Compare
Three ways to get a serious AI assistant.
Rent one from a cloud provider, build a local server yourself, or buy a Spot. The honest version of each, on the dimensions that matter once the novelty wears off.
| Dimension | Spot | Cloud AI subscription | Build your own local server |
|---|---|---|---|
| Where your data lives | On a device you own, on your network | On the provider's servers | On your hardware |
| Who can see your prompts | Nobody but you | The provider, under its policy | Nobody, if you configure it correctly |
| How you pay | Once, for the device; no per-token bills | Monthly per seat, often plus usage | Parts and your time, then power |
| Set-up | Plug it in and say hello | Create an account | Pick parts, install drivers, serve models, keep it running |
| Works offline | Yes: core chat, voice and automations | No | Yes, if you built it that way |
| Several people and models at once | Yes, on one box | One seat per person | Depends on the hardware and your setup |
| Apps on every screen | Desktop, phone, watch and web included | The provider's apps | Whatever you assemble |
| Mail, calendars and the smart home | Built in, processed on the box | Varies; data flows through the provider | Build and maintain each integration yourself |
| Acting on your behalf | Agents run on the box and wait for your yes | Varies by provider | Build it yourself |
| Updates and models | Signed over-the-air updates | The provider decides | You manage them |
| Support | The engineers who build Spot | Varies by plan | Forums and yourself |
What the table does not say
A cloud subscription is the right answer when you want the largest proprietary model for occasional use and your data is not sensitive. A do-it-yourself server is the right answer when assembling and maintaining it is the hobby. Spot is for everyone who wants frontier-class capability, several people using it at once, and data that never leaves the building, without a weekend of setup or a bill that grows with every prompt.
Spot runs frontier-class open-weight models on a compute module with up to 128 GB of unified memory, keeps several loaded at once, and still works when the internet does not. The specifications and the FAQ have the detail.