Workload planner
Where should this AI workload run?
Four questions about the work and the data. No email address, no signup — it runs entirely in your browser and we never see the answers. It is the same rough triage we would do out loud in a first conversation.
The rules behind it
All of it, on one page
There is no hidden model here. These two tables are the entire logic — the tool above just looks up the row your answers land on.
Where a workload runs
| Verdict | When | What it means |
|---|---|---|
| Run it on your own network | The data must never leave your network — or it is internal and the volume is modest. | The data rules decide this one before the technology gets a vote. |
| Managed private processing | The data is internal rather than restricted, and there is a lot of it at once. | The work is heavy now and light later. Buying for the peak is waste. |
| Split it by collection | You are holding both sensitive and already-public material. | You are holding two different kinds of data. Treat them differently. |
| Run it wherever is cheapest | The material is already published or is not sensitive. | There is nothing here to protect, so optimise for cost and speed instead. |
Roughly what it takes to run
| Hardware class | What that means in practice |
|---|---|
| Hardware you already own | An existing server or workstation, working overnight. Very often there is no purchase at all — just something useful to do with a machine that is idle between six and eight. |
| A workstation with a consumer GPU | One desktop-class machine with a gaming-grade GPU in it. This is the single most common answer for a first project, and it surprises people every time. |
| One dedicated GPU system | A purpose-built machine that does this job and nothing else, sized for the model you actually need rather than the largest one available. |
| A dedicated GPU system sized for concurrency | Enough memory and throughput to serve several people or several jobs at the same time without any of them waiting on the others. |
| More than one system | Multiple machines working in parallel — or, where your data rules permit it, short-term rented capacity to clear a backlog with something smaller left behind for the steady state. |
Things that change the answer
- Real time is the expensive choice, and it is chosen far more often than it is needed. If the answer can wait until the morning, the same work usually fits on a much smaller machine.
- A large backlog with a relaxed deadline is the best-value AI project there is. It is heavy once, then finished forever, and it can run on borrowed capacity.
- This is the combination that costs money: everything has to stay inside, and it has to be fast. Worth checking whether the whole workload really needs both, or only part of it does.
- An interactive assistant is the workload where model size shows most, because a person is judging every answer. Budget accordingly, and consider starting with search over your documents before adding a conversational layer.
- At this size the honest answer is often that you do not need to buy anything. We would rather tell you that now than after a purchase order.
What would you use AI for if your data stayed under your control?
Tell us what you are trying to accomplish. We will help work out what should run locally, what can run elsewhere, and what technology actually makes sense.
You do not need an AI strategy. You need a problem worth solving.
Show us the repetitive work, the archive nobody can search, the recordings piling up, or the process that looks automatable. We will work backward from there.