Interview, Fireside Chat
OpenAI CFO Sarah Friar: IPO, AI Rivalries, New Device, and Spending $100B+ on Compute
- Global productivity, affluence, and quality of life are anticipated to rise, with long-term market performance viewed as independent of IPO timing due to a fundamental "weighing machine" dynamic.
- Future growth relies on building durable companies using a single foundational model approach to generate a compounding advantage through scale, data, and personalization.
- Increasing model efficiency is expected to lower per-token costs and improve gross margins, allowing the company to model revenues and costs accurately for 2026 and 2027 based on product shapes.
- A significant compute supply deficit is forecasted for 2026, remaining constrained into 2027, which will drive a shift toward global inference distribution while keeping training centered in the United States.
- External infrastructure capacity, such as the 1-gigawatt Michigan data center, is not expected to provide compute until late 2027 or early 2028, reinforcing the predicted shortages in the medium term.
- Capital expenditure requirements are set to increase through partnerships like SoftBank Energy, shifting the business model toward a "built-to-suit" environment while maintaining multi-chip strategies with Vera Rubens and Cerebrus (AMD) chips.
- Revenue drivers for years beyond 2027 are projected to depend on the volume of compute already acquired rather than current product development, with potential agentic revenue models reaching $2,000 monthly per developer.
- A new consumer earpiece-like substrate is scheduled for unveiling by the end of the current year and general availability early next year, aiming to create a natural, paradigm-shifting user experience.
- Video capabilities are expected to expand significantly toward multimodality, while the company plans to capture larger search market share by leveraging conversational models for high-intent queries.
- Enterprise strategies will focus on integrating memory, context, and intuition to drive efficiency for C-suite operations, while consumer offerings will maintain an ad-free tier alongside ad-supported models.
- Competitive success is predicted to favor entities maintaining direct customer proximity to avoid abstraction, with the infrastructure layer ultimately serving consumers, small businesses, enterprises, and governments globally.
- Historical cost data indicates a 97% price increase for the "five four" model compared to the "five five," resulting in a 20-30% cost reduction per token for customers despite overall price adjustments.