Decart’s Lucy 2.5 unlocks new live video effects, but physical AI is the real winner

This is more of a compute-demand signal than a monetization event. Real-time, continuous video generation is materially more inference-intensive than one-off image or offline video generation, so if the technology is even partially production-grade it increases the ceiling on GPU-hours per user session and pulls demand toward low-latency infrastructure, networking, and memory bandwidth. That is structurally supportive for NVDA, and second-order positive for cloud capacity providers with strong AI attach rates, while it is a mild overhang for traditional post-production software and outsourced VFX workflows if creators can substitute software for labor.

The near-term market risk is that demos get mistaken for scalable economics. If latency, artifacting, or cost per minute remain high, the addressable market stays niche and the public-market read-through fades within days. The more important 1-3 month catalyst is whether we see any evidence of repeat usage, enterprise pilots, or infrastructure partnerships; that would indicate the feature is not just a research showcase but a demand driver. Over 6-18 months, the real question is whether real-time generative video becomes a new class of GPU workload with subscription economics, which would expand the long-duration revenue pool for semis and hyperscalers.

Contrarian view: the consensus may underappreciate how compute-hungry "interactive" AI really is. Even if the product is technically impressive, unit economics can deteriorate quickly at scale, and that often leads to heavy usage caps, thin gross margins, or a pivot toward enterprise-only workflows. If the company cannot show sustained session length without quality decay, the market should treat this as an innovation headline, not a revenue inflection.