Estimation question

Estimate Perplexity's monthly infrastructure cost to serve its free users.

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Estimation combining query volume, cost per query across a mixed model stack, and realistic free-tier usage patterns.

State assumptions: assume Perplexity serves roughly 20 to 30 million monthly active free users.
Estimate queries per user: assume an average of 10 queries per active user per month, giving roughly 250 million to 300 million free-tier queries per month.
Estimate cost per query: assume a blended cost of roughly 0.02 to 0.05 dollars per query, accounting for retrieval, search, and inference across a mix of smaller and larger models.
Multiply: 275 million queries times an assumed 0.03 dollars equals roughly 8.25 million dollars per month.
Sanity check: note this excludes other infrastructure costs like indexing and search retrieval, so the true figure is likely somewhat higher, and label the whole estimate as order of magnitude.

Breaks the estimate into monthly active users, queries per user, and cost per query rather than guessing a single blended monthly figure.
States the cost-per-query assumption as a range and explains what drives it, retrieval plus inference across a mixed model stack.
Arrives at a clear monthly figure and explicitly flags it as excluding other real costs like search indexing infrastructure.
Frames the entire estimate as order-of-magnitude, appropriately hedged given the many underlying assumptions.

Ignoring retrieval and search infrastructure costs entirely and estimating only raw model inference cost.
Skipping the per-query cost assumption and jumping to a final number without showing the reasoning.

How would this estimate change if Perplexity moved more queries to a cheaper, smaller model?
What is the single biggest cost driver you would try to reduce first?