Post by Thomas M.

Those who know me have heard this before...and it still baffles me. I've spent years architecting the software-defined vehicle with OEMs and Tier 1s on three continents. Different brands, different markets — the same recurring failure mode. And it's almost never the software. What quietly kills the SDV isn't the vehicle OS or the zonal architecture. It's a hardware decision disguised as a finance decision, made years earlier: the refusal to put real compute and a common sensing stack in every car — not just the ones whose buyers paid for the ADAS package. No fleet-wide hardware, no fleet-wide data, no flywheel. And without the flywheel there's no L2++, no autonomy, no robotaxi option later. You can't software-define your way out of a hardware-acquisition decision you declined to make. Tesla and BYD understood this and shipped the stack to every car, turning their drivers into an unpaid data flywheel. Much of the West gated it by trim to protect margin — then started renting the missing data from NVIDIA, Wayve and Momenta, feeding a supplier's moat instead of their own. As I extend this work into Physical AI — across mobile robots that fly, roll, walk and swim — the lesson holds without modification. Embodied intelligence is a deployment problem before it's a model problem. The Learn loop only closes if the hardware is already in the field. You can license a model. You can't license a flywheel you never built. New in The Un-Engineering Lens ↓