The options side by side
What each vendor publishes about price, hardware and terms. Sorted from smallest to largest buy.
| Option | Price as published | Hardware as published | Terms | Team size they state |
|---|---|---|---|---|
| DIY: Mac mini or Mac Studio (Apple) | Mac mini from $899 (M6) or $1,699 (M5 Pro); Mac Studio from $2,499 (M5 Max) or $5,499 (M5 Ultra). Memory upgrades cost extra | Your choice; memory can’t be upgraded after purchase | You buy it, install the software (for example Ollama) and maintain it | Not published |
| AdminOnsite | Practice $249/mo + $299 setup; Office $399/mo + $399 setup; Firm $699/mo + $499 setup | Mac mini: M6 32GB (Practice), M5 Pro 48GB (Office), M5 Pro 64GB (Firm) | Managed rental, 12-month minimum; the box is yours after 24 paid months, then optional Care at $99/mo | Up to ~5; ~5–15; ~15–30 |
| PrivateOfficeAI | PrivateOffice Box $4,900 one time; Box Pro $7,500; software only on your own machine $899. Care $99/mo, Remote Access $29/mo, Web Access $19/mo (optional) | Mini server with AMD Ryzen AI Max+, 64GB (Box) or 128GB (Pro), 2TB storage | You own it; 30 days of support and 1 year of updates included | Up to ~15 (Box); 15–40 (Pro) |
| Phoenixtekk AI Box | $5,995 paid in full at order; AI Box Pro bundle $7,995. Other tiers quoted to order | Compact AMD-based server; Phoenixtekk states 104GB of memory available to AI models and 120B-parameter models on one box | You own it; 30 days of support (Pro: 3 months of Care Plus) | Not published |
| Faraday Machines | Starter $299/mo; Growth $599/mo; Scale $999/mo. Setup fee not published | Mac Studio: M5 Max 128GB; M5 Ultra 256GB; M5 Ultra 512GB | 36-month plan, 30-day money-back guarantee, then no early cancellation; return the hardware or renew at the end | 2–9; 10–19; 20–29 |
| MacRun (hosted, not on site) | Agent Mac: M6 16GB $169/mo; M5 Pro 48GB $299/mo; M5 Pro 64GB $349/mo (M5 Pro build fee $199 or $249) | Dedicated Mac mini in MacRun’s facility in New Jersey | Month to month. MacRun says Apple’s licence limits hosted Macs to developer services, so it isn’t an office file server | Not published |
| onprem.ai | Server S1 €19,890; S2 €39,890; Dell D1 €34,490; D2 €49,490; licence €1,490 per GPU per year. Non-binding, excluding VAT | S2: 2× NVIDIA RTX PRO 6000 Blackwell, 192GB GPU memory. D1/D2: Dell PowerEdge R770 with 1 or 2 of those GPUs | Purchase plus annual licence; managed rental €1,490 per GPU per month in the DACH region | 5–45 users (S2) |
| Zylon in a Box | Not published | 1–8 NVIDIA GPUs, 1–2 AMD EPYC CPUs, 128GB+ memory | Pre-configured server, delivered in 6–8 weeks | Not published |
| Zanus AI Server | Not published (request a quote) | Four tiers (Standard to Extreme); GPUs and memory sized with you | Server in your building; software priced per company | Not published |
Each vendor’s figures come from its own public pages, read November 16, 2026 (listed under Sources). “Not published” means we didn’t find it on those pages; ask the vendor. We don’t restate speed or quality claims we haven’t measured. Prices change; check before you buy. AdminOnsite is one of the options listed.
Buy, rent or build: what each model means
Buy a box (PrivateOfficeAI, Phoenixtekk)
One payment and the hardware is yours from day one. Support beyond the included period is an add-on, and someone in the office owns updates and backups.
Rent a managed box (Faraday Machines, AdminOnsite)
A monthly price that includes the hardware, setup and upkeep. Read the term: Faraday’s plans run 36 months; AdminOnsite has a 12-month minimum and the box is yours after 24 paid months.
Buy a GPU server (onprem.ai, Zylon, Zanus AI)
Rack or tower servers with NVIDIA GPUs, priced in the tens of thousands or by quote. Built for bigger teams and heavier workloads than most 5–30 person offices have.
Build it yourself (Mac mini or Mac Studio)
The lowest hardware cost and full control. You pick the models, install them, secure the box and keep it running. See the Mac mini AI server guide.
What three years costs a 10-person office
Simple arithmetic from the published prices above, for the plan each vendor sizes for about 10 people. Optional add-ons, taxes and your own staff time are left out.
| Option | How we got the number | 36-month total |
|---|---|---|
| PrivateOffice Box | $4,900 one time (Care at $99/mo is optional) | $4,900 |
| Phoenixtekk AI Box | $5,995 one time (support after 30 days not priced on the page) | $5,995 |
| AdminOnsite Office | $399 setup + $399 × 24 months, then Care $99 × 12 months | $11,163 |
| Faraday Machines Growth (10–19 people) | $599 × 36 months | $21,564 |
| For contrast: ChatGPT Business, 10 Standard seats | $20 × 10 × 36 months, billed annually | $7,200 |
ChatGPT Business price from openai.com/business/pricing, read November 16, 2026. A DIY Mac isn’t in the table because the configured price depends on the memory you choose.
How to choose
- Size the model first.A 30B-class model needs about 48GB of memory for several people; 120B-class models need 96–128GB. See best Mac for local LLMs.
- Decide who keeps it running.Updates, backups, user accounts and fixes. If nobody in the office wants that job, a managed rental is simpler than a box you own.
- Read the term.One-time, 12-month minimum, or 36 months with no early cancellation are very different commitments.
- Ask what isn’t published.Setup fees, support after the included period, and what happens at the end of the term.
- Compare with cloud seats honestly.For a small team happy with cloud processing, per-seat AI often costs less. Run your headcount in the per-seat cost calculator.
Where AdminOnsite fits
AdminOnsite is the managed-rental option sized for law firms, CPA and tax practices, and insurance agencies with roughly 2 to 30 people: a preloaded Mac mini, set up on your files and maintained remotely, for a flat monthly price. It runs smaller models than a Mac Studio or GPU server, so it suits drafting, summaries and file search rather than the largest open models. AdminOnsite makes no compliance certifications. See pricing, the private AI server guide and the comparison with ChatGPT Business and Copilot.
What is a private AI appliance?
A computer that runs AI models inside your office, so prompts and files are processed on it rather than on an AI provider’s servers. Vendors sell them outright, rent them monthly, or quote GPU servers for larger teams.
How much does an on-premise AI appliance cost for a small business?
From the vendors’ own pages on November 16, 2026: about $4,900–7,500 to buy a box (PrivateOfficeAI, Phoenixtekk), $249–999 a month to rent a managed one (AdminOnsite, Faraday Machines), or €19,890 and up for a GPU server (onprem.ai). A Mac mini starts at $899 if you set it up yourself.
Questions
Is it cheaper to build my own?
On hardware, yes: a Mac mini starts at $899. The cost is your time for setup, security, updates and fixes, and the risk that nobody is on call when it stops.
Which option runs the biggest models?
Among the small-office options, Faraday Machines publishes Mac Studios up to 512GB, and onprem.ai’s S2 publishes 192GB of GPU memory. Bigger memory runs bigger models.
Can I cancel a monthly plan?
It depends on the vendor. Faraday Machines publishes a 36-month plan with a 30-day money-back guarantee. AdminOnsite has a 12-month minimum. MacRun is month to month but is hosted, not in your office.
Do these vendors publish speed figures?
Some do. Phoenixtekk publishes its own measured speeds and onprem.ai publishes preliminary measurements. We don’t repeat speed claims we haven’t measured ourselves.
Why are some prices “not published”?
Zylon and Zanus AI quote each server to order, and some vendors don’t list setup or support fees. Ask them directly.
Sources, checked November 16, 2026
- AdminOnsite: Pricing (plans, setup, terms)
- Apple Store: Buy Mac mini (US), read 2026-11-16
- Apple Store: Buy Mac Studio (US), read 2026-11-16
- PrivateOfficeAI: Build your box (pricing), read 2026-11-16
- Phoenixtekk: AI Box (packages and tiers), read 2026-11-16
- Faraday Machines: Pricing, read 2026-11-16
- MacRun: Pricing, read 2026-11-16
- MacRun: Mac mini server (hosting and licence note), read 2026-11-16
- onprem.ai: Pricing, read 2026-11-16
- onprem.ai: Servers, read 2026-11-16
- Zylon: Zylon in a Box, read 2026-11-16
- Zanus AI: AI Server, read 2026-11-16
- OpenAI: Business and Enterprise pricing