How did you get 3000? That's a wildly inaccurate and out-of-this-world figure. I did some preliminary research recently that I'm not fully recalling (would have to look for it), but I had a baseline of something like 9 minutes of streaming video using the equivalent resource (electricity, etc) of just over 5 average prompts. So 3 hours of streaming video is a bit over 1000 prompts.

I can get an OK tool that still goes a good distance in scratching an itch in say ~3-8 prompts, so with 1000 I'm looking at a small app. With 100000... I'm finding it a bit hard to imagine TBH; somewhere in the region of the Bun port to Rust[0] is my likely-off-base-but-not-as-bad-as-yours guess. Heck just consider that there are devs out there running systems with local models that are probably more than half the size of a frontier model, and how that 3000-hours-equivalent cost would've been eating into their pockets in the form of electricity bills.

[0] https://bun.com/blog/bun-in-rust

Streaming is not computationally expensive at all. The computing and networking cost is so trivial it's difficult to measure, meaning most of the cost is in the cost of electricity your TV, computer monitor, or phone consumes with the display on. Some of this cost may even be rendered redundant if you were leaving the display on anyways; I certainly don't turn my PC monitor off when I'm done watching a video, even if I'm letting it idle for a while. At any rate, the typical device will be in the range of 0.1kwh per hour of streaming. Perhaps substantially less if most people watch on their phones these days.

"Average prompts" is a meaningless measurement. An 8x B200 GPU node, which is capable of running Deepseek V4-Pro, consumes in the ballpark of 300kwh per day while producing over 300 million tokens per day when operating at production loads, so we can get a roughly clean estimate of 1kwh per million tokens from a near-frontier open model. The project in the OP, which we'll take to be "respectably-sized", is 100k loc. How many tokens it takes to arrive at 100k loc is the most hand-wavy, difficult part to quantify here, but looking at my own usage, I've spent about 800 million tokens in the past two weeks on a 25k loc side project[1], and that was with full human-in-the-loop management, single agent usage. If we extrapolate that to 100k loc, we're looking at ~3.2 billion tokens, around 320kwh. Note this is for Deepseek, which is much cheaper to run than the frontier models. Inference on Fable or GPT 5.6 is likely another order of magnitude higher power cost, but those numbers are not disclosed, and we can only make rough inferences from their API token pricing (which also includes undisclosed amount of profit margin).

Anyways, 320kwh for a 100k loc project vs. 0.1kwh per hour of streaming comes out to my ballpark estimate of 3000 hours, but that was at Deepseek rates. It might actually be closer to 30,000 hours of streaming if this was done on Fable.

Note that the Bun project you cited burned $165,000 in tokens at their API rates. Perhaps you can look up the price of electricity in your region and do the math for yourself on how many hours having your TV on that would buy you. I'm estimating around 10 million hours, which sounds in line with my previous math: it's 1m loc written with Fable, so 10x the 30k number for 100k loc, with another magnitude for wasting tokens on swarms of agents checking each other and consuming many more tokens per loc than you would with human management and review. Even if you assume Fable API rate comes baked in with a 50% profit margin on inference, that would still leave you with 5 million hours of streaming for the electricity usage of the Bun rewrite.

[1] To be honest, I'm kind of shocked at how much electricity I've blown on this project, after having done the math! It actually resulted in something I'm shipping and maintaining with hundreds of real-world users, at least. The project would have been easier without LLMs, which have bungled so many things and taken so much effort to correct that it would have been significantly easier to do 90% of the work myself and leave only some data-munging to the bots, but I try to stay in tune with the absolute limits of what frontier models are capable of in real-world scenarios.