01
We bring it in.
A standalone AI system about the size of a desktop tower. It plugs into a normal wall outlet and joins your network the same afternoon.
One box on your network, with local inference.
Runs on your network
Installed in a day
Works with your tools
We keep it running
How it works
One secure box
We bring the machine, connect it, and stay responsible for it. Your team can use AI without starting a data-center project.
02
Your work stays yours.
Documents, code, and customer records get processed in your building. Your existing tools can talk to this box instead of leaving the building.
03
We keep an eye on it.
Uptime, temperatures, load, and drive health, watched around the clock.
04
The software comes with it.
Your team gets a private chat workspace, and an OpenAI-compatible endpoint for everything else. We install it, configure it, and keep it patched.
The app
The chat app your team already loves
Every system ships with the On Premises workspace. It's the same chat you already know, except the answers come from the machine down the hall.
- DocumentsUpload contracts, reports, and private documents, then ask questions across all of them at once.
- Your files and notesPoint it at the share drive and the internal docs your team already keeps.
Included in the monthly. We install it, connect it, and keep it updated.
On Premises
New chat
Today
Vendor contract review
Reply to the Denver office
Rewrite the safety handbook
Tools
Documents
Email
Files
vendor-agreement-v4.pdf
What changed between this version and the one we signed in
March?
Three things. Payment terms moved from net 30 to net 45, the
liability cap dropped to one year of fees, and the auto
renewal clause on page 8 is new. Want me to draft the redline?
Ask anything
Running on your hardware
Data flow
What leaves the building.
ChatGPT, Claude, Copilot
On Premises
Your questions
Leave your network
Stay on the box
Documents you upload
Copied to their servers
Never copied off
Record of who asked what
Their logs
Your logs
What we can see remotely
Your content, per policy
Hardware health only
When the model changes
Whenever they ship
Only when you approve
If the internet drops
Stops working
Keeps working
If you cancel
Deleted on their schedule
You keep hardware and data
Those vendors are careful. Your data still has to leave to reach them.
The systems
Pick the box that fits the room.
The upfront price covers the hardware, burn-in, install, and first configuration. Software and upkeep bill monthly after that.
1 GPU
Focused team
1× RTX PRO 6000
A single tower, sized for a couple of people using it every day.
- GPU memory96 GB
- System RAM128 GB
- Best fit1–2 people
- Ready inTwo weeks
Upfrontfrom $35,000
Software and upkeep$850/month
Start here
2 GPUs
Growing team
2× RTX PRO 6000
Twice the power. Fits bigger models and more people at once.
- GPU memory192 GB
- System RAM256 GB
- Best fitGrowing team
- Ready inTwo weeks
Upfrontfrom $53,000
Software and upkeep$1,400/month
Choose this system
4 GPUs
Company
4× RTX PRO 6000
Enough for a whole department, still one machine in one room.
- GPU memory384 GB
- System RAM256 GB
- Best fit5–10 people
- Ready inTwo weeks
Upfrontfrom $84,000
Software and upkeep$2,400/month
Choose this system
GB300
Frontier
DGX Station GB300
The largest machine we install, for the biggest models you can run.
- Coherent memoryUp to 748 GB
- Best fitLargest models
- Ready in8 weeks
Upfrontfrom $126,000
Software and upkeep$3,200/month
Choose this system
We order hardware once the upfront payment clears. Software and upkeep start at go-live on a 12-month initial term. Final configuration and supplier pricing get confirmed in the written quote.
Try before you buy
Speed up your work
Half a day, one workflow, a demo machine on the table. We show you the speed and power of a private AI system.
See the pilot
No hardware purchase required.
Price$1,500 Free
Half a day of work, plus a written report.
What we testYour real workflow
On your approved samples or synthetic stand-ins.
What you getA go or no-go
What it's for, what could bite you, how big a box, what it costs.
Before the first call
The questions we actually get.
Nothing gets quoted until we've sat down with your IT owner and worked out the workload, who can reach it, and how it should behave on your network.
What can a private AI system do on day one?
Search across your own documents. Draft and summarize from material you've approved. Help your developers. Hand an existing app a private API to call. The pilot is how you find out which one of those is good enough to be worth the hardware.
Does our data ever leave the building?
Inference runs on the box in your office. It does not need the internet to answer a question. Anything the machine does reach out for, like model downloads, telemetry, backups, or our remote support, gets set up to your rules and written into the deployment scope. We won't call a system isolated unless it was built and is operated that way.
Do we need an AI engineer?
No. We install the hardware and the serving stack, hook up the approved tools, write down how to administer it, and pick up the phone when something's wrong. What we do need is one person on your side who can get us network access, handle identity, sign off on the data, and own the workflow itself.
Why start with a pilot?
Because it is a lot cheaper than buying the wrong hardware. Half a day gets you an answer you can defend to whoever signs the check, plus the site requirements and a fixed quote for the build. Right now the pilot is free.
What do the upfront and monthly prices include?
The upfront quote covers hardware, burn-in, on-site setup, and initial configuration. The monthly covers the private chat software, the model-serving stack, monitoring, routine updates, remote maintenance, and support. Taxes, travel, replacement parts, custom integrations, and unusual network work get quoted separately before you buy.
Are you affiliated with NVIDIA?
No. On Premises is an independent integrator. We buy and install whatever fits the workload we agreed on, and manufacturer warranties and license terms stay between you and them.