The AI funding boom: who is holding the risk?

The best indicator of where we are in the AI cycle may sit in the bond prospectuses, not the model releases.

Marc Rubinstein reports that estimates of AI infrastructure capex through 2030 range between US$7.6 trillion and US$8.0 trillion. Bond markets are deep, but not that deep. KKR analysts reckon traditional US investment-grade investors can provide about US$1.7 trillion, leaving a gap of US$6.3 trillion. They expect it to be filled by private investment-grade debt, asset-backed finance, infrastructure debt, structured solutions and, in their words, "financing structures that may not even exist yet.

Rubinstein draws the historical parallel. Between 1870 and 1890, railway capex ran at about 2.24% of GDP. AI capex could reach 3.63% of GDP out to 2032. The railway boom produced first and second mortgage bonds, mortgages on the trunk line and on the branches, and a whole array of instruments that took investors, in the words of the passage he quotes, "potentially, into the drink.

Financial engineering is not a villain. It allocates risk to those best able to carry it, and it is how a continent got railways. But it flourishes precisely when ordinary balance sheets can no longer keep pace, and the structures that keep a boom going are also the ones that make it harder to see where the risk is accumulating.

Rubinstein's point is not that the build is a mistake. It is that new structures get invented to fund booms this size, and the labels on them do not tell you who ends up holding the loss.

For boards and investors, the useful question is not whether AI is real. Both can be true: the technology works and the funding structure is fragile. The question is who is holding the risk on the data centre, in which instrument, and whether they know it.

Source: Marc Rubinstein, Net Interest, "The Art of Doing Financial Engineering" (3 October 2026), https://lnkd.in/gkHBaEPd

#AIInfrastructure #CapitalMarkets #RiskManagement