Estimation question

Estimate the compute cost of hosting ~1M free interactive Spaces.

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

Structured estimation under real infrastructure constraints, not just multiplying arbitrary numbers.

State assumptions upfront: assume 1 million Spaces exist, but only a fraction are actively visited or running at any time, since most free Spaces sleep when idle.
Estimate active fraction: assume 5 percent, or 50,000 Spaces, are awake at a given moment due to a recent visit, based on typical idle-timeout behavior.
Estimate per-Space compute: assume a modest CPU or small GPU instance costing roughly 0.10 to 0.50 dollars per hour depending on tier.
Multiply: 50,000 active Spaces times an assumed 0.20 dollars per hour average equals roughly 10,000 dollars per hour, or about 240,000 dollars per day.
Sanity check against scale: note this is a rough order of magnitude, and the real number depends heavily on idle-timeout policy and GPU versus CPU mix.

Explicitly separates total Spaces from concurrently active Spaces, since billing only for active compute matters, not the full 1 million.
States the idle-timeout assumption clearly, since Hugging Face sleeps inactive Spaces specifically to control this cost.
Shows the arithmetic cleanly and arrives at a daily figure, clearly marking it as an assumption-based estimate, not a real reported number.
Notes the biggest cost driver is the small share of Spaces using GPUs rather than the large share on free CPU tiers.

Multiplying the full 1 million Spaces by an hourly cost without accounting for idle sleeping, which wildly overstates the number.
Skipping the active-fraction assumption entirely and jumping straight to a final number.

How would this estimate change if Hugging Face tightened the idle-timeout window?
What lever would you pull first if this cost became unsustainable?