Reactor
Last reviewed
Jun 3, 2026
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,289 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,289 words
Add missing citations, update stale details, or suggest a clearer explanation.
Reactor is an American artificial-intelligence startup building developer infrastructure for real-time generative video and interactive world models. The company emerged from stealth on May 28, 2026, announcing it had raised a $59 million Series A round led by Lightspeed Venture Partners, with backing that included Jeffrey Katzenberg's holding company WndrCo. Reactor was co-founded in August 2025 by Alberto Taiuti and Bryce Schmidtchen, two former technical leads on the Apple Vision Pro headset, and is based in the South Park neighborhood of San Francisco.[1][2][3]
Rather than train its own consumer-facing model, Reactor positions itself as a platform layer: a single SDK and API that let developers run real-time interactive video models without managing the GPU provisioning, streaming, and session orchestration that those models require. The company's stated mission is "to make real-time world models accessible to developers everywhere."[3]
Most text-to-video systems work in batches. A user submits a prompt, the model renders for seconds or minutes, and a finished clip comes back. Reactor targets a different regime, where frames are generated continuously and respond to user input as it arrives, closer to how a game engine reacts to a controller than how a render farm produces a movie. CEO Alberto Taiuti framed the distinction in the company's launch materials: "World models are redefining what AI can do, moving from systems that generate content in isolation, to ones that perceive and respond in real time."[3]
That shift creates infrastructure problems that conventional inference stacks were not built for. According to Lightspeed, real-time video models need persistent state across interactions, bidirectional streaming with sub-50-millisecond latency, real-time processing of control inputs, and session orchestration to manage many concurrent users. Reactor's platform is designed to handle those requirements behind a simple interface, abstracting away GPU provisioning, infrastructure setup, and streaming complexity so a developer can connect to a video model in the programming language they already use.[4]
The company contrasts its approach with the latency of batch generation. Taiuti has said that some text-to-video systems can take "up to 10 minutes to produce 10 seconds of video," whereas "the time to first frame with our platform is basically nil," adding that the system is optimized to "generate videos instantaneously."[1][3] Lightspeed's account puts a figure on the target: frame latency under 50 milliseconds, which is the rough threshold below which interaction starts to feel immediate to a human.[4]
Reactor describes three broad application areas for the technology: media and entertainment, physical AI, and robotics. In entertainment, the pitch is a form of media that is generated dynamically and shaped by the viewer rather than pre-rendered. In robotics and physical AI, world models can act as simulators that a system uses to predict and respond to its environment.[2][3]
Alberto Taiuti and Bryce Schmidtchen met at Apple, where they worked together on the early Vision Pro team on video models and infrastructure. Both are described in the company's announcement as former technical leads on the Vision Pro AR/VR headset.[2][4]
Taiuti's background is the more publicly documented of the two. Before Apple, he co-founded Luma AI in August 2021 alongside Amit Jain and Alex Yu, serving as the company's chief technology officer. Luma became one of the more widely used 3D and video generation platforms, and Reactor's launch materials credit Taiuti with building the infrastructure behind it. Schmidtchen's prior work focused on kernel optimization and computer vision for augmented and virtual reality. Taiuti is Reactor's CEO and Schmidtchen its CTO.[3][4]
At launch, Reactor reported a team of 16, with engineers and researchers drawn from Apple, Netflix, Meta, Google, Adobe, Replicate, and Microsoft. The company says that group brings experience in graphics, real-time systems, interactive media, and scaling AI infrastructure, the combination it argues is needed to serve video models at low latency rather than in batches.[2][3]
Reactor disclosed $59 million in total funding when it came out of stealth. Lightspeed Venture Partners led the Series A and had co-led the company's earlier seed round. The roster of backers also included WndrCo, Amplify Partners, Sky9 Capital, Abstract Ventures, and FPV Ventures.[2][3]
| Detail | Information |
|---|---|
| Headquarters | San Francisco, California (South Park) |
| Founded | August 2025 |
| Co-founders | Alberto Taiuti (CEO), Bryce Schmidtchen (CTO) |
| Total funding | $59 million |
| Lead investor | Lightspeed Venture Partners |
| Other investors | WndrCo, Amplify Partners, Sky9 Capital, Abstract Ventures, FPV Ventures |
| Employees (at launch) | 16 |
| Website | reactor.inc |
Bucky Moore, a partner at Lightspeed, summarized the firm's thesis in the announcement: "Real-time video models are currently inaccessible to developers due to a lack of infrastructure that can reliably serve them." Lightspeed's investment write-up argued that this infrastructure gap had become "one of the primary bottlenecks preventing these models from being more widely adopted," which is the gap Reactor is trying to close.[3][4]
The most discussed part of Reactor's round was the involvement of WndrCo, the holding company run by Jeffrey Katzenberg, the former DreamWorks Animation chief executive and Quibi co-founder. Variety, which reported the funding as an exclusive, framed Reactor as a real-time AI video startup founded by ex-Apple engineers and backed by Katzenberg. WndrCo participated in the round, and Katzenberg is joining Reactor as a board observer rather than a full board member.[1][2]
Katzenberg's stated reason for backing the company points at the entertainment use case. He described Reactor as "a bridge between the model world and real-world applications."[1] His presence connects a generation of real-time generative tools to Hollywood at a moment when studios are weighing how, and whether, to fold AI-generated content into production. It is worth being careful about how much to read into a board-observer seat, which carries influence and information rights but not a vote.
Reactor enters a fast-moving area often grouped under the heading of world models, systems that learn how an environment behaves from video and let a user move through it in real time. The most prominent example is Google DeepMind's Genie 3, announced in August 2025, which generates navigable environments at 24 frames per second and 720p resolution and keeps them coherent for a few minutes. DeepMind later put a consumer prototype, Project Genie, in front of some Google AI subscribers in early 2026. Reactor's positioning is deliberately different from a flagship model like Genie. Instead of competing to build the best world model, it is selling the plumbing that lets developers run such models in production, an infrastructure play rather than a model play.[4]
That distinction is also Reactor's main risk. The value of a serving layer depends on a healthy supply of real-time video models worth serving and on developers wanting to build interactive experiences on top of them, neither of which is guaranteed this early. The company's bet is that the models are arriving faster than the infrastructure to run them, and that a developer platform can sit in the middle the way earlier inference and hosting companies did for generative AI text and image models.