Who Owns the Bathtub Wins Robotics
NVIDIA already owns the synthetic data, the models and the silicon, which caps how big an independent robotics data infra player can realistically get.
#fw_vertical_op
Robotics Data Infra - Thesis
— the bathtub is already there: NVIDIA Jetson - so for now NVIDIA is very good at synthetic data (data generation) and less good at data management/orchestration. But if I have to be realistic I think they will get there: they own the input (synthetic data), the output (ML models and simulation environments) and the hardware (chips). — I think this is the main attack vectors - I started looking at it by trying to make a case for a $10bn outcome out of Mosaico (or generally robotics data infra) - so I think that what they’d need to do is build vertically for robotics/automation deployments (yes) and then horizontally, in terms of platform capabilities across use cases. My take though remains that of “the array of platform capabilities will never be enough” and you’ll remain capped in terms who you can serve
Other attack vectors include:
- better funded alternatives (already looking at the entire data management stack) which puts pressure on its window of vulnerability (for sure outside Europe)
- Foxglove’s ecosystem lock-in via MCAP (modular container for autonomous purpose) - simplifies the problem to just a matter of formatting data + this makes the data ready for their visualization tool (which positions them one step closer to the bathtub)
- this also gave the feeling of how other robotic devtools founders are trying to build around an MCAP alternative (not good enough)
- Foxglove’s ecosystem lock-in via MCAP (modular container for autonomous purpose) - simplifies the problem to just a matter of formatting data + this makes the data ready for their visualization tool (which positions them one step closer to the bathtub)
- value prop conflict
- Rerun’s value prop is very similar (reducing this to a battle of the most commercial guy)
- open source, built in Rust for performance, targeting the same gap between raw recording and ML training
- Rerun’s value prop is very similar (reducing this to a battle of the most commercial guy)
- convergence scenario
- where Foxglove adds ML data prep features (they’re already moving toward “unified observability”), Rerun adds data management to their visualization SDK, and NVIDIA ships a data service inside Isaac. If all three happen within the same 18-month window, Mosaico (or the likes) gets fucked from above, below, and sideways simultaneously
- So: What capability requires a fundamentally different architectural decision? I believe that decision making here becomes a key node
- where Foxglove adds ML data prep features (they’re already moving toward “unified observability”), Rerun adds data management to their visualization SDK, and NVIDIA ships a data service inside Isaac. If all three happen within the same 18-month window, Mosaico (or the likes) gets fucked from above, below, and sideways simultaneously