What is private AI for business?
The five levels of private AI, what a private AI server is, and what privacy alone does not give you.
Private AI for business means running AI models so a company's prompts, documents and outputs stay in an environment it controls or that is reserved for it: on premises, in a private cloud or in its own cloud account. It ranges from an API with no-training terms to air-gapped servers, and does not make a system compliant by itself.
This page is about private AI for organizations. Consumer privacy apps and the company called Private AI are out of scope.
"Private" borrows its meaning from private cloud. NIST defines a private cloud as infrastructure "provisioned for exclusive use by a single organization", which may be run by the organization or a third party, on or off premises. Private AI applies the same test to models and the data that flows through them.
What counts as private AI?
Anything from a contract that limits what a provider does with your data to servers with no network connection. Five levels cover most setups.
| Level | Where the model runs | Who can see prompts and outputs | Typical controls |
|---|---|---|---|
| 1. Hosted API with business terms | The provider's shared infrastructure | The provider, within its terms | No-training terms, retention limits, regional processing |
| 2. Dedicated hosted instance | The provider's infrastructure, reserved for you | The provider's operators | Single tenancy, private network links |
| 3. Your own cloud account | GPUs in your account with a cloud provider | Your team; the cloud provider runs the hardware | Your keys, your network rules, your logs |
| 4. On premises or private cloud | Servers you own or lease | Your team | Physical control, your identity provider |
| 5. Air-gapped | Servers with no external network path | Your team, on site | Offline updates, no outbound connections |
This page uses "private AI" mainly for levels 3 and 4, where the model and the data path sit inside infrastructure the company controls. Level 1 is privacy by contract, and it can be enough. OpenAI, for example, says data sent to its API is not used for training unless you opt in, and it keeps abuse-monitoring logs for up to 30 days by default. Level 5 is described in the post on air-gapped AI.
What is a private AI server?
A private AI server is a GPU server, or a small cluster, inside your network that runs a model and answers your applications through an internal API.
It usually has five parts:
- GPUs with enough memory for the model and its working state. The answer on how many GPUs you need shows the sizing.
- An open-weight model, chosen for your tasks and licensed for your use.
- A serving engine that batches requests and exposes an API, such as vLLM.
- A gateway that signs users in through your identity provider, applies quotas and logs calls.
- Connections to company knowledge: document search, a vector database or governed access to databases, so answers can use your data.
The steps to build one are in the answer on deploying an LLM on premises.
Why do businesses choose private AI?
Usually because of what goes into the prompts.
- Confidential data. Contracts, health records, source code and customer data may not be allowed to leave the company's control.
- Client requirements. Service providers may have to run AI inside each client's environment, under the client's rules.
- Intellectual property. Fine-tuned models and the data used to tune them are assets worth keeping in-house.
- Control over change. A model you host changes only when you upgrade it.
- Cost at steady volume. Owned GPUs can cost less per token than an API when they stay busy, and more when they don't.
What doesn't private AI guarantee?
Keeping the model inside your walls removes one outside party. It does not secure the system.
- Insiders and over-broad access. If every employee can query every document through the assistant, sensitive data leaks internally. Answers should respect the same permissions as the source data.
- Prompt injection and data leakage. Prompt injection and sensitive information disclosure are the first two entries in the OWASP Top 10 for LLM Applications 2025, and they apply to self-hosted models too.
- Compliance. Rules on personal data, retention and records still apply. Private deployment makes them easier to meet, not automatic.
- Sovereignty. A dedicated instance run by a provider under another country's laws is private but not necessarily sovereign. The answer on sovereign AI vs private AI explains the difference.
Where does Hyperlake fit?
Hyperlake's private AI capabilities cover hosting and serving open and custom models, fine-tuning and evaluating them on selected compute, and grounding them in governed enterprise data such as structured data, documents, vector search and knowledge graphs, with policies on what data and tools each agent or application can reach. It runs in the customer's environment or their clients'. The FAQ covers where it runs and what it includes.
Frequently asked questions
Is a hosted AI API with a no-training policy private AI?
It is the first level, privacy by contract. OpenAI, for example, says data sent to its API is not used to train its models unless you opt in, and keeps abuse-monitoring logs for up to 30 days by default. The provider still processes your data on its infrastructure.
Does private AI have to run on premises?
No. NIST's definition of a private cloud says it may be owned or operated by the organization or a third party and may exist on or off premises. What makes it private is exclusive use, not the building it sits in.
Do I need to train my own model for private AI?
No. A private deployment can run an existing open-weight model and connect it to company documents and data at query time. Fine-tuning is an option for narrow tasks, not a requirement.
How much does private AI cost?
It depends on volume and how busy the GPUs stay. The main costs are GPUs, rented or bought, plus power and hosting, the serving and gateway software, and the people who run it. The cost calculator compares that with paying an API per token, using your own prices.
Sources
Sources for the facts on this page, last checked October 9, 2026.
- The NIST Definition of Cloud Computing, SP 800-145 (NIST, 2011) checked October 9, 2026
- Data controls in the OpenAI platform (OpenAI API documentation) checked October 9, 2026
- OWASP Top 10 for LLM Applications 2025 checked October 9, 2026
Related guides
- Sovereign AI vs private AI: what's the difference?Sovereign AI is about who has authority over an AI system and whose laws apply; private AI is about who can see its data. A side-by-side comparison.
- How do you deploy an LLM on premises?Deploy an LLM on premises in seven steps: model and licence, GPU sizing, serving engine, gateway, security, monitoring and offline updates.
- How many GPUs do I need to serve an LLM?GPU count starts with memory: weights by precision, plus KV cache for every token in flight, plus overhead. Worked examples for an 8B and a 70B model.
- Air-gapped AI: the strictest enterprise standard
Start with a workload. Build the environment around it.
Tell us what you need to deploy, whose environment it must run in, and what it needs to connect to.