Ice Cream Analogy
This is the expensive one-time activity, like buying the giant recipe-learning machine.
Training(machine buidling)
Ongoing small bills that add up each time a user asks the AI something.
Inference cost
Limiting how many requests a user can make to protect your budget.
rate limiting
This is scooping ice cream for each customer — a small cost that adds up fast.
inference
The big occasional bill when you train a large model from scratch.
training cost
A test to stop automated bots from using public endpoints.
CAPTCHA
This is what you do when a sudden rush of customers shows up and you need more scoops and staff.
scaling
The ongoing money spent keeping ingredients and moving them between stores (files and transfers).
storage and data transfer (egress) cost
Spotting strange traffic patterns to block attacks that hike your bill.
anomaly detection
Slide 2 compares these three things: building, scooping, and handling rush hour. Which is building?
training
Money paid for engineers, monitoring, and people who run the shop.
operations or personnel cost?
An attack that floods your API to drive up cloud bills.
billing attack
The strategy to call in more staff only when lines form, so you don’t pay for idle workers.
autoscaling
If one scoop costs $0.0005, how much for 1,000,000 scoops in one day
$500 per day
Charging or capping usage per customer to prevent one tenant from blowing the budget.
per-tenant billing or quotas