What it actually costs to build this yourself

A capable team can build this. That is worth saying plainly, because most pages like this one pretend otherwise. The archives are open, and a capable team can absolutely work with them. rest are public, documented and free to download. The algorithms are in the literature. The libraries are on GitHub. Hire two good remote sensing engineers, give them cloud credits, and within a month you will have a scene on a screen with an index rendered over it. That part is genuinely straightforward, and it feels like most of the problem.

It is roughly ten percent of the problem. The rest sits below the waterline, and it is where projects stall: the atmospheric correction that fails over dark water and steep terrain, the cloud mask that misses thin cirrus, the spectral library entries that turn out to be impure samples, the classifier that is confidently wrong in the one place nobody checked, the validation protocol that has to exist before anyone senior will act on the output. None of these are research problems. They are all grinding, unglamorous engineering, and each one has to be finished before the layer above it means anything.