Scaling world models and real-time video inference

Name: Scaling world models and real-time video inference
Start: 2026-10-15

What if users could change a video as it unfolds—explore a generated world, introduce new events or choose what happens next?

Watch Reactor and Nebius explore what world models make possible for gaming, interactive storytelling and creative applications. See examples and learn how Reactor’s inference stack runs on Nebius to support continuous, interactive sessions.

Reactor CTO Bryce Schmidtchen explains the runtime, streaming and session management behind these experiences. Nebius Product Lead Roman Luchkov discusses the GPU infrastructure and recovery mechanisms that support them.

Watch the recording to learn how to:

Evaluate world models for persistence, physical consistency and responsiveness.
Manage session state, streaming latency and global routing.
Scale concurrent sessions while controlling GPU usage.
Assess infrastructure by cost per successful session.

For technical product leaders, application builders and ML infrastructure teams.

Watch now

Abhishek Ratna

Head of Marketing, M&E, Nebius

Bryce Schmidtchen

Chief Technology Officer, Reactor

Roman Luchkov

Product Lead at Nebius

Register to get the recording

Key takeaways

A world model predicts how an environment changes based on its current state and incoming actions. In the interactive video examples discussed here, users influence what the model generates next by moving through a scene or introducing new events.

The session shows how this can support explorable worlds and generative storytelling. Persistence, physical consistency and responsiveness remain important evaluation criteria: a visually convincing scene does not automatically behave like a reliable simulation.

Batch video generation can return a completed result after processing. An interactive experience needs a continuous loop: receive user input, update the model’s context, generate the next output and stream it back quickly enough for the interaction to feel responsive.

This requires coordination across inference, networking and session management. The session explains why optimizing GPU throughput alone is insufficient when users also depend on consistent frame delivery, low interaction latency and continuity over long sessions.

A session carries the context needed to continue a particular user’s experience. The serving system must manage that context throughout generation and decide how to preserve or restore it when a connection is interrupted.

Reactor describes a runtime with session management and streaming transport built in, alongside an inference engine that manages model context and caches. Preserving this runtime state supports continuity, while the model’s ability to remember objects and maintain a coherent world remains a separate quality challenge.

Responsiveness depends on the complete path between the user’s action and the displayed result. Model execution, scheduling, video encoding, transport and network conditions all contribute to the delay.

Reactor discusses custom inference kernels, streaming transports such as WebRTC, and routing informed by live network conditions. The geographically closest GPU region is not always the fastest network destination. Teams need to evaluate the complete interaction loop for their users and application requirements.

Serving more users requires scheduling work while keeping each session’s context separate and its response times acceptable. GPU utilization must be balanced against the latency requirements of active sessions.

Reactor describes context management, state snapshots and restoration, and scheduling model execution across users. Admission control also matters: when capacity is constrained, the system may need to reject a new session or route it elsewhere rather than degrade every active user’s experience.

Reactor handles the application-facing runtime, streaming and session lifecycle, while Nebius provides the underlying GPU infrastructure and cloud orchestration. The session places Reactor’s serving stack on Nebius compute and managed Kubernetes.

Nebius describes hardware acceptance testing, ongoing health monitoring, fault isolation and automatic node replacement. Infrastructure recovery and session recovery address different problems: restoring compute capacity does not, by itself, restore a user’s model context or guarantee an uninterrupted experience.

GPU-hour pricing is only one part of the cost. Teams also need to account for useful output, performance overhead, setup, maintenance and recovery. The webinar frames this in terms of goodput and cost per successful session.

For an interactive product, that means assessing how much usable experience the infrastructure delivers at the required quality and responsiveness. A lower GPU price may offer less value if slow execution or interruptions reduce the number of sessions it can support successfully.