FastFrames speeds up video generation, while the paper's controlled 1.3B models reach about 50 milliseconds only on eight GPUs. The economics hinge on how much interaction each rented chip can serve.
Four RGB frames once consumed 147 milliseconds just getting out of a video decoder. Apple's FastFrames researchers cut that step to 1 millisecond. Their paper, submitted 17 November 2026 by seven Apple-affiliated authors, one also at Cornell University, puts the smaller model at 36 milliseconds per video response on one Nvidia B200. We would pass on a business priced as though fast video already means cheap interactive worlds. A camera that moves demands another picture, then another, with the machine rented throughout the session. The stopwatch matters because someone has to pay for every response.
The Control Tax is the extra response time and occupied GPU capacity required to make generated video react to camera movements, actions or physics. The 36 ms headline captures only part of that cost. At 832 by 480 resolution, FastFrames-S has 1.3 billion parameters and responds in 36 ms on one B200; FastFrames-L has 14 billion and responds in 45 ms across eight B200s. The authors call the smaller result 28 responses per second, counting each generation unit once regardless of how many video frames it contains. The FastFrames performance figures here are the authors' own B200 measurements. We care about response to a command and the machine time required to supply it.
Consider where the time went before the fix. In the authors' Causal Forcing++ baseline, two diffusion steps took 83.2 ms, cache writes 41.6 ms and decoding 147 ms. With FastFrames, those three jobs take 32.8 ms, 2 ms and 1 ms. The whole response falls from 271.8 ms to 35.8 ms; the rounded 272 ms over 36 ms is a 7.6 times speedup. On the authors' VBench quality measure, the baseline scores 84.11 and FastFrames-S scores 84.23. Speed arrived through changes to decoding and memory operations as well as the model's sampling. Cheaper work per response makes a machine more valuable to anyone selling its time.
A video of a world is easier than a world that listens. In the controlled tests, camera-conditioned LingBot 1.3B takes 224 ms per response on one B200, while physics-conditioned RealWonder takes 153 ms. Eight B200s bring them down to 52 ms and 51 ms respectively. The researchers target under 50 ms, drawing on studies placing human delay detection around 50 to 60 ms. Even eight B200s leave the 14B LingBot model at 233 ms. Those extra sampling and conditioning operations turn the same latency ambition into a much larger hardware purchase. A service that offers steering must charge enough for continuous machine occupation.
The useful financial unit is an hour of live interaction. On 18 November 2026, Lambda listed a B200 at $7.69 per GPU-hour; eight cost $7.69 times eight, or $61.52 per hour. Decart listed Oasis 3 Preview, a real-time driving world model, at $0.02 per active second, or $0.02 times 3,600, which is $72 per hour. In a scenario with one active stream occupying eight GPUs for that full hour, $61.52 over $72 is 85.4% of the selling-price reference. Oasis runs a different workload, so this is hypothetical rental exposure rather than Decart's operating margin. The operator pays for an eight-chip hour long before that hour looks like a profitable product.
Generated worlds were interactive before Apple's 36 ms result. Google DeepMind says Genie 3 allows users to "navigate in real time at 24 frames per second" at 720p for a few minutes. Decart already posts a price for active time in a generated driving world. A buyer of spare hardware need not pay the usual rental price either. On 18 November, CoreWeave listed an eight-B200 spot node at $34.11 per hour. Against the same $72 hourly selling reference, $34.11 over $72 is 47.4%, leaving space for other costs and profit in the one-stream scenario. Buying discounted capacity and selling continuous interaction is a credible path to a category that already has a price.
A spot node is a different bargain from capacity a service can hold. CoreWeave's on-demand eight-B200 node costs $68.80 per hour; $68.80 over the $72 selling reference is 95.6%, while Lambda's on-demand rental absorbs 85.4%. The spot quote saves $68.80 minus $34.11, or $34.69 per node-hour, but spot capacity can be reclaimed. A continuous service has to schedule around that possibility while keeping its controlled model responsive. In the paper's tests LingBot reaches 52 ms only on eight B200s, close to the 50 ms target. Until a service packs more paying interactions into its rented capacity, a low spot quote is a pricing opportunity rather than a durable margin.
The authors' own latency history makes it dangerous to underwrite eight GPUs forever. On one B200, Self-Forcing took 717 ms, Causal Forcing++ 272 ms and FastFrames-S 36 ms. These methods belong to the same autoregressive family, and the hardware stayed at one B200 while the latency collapsed. LingBot itself improved from 1,085 ms to 224 ms on one B200 with FastFrames. The analogy breaks where control adds work: 224 ms over the 50 ms target is still 4.5 times that target, and RealWonder sits at 153 ms on one chip. Software has already shrunk the latency substantially. Paying for the next speedup before it arrives still puts that risk on the investor.
Decart's model documentation describes Oasis 3 as a driving simulator usable for reinforcement learning. Its models are billed for seconds of active generation, turning time spent interacting into the product. A customer pays for responsiveness over a session; the supplier has to keep its expensive machines occupied with billable work. Our Control Tax shrinks when controlled response gets faster on one B200 or paid sessions share a node efficiently. The FastFrames researchers list action conditioning without added latency, higher resolution and lighter hardware as future work. The meaningful product metrics are sustained controlled latency and billable GPU-hours per active stream.
We would pass on valuations that capitalize 36 ms text-to-video latency as if it were already economical real-time control; we would pay for sustained controlled interaction with demonstrated GPU economics per active stream. We would be wrong if a public paper or code-backed benchmark by 31 March 2027 demonstrated camera-conditioned or action-conditioned video at 50 ms or less on one B200, overturning the eight-GPU latency requirement at the center of our Control Tax.