Trained to Imagine
Rendering
Creative Intelligence
AI Video / Image Models

Real-time AI video startup Reactor raises $59M

Source: Variety ·

Reactor, founded by former Apple engineers, raised $59 million from Jeffrey Katzenberg's WndrCo and others to build real-time AI video generation.

TTI Takeaway

Real-time is the threshold that changes the creative process rather than just the cost line. Generation fast enough to sit inside a review session turns iteration into direction — you stop submitting prompts and start giving notes. That is a different craft, and it favours people who can actually give notes.

Analysis

The most telling detail in Reactor’s funding announcement is what the company does not do: it does not train its own foundation models. Reactor takes models built by other labs and optimises them to run in real time, then exposes them through a single SDK and API. That makes it a serving layer rather than a model company, and it is a bet that the next bottleneck in generative video is latency and infrastructure, not image quality.

The specifics, as Variety reported them: a $59 million Series A led by Lightspeed Venture Partners, with WndrCo, Amplify Partners, Sky9 Capital and FPV Ventures participating, and Jeffrey Katzenberg joining as a board observer. The founders, CEO Alberto Taiuti and CTO Bryce Schmidtchen, were technical leads on Apple Vision Pro and started the company in August 2025. It has 16 employees, says it has worked with more than a dozen models from different labs, counts the world-model lab Overworld as a customer, and runs on AWS compute. Taiuti contrasts batch text-to-video systems that can take up to ten minutes to produce ten seconds of footage with a platform where the time to first frame is “basically nil” and output can run for as long as you like.

For production, the interesting part is not speed for its own sake but where speed moves the work. A ten-minute render loop belongs in a queue: someone writes a prompt, waits, reviews, writes another. Generation that answers inside a conversation belongs in the room, alongside a director, a production designer or a client on a call. That has obvious uses in look development, previs and pitch sessions, and less obvious ones in virtual production, where an interactive generated background would have to sit next to tracked cameras and real-time engines. Taiuti’s line that studios “want to train their own models internally” also points to a specific gap: a studio that builds a model on its own library still needs a way to serve it to its teams at interactive speed, and that is the layer Reactor is pitching.

There are reasons for caution. The article describes no output resolution, no frame rate and no named studio deal, only talks. The one demonstration it does describe, turning an AWS executive’s face into Albert Einstein’s live on a Zoom call, is exactly the kind of capability that collides with likeness rules. The SAG-AFTRA agreement settled the same month sets payment and consent terms for digital replicas, and any real-time face or performance transformation used on union work will be judged against those terms regardless of how fast it runs.

Our view is that Reactor belongs to a cluster of signals pointing the same way. Runway’s GWM Worlds 2 in September streamed continuous 720p video at 24fps with no preset session length, and Google’s Genie 4 in October added real-time steering. Runway’s model router and Amazon’s model-agnostic Project Nara make the same structural argument Reactor does: no single model will win everything, so value accrues to whoever routes, serves and integrates them. What to watch is whether real-time output reaches a quality bar that survives a grade and a client review, or whether it settles as a fast sketching layer with final pixels still rendered the slow way.