Hosting options
Where an AI solution runs decides who can see the data, what it costs as usage grows, who fixes it at night and which skills the operator must hold. The hosting choice is a business and regulatory decision as much as a technical one.
The options
- SaaS AI
- Public-cloud AI services
- Sovereign hosting
- Hybrid
- Private cloud
- Self-hosted models
- On-premises
SaaS AI
Inside an application you already license
AI features in a CRM or ITSM product. Fastest to adopt; least control over behaviour and data path.
Public-cloud AI services
Managed services from a hyperscaler
Models and AI tooling consumed as managed services, inside the operator’s cloud tenancy and contracts.
Sovereign hosting
In-jurisdiction by design
Cloud services operated within a jurisdiction, under local legal control and staffing, for national security or residency obligations.
Hybrid
A deliberate split
Sensitive data and retrieval held privately, some model calls to public services, with the boundary enforced centrally.
Private cloud
Your platform, often your data centres
AI workloads on the operator’s own cloud platform, with its own model hosting and retrieval.
Self-hosted models
Open-weight models you run
On any of the above. Full control of version and data; full responsibility for serving, patching and evaluating them.
On-premises
Dedicated hardware in your estate
Typically for network-adjacent or highly sensitive data. Predictable, but hard to scale up quickly.
Control and run burden are relative to each other, not measured; the trade-offs below say why.
Trade-offs
| Option | Cost shape | Data residency | Performance | Security | Operational burden | Skills needed |
|---|---|---|---|---|---|---|
| Public-cloud AI services | Pay per use; grows with volume | Region choice; check where processing happens | Elastic; subject to provider quotas | Shared model; tenancy controls | Low to medium | Cloud and integration |
| SaaS AI | Licence add-on; may be bundled | Set by the vendor | Set by the vendor | Vendor’s controls | Low | Configuration and governance |
| Private cloud | Capital plus run cost; fixed capacity | Fully controlled | Bounded by own capacity | Own controls end to end | High | Platform, model serving, MLOps |
| On-premises | Capital heavy; long refresh cycles | Fully controlled | Predictable; hard to burst | Own controls; physical too | Highest | Hardware, facilities, model operations |
| Hybrid | Mix of both; harder to forecast | Controlled where it matters | Varies by path | Boundary is the weak point | High | Both sets |
| Sovereign hosting | Usually priced above standard cloud | In-jurisdiction by design | Narrower service range | Local legal control | Medium | Cloud plus compliance |
| Self-hosted models | Compute cost replaces per-call fees | Wherever it is hosted | Depends on hardware and tuning | No third-party model access | High | Model serving and evaluation |
Telco considerations
- Network and subscriber data is often subject to national telecoms regulation as well as data protection law. Check which data classes may leave the country before choosing a public service.
- Cost that scales with use looks cheap in a pilot of fifty contact-centre agents and very different across the whole contact centre. Model the full-volume cost before committing.
- Self-hosting is not free control. It moves model upgrades, security patching and evaluation onto the operator’s own teams, permanently.
- SaaS AI inside an existing product inherits that product’s data model. If the CRM’s view of the customer is wrong, its AI features will be too.