The cost floor for video AI just fell. The moat moved up the stack.

The cost floor for AI-generated video just fell, and it fell because of an architecture choice, not a chip. Sand.ai's MAGI-2-preview is a 100-billion-parameter video model that activates only about 6 billion parameters per forward pass. That ratio is the story: 6 billion active parameters can run on commodity GPUs at roughly 5 U.S. cents per 10-second 1080p clip, about one-tenth of what closed APIs have charged. The 5-cent figure is pegged to current 8x H100 cloud spot pricing, not a consumer rate.

Most readers will see Sand.ai's release as another open-source drop. The pattern underneath is sharper. Mixture-of-Experts decouples total model size from per-generation compute, and Sand.ai open-sourced the recipe: weights, code, and a custom MagiMoE kernel. The model itself stops being the moat. The moat moves to the integration layer: the data pipelines, fine-tuning, product surfaces, and distribution that wrap a now-cheap model. (This is a bounded inference: making the model's weights and code freely available structurally shifts competitive advantage toward integration and distribution rather than the base model itself.)

Sand.ai says the model ranks sixth on the AA video leaderboard with only 6 billion parameters active, performance close to closed-source first tier. That is a Sand.ai submission, not an independent rerun. The preview tag, the cloud-cost dependency, and the single-benchmark claim are the real open questions. What is settled: the cost of generating a 10-second clip stopped being a frontier-model budget item. Indie studios, education teams, and non-English-market startups can now build on top of a 100-billion-parameter video model the way they already build on top of open-source text models.