Local AI Models for Business: Cloud, Own Server, or Closed Loop

The CFO has a slide up: connect an AI assistant to contracts, payroll files, and the email thread with the bank. Everyone nods until the general counsel asks one question. “When the AI reads these documents, where do they actually go?” Nobody in the room can answer it, and the project slides to next quarter.

That question is why more companies are looking at local AI models for business: a model that runs on hardware you control instead of on a vendor’s cloud. This article skips the hype and the scare stories. It covers the three ways to host an AI-powered business system, what the hardware really costs, where the honest trade-offs are, and how to pick.

Three options: cloud, your own server, or a closed environment

PRANUS offers all three, and the choice isn’t permanent. Plenty of teams start in the cloud and move later. The difference comes down to where your data lives and where the model runs.

	PRANUS cloud	Your own server	Closed environment with local models
Where data lives	PRANUS servers	your server	your server, in an isolated environment
Where AI runs	cloud model providers, through a single AI gateway	cloud models, your own API keys, or local models	local models on your hardware
Time to launch	fastest	needs a server and an admin	needs a GPU server for the model
Updates	automatic	on an agreed schedule	on an agreed schedule
Best for	most companies getting started	teams that must keep data in-house	banks, healthcare, public sector, manufacturers with trade secrets
Main trade-off	data sits with a provider	you own server upkeep	pricier hardware; local models are slower than top cloud models

Details, timelines, and how a move works are on the PRANUS deployment options page.

What changes when the model runs on your hardware

For your team, almost nothing. AI agents keep working and knowledge-base search keeps answering. What changes is who does the thinking behind the scenes.

Every AI request in PRANUS goes through one gateway. Services don’t ask for a specific model; they ask for a job: “reply to the customer,” “read this document,” “summarize this call.” The gateway decides which model handles it. A local model plugs into that gateway as just another provider. The usual tools for running open models work fine: LM Studio, Ollama, vLLM, and similar.

A few rules matter to a business owner:

• “This organization runs on local models only.” Turn it on, and every conversational AI task for your company goes to your models, even when a service would normally request a cloud model. No need to reconfigure each module.
• You decide what happens when the model is down. Show a clear “AI temporarily unavailable” message, queue the work until the model is back, or fall back to the cloud. Closed environments usually pick one of the first two.
• Health checks about every 30 seconds. Take the server down for maintenance and the system stops sending it requests on its own. When it comes back, traffic resumes.
• A slow answer isn’t duplicated to the cloud. Local models take longer, so the system waits instead of firing the same request at a cloud model “just in case.” You don’t pay twice, and data doesn’t go somewhere you didn’t intend.
• Other organizations can’t see your models, even when a group of companies shares one installation.

Permissions work the same way in every option. In the MEMUS knowledge base, access is checked before search, so the model never receives documents the employee isn’t allowed to open. Your accountant can’t “ask the AI” about sales contracts, no matter where the model runs.

We covered how the gateway itself works, including fallback chains, spending limits, and cost tracking, in one door to every AI model.

The honest trade-offs

A local model isn’t a free copy of a cloud model. Know three things before you decide.

Speed. We tested an open model with roughly 27 billion parameters on a workstation with 64 GB of memory (a Mac). It answers and can use tools, but long tasks take minutes, not seconds, and it handles one or two requests at a time. That’s fine for a pilot, a single executive, or a small team. A whole company needs a GPU server.

Quality. The largest cloud models still beat open models at hard reasoning. For most day-to-day work, like answering from a document, summarizing a meeting, or drafting an email, open models do the job. For complex analysis, keep the cloud in the mix or run a bigger model.

Hardware and people. Someone has to buy or rent the server and keep it running. That’s a budget line, not a dealbreaker.

What the hardware costs

These are street prices from 2025–2026. Treat them as estimates, not a quote. The right setup depends on your workload and headcount.

What you need	Ballpark cost
A ~70B-parameter model: 2–4 H100 cards or 4 RTX PRO 6000 cards	about $70K–150K, or $3.5K–7K per month to rent
235B–400B-parameter models: an 8-GPU high-end server	about $320K–500K
Power and maintenance	add roughly 15–20% per year

How to size it: in a 1,000-person company, usually 5–10% of people use AI at the same moment, so 50 to 100 users. One instance of a ~70B model serves 20–50 concurrent users. That tells you how many cards you need.

A company with 50–200 employees can often get by with one mid-range server or a rented one. And you can start smaller still, with a single machine for a pilot group.

On-premise AI: what the market offers (as of November 2026)

Few business platforms let you run them on your own servers with a local model inside:

• Zoho, Salesforce, HubSpot, monday, ClickUp, and Notion have no self-hosted option at all. They’re cloud-only.
• Bitrix24 sells a self-hosted edition, but its AI features still call Bitrix24’s own cloud.
• Odoo can run local models, but only through third-party modules.

That doesn’t make cloud platforms bad. It means they answer the general counsel’s question with a contract clause, not with the location of the server.

How to choose: a checklist

Answer six questions. The more “yes” answers in the first half, the closer you are to a closed environment.

1. Do data residency rules require you to keep customer data in a specific country or region? Many countries now have localization laws, and regimes like GDPR or industry rules in finance and healthcare add their own conditions. Check with your counsel.
2. Do you hold trade secrets that shouldn’t reach outside AI providers, such as formulas, designs, or terms of major contracts?
3. Do large customers or banking partners audit where their data is processed?
4. Do you have an admin or a contractor who can run a server?
5. Would you rather invest in hardware now than pay monthly?
6. How fast must AI answers be? Seconds, or are minutes acceptable for heavy tasks?

If you answered “no” to the first three, start in the cloud. If you said “yes” to at least one of them, and “yes” to questions four and five, look at your own server or a closed environment.

Hybrid: you don’t have to pick just one

For many companies, the most practical setup is a mix. Model rules can be set per task and per level: the whole organization, a department, or a single service.

A hypothetical example: HR and accounting run only on a local model, because that’s where salaries and ID documents live. Marketing writes copy with a cloud model, because speed matters more there and no customer data is involved. Search across internal policies stays local.

This way you buy hardware for the work that truly needs it, not for the whole company at once. Your AI agents in AGENTUS don’t notice the difference. Each one has its own role, permissions, and budget, and the gateway picks the model according to your rules.

How to make the move

1. Start in the cloud. Set up the system, your org chart, the knowledge base, and AI agents. You’ll quickly see which tasks AI actually handles and how many.
2. Measure the load. How many people use AI at once, which tasks are most common, and where the most sensitive data is. PRANUS shows AI usage by employee and task, so you have real numbers.
3. Draw the sensitive boundary. Decide which departments and tasks must run locally only.
4. Pilot on one machine. Connect a local model for a test group and compare quality and speed on your real work.
5. Size the server. Based on the pilot, buy or rent hardware for the number of users you need.
6. Move. Migrate data to your server or a closed environment on an agreed plan, then switch the right departments to local-only mode.

The first steps are the same as a regular 14-day PRANUS implementation: an AI agent builds your org chart, adds employees, and assembles the knowledge base. The hosting decision can run in parallel without slowing your team down.

The bottom line

Local models aren’t a fad or a cure-all. They let you answer “where do our documents go?” with the location of a server, not just a line in a contract. The price is hardware, upkeep, and somewhat slower answers on heavy tasks.

For most companies, the right path is cloud first, then hybrid: sensitive work local, everything else wherever it runs best. For teams where a closed environment is a condition of the deal, it’s available from day one.

Want to work through your case, including which data goes where, how much hardware you need, and where to start? Request a demo. We’ll show you the system and size a deployment option for your company.