Modular ADAS does not remove validation cost. It moves the cost from one large model into the arbitration between specialized driving agents. That is still a better engineering boundary because teams can isolate scenarios, data sets, and failures, but only if the interfaces are testable.
Autobrains says separately trained agents need two to three times less training data than a unified model and can run only when their scenario requires them. VinFast put the economic pressure behind that architecture at roughly $700 million to start an AI development center, plus $30 million to $50 million each year to retain its data. The warning came from redundant compute: adding a second chip can create millions of arbitration cases and years of testing. Smaller models reduce compute. They do not prove that two decision paths agree under a rare event.
OEMs evaluating this stack should require scenario-level traces, deterministic agent handoffs, and regression coverage before comparing TOPS. Within two years, the winning ADAS platform will be the one that can show why an agent ran and how conflicts were resolved. Vendors selling lower training cost without an arbitration test harness will lose the architecture decision.