The rapid pace of artificial intelligence infrastructure investment is reshaping global capital expenditure patterns at an unprecedented rate, but its potential macroeconomic risks are equally significant and cannot be overlooked.
This week, Torsten Slok, Chief Economist at Apollo Management, analyzed that capital spending by major hyperscale cloud computing companies is projected to rise from 1.4% of GDP in 2025 to roughly 3% between 2027 and 2029. At that peak, it would more than double the level seen during the late 1990s telecom and fiber optic construction boom.
While this investment surge remains below the absolute peak of the 2005 housing boom, it has already surpassed all previous capital expenditure cycles in two key metrics: the cumulative increase as a share of GDP and the speed of its build-up.
Slok warns that it is precisely the velocity of the investment cycle that determines the destructive force of its reversal.
Historical data shows that a sharp decline in housing investment as a share of GDP was a core factor driving that severe recession. In contrast, the relatively smaller telecom bubble burst only triggered the mildest recession in the post-war era. If AI demand fails to meet expectations, a similarly rapid reversal could occur, posing a significant macroeconomic downside risk.
Construction speed: outpacing every historical cycle
Examining this AI capital expenditure boom from three dimensions provides a clearer picture.
In terms of scale, data center construction investment falls between the two previous large capital expenditure cycles. It is more than double the peak of the late 1990s telecom and fiber optic boom (around 1.2% of GDP in 2000) but less than half the peak of the 2005 housing boom (6.6% of GDP).
From a marginal change perspective, the core metric is not the absolute share of investment, but its increase as a percentage of GDP, which directly reflects its marginal contribution to GDP growth.
Data center capital expenditure is expected to rise from 0.6% of GDP in 2023 to 3.1% by 2027, a cumulative increase of about 2.5 percentage points. In comparison, late 1990s telecom investment rose by only 0.4 percentage points, while housing investment increased by about 2.2 percentage points from the mid-1990s to 2005.
By this measure, the AI data center boom is already the largest capital expenditure cycle in history.
In terms of construction speed, the difference is even more pronounced. Data center capital expenditure is forecast to rise from 1.4% of GDP in 2025 to 3.1% in 2027, taking just two years, with an average annual increase of about 0.85 percentage points.
The fastest phase of the housing boom (2002 to 2005) saw an average annual increase of only about 0.5 percentage points, while the telecom construction boom was roughly 0.15 percentage points. The speed of the AI data center build-up is nearly double the fastest phase of the housing boom.
The other side of the risk: the reversal could be just as stunning
Slok's analysis reveals a symmetry logic: the speed of the build-up dictates the potential ferocity of the reversal.
History provides two reference points:
The reversal of the late 1990s telecom bubble was relatively limited in scale, and the economic impact was mild. The US experienced only the mildest recession of the post-war era.
In contrast, the rapid contraction of housing investment—by about 3.2 percentage points from early 2006 to the end of 2008—was a major driver of the severe financial crisis and deep recession.
In the current cycle, with data center capital expenditure climbing at roughly 0.85 percentage points per year, this investment scale could contract at a similar speed if the pace of AI commercial demand realization falls short of expectations.
Slok explicitly states that the real macroeconomic risk is not the construction itself, but the potential for a subsequent investment reversal to impact economic growth if AI demand disappoints.
Comments