We deliver the same models and the same security controls in three ways. The right choice depends less on technology and more on your data and your obligations.

Start with the data

Classify the data the system will touch. If it includes regulated filings, personal records or anything your regulator treats as sensitive, start from air-gapped and relax isolation only for a clear reason.

The three options

  • On-premises: runs on your servers and network, integrated with your identity provider and internal systems.
  • Air-gapped: no network path out; weights, indexes, logs and updates stay inside.
  • Dedicated private GPUs: single-tenant inference in an in-country data center, for teams without spare GPU capacity.

Look at three years, not three months

Hardware, power, licences and the people to run them add up differently in each model. We build a 36-month total cost of ownership model with you before you commit, so the decision is made on numbers.

A simple rule

Choose the most isolated option your team can operate well. Isolation you cannot maintain is weaker than a connected system that is properly monitored.

Related pages