No One Is Seriously Talking About AI Demand (12 min)

AI demand is ultimately labor demand, and distinguishing between Bounded and Unbounded tasks is key. Bounded tasks (e.g., tax filing) aim for cost reduction, so low-cost models dominate, whereas Unbounded tasks (e.g., AI research, space exploration) prioritize 'first' and 'speed,' allowing frontier models to monopolize them.

Approximately 50% of inference revenue for frontier AI labs comes from Unbounded tasks such as AI R&D, software engineering, and trading. These three sectors form a Reflexive feedback loop where token spending leads to revenue and capital raising, which in turn flows back into token investment. This loop has no upper limit, accelerating growth, but it exhibits a Super procyclical character closely correlated with market booms.

This structure is a double-edged sword. It explodes revenue in an upswing, but if exogenous shocks such as regulation or interest rate hikes occur, demand contraction can spread, causing revenue to plummet by up to 50%. Therefore, modeling that considers the vulnerability of demand alongside the supply model is essential.

In the long term, value is expected to accumulate in physical AI supply chains and teams performing Unbounded tasks. While the current market prefers predictable cash flows, teams that wisely allocate long-term Capex (e.g., Elon Musk's Tesla, SpaceX) will be revalued in the future. In the era of Superintelligence, it is anticipated that many such 'Elons' will emerge.