Estimation question

Estimate how many GPUs OpenAI needs to serve 1 billion weekly ChatGPT 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

Tests structured estimation: converting a user base and usage pattern into an infrastructure requirement.

Estimate daily active usage: assume 1 billion weekly users translates to roughly 300 million daily active users, with an average of 10 messages sent per active user per day.
Calculate daily message volume: 300 million users times 10 messages is 3 billion messages a day.
Estimate compute per message: assume each message requires roughly 1 second of GPU inference time on a modern accelerator for a typical response length.
Calculate total daily GPU seconds needed: 3 billion messages times 1 second is 3 billion GPU seconds a day, or roughly 35,000 GPU hours running continuously every hour of the day when annualized per hour.
Convert to GPU count: 3 billion seconds divided by 86,400 seconds in a day is roughly 35,000 GPUs running continuously to handle that steady state load, then add meaningful headroom, roughly 2 to 3 times that, for peak traffic and redundancy, landing around 70,000 to 100,000 GPUs.

Builds the estimate step by step from users to messages to compute time to GPU count, rather than jumping to a final number.
Explicitly accounts for peak versus average load with a headroom multiplier, since steady state average demand understates the GPUs actually needed to handle real world traffic spikes.
Converts total compute seconds into an intuitive GPU count using a clear unit conversion, showing the math transparently.

Estimating only average load and ignoring the need for headroom to handle peak traffic and redundancy.
Skipping the message volume calculation and guessing a GPU count directly.

How would this estimate change if a large share of usage came from longer, more complex conversations?
What would happen to the GPU requirement if model efficiency improved by half?