Table of Contents
On-premise means the servers sit in your own facility and your team or contractor runs them. A private cloud is cloud infrastructure reserved for one organisation, on or off your premises. A sovereign cloud is a provider's service designed to keep data, operations and legal control within one jurisdiction. The three answer different questions: where, for whom, under whose law.
Key takeaways
- The standard NIST definition says a private cloud may be operated by the organisation or a third party, and may exist on or off premises (NIST SP 800-145). "Private" describes exclusive use, not location.
- "Sovereign cloud" has no single technical standard. Providers define it themselves, usually around data access, operations and freedom to move workloads (Google Cloud).
- A local data centre does not settle the legal question: US law can require US providers to disclose data in their custody wherever it is stored (18 U.S.C. 2713).
- On-premise hardware can still depend on a vendor's cloud. Azure Local must normally sync with Azure at least once every 30 days (Microsoft).
- In Algeria, a 2017 ARPT decision requires cloud hosting providers to keep customer data on national territory, according to the CMS guide.
Why do these terms get confused?
Because they describe different things, and vendors combine them. "On-premise" is about where the hardware sits. "Private" is about who shares the infrastructure. "Sovereign" is about which jurisdiction and which people control it.
Product names add to the confusion. Microsoft markets a "Sovereign Private Cloud" that runs on Azure Local, its infrastructure software installed on validated hardware in the customer's own facilities (Microsoft): all three ideas in one product.
For an industrial company deciding where to run an AI assistant or a model trained on process data, the label matters less than four facts: where the data is processed, who operates the system, which law applies to that operator, and what the system needs from outside to keep running. Our guide to sovereign, on-premise AI breaks sovereignty into data, model, infrastructure and operations. This article focuses on the vocabulary.
What does on-premise mean?
On-premise (or on-premises) means the servers are installed in a facility you control: your data centre, a plant server room or a site cabin. Your IT team or a contracted integrator runs them.
Microsoft's shared responsibility model puts it simply: "In an on-premises datacenter, you own the whole stack" (Microsoft). That includes the physical building, network, hosts, operating system, applications, identities and data. In its comparison table, the customer is responsible for every row in the on-premises column.
That is both the strength and the cost: you decide who has access and when to patch, but you also do the patching, buy the hardware upfront and plan power and cooling. For an AI system, that means GPU servers sized to your models and users; see GPU sizing for an on-premise LLM.
On-premise does not have to mean disconnected. An on-premise server can receive updates through a controlled internet connection. An air-gapped system is a stricter case with no network path to the outside at all, covered in our air-gapped AI guide.
What is a private cloud?
The reference definition comes from NIST Special Publication 800-145, published in 2011. A private cloud is infrastructure "provisioned for exclusive use by a single organization". It "may be owned, managed, and operated by the organization, a third party, or some combination of them, and it may exist on or off premises" (NIST SP 800-145).
What makes it a cloud, rather than a rack of servers, are the five characteristics NIST lists: on-demand self-service, broad network access, resource pooling, rapid elasticity and measured service. Users provision capacity through a portal or API, not a ticket.
NIST's companion guide splits private clouds into two variants with different security implications (NIST SP 800-146):
- On-site private cloud. The cloud runs inside your security perimeter. NIST notes that you can protect it "to the same level of security as can be achieved for non-cloud resources", up to physical isolation. The capacity is fixed by what you bought.
- Outsourced private cloud. A provider hosts it. There are two security perimeters, yours and the provider's, joined by a protected link, and security depends on both and on the link. NIST adds that if the provider separates you from other customers only with the same mechanisms used in a public cloud, "this scenario would essentially become the public cloud scenario."
That last point is the one to check in any "private cloud" offer: dedicated hardware and network segments, or virtual machines on shared hosts with a different name?
What is a sovereign cloud?
There is no single technical standard behind the term. Each provider defines it, usually in terms of control rather than location.
Google Cloud described three pillars in 2020: data sovereignty, a mechanism "to prevent the provider from accessing their data"; operational sovereignty, assurance "that the people working at a cloud provider cannot compromise customer workloads"; and software sovereignty, the ability to run workloads "without being dependent on or locked-in to a single cloud provider" (Google Cloud).
Offers built on this idea vary widely:
- A separate region run by local staff. AWS describes its European Sovereign Cloud as "physically and logically separate from other AWS Regions", with its first region in Brandenburg, Germany. It says only EU-resident personnel located in the EU control day-to-day operations, under a new parent company and three German subsidiaries (AWS).
- Provider hardware in your building. Google Distributed Cloud air-gapped delivers integrated racks to the customer's premises and "does not require connectivity to Google Cloud" (Google Cloud).
- A national provider. A cloud operated by a company under national law, in national data centres. This is the meaning most Algerian buyers have in mind.
The legal test is the operator's jurisdiction, not only the data's location. US law requires US communications and remote computing providers to disclose data in their "possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States" (18 U.S.C. 2713). Whether a given corporate structure keeps an offer outside that reach is a question for your lawyers, not for a brochure.
How do the three models compare?
| Question | On-premise | Private cloud | Sovereign cloud |
|---|---|---|---|
| Where is the hardware? | Your facility | Your facility or the provider's | Usually the provider's data centres in the country or region; sometimes provider racks in your facility |
| Who operates it? | Your team or a contractor | You, a provider, or both | The provider, with staff restricted by residency or nationality |
| Shared with other customers? | No | No, by definition | Depends on the offer; many are multi-tenant |
| Which law governs the operator? | Yours | Depends on the provider | Designed to be local, but check the parent company |
| Upfront hardware cost | High | High on-site; low outsourced | Low, except for on-site racks |
| Capacity | Fixed by what you buy | Fixed, but larger if outsourced | Elastic, if GPUs are available |
| Needs an external network? | No, unless you choose updates over a link | On-site no; outsourced yes | Yes, unless delivered as on-site racks |
| Typical industrial use | Sensitive documents, systems near OT | Multi-site groups, shared internal platforms | Pilots, less sensitive workloads, burst capacity |
The rows that most often decide the choice are the operator and the network dependency. For costs over several years, see on-premise LLM vs cloud API: the real cost.
Where do hybrid offers fit?
Many products now put cloud software on your hardware. The question is what still runs outside.
Azure Local is a clear example. Microsoft describes it as infrastructure software that "runs on validated hardware in your own facilities", managed through Azure Arc. For its standard deployments, "Azure Local must sync successfully with Azure once per 30 consecutive days." If it does not, the cluster enters a reduced functionality mode: existing virtual machines keep running, but new ones cannot be created. Microsoft states that Azure Local itself does not store customer data in the cloud, only management data such as resource names, billing and usage. It also offers a separate disconnected operations mode that brings the control plane on-premises (Microsoft).
None of this is hidden, but it shows that hardware location and control plane location are separate questions.
What questions should you ask a provider?
Whatever the label, these questions show what you are buying:
- Where are the data, the backups and the logs stored and processed? Include prompts, answers and the vector index for an AI assistant.
- Which legal entity operates the service, and who owns it? Ask for the parent company and its country.
- Who can access the systems, from where, and with whose approval? Remote support access counts.
- Is the hardware dedicated to you, or shared with other customers?
- What does the system need from outside to keep running? Licence checks, control plane sync, telemetry, model downloads.
- What happens if the external connection is cut for a week or a month?
- Who holds the encryption keys?
- Can you export your data and models and run them elsewhere? This is Google's "software sovereignty" in practical terms.
- Is your data used to train the provider's models? See our list of questions to ask an AI vendor about data training.
Put the answers in the contract, not only in the sales deck. Our guide to choosing an industrial AI vendor covers the rest of the evaluation.
What does this mean in Algeria?
This section summarises published legal guides. It is not legal advice; confirm the details with counsel against the official texts.
According to the CMS expert guide, ARPT Decision 48/SP/PC/ARPT/17 of 29 November 2017 sets specifications for cloud hosting and storage services, and requires providers to "guarantee that customer data is hosted and stored on national territory".
Under Law 18-07, as amended by Law 25-11 of 24 July 2025, a controller may transfer personal data abroad only with the authorisation of the national authority (DLA Piper; CMS). Industrial data often contains personal data: names in maintenance logs, shift reports and HR files.
Policy points in the same direction. When the implementation of Algeria's national AI strategy was launched in August 2026, its stated priorities included "securing the processing and storage of administrative and public-sector data within protected national data centres" (Dzair Tube).
In practice, a foreign public cloud is the option that needs the most legal work. A national provider or your own servers need less.
What we have seen at KABAS AI
We use these terms carefully because our own products span them. DALILE runs completely offline on the client's own servers, with no internet access required. Virtual Lab offers four deployment options: on-premise, local PC with no internet connection, hybrid with local data storage and secure cloud AI processing, and hosted. At ADFC, deployment keeps full data control within Sonatrach's IT environment.
Three practical lessons follow:
- Classify the data before choosing the model. One company may need two answers: on-premise for formulations and drilling documents, something lighter for less sensitive work.
- Write down what crosses the boundary. For each option, list every external connection: updates, licences, support, telemetry. If the list is not empty, decide who approves each item.
- Ask about operations, not only location. Who patches the system and responds to incidents matters as much as where the rack stands.
If you are weighing these options for a specific use case, our consulting team can help classify the data and choose the deployment model. For whether to assemble the stack yourself, see build or buy an on-premise LLM.
Sources
- NIST SP 800-145, The NIST Definition of Cloud Computing, 2011
- NIST SP 800-146, Cloud Computing Synopsis and Recommendations, 2012
- Microsoft Learn: Shared responsibility in the cloud
- Microsoft Learn: Azure Local FAQ
- Microsoft Learn: What is Sovereign Private Cloud?
- Google Cloud: Engaging in a European dialogue on customer controls and open cloud solutions, 2020
- Google Cloud: About Google Distributed Cloud air-gapped
- AWS European Sovereign Cloud
- 18 U.S.C. 2713, Cornell Legal Information Institute
- CMS Expert Guide to data protection and cyber security laws: Algeria
- DLA Piper Data Protection Laws of the World: Algeria
- Dzair Tube: Kamel Baddari launches implementation of Algeria's national AI strategy, August 2026
Frequently asked questions
Is a private cloud always on-premise?
No. Under the NIST definition, a private cloud is reserved for a single organisation, but it may be owned and operated by that organisation, by a third party or by both, and it may sit on or off the organisation's premises. A private cloud hosted by a provider is private in the sense of exclusive use, not in the sense of location.
Is a sovereign cloud the same as a data centre in my country?
Not necessarily. Location is only one part of sovereignty. You also need to know which company operates the service, which country's law applies to that company, and who can access the systems day to day. US law, for example, can require US providers to disclose data in their custody even when it is stored outside the United States.
Is on-premise automatically sovereign?
No. The hardware can sit in your building while the management plane, licences or support access depend on the vendor. Microsoft's Azure Local, for example, runs on your own hardware but must normally sync with Azure at least once every 30 days, unless you use its separate disconnected operations mode. Check what the system needs from outside before calling it sovereign.
Does Algerian law require a sovereign cloud?
We are not aware of a general rule that every system must use one. According to the CMS legal guide, a 2017 ARPT decision requires cloud hosting providers to keep customer data on national territory, and Law 18-07, as amended by Law 25-11, makes transfers of personal data abroad subject to authorisation by the national authority. Confirm your own case with legal counsel.
Which model is best for an industrial AI assistant?
It depends on the data. Documents that must not leave the site, or systems close to operational technology, point to on-premise or air-gapped servers. Less sensitive workloads, pilots and burst capacity can run in a private or sovereign cloud if the provider's operator, jurisdiction and access rules match your requirements.



