Pomegra
Reactor, the San Francisco startup that runs real-time generative video and world models, added about $15M from Nvidia's NVentures and Sapphire Ventures. That brings its Series A to $74M.
- Reactor added roughly $15M to its Series A, lifting the total to $74M from $59M in May 2026.
- NVentures and Sapphire Ventures joined; Lightspeed Venture Partners remains lead. Valuation was not disclosed.
- Reactor's platform streams world models at 60+ frames per second with latency under 40 milliseconds.
Lead
Reactor announced on October 5, 2026 that Nvidia's venture arm, NVentures, and Sapphire Ventures have joined its Series A. The extension adds roughly $15M and brings total funding to $74M. When the company left stealth on May 28, 2026, that figure stood at $59M.
Lightspeed Venture Partners led the original round and keeps that role. Reactor did not disclose the size of the extension, how much each new investor wrote, or the company's valuation.
What Does Reactor Actually Build?
Reactor builds the infrastructure that serves interactive generative video and world models to users in real time. It does not train the models. The platform combines a proprietary inference engine, a global GPU cloud and a low-latency streaming layer, exposed through a single SDK and API.
The company says it runs models at more than 60 frames per second with latency below 40 milliseconds. Pricing is usage-based and billed by model type, and AWS is the preferred cloud partner for the streaming workloads.
CEO Alberto Taiuti, a former CTO at Luma AI, has described the founding question simply: once a model exists, how do you run it? Co-founder and CTO Bryce Schmidtchen shares his background as a former technical lead on the Apple Vision Pro.
Who Is Paying for It?
Reactor lists major Hollywood studios, advertising platforms, video streaming services and world model labs in the US and abroad as customers. It also reports hundreds of individual developers on the platform. Target markets include media, entertainment, gaming, robotics and physical AI.
Hollywood has a direct stake. WndrCo, the firm run by DreamWorks co-founder Jeffrey Katzenberg, invested in the first close, and Katzenberg joined as a board observer. Other earlier backers are Amplify Partners, Sky9 Capital and FPV Ventures.
The customer list is a claim from the company. Reactor has not named individual studios or published revenue figures.
Why Would Nvidia and Sapphire Invest in a Serving Layer?
A model that renders each frame in response to user input is expensive to run, and it cannot batch requests the way a text chatbot can. Each session holds a GPU for its duration and must hit a frame deadline, or the experience breaks. That makes serving cost and latency the practical limit on deployment.
For Nvidia, a startup that drives GPU utilization for a workload that needs sustained, low-latency compute is a natural fit. The chipmaker also builds its own Cosmos world models and Isaac robotics tools, so Reactor sits next to its product lines. NVentures has not said whether the investment carries any commercial arrangement.
Anders Ranum of Sapphire Ventures framed the thesis as an enterprise problem: every enterprise team eventually runs into the same wall, the infrastructure required to serve systems in real time. Ahmed Ahres, Reactor's head of go-to-market, said developers a year ago asked whether real-time world models were possible and now ask how quickly they can reach production.
What Does a $15M Extension Say About the Last Round?
The extension is small next to the $59M first close, about a quarter of the new total. Reactor kept the Series A label rather than opening a Series B, which suggests the company wanted strategic names on the cap table more than a large new check. It also avoids setting a fresh price in public.
The timing is notable. Reactor added investors roughly four months after leaving stealth, and the new money came from a chip supplier and an enterprise software investor. That is a different signal from a pure financial investor re-upping, though without a disclosed valuation, the market cannot judge whether the terms moved up.
How Crowded Is the World Model Field?
Most capital in world models has gone to the labs that build them. Runway has released an open-weight world model, and Nvidia continues to expand Cosmos. Infrastructure for serving those models is a thinner category, and Reactor is betting that labs would rather rent a streaming stack than build one.
The risk is that the largest model developers or cloud providers fold real-time serving into their own offerings. Reactor's counter is that a neutral layer across many models is easier for studios and robotics teams to adopt than separate stacks for each lab.
Outlook
Reactor now has $74M, a strategic chip investor and an enterprise-focused fund alongside a lead investor that has backed it from the start. The open questions are commercial: which studios are paying, how usage-based revenue scales, and whether GPU costs per session fall fast enough to support consumer-facing products. The next disclosed round, or named customer deployments, will show whether the infrastructure thesis holds as more world models reach production.
