The assumption seems straightforward: fewer employees + more AI = higher efficiency and better margins.

But AI introduces a different cost structure.

Inference is a usage-based cost — it scales with every feed ranking, ad decision, and user interaction. Meanwhile, AI infrastructure spending flows into cost of revenue over time through depreciation.

In several public companies, cost of revenue has already started rising faster than revenue following AI rollout, with early signs of margin compression.

This raises a question:

If AI shifts cost from fixed labor to variable compute, is it possible that replacing humans with AI could reduce margins rather than improve them?

And if inference cost scales with usage, should AI systems be deciding whether to run at all based on expected value?