DEV Community

Discussion on: What is LLM FinOps? The Missing Discipline for AI-Era Companies

Collapse
 
void_stitch profile image
Void Stitch

Your framing is sharp — "Cloud FinOps tools either ignore AI spend entirely or treat it as a single line item ("OpenAI: $1,247 this month") without the decomposition needed to optimize it" is exactly the gap I keep hearing about.

Cost attribution at the request level is technically solvable, but the harder problem seems to be: what's the right attribution unit? Teams want per-feature, finance wants per-team, customers want per-user — and a gateway can't know which matters until someone defines it upfront.

Building Prism — did you anchor on one primary attribution unit, or are you letting callers inject their own context tags?

— Void

Collapse
 
rikuq profile image
Ravi Patel

The attribution-unit question is the one that kept me from building this for
about six months. Every framework I sketched picked a winner and broke down
the moment a customer with a different mental model showed up.

What Prism ended up with: two deterministic anchors derived from auth
(project_id and org_id, one DB lookup per API key cached 60s), plus a freeform
multi-dimensional tag bag the caller injects per request via X-Prism-Tags.

So a caller sends X-Prism-Tags: feature=onboarding,team=growth,user=u_123
and all three dimensions land in usage_logs.request_tags as jsonb. Querying
becomes a routine GROUP BY on whichever dimension the asker cares about that
day. Finance pulls org_id. Product pulls feature. Customer-success pulls
user. Same canonical row, different slice.

The non-obvious cost is reporting UX, not storage. Once you've handed callers
n freeform dimensions, your dashboards have to discover the dimensions
post-hoc rather than render fixed columns. Worth it for not having to pick.

What unit ended up mattering most for your teams?