Cost per Request
Total cost divided by units of business value — the metric that makes cost comparable across time, scale and design options.
Definition
Cost per request, per order, per active user, per transaction. It normalises spend against value, so that a rising total bill can be distinguished from a rising inefficiency.
Why the absolute bill is the wrong metric
A bill growing 40% while traffic grows 60% represents improving efficiency, and reporting it as a cost problem produces exactly the wrong response. A bill flat while traffic halves is a serious problem that a total figure hides completely.
Unit cost is also the only form in which cost is comparable to revenue. "£0.004 per order against £28 average order value" is a sentence a business can act on; "£140,000 a month" is not.
How to compute it usefully
- Choose a unit that maps to business value. Requests are convenient and often wrong — a health check is a request. Orders, active users, messages delivered, videos started.
- Allocate all costs, including the shared ones — observability, networking, platform teams, licences. Unallocated shared cost is where inefficiency hides.
- Segment. Cost per request for the search endpoint and for a static asset differ by orders of magnitude. An aggregate hides which endpoint is expensive.
- Track over time, and expect it to fall with scale. If unit cost is flat as volume grows, nothing is amortising and the architecture is linear in a place it should not be.
What it enables
Architectural decisions with a price. "Adding this cache reduces cost per order from £0.006 to £0.004" is a decision anyone can evaluate. It also sets a floor: if unit cost approaches unit revenue, the product is structurally unprofitable and no amount of operational tuning will fix it — which is a finding worth surfacing early rather than at scale.
Capacity forecasting. Unit cost times projected volume is a budget, and it is far more reliable than extrapolating a total.
Failure scenarios
- Only the total tracked, so efficiency changes are invisible.
- A unit chosen for convenience rather than value, so the metric does not move with the business.
- Shared costs unallocated, so the largest inefficiencies are in the unattributed pool.
- Aggregate only, hiding the one expensive endpoint.
Interview question
"Your infrastructure bill rose 40% this quarter. What do you need to know before deciding whether that is a problem?"