Thirty-five times against the official endpoint is a ratio, and a ratio needs a denominator you can inspect [1]. The AINews writeup does not supply one: no resolution, no frame budget, no batch size, no GPU count, no latency target for the reference endpoint [2]. A vendor's default inference path is normally tuned for cost across mixed traffic, which makes it the softest baseline in the building. The post-training and the engine are both fal's, so anyone reproducing this from the public H3 release starts two components short [1].
The relevant number is the point where frame rate crosses into watchable, not the multiplier by itself. Consistency models had already compressed a 30 second generation to about one second, a 30x cut [3][8], and one second per generation buys you one frame per second of wall clock [3]. Take 24 fps as the floor for something a person will sit through, which is my assumption and not the source's, and you need another 24x on top [9]. Fal's claimed 35x clears that with roughly 1.46x to spare [10]. That spare capacity is what "faster-than-realtime" is describing [7]. The two multipliers do not compose: the 30x was consistency-model work on earlier systems, the 35x is engine work on H3 [1][3].
The stream itself is where the product broke down. Once generation is continuous, the artifact is a stream, and a stream needs a player. Fal employees put theirs on Twitch and YouTube, and both platforms removed it immediately [4][5]. The account was removed without a cited policy provision or a stated reason from either platform, and there was no appeal [11]. This limit differs from the throughput ceiling fal had just solved: an engine rewrite can raise a throughput ceiling because fal controls the engine, but fal does not control the platforms' account decisions, and no configuration flag on fal's side reaches across that boundary.
Fal's response was to run its own live video service [5]. That means absorbing what the platforms had been providing at no charge: playback, moderation, abuse reporting, and egress for a feed that never ends. Those costs scale with viewer-hours rather than with generation, so the inference win does not offset them. Anyone planning a continuous generative-media product should price the distribution stack at design time, because it is now the component with a third-party veto attached.
On quality, AINews is blunt: pure slop, a mishmash with no plot and low-quality RL-tuned imagery, and nobody will watch it [6]. I will take the reviewer at their word. The load-bearing claim here is the frame rate, and it holds whether or not the content is watchable [7]. What it does not survive is a platform treating infinite generated video as its own category, which is a policy decision that no inference engine gets a vote in.
