Why “where does our data live?” matters more than it seems
For some companies, the cloud is simply convenient: nothing to install, nothing to update, start tomorrow. For others, the cloud is ruled out by law or security policy. Some countries require citizens’ personal data to stay within their borders. Banks, government contractors, and manufacturers with trade secrets aren’t willing to send customer conversations and financials to someone else’s servers.
Then there’s AI. Many systems keep your data on your servers, yet their AI features still send requests to external cloud models. For regulated industries, that’s the same leak through a different door.
That’s why PRANUS comes in three options, and all three are the same system. Not a stripped-down on-premise edition or a separate product, but the same 18 services, the same AI agents, and the same 14-day implementation.
The three options side by side
| Cloud (SaaS) | Self-hosted | Air-gapped (on-prem with local models) | |
|---|---|---|---|
| Best for | Most companies that want to start fast without their own IT admins | Companies that must keep data in-house or in-country by law or policy | Banks, government contractors, manufacturers with trade secrets: anyone whose data and AI must never leave the premises |
| Where data lives | In the PRANUS cloud; the storage location is specified in your contract | On your servers or in your data center | Inside your perimeter, with no outside connection |
| Updates | Automatic, nothing for you to do | Through one secure connection to PRANUS, in an agreed window | Delivered as packages under contract; fully offline setups have no connection to us at all |
| AI models | Leading cloud models through the PRANUS AI hub, with automatic fallback models | Through the PRANUS AI hub over a secure connection, your own provider keys, or local models | Local models on your hardware only |
| Time to launch | Right away; first working setup in 14 days | Server preparation, then the same 14-day implementation | Depends on AI hardware delivery, then the same 14 days |
| Hardware | None | A server sized to your headcount to our specs; no GPUs needed if AI runs in the cloud | A server for the system plus a GPU server for AI models (estimates below) |
| Price | Starting prices on the pricing page | On request | On request |
Cloud: start tomorrow
The cloud is the fastest path. You request a demo, a sales manager runs it on your own tasks, and once the contract is signed, the implementation agent starts the same day. Servers, backups, updates, and monitoring are on us.
In the cloud, AI runs through the PRANUS AI hub, a single gateway to every model. The hub picks the right model for each task, switches to a fallback if a provider fails, tracks spend, and enforces limits at every level: person, department, company. You can also bring your own model provider key.
Self-hosted: your data stays with you
PRANUS installs on your servers or in your data center. Your data stays with you, and your security team controls access to it.
The standard setup is a closed environment with one outbound connection to PRANUS. That connection carries updates, certificates, and the license. Nothing else leaves. Updates are installed in a window you’ve approved, not whenever it suits us.
Self-hosted AI can work three ways:
- Through the PRANUS AI hub over a secure connection, just like in the cloud.
- With your own keys from model providers: the hub runs inside your installation and stores keys encrypted on your side.
- On local models, which makes it an air-gapped setup.
Air-gapped: data and AI both stay inside
This is the strictest option. The system and the AI models run on your hardware with no calls to outside services. Switch your company to local-models-only mode and no AI request will ever cross your perimeter.
PRANUS works with the common ways to run open models: LM Studio, Ollama, llama.cpp, vLLM, and MLX. You choose models that fit your tasks and budget, and the AI hub routes requests among them just like it does in the cloud.
A fully offline option with no connection to PRANUS at all is also available. Updates and support for it are covered by a separate agreement.
How much hardware your own AI model needs
In an air-gapped setup, the biggest cost isn’t the license. It’s the hardware for the models. Here are estimates based on 2025–2026 market prices for a company of about 1,000 employees, with 5–10% using AI at the same time:
| Model | Hardware | Estimated cost |
|---|---|---|
| Mid-size (~70B parameters): enough for most tasks, such as knowledge base answers, drafts, and conversation analysis | 2–4 H100 GPUs or 4 RTX PRO 6000 GPUs | $70K–150K |
| Large (235B–400B parameters): for complex reasoning and heavy workloads | A server with 8 H200 or B200 GPUs | $320K–500K |
Add roughly 15–20% a year for power and maintenance. GPU prices rose noticeably over the past year, so we put together an exact quote for your case at the time of purchase. A small team can get by with a smaller model and a single powerful server.
These are estimates, not a price list. We size the exact configuration after we learn how many people you have and which tasks you want to hand to AI.
What’s the same in every option
- The same 18 services across the platform and the same AI agents.
- Permissions from your org chart, sign-in with a verification code, audit logs, and the “agent suggests, a person decides” rule. More on the security page.
- Free 14-day implementation with an annual contract.
- You can export your data at any time.
Not sure which option fits? Get in touch. A sales manager will go through your data storage requirements and recommend the right setup.