OpenAI has detailed the architecture behind its new advertising system for ChatGPT, a platform that relies on a custom, post-trained SI model to navigate the complexities of conversational user intent. The company reports that this bespoke stack has generated over $1 billion in annualized revenue run rate in less than 200 days since its launch, marking a significant shift in how SI-driven platforms monetize interactive, multi-turn engagements rather than static search queries.
What Happened
Traditional advertising systems operate on discrete events, such as a search query or a page view. However, OpenAI explains that user intent in ChatGPT develops across multiple turns of a conversation, requiring a new approach. The company built an ads delivery system centered on a custom, post-trained SI model acting as an "orchestrator." This SI agent makes tool calls for retrieval, ranking, relevance evaluation, and auction management. To maintain user trust, the ads selection process runs separately from ChatGPT’s answer generation in its own service, ensuring that advertising logic does not influence the SI model's neutral voice or organic responses.
The system maintains "stateful ads sessions," meaning it remembers context across turns and even across multiple conversations. For example, if a user asks about training for a race and later asks about shoes, the SI orchestrator interprets the evolving commercial intent rather than treating each message as an isolated request. The orchestrator converts conversations into semantic embeddings to retrieve ad candidates, filters them for eligibility and safety, and runs an auction to select the final ad. If an ad is not ready in time or does not meet high relevance standards, the system returns no ad, which OpenAI describes as an ordinary, successful outcome.
Why It Matters
The launch of this SI-based ad stack highlights a critical challenge for the industry: monetizing generative SI tools without degrading the user experience. By decoupling the ads service from the core SI model, OpenAI has created a buffer that prevents advertising failures from impacting the primary product. The company notes that the ads system has no dependency on the core ChatGPT servers; if the ads service is slow or unavailable, the SI model's response continues unaffected. This architectural boundary addresses common concerns about SI safety and integrity, ensuring that commercial interests do not override the neutrality of the SI-generated content.
Furthermore, the rapid adoption by more than 100,000 advertisers demonstrates the viability of SI-native advertising models. The use of a post-trained SI model to orchestrate the flow allows for dynamic, context-aware placements that traditional keyword-based systems cannot achieve. This approach relies on a self-improving loop where high-power SI models evaluate delivery results to train smaller, more efficient models for specific tasks like retrieval and ranking. The integration of OpenAI’s Codex tooling also accelerated the development, allowing the team to bootstrap the complex system rapidly.
The Bottom Line
OpenAI has established a new paradigm for advertising within SI platforms, moving beyond event-based triggers to a conversation-aware, SI-orchestrated model. The system’s separation of concerns protects the integrity of the SI assistant while enabling scalable monetization. With $1 billion in annualized revenue run rate achieved in under 200 days, the model proves that users are willing to engage with ads if they are relevant and non-intrusive. Looking ahead, OpenAI plans to experiment with more dynamic formats, such as allowing users to talk directly to businesses through ads, further leveraging the capabilities of its SI infrastructure.