sarah friar

OpenAI CFO confirms severe compute shortages through 2026-27, with 1GW data centers requiring ~$50B all-in capital; bottlenecks span power, land, memory, talent, and regulatory approval, creating multi-year infrastructure supercycle.

Massive deflationary curve in compute cost per token (97% drop in two years) enables margin expansion; OpenAI is raising model prices 2x while customer per-token cost still falls 20-30%, demonstrating pricing power and improving unit economics.

OpenAI's enterprise business is scaling rapidly with balanced revenue mix; Codex coding agent shows strongest adoption in sales/GTM teams, and enterprises like Thermo Fisher are deploying for R&D acceleration and field productivity.

Frontier model companies are diversifying across Nvidia (priority), AMD, Cerebras, and custom Broadcom silicon to avoid single-vendor lock-in and ensure continuous access to best-in-class compute for training and inference.

OpenAI is transitioning from pure OpEx CSP model to built-to-suit data centers (Oracle Michigan 1GW, SoftBank Texas) for greater control, though still leveraging investment-grade partners for financing until OpenAI achieves investment-grade status.

Contrary to commoditization fears, the agentic harness layer (memory, context, enterprise intuition) makes models more sticky and differentiated, allowing the intelligence layer to capture the largest profit pool in the AI stack.

OpenAI is developing a screenless, voice-first consumer device with Jony Ive's team (earpiece/puck form factor) targeting natural, intimate interaction; could redefine AI interface paradigm beyond smartphones.

OpenAI sees advertising as eventual monetization for free tier, leveraging unique combination of user memory, conversation context, and high-intent queries — potentially more valuable than Google/Meta models — while preserving ad-free paid tiers.