The analysis argues that forecasts for AI infrastructure spending often assume unlimited demand without examining who will use the computing capacity or what value it creates. It frames tokens as units of work time, using benchmarked model capability to distinguish tasks that current or near-future AI can handle from longer-horizon tasks. It also divides work into bounded tasks, whose value and complexity eventually cap out, and unbounded tasks, where more effort can keep producing gains. Bounded work such as routine administration should attract heavy AI use but favor cheap models, since the main benefit is limited cost savings. Unbounded work - such as AI research, software development, and trading - can justify premium frontier models because speed and capability may translate into substantially greater revenue.
The author estimates that these three unbounded uses account for roughly half of frontier labs’ inference revenue, and argues that their spending is reflexive: more tokens can improve models, help startups build and earn faster, or strengthen trading strategies, creating resources for still more AI investment. This feedback loop supports rapid growth and gives frontier models an advantage over cheaper alternatives. But it also makes demand fragile. AI labs, startups, venture investment, financial markets, and quant firms are linked, so a downturn in any could reduce spending across the others. The same cycle that amplifies growth could therefore magnify a reversal, triggered by factors such as regulation, higher interest rates, or an external shock. The central case is optimistic about long-term AI demand but warns that frontier-token demand is unusually concentrated and procyclical.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.