Kyber, the Paris-based startup building real-time robot control infrastructure, has closed a $5 million seed round led by Lightspeed Venture Partners, with co-investment from OVNI Capital and Kima Ventures. The firm managing the cheque oversees over $50 billion in assets and has previously backed Anthropic and Mistral AI. For a $5m seed, that is a fairly well-credentialled cap table.
The company was founded by Jean-Baptiste Kempf, best known in open-source circles as a core contributor to VLC and FFmpeg. His new project takes the same instinct (make media run faster and more reliably, everywhere) and points it squarely at robots, drones, and anything else that needs a human (or an AI) pulling the strings from a distance.
How Kyber’s Real-Time Robot Control Actually Works
Kyber’s core product is an SDK that synchronises video, audio, sensor data, and control inputs with minimal lag. The protocol stack is built on QUIC, using FFmpeg in push mode as a server and a modified VLC in real-time mode on the client side. Rather than transmitting data as standard reliability-guaranteed streams, Kyber sends video and audio as datagrams: you trade the guarantee of delivery for the guarantee of speed.
The practical result: end-to-end latency as low as 8 milliseconds in a 240Hz environment, measured glass-to-glass from camera to display. Kempf reported that figure in a February 2025 interview. ‘If you control things in the real world, every millisecond matters,’ he has said. Eight of them is a number worth paying attention to.
The SDK is deployable on-premise, in the cloud, or in hybrid environments, which matters when you are talking about customers in defence and telco who are unlikely to route sensitive operations through somebody else’s cloud by default.
The Case for Real-Time Robot Control at Scale
The pitch to investors leans into the physical AI moment. Lightspeed wrote in a LinkedIn post announcing the investment that ‘physical AI is only as good as the underlying systems running it.’ That is the kind of line that lands well in a deck right now, and Kyber’s underlying technology gives it some substance.
Kempf frames the platform’s purpose in deliberately broad terms: ‘all the use cases where the person who’s operating is not in the same place as the compute, which is not in the same place as the action.’ Remote driving is the obvious comparison point. He acknowledges that well-resourced companies have already built similar tools internally, but notes that the largest remote-driving fleets today run to ‘maybe 2,000 or 3,000 vehicles.’ Scale that to millions, and bespoke internal solutions stop working. That is the gap Kyber is building into.
The company is already in commercial deployment across defence, telco, robotics, and AI customers. Its near-term focus sits across three segments: robotics, drones of every kind, and remote IT access. That last one is the least glamorous and, Kempf suggests, currently the strongest in terms of demand. The comparison he reaches for is Citrix, which gives a reasonable sense of the total addressable market even if Kyber’s ambitions run further than direct competition with legacy remote-desktop software.
Observability is part of the story too. When AI agents rather than people are managing entire fleets, knowing that systems are actually functioning becomes harder and more critical simultaneously. Kyber’s pitch includes the infrastructure to monitor that at scale.
Open Source Roots, Enterprise Revenue
True to Kempf’s background, the core project is open source. The company monetises through a productised enterprise version and, like Palantir, through forward-deployed engineers who handle custom implementation on-site. Those forward-deployed engineers make up a large proportion of Kyber’s 25-person full-time team.
Headquarters are in Paris, with offices in San Francisco and Singapore to support what Kyber expects will be a genuinely global client base. The name itself is a Star Wars reference: kyber crystals power lightsabers, and the underlying logic is that speed is everything when you are controlling physical objects in the real world.
The open-source model means Kyber’s user base will almost certainly outgrow its paying customer base by a wide margin. That is a deliberate trade-off, and one Kempf has navigated before. The careers page puts the philosophy plainly: ‘The companies that tried to solve it spent years and tens of millions building custom solutions they’ll never share. We’re building the version everyone else can use.’
The next test is whether enterprise customers show up in numbers large enough to fund the infrastructure layer that everyone else gets for free. With Lightspeed’s backing and a latency figure that genuinely stands comparison with consumer video streaming, Kyber has a credible opening argument. The physical AI pipeline it wants to underpin is, for now, still being built.
