It is the right question, and it deserves a straight answer. Cloud AI works today and costs little per seat. This page is the comparison we would want if we were the buyer: appliance, cloud, DIY, enterprise suite — including the rows where the cloud wins.
The difference is not the AI model. It is where your documents live, who can reach them, and what you are still paying in year three.
| Perimeter | Cloud AI | DIY open-source | Enterprise suite | |
|---|---|---|---|---|
| Where your data lives | Your server room | Provider's regions | Yours — once built | Vendor cloud / heavy on-prem |
| Who can reach your documents | Your staff, under your access rules | The provider, under its terms and its jurisdiction | Whoever runs the stack | Vendor admins, per contract |
| Time to working | 30 days | Instant — the cloud wins here | Months | 6–12 month rollout |
| Upfront cost | Hardware + setup | None — the cloud wins here too | Engineering time from day one | A procurement project |
| Who you need to hire | No one | No one | ~3 ML engineers | A project team |
| Cost shape | One box + annual licence | $30–90 per user per month, forever | Salaries, forever | $100k+/year floors |
| Verifiable in person | Yes — pull the cable | No | If you built it right | Rarely |
| Works without the internet | Yes | No | Yes, if fully self-hosted | Usually no |
| Update path | Model updates included in support | The vendor updates on its schedule | Your engineers, ongoing | Vendor roadmap, change requests |
| What happens when you leave | The box is yours. It keeps working | Access ends when payment stops | You keep the stack — and the upkeep | An exit and migration project |
Read the cloud column fairly: it wins on instant start and zero upfront cost. If your documents could be public anyway, it is the rational choice. The comparison changes the moment the documents are the business — bid pricing, client files, supplier terms.
Cloud AI is priced to look small. Multiply it by your headcount and by the years, and the shape changes.
Per user, per month — the going rate for cloud AI assistants and copilots.
List pricing, mid-2026A typical mid-market floor. That is $3,000–9,000 every month — $36,000–108,000 every year, every year, and none of it ever becomes yours.
Arithmetic, not a forecastPerimeter is one box plus an annual licence. The cost does not scale with headcount, and the hardware is yours from day one.
No per-seat meterDIY stacks on open-source are real, and some are good. They are also an engineering project: a server, ingestion pipelines, retrieval tuning, evaluation, monitoring, model updates — and someone on call when it breaks.
Our honest rule of thumb: if you can dedicate about three ML engineers to it, building can make sense — you get exactly what you design, and it is yours. Below that line, the stack becomes a second job for whoever owns it, and it competes with the work the company actually sells.
Perimeter runs the same class of open-weight models a DIY stack would. The difference is who carries the project: we ship the box working, and support keeps it current.
Cloud AI is a very clever stranger. A DIY stack is a second job. An enterprise suite is a mortgage. Perimeter is an appliance: it arrives, it works, it stays inside your walls.
Get the one-page matrix for your IT and your MD — or skip the paper and see Perimeter answer on your own documents, live.