For most workloads a hosted model from a major provider, properly contracted and configured, is the right answer. Self-hosting earns its place where the sensitivity of the data means the processing itself cannot leave your control, and it costs considerably more in operational effort than the licence comparison suggests.

That is the whole decision, and it turns on data classification rather than on architecture preference. The question is not which approach is better. It is which of your workloads genuinely require the harder one.

Should we self-host an AI model or use a hosted service?

Self-host when the data cannot lawfully or safely be processed by a third party, when you need the processing to stay under a single jurisdiction with no external provider able to receive a disclosure order, or when volumes are high enough and workloads stable enough that the economics change. Use a hosted service for everything else.

The mistake is treating this as one decision for the whole organisation. Large organisations generally end up running both, and the useful work is deciding which workload sits where.

Classify first, then choose

The decision needs three inputs and they are all knowable before anybody looks at a model.

How sensitive is the data, judged by what happens if it is disclosed rather than by how it feels. How complex is the task, because a task that a small model handles well opens options that a task needing frontier capability does not. And what is the volume and its shape, because steady high volume and spiky low volume have different answers.

Run those three against each other and most workloads sort themselves. A public-facing summarisation feature on published content is a hosted service. A tool that reads case files containing personal data about vulnerable people is not, or not without a great deal more work. The awkward cases are in the middle, and the middle is where the hybrid pattern earns its keep.

What does self-hosting an AI model actually require?

Hardware with enough memory to hold the model, a serving stack, and somebody whose job includes keeping both current. The licence saving is real and it is not the main cost.

The commitments that are absent from most business cases:

Capacity has a lead time. Accelerator hardware is not something you provision on a Tuesday, and capacity planning has to cover peak rather than average, because a queue at the busiest hour is what users will remember.

Models are updated, and updating yours is your job. A hosted provider improves the model underneath you. Self-hosted, you decide when to move, you carry the migration, and you re-run your evaluations afterwards because behaviour changes.

The serving stack needs patching like any other production software, and it moves quickly. It also needs everything a production service needs and rarely gets on a first deployment: intrusion detection, availability and resource monitoring, log aggregation, configuration management, and off-site encrypted backups. We run that layer on our own platform using Snort, Zenoss, Logstash and Puppet, and the point is not the specific tools. It is that a model endpoint is a production service, and the list does not get shorter because there is a model behind it.

Evaluation is continuous rather than one-off. Any change to the model, the serving stack or the prompt scaffolding can shift output quality in ways that are invisible without a test set built from your own data.

And somebody is on call for it. That is the commitment organisations underestimate most, because a hosted API's availability is somebody else's problem and a self-hosted endpoint's availability is yours.

None of that is an argument against self-hosting. It is the list to cost before deciding, because a business case built on licence savings alone will be wrong.

Is a self-hosted model cheaper than an API?

Not at low or moderate volume, and often yes at high steady volume. Hosted services charge per unit of use and self-hosting charges for capacity whether or not you use it, so the crossover depends on how much of the capacity you actually use rather than on published rates.

The practical implication is that spiky workloads suit hosted services and steady ones suit owned capacity. An organisation with a research tool used by forty people during office hours is buying idle hardware. An organisation processing a continuous document queue is buying a discount.

Model the crossover with your own volumes before treating cost as the reason for either choice. In most of the cases we see, cost is not the deciding factor. Data sensitivity is, and the cost analysis either confirms the decision or tells you what it will cost to do the right thing.

Can a self-hosted model match a hosted one on quality?

For a well-defined task, usually yes. For open-ended work at the frontier of what these systems can do, usually not.

Open-weight models have closed much of the gap on the tasks large organisations actually deploy, and they remain behind on the newest capabilities, which appear in proprietary models first.

Test it rather than accepting either claim, using an evaluation set built from your own material rather than a published benchmark. A few days of that work replaces an argument nobody can win with a result nobody can dispute.

The hybrid pattern, which is where most organisations land

Sensitive workloads run on infrastructure you control. Everything else calls a hosted service. A routing layer decides which, based on the data classification rather than on the user's choice.

The important design decision is to keep the model swappable. Put an interface between your application and whichever model serves the request, so that changing provider, or moving a workload from hosted to self-hosted, is a configuration change rather than a rebuild. Organisations that wired a specific provider's API directly into their application in 2023 have found out what that costs.

The second design decision is to make the routing rule explicit and auditable. A rule that says which classes of data may leave, enforced in code, is worth more than a policy saying the same thing, and it is the thing you show an auditor.

Where jurisdiction comes into it

For the most sensitive workloads, storage location is the smaller half of the question. The larger half is whose law reaches the organisation doing the processing. So a model operated by a company under United States jurisdiction carries that jurisdiction with it, whichever region processes the request.

Where that matters, self-hosting on infrastructure under UK jurisdiction is the only configuration that removes the exposure, because there is no external provider to receive a disclosure order.

This is the part of the market we were built for. Our private cloud is co-located in ex-military data centres rated to Impact Levels 3 and 4, and we are one of only ten providers on the government's Cloud Compute 2 framework accredited to offer Cloud Secure Plus hosting at List X and List N levels. A model deployed on that infrastructure inherits its jurisdiction and its clearance.

We run AI on that platform ourselves rather than only recommending it. GovCloud.hosting, our deployment platform for public sector organisations, uses AI for predictive scaling, automated security scanning and code quality checks, across AWS, Azure, our own secure cloud and on-premise targets. It is a working example of the hybrid pattern above: the AI runs where the workload sits, rather than the workload being moved to wherever the AI is convenient.

We also host on AWS and Azure and recommend hosted models where they fit, which is most of the time. The argument here is narrow and it is worth restating plainly: classify the data first, then choose, and expect the answer to be a hosted service for the majority of what you run.

Before your next AI deployment

List every AI workload you have or plan, and put a data classification against each one before anybody discusses models. Then sort them into three groups: clearly fine on a hosted service, clearly not, and genuinely uncertain.

The first group is the majority and it should proceed. The third group is where the work is, and the question that resolves most of it is what actually happens if this data is disclosed.

If you want a second opinion on where the line falls for your estate, we will give one, and we will tell you when the answer is a hosted service.

Frequently asked questions

Is a self-hosted AI model more secure than a hosted one? Not automatically. Self-hosting removes one exposure, which is a third party processing your data under another jurisdiction, and adds others, including patching, access control and availability, which become your responsibility. It is more controllable rather than inherently more secure.

Do open-weight models keep our data private? Only if you run them yourself. An open-weight model called through somebody else's API is a hosted service with the same data questions as any other. The licence describes the model, not where the processing happens.

How much hardware do we need to self-host a useful model? It depends entirely on the model size and concurrency, and any figure quoted without those is guesswork. The more useful planning step is to size against your peak concurrent requests and your latency tolerance, then test at that load before committing.