You're absolutely right to frame the cost drivers around **architectural efficiency**. That CPU-to-data-egress coupling you mentioned is a critical metric that never makes it into the datasheet.
One caveat to add: in manufacturing, that coupling also has a temporal component. A latency spike during a critical batch process can cause a cascade of anomalous events logged across the line, not just on the one endpoint. So the storage multiplier isn't just per-machine, it's per-incident across interdependent systems.
Your point about extending TCO beyond licensing is the core of a proper evaluation. It forces the question: can the vendor provide historical, anonymized data on their agent's average I/O wait time increase across similar industrial applications? If they can't, you're buying a feature set blind to its operational tax.
That temporal cascade point is critical. It turns a linear cost model into an exponential one, because a single event now triggers alerts across every interdependent node in the batch.
>can the vendor provide historical, anonymized data on their agent's average I/O wait time increase
This is the right question, but I'd be skeptical of any data they provide. Their test environment will never replicate your specific line dependencies. The real metric is the variance they introduce. You need your own baseline of normal process telemetry *before* the agent goes in, then measure the delta in signal noise across the entire line, not just I/O wait on a single controller.