AI inference doesn't have a fixed cost — and that's the problem. Unlike traditional software, where pricing is predictable per seat or per license, large language models are billed on tokens: a unit of cost that fluctuates with prompt length, model choice, context size, and how many times an agent calls the model in a loop. This talk unpacks "token maxxing" — the pattern where token consumption silently expands to exceed whatever budget was allocated, often driven by context window creep and increasingly autonomous, multi-step agentic workflows. We'll break down how tokens are priced, why open-source and open-weight models change the cost equation, and share concrete, benchmarked strategies — prompt caching, model routing, batching, and agent step budgets — that organizations are using to cut effective spend by 50–95% without sacrificing output quality. Attendees will leave with a practical framework for treating token spend as a governed, measurable resource rather than a surprise line item on next month's invoice.