The Economics of AI Development: From Tokens to Accepted Work
Episode participants
A solo episode without invited guests.
What we discussed on the recording
Alexander Polomodov examines AI-development economics: fixed-quality capability gets cheaper while total spend grows with use cases and calls. Cost is tasks × calls × tokens × token price, plus seats, platform services, and capacity; reducing one factor does not stop the others.
Governance should follow the work trace: cost, steps, retries, tool calls, time, and authority. A loop detector, kill switch, and human fallback cap risk and recovery expense. Seats belong to people, exploration to teams, production inference to use cases, and shared platforms to a central owner.
FinOps starts with visibility and showback. The denominator is the accepted task: cost includes models, tools, retrieval, compute, review, rework, and expected loss; output counts after acceptance. For a bug fix that means tests without regression; for an agent, a completed workflow with acceptable consequences.
Routes should be compared on identical tasks and one quality threshold, including repeated runs, fallback, human time, and cost tails. Lock-in lives in prompts, tools, state, and evals, so the portable unit is the task contract. A local model faces the same gate; its full TCO includes idle capacity, high availability, storage, SRE, security, and upgrades.