Azeem Azhar on AI Profitability, Moats, and Infrastructure Challenges

The Crucible for AI Profitability

• 💡 The next 18 months are crucial for AI to prove real ROI, with evidence expected by mid-2026 from companies.
• 📊 Hard evidence means company-level data in earnings reports, showing productivity gains and revenue increases, not just token volumes.
• ⚠️ Many individuals benefit from direct chatbot use (e.g., ChatGPT, Claude) for personal productivity, which is harder for enterprises to track as ROI.

Building AI Moats and Strategies

• 🔑 Companies with structural data advantages (e.g., Waymo's pedestrian data) or those owning critical workflows can build strong AI moats.
• 🎯 Anthropic focuses on coding and software architecture, aiming for developer loyalty and mastery in translating human intent to code.
• 🌐 OpenAI pursues a "ubiquity strategy" like Netscape, aiming for widespread adoption and enterprise deals, but risks spreading resources too thin.

Physical Constraints and Infrastructure

• ⚡ AI's growth faces significant physical constraints including chips, data centers, and power grids, requiring complex two-year advance planning.
• 📈 The demand for compute is unbounded, extending beyond LLMs to physics, biology, and logistics simulations.
• 🌍 America's electric grid is a critical bottleneck, while Gulf countries leverage low energy prices and permissive permitting to become AI data center hubs.

The "Pragmatic Addicts" Dilemma

• 💬 A significant portion of Americans (75%) distrust AI but use it, creating a "pragmatic addicts" scenario.
• 🎭 The current narrative of AGI and job replacement creates dissonance and public resistance to AI development.
• ✅ Product makers should focus on designing tools for human augmentation rather than solely automation, shifting the narrative.

Augmentation vs. Automation

• 🛠️ Azeem's team uses AI for individual augmentation (e.g., chatbots for tasks) and automation for repetitive tasks (doing something 5x/week).
• 🚀 This approach transforms individual contributors into "managers of AI," enabling them to push quality and address higher-value backlog items.

Science Fiction Becomes Reality

• ✨ Modern LLM chatbots on phones, capable of complex tasks like risk modeling or coaching for conversations, are already science fiction made real.
• 💰 While new model training is expensive, the unit economics of existing, amortized models (e.g., Claude 3/4) are likely profitable.