Cloud vs on premise is a decision to make one system at a time. A small business almost never has a cloud estate or an on-premise estate, it has a billing counter, a payroll run, a file share, a website and a backup, and those five have different answers. Three questions sort them faster than any comparison table: what happens to this system when the connection drops, which security work stays with you whatever you sign, and what it would cost to move it back onto your own premises. The rest of this post works through those three and ends with a rule you can apply to your own list.
Start with what must survive a lost connection
Sort every system into two piles. In the first, the ones where an outage means somebody waits. In the second, the ones where an outage means somebody is turned away.
A payroll run that is late by two hours costs nothing. An analytics dashboard that is down for a morning costs nothing. A billing counter that cannot issue a receipt sends a customer home. A ward handover that cannot show who was admitted last night makes a nurse work from memory. A school attendance register that will not open at 9am is a teacher with thirty children and a paper list.
Everything in the second pile needs to keep working on the premises, whatever you choose for the first. That is a design decision about where the data lives and how writes are reconciled afterwards, and it has to be made early. We have written about why offline behaviour belongs in the data model rather than in a fallback screen, and the short version is that a system which treats the network as optional looks very different from one that discovers the problem on the day.
This is also the question most comparison articles skip, because connectivity is assumed. If your premises lose power or the line goes down more than once a quarter, that assumption is doing a lot of work, and it is worth testing against your own experience rather than a vendor's uptime figure.
The provider takes the machines, not the responsibility
The most common misreading of cloud is that it moves work to somebody else. It moves some work. The UK National Cyber Security Centre sets out which parts in its guidance on the shared responsibility model, and three things stay with the customer at every level of service: making sure the service meets your security needs, configuring it securely, and deciding which data goes into it.
Those three are not small. Most publicised cloud incidents come from a bucket left open or an account with more access than it needed, which is configuration, which is yours.
The amount you keep beyond those three depends on what you buy. With infrastructure as a service the provider gives you the raw resources and you retain much of the operating system and application security, which is why the NCSC suggests replacing components with managed services where you can. With platform as a service you supply the application and the service does the rest. With software as a service you carry the least, and the NCSC is direct that this model "allows you to delegate the most responsibility to your cloud provider". Its blunt advice is to "delegate as much responsibility for security to your hyperscale cloud platform as you can".
For a small business this changes the question. A ten-person company buying infrastructure as a service has quietly hired itself a part-time systems administrator it has not budgeted for. The same company buying the equivalent software as a service has not. If nobody in the building patches servers today, choosing a model that requires patching is choosing an outcome.
There is a second clause worth reading. If a managed service provider administers the service for you, the NCSC treats that provider as a third party with privileged access to your systems, and expects least privilege to apply to its accounts as much as to your own.
Hybrid is a named model, not a failure to decide
Teams who split their systems sometimes describe it apologetically, as though the real answer were one or the other. The standard disagrees. The NIST definition of cloud computing sets out five essential characteristics, three service models and four deployment models, and hybrid is one of the four, alongside private, community and public. A mixed estate is an ordinary outcome, not a hedge.
What makes a hybrid estate go wrong is rarely the split. It is the part nobody owns: the copy of the data that now exists in two places, and the rule for which one wins when they disagree. Decide that rule per system, write it down, and the split stops being a liability. Leave it undecided, and you have two sources of truth and a reconciliation argument every month.
Count the costs that do not appear in the comparison
Microsoft's cloud adoption guidance lists the honest reasons organisations move, among them datacentre modernisation, described as reducing hardware and staffing, and cost flexibility, described as using a pay-as-you-go model to run short-term projects without owning anything once they are shut down.
Both are real, and both point at the same thing. Cloud is good value where demand is uneven or temporary. A month-end reporting job, a seasonal spike, a pilot that might be cancelled, a workload that runs for six hours a week: renting these is almost always cheaper than owning hardware that idles the rest of the time.
The inverse holds too. A workload with flat, predictable demand running every hour of every day is the case where owned hardware competes well, because you are paying for utilisation you actually achieve.
Four costs rarely make it into the table, and all four fall on small teams hardest:
- The person who patches. On-premise requires someone to apply updates, test restores and replace a failed disk. If that person does not exist, the saving is theoretical.
- The cost of leaving, in both directions. Moving data out of a provider, or back out of your own racks, is a project with a timeline. Ask what it would take before you commit, not when you are unhappy.
- The second environment. Development and testing are easy to rent and awkward to own.
- The audit. Where regulated data lives is a question you will have to answer in writing eventually. Answering it is cheaper when the answer was designed rather than inherited.
A rule you can apply to your own list this week
Write down every system the business depends on. For each one, answer four questions in order, and stop at the first that decides it.
- If the connection drops for an hour, does somebody get turned away? If yes, it runs on the premises, or it runs locally and syncs. This outranks cost.
- Does it hold data whose location you will have to justify to a regulator or a parent or a patient? If yes, decide the location deliberately and record the decision.
- Is demand uneven, seasonal or temporary? If yes, rent it.
- Is demand flat and constant, and do you already employ somebody who maintains servers competently? If both, owning is defensible. If you had to think about the second half, rent it.
Most small businesses who run this exercise end up with three or four systems on the premises and everything else rented, which is the hybrid model and perfectly ordinary. The value is not in the ratio. It is in having a reason for each line, so that the next outage or audit meets a decision somebody made on purpose.
The same discipline applies when the system in question is an AI feature rather than a database, and we have written separately about what safeguards should control an AI agent in production.
If you are weighing this for a school, a hospital or an institution where an outage reaches people rather than dashboards, tell us about the project.
