Free tool · sizing
What does private AI actually take to run?
Answer four questions and get a straight spec: what model class fits, what hardware runs it, what it costs, and how fast it really goes. The numbers come from machines we run every day, and we round them down rather than up.
our recommendation
- Model class
- Typical hardware cost
- Speed you can expect
- Power under load
Want to see it before buying anything?
The private AI engagement starts with a free demo on our own hardware, so you can judge the speed and quality with your own eyes. How private AI works →
Book the free call →answers
Sizing FAQ
Where do these numbers come from?
From hardware we run ourselves: a 32 GB workstation GPU serving a 35B-class model to real users daily, and an 8 GB card running a 9B model beside it. The speeds and power figures are what we observe, rounded conservatively, and prices are typical street prices that drift over time.
Why would a small business run AI on its own hardware?
Privacy is the usual reason: patient records, client files, and financials that must not leave the building. Cost is the second: past a certain usage level, a one-time machine beats per-seat subscriptions. Our frontier vs local guide covers when each makes sense.
Is a used or refurbished machine really fine?
For this workload, usually yes. Inference wants GPU memory and steady cooling more than the latest CPU. A clean refurbished workstation with a current GPU is the best value per dollar we know of, and it is what we would build for ourselves.