The same IT department can operate, on the same day, a payroll package on its own servers, a customer portal with a public cloud provider and an email service consumed as software as a service. The question “cloud or on-premise” therefore does not call for a single answer for the whole information system: it arises for each application, and the answer may differ from one to the next.
Five criteria decide between the options: the variability of the load, the application’s dependencies, the location constraints on its data, the skills available and the cost structure. One last point is settled before signing, whichever way the decision goes: how to exit. Between public cloud and on-site hosting, intermediate forms exist (private cloud, dedicated hosting with a provider, hybrid cloud), and they are judged on the same grid.
Cloud or on-premise, and the forms in between
The term on-premise, or “on-site”, refers in common usage to an infrastructure that the organisation operates itself on its own premises. Here, as a scoping choice, we also include infrastructure it installs in a data centre where it rents space. For the cloud, one public definition is that of NIST, in its Special Publication 800-145 of September 2011. It retains five essential characteristics, two of which weigh on the choice: rapid elasticity, meaning capabilities provisioned and released according to demand, and measured service, whose usage is metered.
The same publication distinguishes deployment models. The public cloud is open to all and exists on the provider’s premises. The private cloud is reserved for a single organisation; it may be owned and operated by it, by a third party or by both, and may exist “on or off premises”. The alternative thus has more than two boxes: a private cloud installed on the organisation’s premises counts as both cloud and on-premise.
Long-term coexistence of the two worlds is covered in our article on hybrid cloud migration, and the use of several providers in the one on multicloud strategy.
Load variability is measured before it is debated
On site, capacity is bought for several years and sized for the peak: what remains unused is paid for all the same. In the public cloud, it is adjusted and billed by usage. The benefit of this second model therefore depends on the load profile, which operating records make it possible to chart over a whole year, seasonal peaks included.
A highly variable, seasonal or still unknown load (a new service whose audience nobody can predict) takes advantage of elasticity. A stable and predictable load gains little from it: the application consumes the same capacity all year round.
The application must still be able to scale out and scale in. Designed for a single server, it does not become elastic by changing hosting provider. Transforming it is another workstream, which our page on cloud migration places among the possible strategies for each application.
Dependencies and data location point to groups of applications
An application is not decided on its own. One that, at every transaction, exchanges with a system that has stayed on site, or that responds to local equipment within a tight time frame, follows the fate of what it depends on. A software licence agreement can also tie its use to a hosting environment. Mapping these links reveals groups of applications that are classified together; the article on hybrid cloud migration, cited above, details how to read them.
Data location calls for a precaution. The requirement is first written in precise terms (where the data reside, which law the provider is subject to, which qualification the service must hold), then tested against each option. Depending on the requirement, an internal server room can meet it, but so can a service qualified under SecNumCloud, ANSSI’s qualification framework for cloud service providers. Our page on sovereign and hybrid cloud describes the data classification that precedes this choice, and our article sovereign cloud or hyperscaler compares the offerings on this ground.
Skills change in nature from one option to the other
Operating on site requires knowing how to buy, install and repair hardware, and how to run virtualisation, storage, networking and physical security, with an on-call rota. The public cloud removes these tasks and brings others: describing infrastructure as code, setting identities and permissions, monitoring spending that varies every day, knowing what the provider handles and what it leaves to its customer. A managed service moves the boundary again: the provider operates the database or the message queue, the team keeps the configuration and the data.
The question then concerns people. Which skills does the organisation have, which can it train or recruit, which is it willing to entrust to a service provider? A small team, at ease with a recent internal platform, does not have the same interest in moving as a team whose hardware specialists will not be replaced. In a migration, training the operations teams is part of the plan, and the foundation environment, or landing zone, is prepared before the first application.
Cost structures to compare over the application’s lifetime
The two options do not spend in the same way. On site, spending is an investment depreciated over several years, to which fixed costs are added: premises, energy, maintenance, renewal, operations staff. In the public cloud, it becomes a variable operating expense, incurred by each resource deployed. Neither model is cheaper in principle, and we give no order of magnitude here: the result depends on the load profile and the period chosen.
The comparison covers all items, over the same period. On the site side: those mentioned above. On the cloud side: resources sized according to measured usage rather than copied from the existing set-up, outbound data transfers, retention of logs and backups, licences, and the period of parallel running during cutover. After the move, monitoring spending becomes a continuous practice, described in our FinOps offering.
Some situations invite you to hold back. An application whose withdrawal or replacement is scheduled does not justify an intermediate move. A stable application, hosted on recent hardware, can wait for the renewal date: the decision then receives a date.
The exit is prepared in both directions
Choosing the public cloud entails a dependency to be understood: formats and volumes of the data to extract, provider-specific services with no equivalent elsewhere, exit clauses. Chapter VI of Regulation (EU) 2023/2854, known as the Data Act, deals with switching between data processing service providers, switching charges included. Our entry on the Data Act presents its articles and timetable.
Staying on site is also a commitment: a lease or a hosting contract, hardware to depreciate, maintenance and licence contracts, skills that may become scarce. These commitments have an end date, which sets the moment when the question will reopen.
For each application, the decision is written with three entries: the option chosen and the criterion that prevailed, what it would take to exit it, and the date on which the choice will be reviewed. The review can start with the applications for which a deadline is approaching (end of support for hardware, renewal of a hosting or licence contract): their classification cannot wait for the end of the inventory.