An architecture review must estimate the cost of a proposed design before it is built. What should the model contain, and what makes such estimates useless?

What the model should contain

1. The cost drivers, named explicitly. Not "compute, storage, network" but the quantities that scale: requests per second, gigabytes stored per month, events per day, cross-region transfer per request, supplier API calls per search. The model is a function of these.

2. The scaling relationship for each. Linear with users, linear with data, super-linear because of fan-out, or fixed. This is more valuable than the absolute number, because it tells you which decision matters at ten times the volume.

3. A range, not a point. Low, expected and high volume scenarios. A single number invites false precision and will be wrong.

4. The expensive lines identified. Usually two or three items dominate. Identifying them focuses the design conversation on what matters rather than distributing attention evenly.

5. The comparison. Cost of this design against at least one alternative, including the do-nothing or simpler option. An absolute number without a comparison cannot inform a decision.

What makes estimates useless

1. Precision without accuracy. A number to four significant figures built on an unexamined traffic assumption. The assumption is the estimate; the arithmetic is decoration.

2. Omitting the dominant term. Cross-region transfer, egress, per-request charges on managed services, and observability cost are the lines most often missing — and frequently the largest. A model omitting telemetry cost for a high-cardinality system is wrong by a large factor.

3. Ignoring the cost of not serving — idle capacity, headroom, redundancy, non-production environments. These are real and are usually absent.

4. Excluding operational cost. Engineering time to build and run, on-call burden, and the cost of each additional system to operate. A design that is cheaper in infrastructure and requires a new stateful system to operate may be more expensive in total.

5. No revisiting. An estimate produced for approval and never compared against actuals teaches nobody anything, and the next estimate is no better.

The most valuable output

The sensitivity analysis: which assumption, if wrong by a factor of two, changes the answer?

That question identifies what to measure early and what to design against. A design whose cost is dominated by one uncertain quantity should be built so that quantity can be changed cheaply — which is an architectural requirement discovered by the cost model, and the reason the model earns its place in a design review.