2027 planning guide · Private deployment

Private AI in 2027: sovereign, VPC, on-prem and air-gapped compared

Private AI is a set of deployment choices, not one product. Here is how the models differ, where each one fails, and what to put in the contract.

Short answer. Private AI means inference runs inside a boundary dedicated to one customer, with a defined data path and access control. It can be a private VPC, dedicated endpoint, on-premises deployment or air-gapped network. Sovereign AI adds a legal and jurisdictional promise. Private is not automatically sovereign, and on-premises is not automatically secure; the contract and the operating model decide.

Quick answers to the questions buyers ask first.
QuestionOne-sentence answer
What is private AI?Inference inside a customer-specific boundary with a defined network, identity and data path, rather than a shared public endpoint.
Is private AI the same as sovereign AI?No. Private describes isolation; sovereign describes jurisdiction, ownership and legal control.
Is on-premises always more private?No. It can be, but only if identity, logging, patching and physical access are controlled.
What is a private VPC deployment?The model runs in an isolated network you control, with private connectivity and no public inference endpoint.
What does air-gapped mean?The environment has no internet path; model and software updates arrive through a controlled offline process.
Who should manage encryption keys?Prefer customer-held keys in your KMS or HSM, with documented break-glass and rotation procedures.
Can we use our own identity provider?Ask; SSO, SCIM and role mapping are contract and configuration questions, not default assumptions.
What is the main trade-off?Control versus operating burden: private deployments shift patching, capacity, monitoring and upgrades to you or a named support team.

Private AI is sold as a single idea and delivered as several different architectures. A dedicated endpoint, a private VPC, a rack in your data centre and an air-gapped cluster solve different problems. They also carry different operating costs.

This guide separates the claims. It defines what each model isolates, where data actually travels, what to ask about keys and support access, and which clauses make a private deployment enforceable rather than aspirational.

Private, sovereign and on-premises are different claims

The words are often used interchangeably in sales conversations. In procurement they mean different things, and one does not imply the others.

What does private AI actually isolate?

Private AI isolates the inference environment at the network, identity and data-path layers. Compute and storage may be physically dedicated or logically separated, so the contract should say which. It does not necessarily mean dedicated model weights, dedicated staff or a separate legal jurisdiction. A managed private endpoint can still share a control plane, logging pipeline or support team with the provider's public service. Ask for a data-flow diagram that marks every hop: request, model server, log store, backup, support access and deletion path. The isolation claim is only as strong as its weakest shared component. See how private VPC deployment is scoped.

How is sovereign AI different from private AI?

Sovereign AI adds jurisdiction and control to isolation: where data and metadata are stored, who owns the operator, which courts can compel access, and whether keys remain under local control. A local region is not the same as a sovereign service; a foreign-owned operator can run in-country infrastructure and still be subject to another country's laws. Ask who owns the entity, where the control plane sits, who can access keys and what government-access history applies. For GCC buyers, Riyadh is available on Enterprise, not as a self-serve region. The sovereign AI page sets out the deployment options.

Is on-premises always more private?

No. On-premises moves the boundary but does not remove the need for identity control, patching, audit and physical security. A poorly operated on-premises deployment can be less private than a well-run dedicated cloud endpoint, because insider access and unpatched services are common failure paths. The advantage is control: you decide who enters the building and which network path exists. The cost is operational: you own capacity, upgrades and incident response. Air-gapped deployment goes further by removing the network path entirely, but it also makes updates slower.

The five deployment models compared

The right model depends on the threat you are defending against, not on the label a vendor prefers.

Deployment models differ in isolation and operating burden. Confirm the exact scope in writing.
ModelCompute locationOperated byData pathBest fitMain burden
Private VPC or dedicated endpointIsolated cloud networkProvider or jointPrivate connectivity; no public inference endpointRegulated SaaS, latency-sensitive workloadsNetwork, IAM and egress design
On-premisesYour data centreYour team, vendor supportStays on your networkHigh-control, existing GPU capacityPatching, capacity and upgrades
Air-gappedIsolated network, no internetYour teamOffline transfer onlyDefence, classified, operational technologyUpdate logistics and model freshness
Sovereign cloudIn-country regionLocal operatorIn-region, local lawPublic sector, banks, critical infrastructureService breadth and operator maturity
Shared cloud APIMulti-tenant regionProviderPublic endpointLow-sensitivity, fast startNot private in the procurement sense

What is a dedicated endpoint?

A dedicated endpoint reserves model capacity for one customer, usually with a private network path and its own rate limits. It is managed by the provider, so you do not patch the model servers or own the GPU fleet. Isolation varies: some endpoints are logically separated in a shared cluster, others run in a dedicated account or tenancy. Ask whether weights are shared, whether logs are segregated, whether support access is scoped and whether the endpoint can burst into shared capacity. The answer belongs in the architecture document, not the sales deck.

Keys, identity and access

Encryption is not a deployment model. It is a control that only helps when key custody, access and logging are designed together.

Who can see prompts in a private deployment?

Only the roles you authorise should be able to see prompts, and the provider should have no default access. In practice, support engineers may have break-glass access, observability systems may capture request metadata, and logs may contain prompt fragments. Ask for the access matrix, the approval path for support access, the retention window and whether access is logged. Encryption helps only if keys are under your control and the provider cannot decrypt at will. Put the answer in the DPA and the security schedule.

What does bring-your-own-key change?

Bring-your-own-key moves custody of the encryption key to your KMS or HSM, so the provider stores ciphertext and requests decryption through a controlled grant. It reduces the blast radius of a provider-side breach and gives you a revocation lever. It does not control model behaviour, stop a provider from processing plaintext during inference, or replace access logging. Ask whether the key is single-tenant, whether rotation is automatic, what happens when the key is unavailable, and whether the provider can access the key under a legal order. See BYOK for the control model.

How should access be audited?

Every administrative action and data-plane access should produce an audit record you can export: who, what, when, from where and under which role. Separate control-plane changes from inference requests, because they have different retention and privacy needs. Ask whether logs are immutable, how long they are retained, whether they include prompt content, and how they reach your SIEM. Time synchronisation matters for incident reconstruction. A private deployment without an audit trail is a private network, not a controlled one.

How to choose without over-buying

The most private option is not automatically the right one. Match the architecture to the obligation, then cost the operating model.

When do you need on-premises or air-gapped?

Choose on-premises when data cannot leave your facilities, you already run GPU capacity, or you need custom hardware or network integration. Choose air-gapped when no network path is acceptable at all, such as classified, defence or operational-technology environments. Both shift operating burden to you: patching, capacity planning, monitoring, incident response and model updates. Air-gapped also slows model refresh, because updates arrive through a controlled offline process. Be honest about whether your team can run that lifecycle before choosing the architecture.

What does a private AI deployment cost to operate?

The licence is one line. GPU capacity, redundancy, storage, network, monitoring, patching, security review and staff time usually dominate the three-year cost. A private deployment also needs spare capacity for failover and a plan for model upgrades. Ask for a total-cost model that includes support and the engineering hours your team must supply. Do not compare a private quote with a self-serve plan without adding those lines; the self-serve plan carries the provider's operating cost inside its price. Use the pricing page for the self-serve baseline.

What to verify in a private AI contract

The architecture diagram and the contract should tell the same story. If they do not, the contract wins in a dispute and the diagram wins in a sales meeting.

  • Data-flow diagram. Every hop, including logs, backups, support access and deletion.
  • Subprocessors. A current list, change notice and a right to object.
  • Key custody. Who holds the key, how rotation works, how revocation is handled and what happens under legal access.
  • Access matrix. Named roles, break-glass approval and audit export.
  • Deletion. Data sets, deadlines, backup handling and a certificate of deletion.
  • Incident notice. A notification window, forensic cooperation and breach communication responsibilities.
  • Evidence. Audit rights, a penetration-test summary and the current attestation period. Request the SOC 2 report under NDA and verify the period; do not treat “SOC 2” as a certification claim.
  • SLA and support. The uptime target, exclusions, support boundaries and escalation path. The 99.9% headline is a target; the contract schedule governs.
  • Model lifecycle. Version pinning, deprecation notice and the update process for on-premises or air-gapped installs.
  • Exit. Export format, transition assistance, key return and deletion after termination.

“Private” is not a product SKU. It is a set of boundaries you can draw on a diagram and enforce in a contract.

What we did not verify. We did not inspect Plugsky's private deployment environments, review a signed contract or test isolation. This page describes procurement questions, not a certification. Request the current security pack, and have counsel review the terms that govern your deployment.

Frequently asked questions

Is private AI the same as a private cloud?

No. A private cloud describes infrastructure tenancy: compute and network reserved for one organisation. Private AI adds model serving, inference data paths, prompt and output handling, and the access controls around them. You can run private AI on a private cloud, but a private cloud alone does not define who can see prompts or where model logs go. Ask for the AI-specific data-flow diagram, not only the cloud architecture. That diagram is the document to test in a security review.

Does a private VPC guarantee data residency?

No. A VPC can keep inference traffic in a region while logs, backups, support access or the control plane remain elsewhere. Residency depends on the full data path and the contract, not on the network topology alone. Ask which region holds prompts, outputs, metadata, logs and backups, and whether support staff can access them from another country. The data residency cluster covers the verification steps in detail. Ask for the diagram, not a region name.

Can we use private AI without a DPA?

You should not. A data processing agreement defines roles, permitted purposes, subprocessors, transfer terms, deletion and breach obligations. Without it, you are relying on general terms that may not match your regulatory position. A private deployment does not remove the need for a DPA, because the provider may still process metadata or provide support. Ask for the current DPA, the subprocessor list and any bespoke addendum your sector requires, and confirm the signatory.

What is the difference between air-gapped and offline?

Air-gapped means there is no network path between the environment and the internet or another untrusted network. Offline can mean the system is disconnected most of the time but still syncs periodically, or that updates are shipped on removable media. The distinction matters for update cadence, monitoring and incident response. Ask how model weights, security patches and licence renewals reach the environment, and who approves each transfer. Define the allowed transfer path in writing.

Can we bring our own model weights?

That is a contract and platform question, not a default feature. Some private deployments accept customer-supplied weights; others require the provider's supported catalogue for security, licensing and support reasons. Ask about weight format, licence obligations, validation, patching responsibility and whether the provider will support the model after deployment. If the model is your differentiator, confirm in writing that you can export it and continue using it after the contract ends.

Who patches the model in a private deployment?

The contract should name one owner. In a managed private VPC, the provider usually patches the model servers and the customer owns the network and identity configuration. On-premises deployments often split responsibility: the vendor supplies updates and the customer applies them. Air-gapped installs need a documented offline update process. Ask for the patch cadence, the severity timeline and the consequence of running an unsupported version. An unnamed owner is an unpatched system.

How do we test isolation?

Start with the diagram and the contract, then test the path: place canary data, attempt egress from the inference network, inspect audit logs, and confirm that support access requires approval. Ask for a recent penetration-test summary and the scope it covered. A test that only checks the public endpoint is not an isolation test. Record the result and the date, because the environment changes when models, network rules or support tooling change.

What SLA applies to private deployments?

The contract schedule governs, not the marketing headline. Plugsky publishes a 99.9% uptime target, with plan-level terms and exclusions; Hobby is not in the SLA table. Private deployments may have different availability terms because the customer controls part of the stack. Ask how uptime is measured, which components are in scope, what excludes a credit, and how support response times differ from the self-serve plan. Put the measurement method in the schedule.

Does private deployment include a HIPAA BAA?

A HIPAA business associate agreement is available by contract, subject to review. It is not a certification, and a private deployment does not automatically include one. Ask whether the BAA covers the specific services and data flows you need, who signs it, and how subprocessors are handled. Counsel should review the BAA together with the DPA and the security schedule, because the obligations interact. The BAA is not a substitute for a security review.

What happens to our data at exit?

The contract should define the export format, the export window, transition assistance, key return and deletion deadlines for primary data, backups and logs. Ask for a certificate of deletion and confirm which subprocessors must also delete. Test the export before you need it; a theoretical right to export is not the same as a working process. If the data is in a proprietary format, negotiate a machine-readable export in the schedule.

Is open-weight model hosting more private?

Not inherently. Open weights let you run the model yourself, which can improve control and auditability, but privacy still depends on the operator, the network, the identity system and the logging. A managed open-weight endpoint can be less private than a well-run proprietary endpoint if support access is broad. Ask who operates the inference, where the data goes and who can see it, regardless of whether the weights are open.

Sources

Primary sources and official documentation for this 2027 planning guide — evidence checked 20 November 2026.

  1. NIST SP 800-145, The NIST Definition of Cloud Computing — checked 20 November 2026.
  2. NIST SP 800-207, Zero Trust Architecture — checked 20 November 2026.
  3. CISA, Cloud Security Technical Reference Architecture — checked 20 November 2026.
  4. ENISA, Cloud Security — checked 20 November 2026.
  5. European Commission, Cloud Computing — checked 20 November 2026.

Last checked 20 November 2026.

Related reading

Draw the boundary before you sign

Bring your data-flow requirements and we will map them to VPC, on-premises or air-gapped deployment.

Request architecture review See private deployment