OpenAI is openly rethinking how it makes money, and its chief financial officer, Sarah Friar, has just offered one of the clearest hints yet about what’s coming next. In a recent episode of The OpenAI Podcast, Friar introduced the idea of “licensing models” as a potential new revenue stream — a move that reflects the company’s growing need to fund massive computing expenses.
Under this model, OpenAI wouldn’t just charge upfront fees or subscriptions. Instead, it could take a percentage of downstream revenue when customers use its AI to build successful products. Friar gave the example of drug discovery: if a pharmaceutical company licenses OpenAI’s technology and develops a breakthrough drug, OpenAI could receive a share of the drug’s sales. According to Friar, this approach creates strong alignment between OpenAI and its customers, rewarding both sides when innovation succeeds.
This marks another step in OpenAI’s steady evolution from a single-product company into a diversified AI platform. What began with a basic ChatGPT subscription has expanded into multiple pricing tiers, enterprise SaaS plans, and usage-based credit systems for customers who want higher performance or scale.
At the same time, OpenAI is cautiously exploring advertising. Friar emphasized that any ad-supported approach would preserve trust: the AI should always provide the best answer, not a sponsored one, and an ad-free tier would remain available. This represents a noticeable shift from CEO Sam Altman’s earlier position, when he called advertising a “last resort.” As compute costs balloon — reportedly reaching long-term commitments of around $1.4 trillion — the company’s stance has softened.
These changes come after OpenAI’s restructuring into a more traditional for-profit organization, completed in October. Altman has said this move will make it easier to raise capital and sustain long-term growth.
Taken together, licensing, ads, and flexible pricing signal a pragmatic turn: OpenAI is no longer just building powerful AI — it’s building a business model capable of supporting it.