Glossary
Cloud AI
Cloud AI means the model runs on a provider's servers and reaches you over the internet, which is how nearly every widely used assistant works and is the arrangement people mean when they say nothing at all.
In plain terms
Your text travels to somebody else's computer, the answer travels back, and you never see the machine. It is the default, so nobody announces it. Most of the questions people ask about privacy, cost and reliability in this subject are really questions about this arrangement, asked without naming it.
Why it matters
It is worth naming precisely because it is invisible. Once you know it is the default, three things become questions you can ask rather than assumptions you are making. Where your material goes, what happens when the provider has an outage, and whether the model behind the product can change without anybody telling you. All three have answers, and none of them is obvious from using the tool.
How it works
You send a request over the internet, it is handled on hardware the provider owns and operates, and the answer comes back. Nothing is installed and nothing is stored on your side beyond whatever the application keeps. The whole arrangement is rented, including the part that would otherwise be a capital decision.
You get the largest and newest models this way and rarely any other. The most capable systems need more hardware than most organisations would buy, so the choice of running elsewhere is generally a choice of something smaller. That gap has narrowed and it has not closed.
Cost follows use rather than sitting still. Nothing is charged when nobody asks anything, which suits uneven demand well, and the bill rises with activity in a way subscription software does not. Forecasting is genuinely harder here and that is a property of the arrangement rather than a failure of anybody's spreadsheet.
The provider's availability becomes yours. An outage at their end stops your work, a change to their model changes your outputs, and a change to their pricing changes your budget. Those are ordinary supplier risks and they are worth writing down before something depends on them.
What you rent, and what you accept with it
Seen in the wild
Use any well-known assistant in a browser, which is this arrangement with none of it visible to you.
ChatGPTSend a request through a routing interface that reaches several providers' servers, where the arrangement is the same and the destination varies.
OpenRouterRun the same sort of thing on your own machine for an afternoon and notice what changes: the wait, the model size, and where your text went.
Ollama
Common misconceptions
People assume
Cloud means our data is stored there.
In fact
It means your request travels there to be handled. Whether anything is retained afterwards, for how long, and who may look at it are separate questions answered by the provider's terms rather than by the arrangement. Plenty of business agreements retain very little, and the only way to know is reading the plan you are on.
People assume
It is the expensive option.
In fact
It is the variable one. You pay nothing when idle and more when busy, which is cheaper than owned hardware for the intermittent use most organisations have and dearer for constant heavy load. Which is expensive depends entirely on your pattern of use rather than on the arrangement.
Telling them apart
Cloud AI vs Self-hosted
Cloud AI
The provider's servers. Their operational burden, the newest and largest models, and a bill that tracks your activity.
Machines your organisation controls. Your operational burden, a narrower choice of models, and a cost you pay whether busy or idle.
The deciding question is usually whether anything is contractually forbidden from leaving your boundary. Where nothing is, this becomes a comparison of cost and capability rather than a matter of principle.
Questions
- Is our material safe in this arrangement?
- That depends on the provider's terms and your plan rather than on the arrangement itself. The questions to answer are how long material is retained, who may access it and whether anything is trained on it, and business agreements frequently differ from consumer ones on all three within the same product.
- What happens when the provider has an outage?
- Your work stops, and you find out at the same time as everybody else. Anything that must keep running needs either a second provider you can switch to or a way of failing usefully rather than silently. This is a supplier-risk question of the sort your organisation already knows how to ask.
- Can they change the model without telling us?
- Frequently yes, and a product can move to a successor without changing its name. That is why an evaluation set you can rerun matters: it turns a silent change into something you notice. Ask a vendor whether you can be pinned to a version and how much notice you get.
Key takeaways
- This is the default arrangement, which is why nobody names it and everybody has questions about it.
- The largest and newest models are generally available only this way.
- Cost tracks activity, so nothing is charged when idle and forecasting is harder.
- The provider's outages, model changes and price changes all become yours.
- What happens to your material is a question about the plan you are on, not about the cloud.
Tools that use this
- ChatGPT
The arrangement in its most familiar and least visible form.
- OpenRouter
Several providers' servers behind one route.
- Ollama
The contrast case, for an afternoon, on your own machine.
Last checked July 2026