You do not need a workstation, you need 21 GB of RAM for an hour

The economics of buying hardware for photogrammetry, against the economics of renting it for the length of one job.

The standard advice for an operator who wants to process their own imagery is to buy a workstation. It is not bad advice. It is just expensive advice for a machine that will idle most of the month.

What one job needs

From a measured 167-image run: 21 GiB peak memory, 10.25 GB of scratch, 51 minutes at 8 vCPU. That is the shape of the requirement — a large, brief spike.

Hardware has to be bought for the peak. If you fly twice a month, you are buying 32 GB of RAM so it can be busy for under two hours and idle for the other 718.

The failure mode nobody mentions

An undersized machine does not refuse the job. It accepts it, runs for forty minutes, and then dies during meshing or orthophoto generation. You lose the time, learn nothing about your data, and have no output to show. Repeat that twice and the workstation looks cheap by comparison — which is how people end up buying one.

The honest comparison

Renting the peak for the length of one job costs a few dollars of compute. Buying it costs four figures once, plus the electricity, plus your time as the systems administrator when a dependency breaks.

Buy the workstation if you process constantly, want your imagery to stay on your own disks, or enjoy running the toolchain. Those are all good reasons. Just do not buy it because you think processing is cheap on hardware you already own — the measured numbers say otherwise.

This is, transparently, an argument for the thing we sell. The numbers above are the ones we measured, and you can reproduce them with OpenDroneMap for free if you would rather check than take our word.