Bringing AI to a network where ChatGPT won't load

#On the business network, a question has nowhere to go

Type the ChatGPT address on a PC on your business network, and the page doesn't load. Many companies in finance, the public sector, and manufacturing keep their business networks cut off from the internet, and most AI services work only by sending each question across the internet to the AI vendor's servers. For these companies, generative AI has always been someone else's story. Asking for an exception to network separation means getting through security review; doing nothing means falling further behind.

You don't have to give it up. Put the AI models on your own servers, and a question no longer needs to leave. OwlAI installs and runs on a network with no internet, and business features such as chat, database queries, document search, and meeting minutes run as they would anywhere else.

#With the models in-house, nothing leaves the network

Publicly released open models run on your in-house GPU servers, and OwlAI calls them there. The chat model that reads questions and writes answers, the embedding model used for document search (it turns the meaning of text into numbers so similar documents can be found), and the speech recognition model that transcribes meeting audio all run on in-house servers. Questions, company data, and meeting audio do not leave the network.

Even the steps that sound as if they need an outside connection finish inside the server. The server checks the license file's signature itself, so there is no online activation step, and you install the license file the same way on an air-gapped network. The desktop app on employees' PCs gets its installer and updates from your in-house OwlAI server, not from outside.

#Business features work the same on an air-gapped network

Asking questions in chat, querying internal databases, searching documents, meeting minutes, creating Excel, Word, PowerPoint, and Hangul documents, and playbooks (automations that run a routine task in a set order) all work on an air-gapped network. Other systems inside the network connect through connectors. Research-style answers such as deep research are built the same way, with internal databases and documents as the scope instead of the web. Even the interface fonts ship in the installer, so the screens look exactly as they do with internet access.

That leaves the question of whether an open model running in-house gives answers you can use. Pre-analyzed database descriptions, the glossary, and table notes leave the model less to guess about which tables to look at and how to count. How well an open model handles your company's questions is something to check with real business questions during a pilot.

#A brought-in file is applied only if its signature checks out

For the first installation, you bring the image bundle (the package that holds the programs) and the installation kit into the network. The tools that come with them check the received files' checksums, load the images, and then compare what was loaded against the originals.

To update, bring in a single release file and upload it in the admin console. Everything that version needs, from programs and installation files to desktop installers, is in that one file. The server verifies the file's signature before it opens the confirmation dialog, and once an administrator approves, it backs up the database and then applies the update. A file whose signature does not match is rejected and deleted on the spot. Large files upload in chunks, so if the connection drops partway, only the missing chunks are sent again.

#You need two kinds of servers, plus internal addresses and certificates

You need one OwlAI server and a GPU server to host the AI models. The GPU server is sized by the models you'll use, the number of concurrent users, and question length, and there are recommended configurations for each installation size. As an example, as of August 2026 our in-house production environment runs chat on two 24GB GPUs hosting a 35-billion-parameter-class open MoE model quantized to 4 bits (compressed to use less memory), tuned to process four requests at a time. We set that concurrency by measuring with real question lengths, because short test questions make it look as if the server can take far more requests than it really can.

When you review adoption with your IT team, check these conditions first.

• OwlAI server minimum specs: 16GB of memory, 4 CPU cores and 80GB of disk or more, running a Linux distribution such as Ubuntu 22.04 or 24.04, or RHEL or Rocky 8.10 and later
• Internal DNS that resolves the addresses of OwlAI and the model servers
• HTTPS with an internal certificate (needed for voice input and the desktop app)
• Images built in advance and brought in (images are not built inside the network)
• Results from running the install check tool in air-gapped mode before installing (it confirms that every brought-in image is loaded on the server)