Reuters’ October 21 market column assembles several measures that should not be collapsed into one AI-demand number. JPMorgan Private Bank’s Michael Cembalest estimates that 41 AI-related stocks have produced 75% of S&P 500 returns, 80% of earnings growth and 90% of capital-spending growth since ChatGPT launched. At the same time, CoreWeave’s October 2031 bond trades below 80 cents on the dollar, more than 20% below its April launch. Reuters
Usage and revenue are also separating. Token consumption is still rising strongly, while spending growth has slowed as token prices fall. That is good for a buyer’s unit economics and ambiguous for a model provider’s revenue trajectory. It is possible for useful work to increase while the dollar value of each unit declines.
Power is the third constraint. Morgan Stanley analysts estimate a 30–40% US data-center power shortfall through 2028, equivalent to 34% of total chip demand. Some purchased chips may therefore depreciate before grid connections let them earn. A CTO choosing capacity should distinguish contractual compute, delivered hardware and energized, schedulable capacity.
The long-horizon arithmetic is sobering but assumption-heavy. McKinsey and S&P Global estimate more than $7 trillion of top-six US hyperscaler AI capex from 2025 through 2030. Columbia’s Stijn Van Nieuwerburgh estimates $10.3 trillion of financing through 2032 and roughly $3.55 trillion of annual revenue by then to earn a 10% return. Those figures frame a hurdle; they are not a forecast that the revenue will arrive.