Nah, local LLMs are easily in the range of transcribe/summarize. I bet you could do that nicely with llama 8B without even needing a gpu.

By that point I'm pretty sure we'll have an effective compact model that can run locally and transcribe downloaded videos on reasonable hardware. Or you can just sic a paid model like chatgpt on the task. The corporate Internet is entirely focused on subscription service models now, unless you run the model yourself on local hardware you're going to end up paying someone somewhere a service fee.

Edit: y'all need to learn about minified models designed to run on edge hardware, they're a thing and often work shockingly well.

Local and open source