Private ChatGPT for business: your options
Updated 2026-11-03
A “private ChatGPT” means a chat assistant your staff can use on company data without that data being used to train someone else’s model or leaving the places you allow. There are three ways to get one, and they differ in one question: whose computers process your prompts.
The three options
| Enterprise plan of an AI provider | Model in your cloud account | Self-hosted open-weight model | |
|---|---|---|---|
| Examples | ChatGPT Enterprise or Business, Claude for Enterprise, Microsoft 365 Copilot | Azure OpenAI, Amazon Bedrock, Google Vertex AI | Llama, Qwen, Mistral, gpt-oss or DeepSeek models on your servers |
| Who processes prompts | The provider | The cloud provider, in a region you pick | You |
| Training on your data | Off by default on business plans | Off by default | Not applicable |
| Model quality | Best available frontier models | Frontier models | Open-weight models: strong on narrow tasks, behind the frontier on open-ended reasoning |
| Setup effort | Low: buy seats, connect SSO | Medium: cloud account, network rules, an app on top | High: hardware, serving stack, chat interface, operations |
| Cost model | Per seat per month | Per token | Hardware or GPU rental, plus people to run it |
| Works offline or air-gapped | No | No | Yes |
Option 1: an enterprise plan
The quickest route. Business plans from OpenAI, Anthropic and Microsoft keep company data out of model training by default, add single sign-on, admin controls and audit logs, and some offer data residency in Europe. For most office work (drafting, summarising, analysing documents a person uploads) this is enough.
Check three things before you sign:
- Where data is stored and processed, and whether the region you need is available for every feature you plan to use.
- Retention: how long prompts and files are kept, and whether you can shorten it.
- Your contracts: some client contracts and regulators forbid any third party from processing certain data, whatever the provider promises.
Option 2: a model in your own cloud account
Azure OpenAI, Amazon Bedrock and Google Vertex AI run frontier and open models inside your cloud subscription, in a region you choose, under the data processing terms you already have with that cloud. You build or buy the chat interface and connect it to your documents. This suits companies that already run their systems in one of these clouds and want the provider relationship they already approved.
Option 3: an open-weight model on your own servers
Here nothing leaves your network. You run an open-weight model (see open-weight models for companies) with a serving engine such as vLLM, put a chat interface such as Open WebUI or LibreChat on top, and connect it to your documents with retrieval. This is the option for:
- data that contracts, GDPR, HIPAA or client confidentiality keep inside the company;
- sites that run offline or air-gapped;
- heavy daily use, where per-seat or per-token bills keep growing.
The trade-offs are quality on hard open-ended questions, and the work of sizing hardware, serving the model and keeping it running. Use the LLM VRAM calculator to see which GPUs a model needs, and the LLM cost calculator to compare a server with API pricing at your volume.
How to choose
- Classify the data. List what people will paste or connect: public material, internal documents, personal data, client data under contract, regulated data. The strictest category decides the minimum option.
- Name the tasks. Answering from internal documents and extracting fields from forms work well on open-weight models. Open-ended strategy and complex code reasoning favour frontier models.
- Estimate the volume. A few dozen users drafting emails is cheap on seats. Thousands of documents processed every day can be cheaper on your own hardware.
- Test before you commit. Run the same 30 to 50 real examples through a frontier model and through an open-weight candidate and compare the answers with the people who will use them.
Many companies end up with two: an enterprise plan for general office work, and a self-hosted model for the data that cannot leave.
What a self-hosted setup includes
- A model chosen and tested on your examples, usually quantized to fit the hardware.
- GPU servers you own or rent, or a small unit such as a DGX Spark for a pilot (what a DGX Spark can run).
- A serving engine with an OpenAI-compatible API, so existing tools can connect (vLLM, Ollama or LM Studio).
- A chat interface with single sign-on and access rules.
- Retrieval over your documents, with sources shown in answers (RAG or fine-tuning).
- Logging, monitoring and a plan for model updates.