Box 6.1 Boom, bang, bust in technology investments: Does history repeat itself? (continued) building ahead to meet anticipated demand. An analysis from 2018 shows that demand for compute has been doubling every 3.4 monthsc—a figure that is remarkably close to the 100day doubling cited in the early days of the internet. Elevated levels of capital investment. In anticipation of future demand, in past technology booms as well as the current AI wave, levels of capital investment have soared. The belief that the use of railroads and the internet would grow exponentially led to frenzied building of infrastructure. Between 1868 and 1871, rail companies in the United States built 33,000 miles of tracks, often in remote areas.d During the telecommunications boom, capital investment by publicly traded telecommunications service companies nearly tripled from $47 billion in 1995 to a peak of $121 billion in 2000.e The increase was so great that only 2.7 percent of fiber-optic lines were being used in 2002 when the sector crashed.f Strikingly, capital spending on AI by hyperscalers in the United States is expected to be about 2 percent of GDP in 2026, more than double the figure at the height of the telecommunications boomg and about the same as that on railroads in the 1860s.h High level of debt. Elevated levels of capital investment that were mainly funded by debt have characterized past technology booms and the current AI wave. When financial crisis in Europe led to investors selling off railroad bonds in the 1870s, the market was flooded with more bonds than buyers were willing to absorb.i Telecommunications firms had run up total debts of about $1 trillion at the time of the crash.j Twenty-three telecommunications companies went bankrupt in a wave capped off by the collapse of WorldCom in 2002, the single largest bankruptcy in the history of the United States at that time.k Today, concerns are growing that AI hyperscalers are taking on excessive debt, as seen in their increasing share in the public investment-grade bond index and the increasingly complex structures of deals between AI companies.l Source: WDR 2026 team. a. Economist (2002b). b. Couper et al. (2003). c. OpenAI (2018). Compute is defined in this article as petaflop/s-days. A petaflop/s-day consists of performing 1015 neural net operations per second for one day, or a total of about 1020 operations. d. Wooldridge (2025). e. Doms (2004). f. Dreazen (2002). g. Slok et al. (2026). h. Wooldridge (2025). i. Refer to “Financial Panic of 1873” (web page), US Department of the Treasury, https://home.treasury.gov​ /about/history/freedmans-bank-building/financial-panic-of-1873. j. Economist (2002a). k. Starr (2002). l. Sam et al. (2026). 210 World Development Report 2026