-usage terms: whether prompts are stored, for how long, and whether they are used for training.

Running a model yourself keeps prompts on the server you rent, but the hosting provider is still a processor of that data. Encryption with your own keys (BYOK) and a signed DPA are the things to check.

AIHostCompare records the regions and data terms that providers publish. It is not legal advice: for regulated data, ask your privacy officer or lawyer.