A new AI agent is like a new hire; it knows little. But over time, by soaking in context, memory, and skills, its value compounds dramatically. Like a tenured employee, it can make excellent assumptions on the user's behalf, creating product stickiness and pricing power.
Labs are focusing on integrating downward into inference and compute, which are homogeneous and highly scalable workloads. They are avoiding the application layer because it is fragmented, idiosyncratic, and requires heavy operational spending on product, pricing, and packaging for diverse markets.
Most moats like brand and network effects are unaffected by AI. However, the difficulty of integrating with complex enterprise systems like SAP, a historical moat for incumbents and System Integrators, is being dismantled by AI coding agents that can automate and simplify these connections.
Non-technical entrepreneurs, who previously might have become YouTubers to build an internet business, can now use coding agents to create small but profitable SaaS products. This is creating a new wave of digitally native entrepreneurs and small business formation, distinct from venture-backed startups.
The current wave of AI founders are predominantly researchers and engineers, in contrast to previous cycles that saw more product managers and MBAs. While they may lack initial business sophistication, their deep technical expertise is the critical, hard-to-teach ingredient for success in this product cycle.
It's economically rational to use expensive, high-IQ frontier models for functions with unlimited upside, like sales or product development. For functions where the goal is precision rather than unbounded creativity (e.g., accurately closing financial books), cheaper, fine-tuned open-weight models are more efficient.
Models possess unique traits, much like human personalities (e.g., 'neurotic' and literal vs. 'open' and creative). This, combined with domain-level specialization (e.g., OpenAI for knowledge work), means a multi-model strategy is essential for building robust applications, as no single model is best for all tasks.