On-premise AI

Local AI, under your control.

Three architectures to run AI models inside your own infrastructure. No data, prompt or response leaves your network: all three options give the same privacy guarantee and differ in scale, availability and governance.

Book a discovery session →Compare the three options
Open-weight modelsInference in your data centerA log of every interaction

Why on-premise

When data is sensitive, the model goes to the data

Not the other way around. Four reasons companies bring AI into their own infrastructure.

Privacy and sovereignty

Prompts, documents and responses are processed on your servers, with no third parties and without your data training anyone else’s models.

Regulation

Data protection laws (in Mexico, the 2025 LFPDPPP) and your industry’s regulators require knowing where data lives and who touched it.

Predictable cost

You invest in your own capacity instead of paying per token, with a cost you can budget.

Control and continuity

You decide which model version to use and when to update it. The platform keeps running without an internet connection.

Three options

Three architectures, one principle

The same privacy guarantee, with different levels of scale, availability and governance.

Three architectures, one principle
CriterionOption 1 · Local NodeOption 2 · Private PlatformOption 3 · Governed Sovereign AI
Who it is forMid-sized company with a first use caseUpper mid-market or enterprise with several teamsRegulated enterprise: banking, insurance, healthcare, government
Technology baseOpen source: vLLM or Ollama on Docker Compose, LangGraph, pgvector, LibreChat, LangfuseRed Hat OpenShift AI 3: llm-d + vLLM, Models-as-a-Service, EvalHub, NeMo Guardrails, GitOpsIBM watsonx (watsonx.ai, watsonx.data, watsonx.governance) on Red Hat OpenShift AI
AvailabilityOne server with cold backupHigh-availability clusterTwo sites with disaster recovery, disconnected (air-gapped) environment
AI governanceTraces and evalsAdds per-team metered usage, evaluation and platform guardrailsAdds model inventory, risk and regulatory evidence
Time to launch4 to 6 weeks8 to 12 weeks12 to 20 weeks

Timelines and capacities are reference ranges: they are confirmed during discovery with your use cases, users and real volumes.

Option 1

Local Node: private AI on a single server

The shortest path to production for a mid-sized company with a first use case: one GPU server inside your network, with open-source components deployed with Docker Compose.

Launch: 4 to 6 weeksAvailability: cold backupGovernance: traces and evals
Diagram of Option 1, Local Node: Your corporate network · no internet egress
UsersBrowser on the internal network
Internal systemsERP · CRM · help desk
DirectoryAD / LDAP · SSO
Data sourcesDocuments · databases · file server
GPU server · Docker Compose
InterfaceLibreChat · REST API
AgentsLangGraph · tools · memory
InferencevLLM or Ollama · LLM and embeddings
Knowledgepgvector · document search (RAG)
ObservabilityLangfuse · traces and evals

Limit: no high availability. The server is a single point of failure and is protected with cold backup.

Option 1 components and their licenses

ComponentRoleLicense
vLLMInferenceApache 2.0
OllamaInferenceMIT
LangGraphAgentsMIT
pgvectorVector storePostgreSQL License
LibreChatChat interfaceMIT
LangfuseTraces and evalsMIT (core); enterprise features under a commercial license
Docker ComposeDeploymentApache 2.0

Licenses verified in November 2026; they are confirmed at the start of each project. Each open-weight model has its own license, which we review for your use.

Option 2

Private Platform: AI as an internal service

For upper mid-market and enterprise companies with several teams: a cluster that serves several models to the whole organization, with high availability, per-team metered usage and declarative GitOps operations.

Launch: 8 to 12 weeksAvailability: HA clusterGovernance: + metered usage, evaluation and guardrails
Diagram of Option 2, Private Platform: Your data center · segmented network
UsersChat and copilots
SystemsERP · CRM · core systems
DirectorySSO (OIDC)
Red Hat OpenShift AI 3 cluster
AccessGateway API · SSO with your directory
Apps and agentsLangGraph · copilots · APIs
Models-as-a-ServiceSubscriptions · quotas · per-team usage
Inferencellm-d + vLLM · autoscaling
ModelsValidated catalog · several at once
Datapgvector or Milvus · S3 · PostgreSQL
Quality and controlEvalHub · NeMo Guardrails · GitOps
Control plane nodesCPU workersGPU workersStorage

Requires a platform team, your own or managed by SISCON.

What we do with Red Hat →

Option 3

Governed Sovereign AI: evidence for the regulator

For regulated enterprises —banking, insurance, healthcare, government—: end-to-end model governance (inventory, risk, lineage and audit) across two sites with disaster recovery and disconnected operation.

Launch: 12 to 20 weeksAvailability: two sites with DRGovernance: + inventory, risk and evidence
Diagram of Option 3, Governed Sovereign AI: Disconnected (air-gapped) environment · two sites
Primary site
watsonx.governanceInventory · risk · drift · audit
watsonx.aiGranite and open models · tuning
watsonx.dataLakehouse · vectors · lineage
OpenShift AIllm-d · vLLM · MLOps pipelines
SecurityAdvanced Cluster Security · KMS/HSM · mirror registry · SIEM
InfrastructureHigh-availability GPUs · replicated storage
Disaster recovery (DR) site
StandbyCluster ready to fail over
ReplicaData and models
FailoverAdvanced Cluster Management
ContinuityDefined RPO and RTO

Higher investment and change management; justified by regulation or criticality. Disconnected operation is confirmed per product and version during design.

What we do with IBM →

Security

Controls common to all three options

Present in all three; depth and evidence grow from Option 1 to Option 3.

No internet egress

Models and updates come in through a controlled internal registry, after their integrity is verified.

SSO with your directory

Access always goes through your corporate directory, with role-based permissions.

Permissions inherited from the user

The agent only sees what the user can see; document search is filtered by their permissions.

A log of every interaction

Every prompt, response and agent action is recorded and available for audit.

Guardrails

Filters for personal data and injection attempts, usage policies and escalation to a person in edge cases.

Encryption

In transit and at rest, with the keys under your control.

Reference frameworks we design the controls against: ISO/IEC 42001 · NIST AI RMF · OWASP Top 10 for LLM Applications · OWASP Top 10 for Agentic Applications. The platform provides controls and evidence for your compliance program; it does not replace the assessment of your compliance teams or auditors.

Decision

How to choose and how to grow

Start where you are today; the architecture grows when the business asks for it.

1

Start with Local Node if…

you have a clear, well-scoped use case and want results in weeks with a contained investment.

2

Scale to Private Platform if…

several teams are asking for AI, you need high availability and want to offer models as an internal service with metered usage.

3

Go to Governed Sovereign AI if…

a regulator or an audit requires evidence: model inventory, lineage, risk control and two-site continuity.

Evolution path: nothing you build is thrown away

Agents, document search (RAG) and evals move from one option to the next, because everything runs in containers and uses OpenAI-compatible APIs. You can start with one server and grow without rebuilding.

  1. 1 · Local Node
  2. 2 · Private Platform
  3. 3 · Governed Sovereign AI

Methodology

Six phases, from discovery to operations

The same path for all three options: measure before buying hardware and validate with a real case.

  1. F1

    Discover

    Use cases, sensitive data, regulatory framework and infrastructure. The output is the scope and the recommended option.

  2. F2

    Design

    We evaluate open models with your data, size GPUs, storage and network, and the architecture and its controls are approved.

  3. F3

    Develop

    We install the platform, integrate it with your directory and build the first use case.

  4. F4

    Validate

    Quality tests against a reference set, plus security, load, and verification that the network has no internet egress.

  5. F5

    Deploy

    Staged go-live, training, and handover of the runbook and the evidence of controls.

  6. F6

    Optimize

    Measured operations with KPIs, continuous evals and new use cases.

Evals are the exit criterion: nothing goes to production if it degrades measured quality.

IA Factory inside your network. The IA Factory software production line can run on any of the three options, with models and logs inside your infrastructure.
Related services: Fine-tuning & Custom Models to adapt an open model to your domain, AI Evaluation & Audit to choose the model with evidence, and our IBM and Red Hat partnerships. Read also: how to choose between the three architectures.

FAQ

Frequently asked questions

Are open models good enough for my use case?

It depends on the case, which is why it is measured before committing scope: in the Design phase we compare open models with your own data, not with public benchmarks.

Do I need to buy GPUs before starting?

No. Models are evaluated first and hardware is sized with your real users and concurrency; the purchase comes after design.

Does an on-premise platform make me compliant with data protection law or my regulator?

Not on its own. It provides controls —data inside your network, permissions, logs and encryption— and evidence for your compliance program. Compliance is determined by your organization and its auditors.

Can I start with one server and grow later?

Yes. Agents, RAG and evals are built in containers with compatible APIs, so they move to the Private Platform or to Governed Sovereign AI without being rebuilt.

Does Option 1 have no license costs?

Its components are open source and each has its own license (Apache 2.0, MIT, PostgreSQL License); open-weight models have theirs too, which we review for your use. Options 2 and 3 include Red Hat and IBM subscriptions.

Can IA Factory run inside my network?

Yes, on any of the three options. With local models, scope is committed after measuring their quality with evals on your own code.

Next step

Choose the option with your numbers

A free 60-minute discovery session to review your use cases, sensitive data and infrastructure, and recommend where to start.