Unit Economics (Cloud Infrastructure)
Unit economics connects infrastructure cost to business-relevant outputs. The right question is not “how do we reduce AWS cost?” but “which spend is not producing enough value, and who can safely change it?”
Defining unit metrics
unit cost = infrastructure cost / business-relevant unit
The denominator must be something the business cares about: cost per ad impression, per search query, per prediction, per TB processed, per active user. Not per instance or per container — those are meaningless to leadership and don’t connect cost to value.
Computing unit economics
Direct services
For a service with dedicated infrastructure:
Unit cost = monthly infra cost of service / monthly business units produced
Example:
- Search service costs $400K/month
- Serves 2 billion queries/month
- Unit cost = $0.0002 per query
Shared platform services
For tenants on shared platforms, use attributed cost (see Cloud Cost Attribution):
Unit cost = attributed platform cost / tenant's business units
Example:
- ML team’s attributed GPU cost: $800K/month
- Serves 50 billion predictions/month
- Unit cost = $0.000016 per prediction
Using unit economics
Trend analysis
Track unit cost over time. Healthy patterns:
- Flat or decreasing — efficiency improving with scale
- Proportional to growth — cost growing because business is growing (acceptable)
Unhealthy patterns:
- Unit cost increasing — efficiency regressing, investigate
- Cost growing faster than business units — waste accumulating
Investment decisions
When a team requests more capacity:
Additional cost / additional business units enabled = marginal unit cost
If marginal unit cost is much higher than current average, the request may indicate inefficiency. If it’s lower (economies of scale), it’s a good investment.
Optimization prioritization
Rank optimization opportunities by:
Potential savings × confidence × (1 / risk to business units)
A 20% savings on a service with zero business impact is better than a 5% savings on a revenue-critical path — even if the absolute dollars are similar.
Common pitfalls
- Missing the denominator change — “unit cost dropped 30%!” because traffic doubled, not because anything got more efficient. Always decompose: was the improvement from volume (denominator grew), efficiency (numerator shrank per unit), or rate (cheaper pricing)?
- Ignoring reliability cost — headroom, failover capacity, and redundancy are part of unit cost. Removing them “improves” the metric while increasing risk. A correct unit cost includes the insurance premium for availability.
- Comparing incomparable units — cost-per-query for a simple lookup vs a complex ML-powered search. Unit economics only works for benchmarking across time for the same service, or across services doing genuinely equivalent work.