The Business of Building God

This piece analyzes AI frontier labs as businesses, arguing their real moat is being one or two model-generations ahead rather than sheer investment or margins. That lead arises from superior talent, compute, and privileged access to user data, but open-source projects and competitors are steadily closing the gap. As CFOs and customers push back on rising token and inference spend, enterprises are rationing access to top models and routing workloads to cheaper alternatives, exposing a fragility in a strategy built on superiority alone. To sustain revenue, frontier labs are diversifying - ads, robotics, wet labs, bespoke vertical products - but turning frontier intelligence into multiple $100B-scale businesses is extremely difficult and will invite aggressive competition and market fragmentation.

The true existential prize is recursive self-improvement (RSI): automated researchers that can iterate successors and accelerate generations. That outcome is deeply uncertain; recent breakthroughs show models can conquer narrow, impressive tasks while still failing at many practical ones, challenging a simplistic scalar view of intelligence. The likely equilibrium is continued expensive, incremental advances, specialized superintelligences for particular domains, and a competitive ecosystem resembling cloud markets rather than permanent monopolies. Regulatory attempts to lock models down are possible but costly and awkward; overall, dominance depends on sustaining practical, monetizable advantages, not on any guaranteed path to universal ASI.