2 min lesson
Allocating shared and contested cost
Give a plain answer to "Why is launching chargeback before showback a mistake, even when leadership wants budget accountability immediately?" Then ground it in one lesson detail.
Step 1 of 2
Allocating shared and contested costwhere the fights happen
- Cost type
- Single-feature inference
- Fair allocation rule
- Direct, by request tag
- Why it holds up
- Clear cause, clear owner
- Cost type
- Shared base model serving
- Fair allocation rule
- By usage share (token or request volume)
- Why it holds up
- Proportional to draw on the pool
- Cost type
- Idle / reserved baseline
- Fair allocation rule
- Spread by committed capacity, not actual use
- Why it holds up
- You bought the floor for everyone
- Cost type
- Platform overhead
- Fair allocation rule
- Flat or by headcount, stated openly
- Why it holds up
- Cheap to compute, hard to game
| Cost type | Fair allocation rule | Why it holds up |
|---|---|---|
| Single-feature inference | Direct, by request tag | Clear cause, clear owner |
| Shared base model serving | By usage share (token or request volume) | Proportional to draw on the pool |
| Idle / reserved baseline | Spread by committed capacity, not actual use | You bought the floor for everyone |
| Platform overhead | Flat or by headcount, stated openly | Cheap to compute, hard to game |
Every rule must be explainable in one sentence to the team being charged.
Anticipate gaming. If you allocate shared cost by request count, a team will batch requests to look cheaper. If by tokens, they'll trim outputs in ways that hurt users. The defense isn't a perfect rule, it's a rule that's transparent and a forum where disputes get resolved out loud.
Your job is the model and the operating forum, not policing every dollar. You design the allocation rules, stand up the review where ML, Infra and Finance look at the numbers together and keep the rules fair and explainable. You are not the spend cop chasing individual line items - that doesn't scale and it burns the trust you need to influence without authority.
If asked "showback or chargeback," never pick one outright. Say: showback first to earn trust in the data, chargeback later only where ownership is clean and shared costs stay in a fair-split pool with an explainable rule. Then name a specific shared-cost case - base model serving - and how you'd allocate it. The phased answer shows judgment; the worked example shows you've done it.