When Alberto Taiuti and Bryce Schmidtchen left their technical lead roles on Apple Vision Pro, they weren't chasing another hardware moonshot. Instead, they saw a gap: generative video models could conjure impressive clips, sure, but making them respond in real time—the kind of instant feedback loop that gaming or interactive media demands—remained largely unsolved.
On May 28, 2026, their answer emerged. Reactor Technologies stepped out of stealth with $59 million in combined seed and Series A funding, positioning itself as the infrastructure layer for what it calls "interactive world models." Lightspeed Venture Partners led both rounds, joined by WndrCo, Amplify Partners, Sky9 Capital, FPV Ventures, and Abstract Ventures. AWS signed on as the preferred cloud provider, a technical endorsement that signals real enterprise ambition.
Then there's the Jeffrey Katzenberg piece. The DreamWorks Animation founder and WndrCo partner is taking a board observer seat—hardly a standard move for early-stage infrastructure plays. But Reactor isn't pitching itself as just another developer tool. The San Francisco startup is threading a peculiar needle between gaming engines, film production pipelines, and foundational AI infrastructure, a convergence that's attracted attention from corners of the industry that don't always overlap.
The Latency Problem No One Else Is Solving
Most generative video tools work like this: you prompt, you wait, you get a result. Pre-rendered. Static, even if it's dazzling. What Reactor's backers keep returning to is a harder technical challenge—getting interactive video generation down to sub-50-millisecond frame latency. That's the threshold where interaction stops feeling delayed and starts feeling immediate.
Lightspeed's Bucky Moore calls it "foundational infrastructure for a new class of applications," the sort of phrase venture capitalists deploy when they're betting on a category shift rather than an incremental product improvement. His colleague Amber Yang framed the investment around that specific latency hurdle, one that becomes more acute as world models—AI systems that don't just generate content but simulate entire interactive environments—move from research labs into production use cases.
Whether developers actually want another abstraction layer in an already crowded generative AI stack remains to be seen. But Reactor's pitch is straightforward enough: it handles GPU provisioning, streaming complexity, and session orchestration so developers don't have to. The company launched with SDK and API access, supporting JavaScript, React, and Python. Documentation suggests a session-based architecture with bidirectional streaming, though the real test will be how well it scales under production load.
A Team Pulled From the Usual Suspects
Taiuti brings more than his Apple pedigree—he co-founded and served as CTO of Luma AI before starting Reactor. Schmidtchen shares the Vision Pro background. Together they've assembled a team from the expected roster: Apple, Netflix, Meta, Google, Adobe, Replicate, Microsoft. The company hasn't disclosed headcount, though its careers page lists openings for ML inference engineers, research scientists focused on real-time interactivity, DevOps/SRE roles, and forward-deployed engineers—all based in San Francisco, with some European timezone coverage added.
"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," Taiuti said in the announcement. It's the sort of statement that sounds visionary until you start asking how many customers actually need what he's describing, and whether they'll pay enough to justify the infrastructure costs.
The Model Catalog and the Pricing Question
Reactor's model catalog as of mid-2026 includes Helios (interactive real-time video generation), LingBot (action-controlled world generation), LongLive 2 (multi-shot video generation), and SANA-Streaming (real-time video editing). Pricing follows usage-based compute billing; the public playground interface shows rates around $6 per hour for some models, though that number will likely shift as the platform scales and competition heats up.
The company claims sub-1-second round-trip latency and access to thousands of GPUs on demand—a claim that's easy to make in a launch announcement, harder to maintain when paying customers start stress-testing the system. Lightspeed's investment memo cited recent research benchmarks as validation: DeepMind's Genie 3 hitting 720p at 24 FPS, StreamDiT reaching 16 FPS on a single H100. The technology, in other words, is moving from academic curiosity to production viability.
Or at least, that's the thesis.
Early Customers and the Entertainment Angle
Reactor has named Overworld as an early customer—a world-model developer that raised $4.5 million in its own pre-seed round earlier this year. Overworld CEO Louis Castricato framed the challenge neatly: making world models "usable and responsive in real time" matters just as much as developing the underlying AI. The company also reports interest from film and television studios, plus robotics companies exploring simulation use cases.
That entertainment studio interest—and Katzenberg's involvement—hints at where Reactor thinks the market opportunity sits. Not just in gaming or virtual production, but in a broader reimagining of how content gets made when interactivity and generation collapse into the same moment. Whether studios actually adopt this approach, or whether it remains a technical curiosity for researchers and hobbyists, will determine if the founders' bet was prescient or premature.
Jason Bennett, AWS's VP and global head of startups and venture capital, positioned the partnership around "speed of interaction" requirements that traditional cloud architectures weren't built to handle. It's a compelling narrative. The infrastructure exists, after all, to support existing workflows—not necessarily the ones that don't quite exist yet.
What Comes Next
The $59 million gives Reactor runway to prove the thesis, build out the team, and land enough design partners to demonstrate that real-time generative video infrastructure is a category worth creating. The hiring push suggests the company plans to scale both its technical stack and its go-to-market motion quickly.
But questions linger. The generative AI landscape is littered with promising infrastructure plays that solved problems developers didn't know they had, or couldn't afford to care about yet. Reactor's challenge isn't just technical—it's convincing a market that's already drowning in tools, models, and platforms that this particular abstraction layer is the one worth building on.
If they're right, the company could become foundational to a new class of interactive experiences. If they're wrong, it'll be an expensive lesson in the difference between what's technically impressive and what developers actually need. Either way, it's the kind of bet that makes sense when frame latency becomes the bottleneck, and when the line between generating content and simulating worlds starts to blur.
