Inference Cost Attribution for Multi-Tenant

For SaaS products and internal multi-department AI deployments, you need to attribute infrastructure cost to tenants / customers / departments. Self-hosted has a different attribution model than per-token API billing — predictable monthly cost gets allocated based on usage proxies.

Attribution models

• Per-token: most accurate; needs token counting per request; matches hosted-API billing model
• Per-request: simpler; works when token counts are similar across tenants
• Per-active-user: simplest; useful for internal departmental allocation
• Time-based: GPU minutes consumed per tenant; useful for batch workloads

Data needed

Your structured JSON logs already capture this:

• tenant_id per request
• prompt_tokens + completion_tokens
• model + model_version
• request_id + user_id

Aggregate via ClickHouse / Postgres window functions, output per-tenant per-month tokens consumed.

Billing

For a SaaS billing model on top of self-hosted AI:

• Compute monthly cost: server (£279) + ops engineering pro-rated + frontier API fallback
• Compute monthly tokens: from logs
• £/M tokens this month: total cost / total tokens
• Per-tenant bill: their tokens × £/M tokens (or with margin for SaaS pricing)

For internal departments: use the same per-token rate, charged back at cost.

Verdict

Per-token attribution from structured logs is the right default for multi-tenant self-hosted AI cost allocation. Matches hosted-API mental model; transparent to customers; supports usage-based pricing tiers. Build this from day one — retrofitting attribution after launch is painful.