Bank of America warns AI’s investment boom could outgrow its financial rewards

AI does not have to fail for its investors to disappoint. In a client note reported on October 16, Bank of America strategist Sebastian Raedler warned that competition and falling prices could erode financial returns even as the technology delivers. His concern is the widening gap between infrastructure spending expectations and the money AI usage generates.

A spending forecast larger than earlier booms

Bank of America said consensus forecasts for capital expenditure by hyperscalers—the largest cloud operators—over the following 12 months have climbed to approximately $1 trillion. In early 2025, the corresponding forward-looking estimate was $330 billion. These are spending expectations, not a completed investment total.

The bank puts the latest forecast at around 3% of US gross domestic product, a measure of economic output. Its historical comparisons are smaller: capital spending during the 1870s railroad expansion reached an estimated 2.2% of GDP, while investment associated with the 1990s technology expansion represented approximately 1.1%.

Raedler’s comparison concerns the economics of building capacity, rather than whether the underlying inventions were useful. He argues that earlier investment cycles often produced disappointing financial returns as competition increased, even when the technologies continued developing and gained adoption. Technological progress and attractive investor returns were not the same outcome.

Competition changes the cloud comparison

The mechanism Raedler describes starts with strong profits attracting more investment. As companies add capacity and compete for customers, prices come under pressure. That can reduce returns on the money committed to the buildout. The risk is not simply that customers reject AI; suppliers may struggle to earn enough from growing demand.

Raedler contrasts that pattern with cloud computing in the 2010s. In his account, barriers to entry limited competition and helped sustain industry profitability. He sees a different competitive environment in AI models, including pressure from Chinese developers offering open-weight models, whose underlying model parameters are available to others.

The bank says average AI token prices have fallen approximately 50% since summer. Tokens are the pieces of text AI models process and a unit providers use to charge customers. Lower prices mean a provider must sell more usage just to offset the decline in revenue earned from each unit.

More usage has not meant more spending

There is a counterargument: cheaper AI could encourage so much additional use that total spending keeps rising. The bank discusses this through Jevons Paradox, the idea that greater efficiency and lower costs can stimulate consumption. Supporters of continued investment growth argue that this demand response could offset falling token prices.

Raedler points to a less reassuring recent signal. Ramp AI data cited by the bank show US business spending on AI declining slightly since August, despite a substantial increase in token usage. That is evidence about one recent spending trend, not proof that lower prices cannot eventually unlock a larger market.

For Raedler, the divergence is a warning about turning consumption into revenue. The bullish case needs demand to grow enough to sustain total spending, not merely a rising count of tokens processed. His note identifies the recent mismatch as a possible sign of the pricing pressure that damaged returns in earlier capital cycles.