The private AI landscape
What published deployment guidance tells us about private AI: where the model runs, what crosses the boundary and who keeps the system working.
What the evidence suggests
Private AI is a complete operating arrangement. The useful question is where each part of a request goes, not simply where the application is installed.
A review of published research and primary documentation, with our interpretation and relevant Exemplary capabilities. This is not an original benchmark or customer-results report.
- FINDING 01
Hosting and model location are separate choices.
- FINDING 02
Offline AI needs an operating plan as well as a local model.
- FINDING 03
Clear ownership makes privacy commitments easier to verify.
The boundary is bigger than the upload
Figure 10Source files and prepared knowledge.
Assembles the request and checks access.
Processes the selected context here.
Selected prompt and context cross the boundary if an external model is enabled.
Its own service, location and terms.
1. “Private” describes several different arrangements
Cohere documents deployment options spanning its platform, cloud services, private cloud and on-premises infrastructure. That range is evidence that enterprise model delivery is not a single hosting choice. The controls and operating work change with the arrangement.
Our reading: a useful comparison separates three questions. Who controls the environment? Where does computation happen? Who operates the service? A private cloud in the customer’s account and a vendor-operated environment may both offer isolation, but they do not give the same people responsibility for access, updates or incident handling.
The same distinction applies inside a product. A company can keep its application and source files on its own equipment while sending selected text to an outside model. That may be an intentional choice. It should be described as such, rather than being hidden behind a general claim that the application is private.
Primary source: Cohere: Deployment options — overview
2. An offline model still needs an operating system around it
NVIDIA’s air-gap guidance describes preparing model-serving assets for an environment without direct internet access. The relevant lesson is that offline execution requires advance preparation; the installation cannot fetch a missing dependency on demand.
Our interpretation extends to the surrounding workflow. A local model does not make an internet-based search tool local. A document service, identity dependency or update process can reintroduce a connection the organisation intended to remove. Each dependency needs an internal route, an approved transfer process or a deliberate decision to leave the capability out.
Consider the day after a software update, rather than only the first successful answer. Operators need to know which version was installed, how to verify it and how to return to a working version if a critical task breaks. Offline operation moves that work into a planned process. It does not make maintenance disappear.
Primary source: NVIDIA: Air-gap deployment
How Exemplary helps
Install around the agreed boundary
Exemplary supports on-premises, private-cloud and air-gapped deployments. We help assess the equipment, install the platform, connect the agreed systems and test the setup before handover. You bring the business requirements; leave the AI setup to us.
See the deployment approach3. The request is the right unit to follow
A privacy review often starts with the uploaded file. An AI request can create several other forms of the same information: extracted text, search entries, selected passages, a prompt and an answer. A connected tool may add another record or send part of the result to another system.
Our analysis: a clear data journey should show these transformations and destinations. For each stage, record where the information is stored or processed, who can access it and whether another provider receives it. The answer can vary by configuration, so a single diagram should state which configuration it represents.
This also makes commercial discussions more concrete. Data region, backup arrangements, support access and deletion behaviour can be agreed against actual components. A generic “secure AI” label cannot answer those questions on its own.
Putting it in context
The same contract question, two different routes
Imagine a legal team asking for the notice period in a supplier contract. In one arrangement, the file, extracted text, search and language model all run within the customer’s environment. The request uses a local model and internal document access.
In another arrangement, the file remains in the same place but the application sends relevant clauses and the question to an external provider. Both routes can produce a similar-looking answer. Their data handling is different. The team should choose with that difference visible, including the provider’s terms and any connected tools. This is an illustrative scenario, not a customer deployment result.
How Exemplary delivers the relevant pieces.
The research explains the idea. Here is how our platform and setup work connect to it.
How Exemplary helps
Make model routes and ownership visible
Our data-location guide separates hosting from model choice and explains the information each stage handles. On-premises and private-cloud installations are handed over to customer operators; managed cloud has an agreed operating scope. External model routes remain explicit choices.
Trace where your data goesScope & method
This is an editorial review of primary vendor documentation, not a survey or an audit of any deployment. The sources demonstrate available approaches; they do not establish that every feature or model works in every environment. Exact region, retention, backup and operating commitments depend on the agreed setup.
We selected primary sources that explain the underlying approach, cover an important failure mode or describe an available deployment choice. Findings are attributed to their authors. Scenarios and business implications are our analysis; they are not findings from an Exemplary customer study.
Sources & further reading
Primary publications and documentation. Accessed 22 September 2026.
- 01Deployment options — overview
Cohere · Product documentation
- 02Air-gap deployment
NVIDIA · Product documentation