Reading time: 5 min
Table of Contents
Key Takeaways
- Production-grade remote infrastructure: Kyber provides an SDK that synchronizes video, audio, sensors, and control inputs with millisecond latency — built for real devices, not demos.
- Open-source core: The platform is open source, with a paid enterprise version plus on-site deployment via forward‑deployed engineers for companies that need it to just work.
- Three focused segments: Robotics, drones, and remote IT access — each one a use case where the person operating is not co‑located with the compute or the action.
The Gap Between Demo and Production Is Milliseconds
Here’s what actually happens in production: you push an update to 2,000 field robots and half of them drop their video feed because the SDK was tuned for a lab’s local network. That’s not automation — that’s a liability. The same guy who built VLC, Jean‑Baptiste Kempf, knows this pattern well. He’s now behind Kyber, an infrastructure layer that’s been architected for the scale where things break. The demo worked. Production didn’t. Here’s why: remote control isn’t hard with one device in a clean room. It’s hard when an AI agent orchestrates tens of thousands of drones across geographically dispersed sites with variable bandwidth and compute.
From Video Streaming to Device Control
Kempf built Kyber as a side project while CTO at a cloud gaming startup. The core insight is straightforward: streaming game video with minimal latency is the same problem as streaming a drone’s camera feed while sending control commands back. Most people get this wrong — they bolt a WebRTC library onto a device and hope. Kyber’s SDK isn’t a general‑purpose streaming wrapper. It’s tuned to a device’s actual compute capacity and network profile, the same way VLC adjusts to pixel formats and codec availability at runtime. This isn’t theory. I’ve seen similar approaches fail because the abstraction layer was too thick — each added hop became a latency multiplier. Kyber treats every millisecond as a constraint, not a requirement.
Scale Changes Everything — Especially Observability
“The largest fleets today have maybe 2,000 or 3,000 vehicles,” Kempf said. “Imagine you need to manage millions of them; that’s not the same thing.” The real cost is: when you scale that far, manual intervention stops being an option. If a sensor stream dies on unit 47, no human is there to reboot it. Observability becomes the product — knowing not just that a device is online, but that its control packet has actually arrived and the actuator moved. Kyber’s architecture reflects this: it’s built so that AI agents managing entire fleets can trust the data pipeline. At even smaller scale, the benefit is tangible: no more sending a technician to a radio tower just to push a config file.
Three Segments, One Stack
Kyber is already in commercial deployment with defense, telecom, robotics, and AI clients. Kempf has prioritized three segments: robotics, drones, and remote IT access. Remote IT access isn’t glamorous, but it’s a massive TAM that’s currently held together by clunky VPNs and Citrix workarounds. “The companies that tried to solve it spent years and tens of millions building custom solutions they’ll never share,” Kempf said. “We’re building the version everyone else can use.” That’s the same mindset that made VLC ubiquitous — pick a boring, painful problem and solve it so well it becomes invisible.
Open Source Core, Enterprise Delivery
True to Kempf’s open‑source roots, Kyber’s core SDK is open source. The company sells a productized version and offers hands‑on deployment through forward‑deployed engineers (FDEs) — a model that Palantir made famous. If you need remote infrastructure that holds in production, you don’t want a black box; you want a team that understands your failure modes. The $5M round led by Lightspeed is a bet that physical AI will depend on this kind of infrastructure. Lightspeed knows the pattern: “Physical AI is only as good as the underlying systems running it.” They’re right. We built Rebirth Distribution on that same principle — automation has to earn its place in production by surviving real load, real latency, and real friction.