On-device vs cloud AI is routing, not religion
Local analysis is free and private by construction, cloud is faster and reads harder frames, and the right pick changes per batch.
The local-versus-cloud argument usually gets run like a church schism, and it deserves better. It is a routing decision. Some batches belong on the Mac in front of you, some belong on bigger hardware, and the same person will make both calls in the same week for reasons that have nothing to do with ideology. Full disclosure: I build ClipLogger and we sell cloud compute, so you know where my incentives point. Judge whether the reasoning stays honest anyway.
What local buys you
Two things, and both are real. It is free: no meter runs, so the 3,000-clip drive from 2019 that nobody budgeted for can grind away overnight, every night, for the cost of electricity. And it is private by construction rather than by policy: the frames never leave the machine, so there is no retention window to read up on and no promise to take on faith. The construction is the promise.
The price is time and ceiling. Your Mac works at your Mac's pace, and the models that fit on it are smaller than the ones that do not. For most frames on most days that is fine, which is exactly why ClipLogger's local pipeline is free forever with no account: for a large share of real work, local is the correct route. Overnight is a feature here, not a workaround; the Mac was going to sit idle anyway.
What cloud buys you
Speed and reach. Rush, our cloud, runs a batch through parallel workers on larger models, so the 400-clip Saturday is reviewable after lunch instead of tomorrow morning. Those larger models also read hard frames better: the scoreboard behind glare, the number half-hidden by a shoulder pad, the whiteboard at the back of a wide shot. When an editor is waiting on selects tonight, that difference is the job.
There is also a middle route: your own API key, with frames going straight from your Mac to your chosen provider on your own account. No middleman, and you pay the provider directly. It is the route for people who already carry provider accounts and want the bill to land on them.
None of this makes cloud the villain, and I would say that even if we did not sell it. The principle that keeps it honest: you pay to go faster, never to get your own work back. Every engine drops its results into the same review, and confirmed values land in the same open sidecars either way.
The word that matters is silently
Here is the line I would hold any tool to, ours included: frames may leave the Mac, they must never leave it silently. Some footage carries stakes. A documentary subject with safety concerns, an unreleased product, a minor in the frame. The rule for that footage is not "never cloud", it is "never cloud by accident".
Concretely, that means no silent fallback from a local engine to a cloud one, an engine and a cost disclosed before every run, and provenance stamped on every value afterward, so a cloud-derived value stays identifiable per clip forever. ClipLogger shows the routing before you run and records it after. The standard I would apply anywhere: you should be able to name, from memory, every route by which frames leave your machine. If you cannot, the tool knows something you do not, and if a tool cannot tell you which engine produced a value, the routing decision was made without you in the room.
Choose per batch, not per ideology
The routing call takes ten seconds and four questions. Who is waiting: an editor who needs selects tonight justifies credits, an archive nobody has opened since 2019 does not. What is in the frames: anything you would hesitate to put in an email stays local. How hard are the frames: clean interviews read fine locally, busy scoreboards and fast glass earn the bigger model. What does the budget say: a number you check, not a mood.
Ask them per batch and the week sorts itself. Rush the game cards while the editor waits. Let the archive drive churn locally overnight for free. Keep the sensitive interview on the Mac, by construction, no policy required. Most real days are mixed, and that is the argument in miniature: the A-camera selects earn credits because someone is waiting on them tonight, and the B-roll runs locally while you sleep because nobody is. The rubric works in whatever software you run; the only requirement is a tool that lets you route per batch and shows you afterward where each value came from.
Route each batch on who is waiting, what is in the frame, and how hard it reads, and leave the religion out of the pipeline.