Take the promotional sheet at face value and the arithmetic is unusually clean: 12.5 cents buys five seconds at 480p [10], which is two and a half cents per second of finished footage [1], a dollar fifty a minute [2], ninety dollars an hour [3], and $2,160 for a channel that runs a full day without stopping [3]. Anything above 480p doubles [10], so the same day costs $4,320 [4]. Hold the lower tier for a year and you have spent $788,400 [5] on footage nobody commissioned.
The latency number does something different to the cost number. Fifteen seconds arriving every nine [1] means the renderer is occupied 60% of wall-clock time [6], so a single always-on channel leaves four tenths of a pipeline idle, the same capacity could in principle carry about 1.67 concurrent streams [6], and a full day of programming takes 14.4 hours to manufacture [8].
Which is where the source material argues with itself. The same write-up that reports 15 seconds in 9 also gives the window as 9 to 16 seconds depending on resolution and settings [7], and at 16 the model returns 0.94 seconds of footage per second of compute [7], so the day needs 25.6 hours [7] and the channel falls behind its own schedule. Faster than playback is a configuration, or rather the fast end of a range with a demo attached. The short-clip figure is the sturdier evidence: under three seconds for five seconds at 720p or 768p, roughly 2.5 in some tests [6], which is two seconds of output per second of compute [9].
Worth naming what fal did not spend on. H3 Max is a closed, post-trained iteration of MiniMax H3 [4], an open-weight multimodal model that appeared in July [5], so the budget went into inference optimisation over somebody else's weights rather than a pretraining run, and Crypto Briefing's reading of the 35x throughput claim [8] is that architecture rather than additional hardware produced it [13].
Three ways this goes. The promotional rate holds and continuous video becomes a metered utility that anyone with a card can buy by the hour; or the usable resolution sits at the doubled tier and nearer the 16-second end, in which case the crossing evaporates outside 480p; or demand caps out because acceptance does, since the render cost fell to cents while a platform's decision to carry the stream stayed binary and both Twitch and Kick said no [11].
This is probably wrong, but the always-on channel looks like the least valuable of the three uses on offer [14], and the one with money in it is the advertising pipeline with no rendering queue, because latency under playback means bespoke creative gets made inside a session rather than booked against a farm. I would drop that view if the first-place human-preference result [9] turns out to hold only at the resolution tier that costs twice as much, which would put the interesting quality and the interesting speed on opposite sides of the price list.
