The current AI breakthrough is more analogous to the railroad than the PC. The leap forward came from massive scale and resource investment, not just a new algorithm. This infrastructural build-out will enable entirely new business models, much as railroads enabled mail-order catalogs.

Historical tech cycles like the cloud and mobile demonstrate a consistent pattern: the application layer ultimately generates 5 to 10 times the value of the underlying infrastructure capital expenditure. With trillions being invested in AI infrastructure, future value creation at the application layer will be astronomically larger.

The focus in AI has evolved from rapid software capability gains to the physical constraints of its adoption. The demand for compute power is expected to significantly outstrip supply, making infrastructure—not algorithms—the defining bottleneck for future growth.

Massive investments in AI hyperscalers are not the end game. They are laying foundational infrastructure, like the 19th-century electrical grid, which will enable a future explosion of derivative applications across all industries.

While model performance gains headlines, the true strategic priority and bottleneck for AI leaders is the 'main quest' of securing compute. This involves raising massive capital and striking huge deals for chips and infrastructure. The primary competitive vector has shifted to a capital war for capacity.

The current AI boom focuses on GPUs for "thinking" (Gen AI). The next phase, "Agentic AI" for "doing," will rely heavily on CPUs for task orchestration and memory for context, creating new investment opportunities in this previously overshadowed hardware.

In the rapidly advancing field of AI, building products around current model limitations is a losing strategy. The most successful AI startups anticipate the trajectory of model improvements, creating experiences that seem 80% complete today but become magical once future models unlock their full potential.

Karpathy identifies the AI community's 2010s focus on reinforcement learning in games (like Atari) as a misstep. These environments were too sparse and disconnected from real-world knowledge work. Progress required first building powerful representations through large language models, a step that was skipped in early attempts to create agents.

AI is currently a challenging business because it's in a heavy infrastructure investment cycle, similar to the early days of the web or cloud. Significant value creation typically occurs years after this initial investment phase, and the market isn't there yet.