What happens when the AI capex cycle slows?

AI-related capital spending is expected to grow at around 40% in 2026, according to Wolfe Research, with another 30% increase forecasted for 2027. Despite a slowdown in spending pace, strong hyperscaler balance sheets, a robust project pipeline, and competitive dynamics continue to fuel elevated expenditures. Quarter-over-quarter growth in AI capex has exceeded 80% since early 2022, before sharply decelerating through 2023 and stabilizing at around 40%.

Wolfe's base case anticipates further deceleration towards 10%-15% annualized growth by 2028, marking a soft landing for the capex cycle. Domestic AI-related capital spending, encompassing equipment, software, R&D, data centers, and CHIPS facilities, has surged from near-zero in 2021 to roughly 2% of GDP in the second quarter, representing one of the fastest investment booms in recent history on a capex-to-GDP basis.

The direct impact of AI investment has contributed to 10%-20% of nominal GDP growth in recent quarters, excluding the negative accounting effect from imports. Over 700 data centers are currently in the pipeline, equivalent to adding 16% to the existing stock. These data centers typically take three to six years to become fully operational, creating employment gains of approximately 3.5% and wage increases of 5% in the host countries.

However, employment in operational data centers is limited, ranging from 50 to 400 permanent workers, compared to the thousands employed during construction. Despite high hyperscaler capex relative to operating cash flow, the gap between AI services revenues and capex estimates is expected to narrow in the coming quarters. The risk of an orderly deceleration turning disorderly arises from an equity repricing among hyperscalers, which hold a dominant share of S&P 500 market capitalization.

This repricing could accelerate capex pullbacks, giving rise to a circularity risk. Wolfe notes that conditions for this disorderly scenario, such as disappointing AI revenue growth and multiple compression in a risk-off environment, are not its base case, emphasizing that the risk pertains to market dynamics rather than the technology itself.