02 / COLUMN — AIR-GAPPED LLM
On the claim that air-gapped means no AI
Air-gapped LLM infrastructure means running large language models entirely inside an isolated, secure network. Most of the generative AI wave happens on external clouds, which makes it sound like someone else’s story to organizations behind a security perimeter. Many stop evaluating right there.
The cloud-or-nothing dichotomy
Security-bound organizations tend to end their AI review one of two ways: a doomed attempt at cloud exception approval, or indefinite postponement. Either way, nothing changes inside while AI-driven work becomes the standard outside. Some try token gestures, a small model formally installed, but when performance misses the bar, users leave and the "air-gapped AI does not work" conclusion only hardens.
The facts have changed
That dichotomy rests on facts from years ago. Open-weights models now approach frontier-class performance, and inference serving has matured to production grade. Bringing the model in, instead of sending data out, is now a real option. What remains is not a question of possibility but of engineering: which model, on what hardware, at what throughput, operated by whom.
Data stays in. Models come in.
How KDX Labs does it
KDX Labs owns that engineering end to end. We size and design the servers, select and verify open-weights models against your work standards, and bring them through your import procedures. Inference optimization stretches the same hardware to more concurrent users, and we hand over an operable system with key issuance, usage metering, and log management included.
This is the process we built and still operate inside actual air-gapped defense-supplier networks. Respecting import procedures and security requirements is as much a part of the craft as the technology.
Standard infrastructure, inside the perimeter
Air-gapped LLM will not stay a special experiment. It is becoming standard infrastructure for secure organizations, as unremarkable as the internal network itself. KDX Labs is building that foundation starting from the strictest environments. Data stays in. Models come in.
If this maps to your challenge, let’s start from your constraints.
Ask about this capability