AI capex spending is growing at twice the rate of the 2000s housing bubble, signaling unprecedented infrastructure investment intensity.
The U.S. economy has not seen capital pour into a single sector as rapidly as it currently is in artificial intelligence infrastructure. That comparison to the last such moment—the housing boom—is the framing Torsten Slok, Partner and Chief Economist at Apollo Global Management, is using to assess the current AI infrastructure buildout. He is not predicting a crisis, but the historical parallel warrants scrutiny.
Slok's analysis tracks hyperscaler data-center capital expenditures as a share of U.S. GDP. The trajectory is steep: from 0.3% of GDP in 2019, to 1.4% in 2025, and a projected 3.1% by 2027. That two-year jump of 1.7 percentage points represents roughly 0.85 percentage points annually. By contrast, during the housing boom from 2002 to 2005, residential construction added only about 0.5 percentage points to its GDP share each year. AI capex is running at nearly double that pace.
The absolute scale, however, differs significantly. Housing's GDP contribution peaked at 6.6% in 2005. Even at its 2027 projection, data-center capex sits at less than half that level.
A useful historical comparison is the telecom boom of the late 1990s, which peaked at around 1.2% of GDP in 2000. The AI buildout is on track to more than double that figure—and it is doing so backed by far more established revenue bases. Amazon, Meta, Microsoft, Alphabet, and Oracle are not speculative startups; they are profitable businesses making enormous forward bets on infrastructure they believe necessary.
Apollo's analysis projects data-center capex will hold around 3% of U.S. GDP annually from 2027 through 2029—a sustained plateau that no technology infrastructure sector has previously maintained. To contextualize this: the U.S. economy produces roughly $28 trillion annually. Three percent of that is approximately $840 billion per year, directed almost entirely at a single category of technology infrastructure.
Slok's warning is precise: rapid build cycles tend to unwind rapidly. Companies spend ahead of demand based on their projections of what they will need. If demand materializes, the capex looks prescient. If it does not, assets sit underutilized and companies pull back hard and fast.
AI infrastructure differs from residential housing in its financing structure—there are no mortgage-backed securities tied to GPU clusters. But the macroeconomic impact of a sudden pullback in spending worth several percentage points of GDP would still be significant. The pressure would concentrate not at the top of the stack but among the more exposed suppliers: chip manufacturers, power companies, data-center real estate investment trusts, and smaller software companies betting that hyperscaler infrastructure translates into enterprise AI demand. If the capex plateau arrives without corresponding revenue growth, that ecosystem layer absorbs the shock.