Reactor Emerges From Stealth With $59M to Build the Developer Platform for Real-Time AI Worlds

There is a pattern that recurs every few years in technology infrastructure: a powerful new computational capability arrives, it works well in research but proves brutally difficult for developers to deploy at scale, and then a startup builds the abstraction layer that makes the hard parts disappear. AWS did this for servers. Stripe did it for payments. Twilio did it for telephony. Reactor wants to do it for real-time world models.

The company emerged from stealth on May 28 with $59 million in seed funding led by Lightspeed Venture Partners. WndrCo — the venture fund and production company co-founded by Jeffrey Katzenberg, former chairman of Walt Disney Studios and DreamWorks Animation co-founder — participated alongside Amplify Partners, Sky9 Capital, and FPV Ventures. Amazon Web Services is Reactor’s preferred cloud partner, providing the backbone compute infrastructure for a category of workload that is orders of magnitude more demanding than static inference.

What World Models Actually Are

A world model is an AI system that builds and maintains a dynamic internal representation of an environment, then uses that representation to simulate how the environment evolves in response to actions. Unlike generative models that produce a discrete output — an image, a text response, a code snippet — world models generate experience: continuous, interactive, physically plausible streams of perception that respond to inputs in real time.

The clearest current applications are in video gaming, simulation, and robotics. In games, a world model can generate an interactive environment on the fly rather than rendering pre-authored geometry. In robotics training, world models create realistic simulation environments where robots can train for millions of hours without hardware wear or safety risk. In physical AI research, they enable models to learn causal relationships between actions and outcomes at a scale that pre-recorded data cannot provide.

But Reactor’s founders see the application space as far broader. A world model that streams at low latency and responds to user inputs in milliseconds becomes a new kind of medium — not video you watch but video you inhabit. Interactive entertainment, training simulations, architectural walkthroughs, digital twins of physical environments — all become possible when the generation is fast enough to feel real.

The Infrastructure Problem Reactor Is Solving

The challenge is that real-time world models are computationally ferocious. A static image generation model runs inference once and returns a file. A world model runs continuous inference at whatever frame rate the experience requires — 30, 60, 120 frames per second — while incorporating user inputs that can change the generated state at any moment. The latency requirements are unforgiving: perceptible lag above roughly 100 milliseconds breaks the illusion of interactivity.

Managing this at production scale requires specialized infrastructure: hardware optimized for streaming inference rather than batch jobs, networking architected for low-latency delivery to global users, and orchestration systems that can spin up capacity instantly for unpredictable demand spikes. Building this from scratch requires expertise in GPU infrastructure, distributed systems, networking, and generative AI — a combination that few teams possess and that most application developers have no interest in acquiring.

Reactor’s SDK and API abstract all of this. A developer integrating Reactor’s platform can build a real-time interactive AI application in what the company describes as “a few lines of code,” without managing the underlying infrastructure. Reactor handles provisioning, scaling, latency optimization, and the model serving layer, exposing a simple interface that the application calls to generate or update the world state.

The Founders: Apple’s Computer Vision Alumni

Reactor was founded in August 2025 by two former Apple engineers identified in reporting as Taiuti and Schmidtchen. Neither has a long prior public profile, which makes their institutional background at Apple the clearest signal of technical credibility. Apple’s computer vision and machine learning teams have historically operated on problems at the intersection of real-time performance and model quality — Face ID, object detection in camera systems, spatial computing for Vision Pro — precisely the combination of constraints that define the world model deployment problem.

The choice to target developer infrastructure rather than end-user applications is a deliberate positioning decision. End-user AI applications are high-profile but require sustained marketing investment and consumer acquisition costs. Infrastructure companies tend to grow through word of mouth within developer communities, scale with their customers’ success rather than fighting for attention alongside it, and build switching costs that accumulate with every line of customer code written against their API.

Jeffrey Katzenberg’s involvement through WndrCo adds a dimension worth examining. Katzenberg built two of the defining media companies of the late twentieth century. His fund has consistently bet on the intersection of technology and entertainment. WndrCo’s participation in Reactor suggests a conviction that world models will reshape the economics of entertainment content production — not as a curiosity but as the primary medium through which interactive entertainment is created.

The AWS Partnership

Amazon Web Services is not merely a cloud vendor for Reactor — the press release explicitly names AWS as Reactor’s “preferred cloud provider, supporting the compute infrastructure and distribution required for real-time generative video workloads at global scale.” This language implies a deeper commercial relationship than a standard cloud credits arrangement.

For AWS, world models represent a new category of continuous, high-throughput GPU workload that could become a significant revenue driver as applications scale. Unlike batch training runs that spike and subside, real-time world model serving generates consistent, predictable GPU demand that maps well to AWS’s capital-intensive data center investments. Being the preferred provider for a platform that could underpin a broad category of next-generation applications gives AWS a structural advantage over Azure and Google Cloud in this emerging workload type.

The Competitive Landscape

Reactor enters a world model space that is suddenly crowded with well-funded competitors. Runway raised $5 billion in a round reported in May at a $35 billion valuation, with a product portfolio that spans video generation, world model research, and creative AI tools. Decart AI raised $300 million in May to build world models for physical AI training. Google’s Gemini Omni family, announced at I/O 2026, targets the world model category from a frontier lab perspective.

What distinguishes Reactor’s position is the developer infrastructure layer. Runway builds for creative professionals. Decart focuses on physical AI training data. Google and other frontier labs build the underlying models. None of them have positioned themselves as the API layer that lets any developer integrate real-time world model capability into any application, regardless of which underlying model they run on top.

This agnostic infrastructure position is strategically significant. If world models become a standard component of interactive applications — as Reactor’s founders clearly believe — then the platform that makes them accessible will accrue substantial structural leverage over the ecosystem, similar to how Stripe became the financial infrastructure layer for e-commerce rather than building any particular e-commerce application itself.

What the $59M Buys

Seed stage for a company building infrastructure at this level of technical complexity is more analogous to Series A for a typical software startup. The $59 million will support continued research and engineering investment in the core streaming inference platform, expansion of industrial deployments and customer pilots across gaming, media, and physical AI verticals, and the team growth needed to support enterprise customers requiring reliability guarantees and dedicated support.

The AWS partnership likely provides meaningful leverage on infrastructure costs, allowing Reactor to deploy at scale without absorbing the full cost of GPU capacity during the customer growth phase. This matters because world model inference is expensive — per-session costs that are manageable at modest scale can become significant at the volumes required to serve a large consumer application.

Reactor’s bet is that by the time those volumes arrive, the market will have made clear that real-time world models are a foundational technology layer, and that the developer platform for that layer is worth the infrastructure investment that gets you there.

---

Reactor launched from stealth on May 28, 2026. The company is available in early access for developers building real-time AI applications.