Thiyagarajan Maruthavanan examines how token use, repeated context, memory movement and agent loops shape AI costs. He presents spending anecdotes to argue for measuring the actual workload rather than assuming model prices alone predict the economics. The anecdotes are his reported examples, not independently verified cost studies.
Thiyagarajan Maruthavanan contrasts hosted inference with owning deployment infrastructure. He argues that greater control over capacity, access and reproducibility can matter once a product has established demand, while hosted services can be sensible for experimentation before that point. Ownership brings operational responsibilities, so the talk does not establish a universal cost advantage for self-hosting.
Watch on YouTube
