Nvidia's GEAR lab released Flora on October 20, a suite of open models that turns raw manipulation video into the labels needed to train its GR00T robotics stack. Nvidia's own numbers put the annotation cost at roughly 72 cents per video hour, a figure it says is about 125 times cheaper than running the job on GPT-6 Astra. The release landed in the middle of GTC Berlin, a show where Jensen Huang spent the keynote touting physical AI rather than making the data engine the headline. That decision is itself a tell.
Flora is three models in one pipeline: TempSeg splits episodes into timestamped action segments, Captioner writes the timestamped descriptions, and Rewarder produces outcome and process scores for reinforcement-learning post-training. Nvidia claims it can label up to 200,000 hours of video per day across 512 GB300 GPUs while training GR00T in parallel. The weights are on Hugging Face, but the Captioner and Rewarder sit under a one-way noncommercial license, which matters more than it sounds. It means the cheapest way to get robot labels at scale is open, but the commercial door stays locked to research.
The bigger implication is that the bottleneck in robotics was never inference. It was data. A manipulation episode is easy to record and hard to annotate. If you can drop the annotation cost by two orders of magnitude, the math on building a physical AI stack changes. The question is not whether Nvidia's numbers hold under a real production bill. It is whether the license terms will keep the data engine inside the wall, and who gets to scale it outside. That is the line the next round of robotics funding will have to draw.
Image source: i.ibb.co