OpenAI CFO Sarah Friar on the race to build artificial general intelligence

Leadership and Strategic Context

• Dan Dees (Goldman Sachs Global Banking & Markets Co-Head) and Sarah Fryer (CFO, OpenAI) discuss the urgency and scale of AI infrastructure investment.
• Fryer notes that OpenAI is pivoting from a pure "model company" to owning the full stack: infrastructure, API layer, and applications.
• OpenAI currently manages 400 million weekly active users, with revenue tripling annually for the third consecutive year.

Product Evolution and Breakthroughs

• OpenAI is shifting from predictive real-time models (ChatGPT series) to reasoning models (O-series) capable of complex "chain of thought" problem solving.
• 2025 Focus: The company is aggressively deploying "agents" (independent AI workers), citing three active tools:
  • Deep Research: An agentic tool for comprehensive reporting.
  • Operator: A task worker capable of executing web-based tasks (e.g., booking travel/dining).
  • A-Suite: An agentic software engineer that builds apps, conducts QA, and writes documentation without human intervention.
• Model Performance Benchmarks:
  • O3 ranks as the 175th best competitive coder globally and achieves PhD-level proficiency in physics, chemistry, and biology.
  • O3 Mini is already the #1 competitive coder in the world.
  • GPT-4.5 is optimized for emotional intelligence (EQ), creativity, and design rather than pure hard science.

Infrastructure and Capital Expenditure (Stargate)

• Stargate Investment: A $500 billion, four-year investment plan to secure 10 gigawatts of compute power.
• Power Constraints: 10 gigawatts exceeds the total IT/utility load of the entire island of Ireland (approx. 7 gigawatts).
• Strategic Necessity: Fryer cites compute scarcity as a primary bottleneck, noting that features like Sora and Deep Research were delayed due to insufficient hardware capacity.
• Geopolitical Implications: The investment is driven by a race against China (referencing DeepSeek) and involves national security considerations.
• Construction Challenges: Scaling requires overcoming labor shortages in specialized trades (electricians, HVAC) and securing renewable power sources.

Scaling Laws and Research Methodology

• Fryer identifies three concurrent "laws of scaling" driving AI advancement:
  1. Pre-training: Increasing general model intelligence via data and algorithmic efficiency.
  2. Post-training (Fine-tuning): Specializing models for specific domains (e.g., medical diagnostics for cancer vaccines) using niche datasets.
  3. Test-time Compute: Allocating additional inference compute to improve answer accuracy during specific tasks.
• Reinforcement Fine-Tuning: Research indicates that small amounts of specific data can yield massive utility uplifts in niche areas.

Market Adoption and Use Cases

• Education: Massive deployment scaling, including Arizona State University (181,000 seats), California State University system (500,000 seats), and a national rollout in Estonia's secondary schools.
• Enterprise Banking: Specific use cases include credit fraud detection, KYC/AML compliance, wealth management research, and client relationship management.
• Internal Efficiency: OpenAI utilized internal "hackathons" to identify routine tasks, resulting in custom GPTs that reduced repetitive diligence work for the Investor Relations team.

Future Outlook and AGI

• AGI Definition: Defined as the point where AI systems can perform the majority of real, value-added human work; Fryer suggests this is "imminent" and that models have already surpassed human capability in specific technical domains (physics, coding).
• Next Frontiers: Post-AGI development is expected to shift toward robotics, 3D interaction, and physical world tasks (factory work, farming).
• IPO Consideration: OpenAI is not currently prioritizing an IPO; Fryer states a public listing is "not on the immediate horizon" but acknowledges it as eventual "good hygiene" for capital access and operational rigor.
• Long-term Goal: The company intends to sustain "big, ambitious" cycles (90-day cycles may be insufficient) and continue aggressive capital deployment in compute and research.