Sora Is Already the Answer to a Question Nobody Finished Asking
The 4D world generation moment has arrived, and the compute bill is probably what's keeping Sam Altman up at night.
the 'scale compute and world model emerges' thesis was too strong. sora's launch proved that demo quality and product quality are different things. world models need more than scale — runway's GWM-1 took deliberate architecture work, not just bigger compute.
The Gaussian splat demos have been quietly stunning for a while now — radiance fields reconstructing 3D environments from a handful of video frames, photorealistic, navigable, the kind of thing that used to require a film studio's worth of LIDAR rigs. And then Sora arrived and the question "when will we have generative photorealistic 4D worlds" collapsed into "oh, we have those now."
Not by end of year. Now. If you have access, you're generating them today.
What Sora actually is, under all the discourse about Hollywood jobs and deepfakes and responsible deployment, is a very loud proof of a very simple thesis: scale compute and the world model emerges whether you asked for it or not. The videos don't just look like the world — they simulate it. Physics, occlusion, the way light behaves when something moves through it. Nobody programmed that in. It arrived with the FLOPs.
This is the Q* thing. Not some secret reasoning module, not a hidden capability someone leaked — just the observation that past a certain threshold of compute, the model stops being a pattern matcher and starts being something that has, functionally, a theory of the physical world. Sora is the clearest demonstration of this we've had, and it's sitting right there in the sample videos for anyone who wants to see it.
Which is why Sam Altman asking for seven trillion dollars in semiconductor infrastructure suddenly reads less like a PR stunt and more like arithmetic. Someone ran the numbers on what it costs to generate a second of Sora output — and whatever that number is, it's not the kind of thing you cover with a Series B. The gap between "we can demonstrate this" and "anyone can use this at scale" is measured in fabs and power plants, apparently.
The step that nobody has quite assembled yet is combining the generation side with the simulation behavior — taking what Sora already does implicitly, the world-modeling that falls out of the training, and making it explicit. Generative environments you can actually inhabit, that evolve based on physical priors the model built up by watching enough video of the world.
We are, as of this week, approximately one architectural insight and an enormous amount of money away from that being real.
The money is apparently the harder part.
Counterpoints
Push back, extend the argument, or sharpen it. New counterpoints go through review before they show up here.
No approved counterpoints yet.