The short answer
Private AI means the AI model runs on a machine your company owns, inside your building, connected to your own files and systems. Your data never goes to an outside company, and nobody pays a monthly fee per employee. It's the right call for companies with confidential data and real document volume. It's the wrong call for companies that haven't organized their files, or that only need AI for light, occasional work. Under about ten daily users, stay on subscriptions.
The model runs in your building. That's the whole idea.
When most people say "AI" they mean a website. You type into a box, the text goes to a data center owned by OpenAI or Anthropic or Google, an answer comes back. That works, and for a lot of tasks it works well. But every question, every pasted contract, every uploaded spreadsheet makes a round trip through somebody else's computers.
Private AI flips that. The model, the same kind of technology behind those websites, runs on a machine in your office. Ask it about the warranty terms on a job from 2022 and it reads your contracts, on your network, and answers. Nothing crossed the internet to get you that answer.
Your team doesn't experience anything exotic. They get a chat box that looks like the one they already use. What sits behind it is the difference.
There are four ways to run AI in a business, and most companies under ten people belong on the second one. Past that, keep reading.
- Public tools on a personal account. Cheap, and fine for questions with nothing sensitive in them. This is where most of your employees already are, whether you know it or not.
- Cloud AI with business controls. A company account, or better, custom software between your business and the model, where you decide exactly what data leaves and what doesn't. The provider promises not to train on your data.
- Hybrid. Sensitive work stays on a local machine. Hard problems go to the big cloud models with the sensitive parts stripped out.
- Fully private. The model runs on hardware you own. Nothing leaves. That's what this article is about.
We sell three of those rungs, and we still tell plenty of companies to stay on rung two. The point is picking a rung on purpose instead of drifting into one by accident.
Why companies go private
The data is genuinely confidential. Client files under NDA, financials, medical or legal records, pricing your competitors would love to see. I've been on record about this for years: when a business owner tells me they won't paste bookkeeping data into a chatbot, I agree with them. My own rule when I used cloud tools for reconciliation was to strip the account numbers and names and send only the dates and transactions the task needed. That discipline works. It also depends on every employee doing it right, every time. A private system removes the question. There's nothing to leak because nothing leaves.
The second reason is the subscription math. AI arrived in most companies the way software always does: one seat at a time. A chat assistant for whoever asks. A meeting notes tool billed per attendee. A document search product on top of your file storage. Nobody chose that stack. It accumulated. At current per-seat prices a 40-person company can be into five figures a year before anyone's added it up. A private system has no seats. The whole company uses it, including the new hire on day one.
And then somebody asks "where does this data actually go?" Usually a client, an auditor, or a lawyer. The cloud providers publish real policies about business data, and the good ones mean something. But "our vendor promises retention controls" is a very different sentence in a security review than "it runs on that machine, in that room."
A deployment is mostly not about the AI
The first thing we test on any deployment is whether the system can answer a question about a four-year-old change order without somebody telling it which folder to look in. That test fails more often than you'd expect, and when it fails, the model is almost never the problem. The file storage is. Twenty years of documents named FINAL-v3-revised, in folders only one person understands.
So here's what actually gets built. A machine with real GPU hardware, 32 to 96 gigabytes of video memory, sized to your team and installed on your network. An open model loaded onto it: these are models whose weights you can legally download and run, and the good ones now handle document and drafting work fine. Connections into the places your information already lives: file storage, email, CRM, project management, accounting. Access rules, so the system answers each person only from what that person's allowed to see. The estimator doesn't get payroll answers.
Then the unglamorous part, which is the part that decides whether the thing survives: backups, monitoring, patching, documentation. The box gets treated like the rest of your network, not like a science project in a closet. A private AI machine nobody maintains becomes shelfware in six months.
The hardware costs less than a year of the subscriptions it replaces
Prices below are what real retailers were charging for complete systems as we write this, in August 2026. They move. The shape of the math doesn't.
Three things drive the price: how many people use it at once, how big a model you need, and how much wiring it takes to connect your systems. The hardware is the visible part. The integration work is usually the bigger variable.
A serious single-GPU workstation, the kind that runs a capable open model for a small team, sells complete for roughly $4,500 to $10,000 right now. NVIDIA's DGX Spark, a small machine built for exactly this, lists at $4,699. A Mac Studio handles this work better than most people expect and runs $2,500 to $8,300 depending on memory. A dual-GPU machine for more simultaneous users puts you in the low twenty-thousands. The 96GB professional-GPU class, for companies that want large models on site, reaches $45,000 configured. One warning from shopping this market: a lot of quoted prices are for the graphics card alone, not a working computer. Compare complete systems to complete systems. The full tier-by-tier breakdown, including what we paid for our own machines, is in the private AI server cost guide.
Electricity is a rounding error. At the national average commercial rate, about 13.5 cents per kilowatt-hour, a single-GPU box burning full power around the clock costs about $99 a month, and no office box actually runs full draw all day. If someone tells you the power bill is the reason not to own AI hardware, they haven't done the arithmetic.
Now the other column. Current per-seat prices from the vendors' own pricing pages: ChatGPT Business at $20 to $25 per user per month. Claude Team at $20 to $25, or $100 and up for the heavy tier. Microsoft Copilot at $18 to $25. A meeting notetaker adds $10 to $39 per user. A 40-person company running a chat assistant plus a notetaker lands somewhere between $14,000 and $26,000 a year, every year, before the document-search add-on.
A $9,000 machine plus a support arrangement, against a pile that costs more than that annually and never ends. Under about ten daily users the subscriptions win and we'll say so. Past fifteen or twenty, owned hardware wins, and the gap widens every year you run it, because the models keep improving and the good ones are free to swap in.
One caveat on the studies you'll find on this topic: most published owned-versus-cloud cost analyses come from hardware vendors, and their scenarios assume the machine stays busy. A lightly used machine is the worst of both worlds. That's why we count your hours before we quote anything.
Skip it if your files are a mess
We turn companies away from private AI regularly, and it's almost always for one of the same few reasons.
The most common one is the files. If your documents are scattered across personal drives, old laptops, and a share nobody's cleaned since 2019, a private AI system pointed at that gives you fast, confident answers drawn from chaos. Fix the storage first, deploy second. We say this in the first meeting, and it's not what people want to hear, and it saves them from buying a machine that answers wrong.
Close behind: the workload's too light. If AI at your company means occasionally drafting an email or summarizing a meeting, a business-tier cloud account is cheaper and better for that use. The frontier cloud models are still ahead of what runs on local hardware. For casual use, the difference shows. That's also why most of our private deployments end up hybrid, with a written rule about which work may leave the building and which never does.
Then there's ownership. Hardware in a closet with no owner has a well-known life story, and it doesn't end with a return on investment. The machine needs somebody responsible for it. In-house IT or a support contract, either works. And occasionally we meet a company that wants private AI because it sounds safer, while everyone in the building can already open every file on the server. A private AI just makes that problem faster. Permissions first.
Who's left? Companies with real document volume, data that has to stay home, and enough daily users for the math to work. If that's you, the system pays for itself answering questions that used to mean twenty minutes of digging, times everyone, times every day.
How a deployment runs
We don't start with hardware. We start with a count.
- Consult. A conversation about what you handle, what's confidential, and where the hours go. Twenty minutes tells us whether this is worth pursuing.
- Assessment. We map the repeat work and the document load, time it, and put the current cost on paper. If the numbers don't justify a system, we tell you, and you keep the numbers.
- Build. Hardware installed, model deployed, systems connected, access rules set. Your team gets trained on their actual work, not demo work.
- Run. Monitoring, model updates, support, and a monthly look at what it's doing and what it should do next.
Questions owners actually ask
Is a private model as smart as ChatGPT or Claude? Not quite, and for most business work it doesn't matter. Answering questions from your documents, drafting from your templates, summarizing, extracting, reconciling: current open models handle all of it well. For the hardest problems, run hybrid and send a sanitized version to the frontier.
What happens when better models come out? You download one and load it. The hardware is the durable purchase. The models are replaceable, and the good ones cost nothing.
Can it still leak our data some other way? Yes, the ordinary ways: an employee pasting output into a personal email, a wide-open network share, a missing backup. Running locally closes the biggest door, but access control, logging, and backups still have to be done right. That's infrastructure work. That's the business we came from.
Do we need our own IT department? No. But the machine needs an owner, and that can be us.