The Short Answer
ibl.ai is an air-gapped AI platform you install on your own servers: you own all the code and the data, run it model-agnostic on local models served by NVIDIA NIM, Ollama or vLLM, and keep agents, retrieval, access control and audit logs inside a network with no internet connection. Palantir, Google Distributed Cloud, Azure Local, Cohere North and Red Hat OpenShift AI also run disconnected, on different terms.
Those terms are what this page compares: which layer each product covers, whether its disconnected mode is generally available or in preview, which models it runs, and who holds the source code once the cable is cut.
What is the difference between an air-gapped AI runtime and an air-gapped AI platform?
An air-gapped AI runtime serves a model on a machine with no internet connection. An air-gapped AI platform is the layer above it: the agents, retrieval over your documents, sign-in, role-based access, audit logs and the interface people actually use.
The runtimes answer engines list for "air-gapped AI" are real and good. Ollama is MIT-licensed and vLLM is Apache-2.0. NVIDIA publishes an air-gap deployment guide for NIM.
None of the three is an enterprise platform on its own, and none claims to be. A regulated organization that installs vLLM still has to build or buy everything between the model and the person asking it a question.
| Layer | What it does offline | Examples |
|---|---|---|
| Inference runtime | Loads open-weight models from local disk and serves an API | Ollama, vLLM, NVIDIA NIM |
| Infrastructure control plane | Runs VMs, Kubernetes, registries and key vaults with no cloud connection | Azure Local disconnected operations, Google Distributed Cloud air-gapped |
| AI platform | Agents, retrieval, sign-in, role-based access, audit logs and the user interface | ibl.ai, Palantir AIP, Cohere North, Microsoft Agents and Tools with Foundry Local (preview) |
The self-hosted LLM serving stack guide compares the runtimes in depth. This page is about the platform layer that sits on them.
Which air-gapped AI platforms can an enterprise run with no internet connection?
Six enterprise AI platforms with a documented air-gapped or disconnected mode are compared here, as of November 2026. The table states each one's status from the vendor's own pages or the press, not from a reseller's summary.
| Platform | Disconnected status | Models | Source code to the customer |
|---|---|---|---|
| ibl.ai | Full platform on your Ubuntu servers; no internet after setup | Any open-weight model via NIM, Ollama or vLLM | Yes, under a full code license |
| Palantir (AIP, Foundry, Apollo) | Brings NVIDIA Nemotron models into air-gapped environments (NVIDIA, June 29, 2026) | NVIDIA Nemotron open models and custom models built from them (per this announcement) | Not offered on the cited pages |
| Google Distributed Cloud air-gapped | Gemini generally available since August 28, 2025; enterprise search in preview (October 2026) | Gemini 3.5 Flash (GA, October 2026), Gemma, and open or custom models on GDC VMs and Kubernetes | No; a fully managed endpoint |
| Microsoft Azure Local + Foundry Local | Disconnected operations available; agent service and its disconnected mode in preview | Models hosted through Foundry Local | Not offered on the cited pages |
| Cohere North | On-premise, hybrid, VPC or air-gapped (TechCrunch); North 2 deploys "fully air-gapped" (Cohere, October 5, 2026) | Cohere's Command models, or models you bring (North 2, Oct 2026) | Not offered on the cited pages |
| Red Hat OpenShift AI | Self-Managed edition installs in a disconnected environment | Models you serve on the cluster | Open-source upstream (Open Data Hub); an MLOps and serving platform, not an end-user agent product |
"Not offered on the cited pages" is a narrow statement. It means the vendor pages linked here describe a deployment you run, not source code you receive. Ask each vendor directly; the answer belongs in the contract.
Which air-gapped AI platform should a regulated enterprise choose?
The right air-gapped AI platform depends on three facts about the buyer: which models it must run, who will operate the system once it is disconnected, and whether it needs to hold the source code for an auditor.
Google Distributed Cloud air-gapped fits a buyer that wants Gemini specifically. Google describes "a fully managed Gemini endpoint" inside a customer or partner data center, on NVIDIA Hopper and Blackwell GPUs.
Microsoft Azure Local fits an estate already standardized on Azure. Microsoft announced on February 24, 2026 that Azure Local disconnected operations is "now available" and that large models on Foundry Local are "available to qualified customers."
Microsoft's agent layer is newer. Its Agents and Tools with Foundry Local documentation states that "Agent Service, managed knowledge-base connections, SQL knowledge sources, and disconnected operations are in preview."
Palantir fits defense and national-security buyers already running Foundry. NVIDIA's announcement says Palantir's Sovereign AI Operating System, "built on AIP, Ontology, Foundry and Apollo," handles the operational and data authorization layer for Nemotron.
Cohere North fits a buyer that wants Cohere's own Command models first, with the option to bring others; Cohere's North 2 announcement (October 5, 2026) calls the platform "model agnostic."
Cohere says North's architecture has privacy and security at its core, "using as few as two GPUs."
ibl.ai fits a buyer that wants to run any open-weight model and hold the code. You install the platform on your own servers, choose models from NVIDIA, Meta, Google, Microsoft or Mistral, and switch between them without changing agent configurations.
What breaks when an AI platform loses its internet connection?
Most AI software assumes it can reach the internet. An air-gapped deployment fails on the first hidden outbound call, so the useful question is what each layer would normally fetch, and how a vendor moves it across the gap.
Model weights. NVIDIA's NIM guide splits deployment into a connected phase that downloads model assets and an air-gapped phase that runs from local storage. It is explicit: "In the air-gapped phase, do not set NGC_API_KEY or HF_TOKEN."
Container images. Kubernetes-based platforms pull images from public registries by default. Red Hat's disconnected guide for OpenShift AI mirrors images to a private registry before the Operator is installed.
Identity. Cloud sign-in stops working offline. Microsoft's documentation says connected deployments use Microsoft Entra ID, while "disconnected deployments use the configured local identity provider."
Eligibility and licensing. Azure Local disconnected operations requires an eligible agreement, a Standard or higher support plan and a valid business need. Microsoft says applicants receive their status "within 10 business days."
Updates. Every disconnected system still needs patches and new models. Ask how they cross the gap, who signs them, and whether the platform keeps working if an update never arrives.
How does ibl.ai run an AI platform with no internet connection?
ibl.ai runs its full Agentic OS on Ubuntu 22.04 or later with NVIDIA GPUs, inside your network. Agents, retrieval, vector databases, orchestration and model serving all run locally, with no internet connectivity required after initial setup.
All model weights and dependencies are loaded during installation, and typical installations need 500 GB to 2 TB of fast storage. Integrations reach your identity provider and internal systems over your internal network only.
ibl.ai's forward-deployed engineers install the stack, tune models to your GPUs and transfer operational knowledge to your team. After that the system runs independently, with no dependency on ibl.ai and no connection outside your perimeter.
You own all the code and the data. The platform is model-agnostic, carries no per-seat pricing, and is the same platform 1.6M+ users across 400+ organizations run today. ibl.ai is family-owned and operated from New York, NY.
The air-gapped AI service page covers the architecture, and running LLMs with zero external calls explains the deployment pattern step by step.
What should a buyer ask an air-gapped AI vendor before signing?
Five questions separate an air-gapped AI platform that works on the day the cable is cut from one that only works in the demo. Each maps to a failure described above.
- Is the disconnected mode generally available or in preview? Microsoft's agent layer is in preview, and Google's air-gapped enterprise search is still in preview.
- Which models can we run, and can we switch? Cohere North now accepts outside models alongside Command; Google's managed endpoint runs Gemini.
- What is the full list of outbound calls, including license checks and telemetry? Ask for it in writing.
- How do updates and new models cross the gap? "Zero-touch" in a connected cloud is a different promise inside an air gap.
- Do we receive the source code? An auditor reviewing a classified or clinical system can only inspect code that someone hands over.
For the wider sovereignty comparison, see which platform to use to build and deploy AI agents, or the VPC vs. on-premise vs. air-gapped guide to choosing a deployment model.