OpenAI details ChatGPT ad stack architecture achieving $1B annualized run rate

Details

OpenAI built a custom ad stack for ChatGPT that reached a $1 billion annualized run rate and attracted over 100,000 advertisers in under 200 days.
The architecture separates ad selection from organic answer generation to ensure ads do not influence response quality or cause latency.
A custom post-trained LLM orchestrator manages stateful ad sessions, interpreting conversation context to retrieve, rank, and deliver relevant ads.
The pipeline uses semantic embeddings and Approximate-Nearest-Neighbor (ANN) search to generate candidates, followed by eligibility checks and auction nomination.
High-power models evaluate system outcomes to train smaller, more efficient orchestration and retrieval models in a self-improvement loop.
Human feedback from users and advertisers steers product direction and taste, complementing model-based intent understanding.
Future plans include experimenting with more dynamic ad formats and direct user-business conversations.

Key point