New Co.

Model protection & device-bound licensing

Ship the model.
Keep the weights.

AI‑DRM makes a trained network useless as a file. A protected model runs only on a device you have licensed — and the secret that makes it run never exists in the clear outside that device's secure element.

Request a demo How it works

The short version

The problem, and what we do about it.

AI‑DRM · 1 min 8 s

A model you ship is a model you have given away

Once the file reaches a customer site, an edge box or a partner's server, it can be copied, inspected and redeployed — and you will not know. Encrypting the weights at rest does not close it: something has to decrypt them to run them, and that something is running on hardware you do not control.

AI‑DRM moves the part that matters somewhere the host cannot reach, and ties permission to run to a specific piece of silicon.

Bound to one device

Every licence is minted for a single enrolled device and is worthless on any other. Approval is explicit, per device, and revocable.

Your model stays yours

Protection runs against infrastructure you operate. The trained model is not uploaded to us, and we never hold a copy of it.

Air-gapped by design

Disconnected sites are a first-class path, not a workaround: signed request and response bundles travel by courier, with no network at either end.

Hardware root of trust

Protected material is only ever in the clear inside a trusted execution environment — or, on hardware without one, a USB dongle that provides it.

How it works

  1. Protect

    You upload a trained model to your own orchestrator and get back a protected one. That protected file is what ships.

  2. Enrol

    A device registers itself and waits. Until an operator approves it, it holds no key and can download nothing.

  3. Licence

    An approved device is issued a small licence bound to it alone. Nothing secret travels in either direction to produce it.

  4. Run

    The protected model loads and runs at full speed — but only where a valid licence and its secure element agree that it may.

Where it runs

Model formats
ONNX, TorchScript
Runtimes
CPU, CUDA, TensorRT
Trust anchor
OP-TEE, or an ESP32‑C6 USB dongle
Deployment
Connected or fully air-gapped

Request a demo