The boom in AI related capital expenditures is generating questions about its sustainability. We believe that quantifying the size of the capex surge may offer hints as to its sustainability. We believe the current AI capex is around 300-350 BN USD in CY25 roughly 1% of the US GDP. This is not much if we look at historical figure of 1.5-2% of US GDP being spent on telecom investments way back in 2000 or with the 6% of GDP being spent on railroads in the late 1800s. The railroad capex averaged 2.4% of GDP in the 1870s and 1880s but fluctuated wildly, approaching 6% near the beginning of the period and collapsing to 0.3% during the long depression that followed the financial crisis of 1893 before beginning a long era of steadier but lower investment relative to GDP. So, overshooting and correction is in the nature of investment booms. Also continued spending increases by frontier model developers will be necessary to remain competitive. It is hard to know at what point the technological incentives to invest will diminish, but the combination of promising early productivity uplifts and continued improvements in model performance from increased computational resources suggest we are not at that point yet. So, while investment should eventually moderate as the AI investment cycle moves beyond the build phase and declining hardware costs dominate, the technological backdrop still looks supportive for continued AI investment. From a financing point of view, current levels of operating cash flows for listed tech companies (1.6 TN USD) is sufficient to cover current AI capex levels. Also a substantial share of the AI data centre capex expansion over the next five years could in principle be funded through internally generated cash flows. In particular, taking into account our capex and cash flow projections, the listed tech sector globally would swing from a financial surplus of $300bn this year to a deficit of $600bn by 2030. But they could get support from debt financing as well as private funds during this period. So, while at face value Nvidia’s CEO’s prediction that annual global AI data center expenditure would reach $3tr-$4tr by 2030 might appear excessive, our view is that this is manageable from a financing perspective. That said, there are other potential constraints that could make it more difficult, such as power generation and transmission capacity. We also conclude that current AI capex is still not a huge no as implied by nominal numbers with respect to historical capex cycles of telecom and railroads. But what might act as the only headwind to the AI capex cycle is profitability concerns around AI projects. Will the size & cost of financing be as benevolent as it is now? Only time might tell.
