NVIDIA and Sapphire join Reactor’s Series A: the bet that world models need real-time infrastructure
Reactor, the developer platform for running generative video and world models in real time, announced on Monday that NVIDIA and Sapphire Ventures have joined its Series A round, led by Lightspeed Venture Partners. The company says the new investors are coming in as demand grows for infrastructure capable of running increasingly powerful world models interactively — the last mile between a demo video and a deployable product.
The news came via a Business Wire release on Monday morning. Reactor describes its platform as three layers stacked together: a proprietary inference engine, a global GPU cloud, and a low-latency streaming layer that let developers run models at more than 60 frames per second with latency under 40 milliseconds, all through a unified SDK and API that abstracts away the complexity of deploying and scaling the models.
The bottleneck nobody demos
World models are moving beyond video generation into systems that can simulate, understand and interact with dynamic environments, the company says — but as these models grow larger and more computationally intensive, running them interactively demands significant GPU resources and infrastructure purpose-built for low-latency inference. A robot control loop that must respond in tens of milliseconds, or a live video pipeline that needs a stable 60 FPS, sits on a different order of magnitude from batch generation that can tolerate seconds of delay.
Lightspeed partner Bucky Moore put it bluntly when the round was first announced in May: “Real-time video models are currently unusable by developers due to a lack of infrastructure capable of reliably serving them,” as reported by RecodeX Pro. Reactor’s whole pitch is to own that gap.
Studios, labs, and robots
Reactor says the platform is already in use across media and entertainment, where major Hollywood studios, advertising platforms and video streaming services have active projects on it. Leading world model labs in the US and internationally also use its infrastructure, and hundreds of developers are building real-time applications on the platform, according to the announcement.
Physical AI and robotics are the other natural market. As robot policies and world models grow more computationally demanding, many require more compute than can practically run on a robot or autonomous system — so Reactor runs those models in the cloud while streaming observations and actions between the model and the physical system in real time. The platform can also operate world models as closed-loop simulated environments, Pulse2 reports — the synthetic training grounds that robot learning increasingly depends on.
What the investor names signal
Reactor did not disclose the size of the new investment; RecodeX Pro reports the company’s combined seed and Series A financing totals $59 million. The names on the round may matter more than the number: NVIDIA joining puts the company whose GPUs underpin the entire real-time inference stack onto the cap table of a startup whose pitch is that inference — not training — is the bottleneck that matters now.
The timing is telling. The world-model field has spent 2026 racing on generation quality — more realistic video, longer coherence, richer physics — while the infrastructure to actually deploy those models interactively has lagged behind. Reactor’s bet, now backed by two of the most watched names in AI infrastructure investing, is that the next constraint — and the next margin — is in serving.
For developers, the pitch is simple: one SDK and API that abstracts GPU supply, inference scheduling and streaming, so a world model can respond in milliseconds instead of seconds. Whether that abstraction holds up at real scale is the question the new capital is meant to answer. The full announcement is on Business Wire.