Jean Baptiste Kempf's open source video player has 6 billion downloads. He says physical AI will fail on network lag, not on model quality, and Lightspeed, the venture firm behind Anthropic and Mistral, is betting $5 million he is right.
Six billion downloads of VLC gave Jean-Baptiste Kempf two decades of credibility in open-source media. The Paris developer's next chapter is a $5 million bet that the same real-time plumbing that keeps a video stream from tearing apart is the invisible wall physical AI is about to hit, and that solving it is a standalone business, per TechCrunch's profile of his new startup, Kyber.
The thesis is unglamorous on purpose. Kyber sells a software toolkit, an SDK, that synchronizes video, audio, sensor data, and control inputs across a network with minimal delay. That sounds like the kind of problem media companies solved in the 2000s. It is also, Kempf argues, the problem the physical-AI industry has not publicly solved, because the labs building robot brains are busy building robot brains. "Physical AI is only as good as the underlying systems running it," Lightspeed wrote in a LinkedIn post announcing the investment, according to TechCrunch reporter Anna Heim. Lightspeed, the venture firm that has also backed Anthropic and Mistral AI, led the $5 million raise.
The structural question is who owns that layer. If Kempf is right, synchronization and observability become shared infrastructure for robots the way content delivery networks became shared infrastructure for the web. That would give fleet operators and developers a new place to buy capabilities they currently assemble themselves. It would also make the layer a target for the parties already selling adjacent pieces: chip vendors with controller stacks, hyperscalers with edge compute, and robot makers with proprietary middleware. The $5 million buys time to find out which way the consolidation goes.
Kempf is also the person making the loudest claim about the prize. He forecasts "hundreds of millions of robots and drones" on city streets within a few years, a projection that comes from the founder, not the publication, and that has no independent benchmark. The largest robot fleets he cites, per TechCrunch, run 2,000 to 3,000 vehicles. The gap between the current state and the forecast is the bet. Five million dollars is what the rest of the market will pay to watch the founder try to close it.
Three things will move the story from here. The first is whether Kyber publishes latency benchmarks against ROS 2, NVIDIA Isaac, and OEM controller stacks on fleets of more than a few dozen devices, the kind of head-to-head that turns an SDK into a reference implementation. The second is whether the model companies Lightspeed has already backed treat Kyber as a partner or as a feature to absorb. The third is whether Kempf's open-source instincts survive a Series A conversation with a board that wants a moat, and what that moat ends up being.