Category: Ch20: Economic Growth, the Financial System, and Business Cycles

Image created by ChatGPT, as are the other images in this post.

Technological breakthroughs are often embodied in machinery, equipment, or, as in the case of artificial intelligence (AI), software. For example, the Industrial Revolution of the late 1700s, which involved mass production of low-priced cotton textiles, was made possible by the development in England of large cotton spinning machines that were often powered by steam engines. 

Other technological breakthroughs are disembodied because they involve reorganizing production rather than employing new machinery or software. For example, the just-in-time (JIT) system that Toyota developed beginning in the 1940s, lowered the costs of assembling automobiles by, among other things, having parts arrive at a factory just when needed.

Across U.S. history, there have been a number of technological revolutions that required massive capital investments.  For instance, the transportation revolution of the early 1800s that significantly reduced the costs of shipping freight, relied on substantial investments in canals.  Later in the 1800s, transportation costs were reduced further as firms invested in the largest network of railroads in the world. Electrification of businesses and private homes in the late 1800s and early 1900s required the building of power plants and transmission lines. In the late 1990s, many firms competed to build networks of fiber-optic cable to carry the rapid increase in digital traffic resulting from the development of the internet.

Each of these technological revolutions succeeded in providing consumers with new or lower-cost goods or services. But building the physical capital that embodied these technologies could result in substantial losses or bankruptcy. For example, in the 1830s, several states, including Pennsylvania, Illinois, and Michigan, defaulted on the bonds they had issued to finance canal instruction.

Similarly, in the 1890s, some of the largest private railroad companies in the United States, including the Union Pacific, the Northern Pacific, and the Atchison, Topeka & Santa Fe, went bankrupt. In the early 2000, investors realized that internet commerce wouldn’t expand as rapidly as they had expected. As a result, several prominent internet firms, such as Pets.com and eToys.com, saw their stock prices collapse in what was called the dot com crash, and eventually declared bankruptcy.  Several firms that had invested in building large fiber-optic networks, including Worldcom and Global Crossing, also failed.

Why did the companies making large capital investments in new technologies sometimes fail even though the technologies themselves proved to be revolutionary? The answer in part is that forecasting accurately the demand for a new product can be difficult. For example, ultimately the level of internet traffic was more than great enough to justify the investment in fiber-optic networks, but the increase came too late for some companies to avoid bankruptcy. Competition among firms to take advantage of a new technology can lead to overbuilding, as when railroad companies built more miles of track than were needed to meet the demand to transport freight and passengers.

A number of economists and investment analysts have noted that the very heavy investment in data centers used to power AI models resembles the large capital investments firms made during earlier technological revolutions. As the following figure shows, beginning in early 2023, shortly after the public release of ChatGPT, the first widely used AI chatbot, investment in information processing equipment and software as a percentage of all nonresidential fixed investment has increased dramatically.

Do the firms building data centers run the risk of overestimating demand for AI and overbuilding data centers? More generally, how does the build-out of AI data centers compare with the earlier build-outs of railroads and fiber-optic networks? Who better to answer this question than ChatGPT!

The following (very long!) answer was ChatGPT’s response to the following prompt: “In what ways is the current build-out of AI data centers similar to the build-out of the railroad network in the United States in the late nineteenth century and the build-out of the fiber-optic network in the United States in the 1990s and in what ways is it different?” (A quick summary is given in the brief “Bottom Line” section at the very end.)

AI Data Centers, Railroads, and Fiber-Optic Networks

The best way to frame the comparison is this:

The railroad analogy is most useful for understanding the physical scale, bottlenecks, public subsidies, and geographic consequences of the AI buildout. The fiber-optic analogy is most useful for understanding its financial risks, uncertain demand, rapid technological improvement, and potential for overcapacity.

AI data centers nevertheless differ importantly from both. They are not simply links in a transportation or communications network. They are more like factories that produce computing services, and the extremely expensive machinery inside them becomes obsolete much faster than either railroad rights-of-way or buried fiber.

The scale is already extraordinary. The International Energy Agency estimates that the largest technology companies spent more than $400 billion on capital investment in 2025 and expects that figure to rise about 75 percent in 2026. U.S. data centers consumed approximately 176 terawatt-hours of electricity in 2023—4.4 percent of U.S. electricity use—and Berkeley Lab projects consumption of 325–580 terawatt-hours by 2028.

Ways the three buildouts are similar

1. All three are investments in enabling infrastructure

Railroads did much more than generate revenue from tickets and freight charges. They brought previously isolated regions into a national market, encouraged settlement, changed the location of factories and cities, and made complementary investments in farms, mines, warehouses, and factories more valuable. The United States added roughly 170,000 miles of railroad track between 1871 and 1900. 

Fiber-optic builders similarly expected abundant bandwidth to generate applications that could not yet be foreseen. More bandwidth would encourage new internet services, which would increase bandwidth demand, which would justify still more fiber.

The AI argument is nearly identical: companies are building computing capacity before all its profitable applications have been discovered. The value of the data centers will depend partly on complementary developments—better models, proprietary data, new software, organizational changes, and the willingness of companies and consumers to pay for AI services.

In all three cases, therefore, infrastructure and applications are jointly determined. Waiting until demand is plainly visible can mean waiting too long, because the infrastructure itself helps create the demand.

2. High fixed costs encourage companies to build ahead of demand

A railroad requires land, grading, bridges, tunnels, stations, track, and rolling stock before it carries its first paying customer. A fiber network requires rights-of-way, trenches, cable, switching equipment, and connections before it can sell substantial bandwidth.

AI data centers likewise require sites, grid connections, substations, transformers, cooling systems, generators, networking equipment, and large orders for advanced chips. Many of those expenditures are sunk: once a company has built a substation or a specialized data-center shell in a particular location, much of the investment cannot readily be recovered.

Companies also face a strategic incentive to build early. An individual hyperscaler may reasonably conclude that the loss from having insufficient computing capacity—lost customers, inferior models, or technological dependence on a competitor—would be greater than the loss from temporarily having too much. But when every major company reasons this way, the industry can collectively build more capacity than it needs. The IEA reports increasingly severe bottlenecks involving electricity, grid connections, chips, memory, transformers, and capital, even while acknowledging that many announced projects will never be completed.

3. Each boom rests on highly uncertain projections of future demand

Railroad promoters projected the future settlement, agricultural production, and mineral output of regions that were sometimes scarcely populated. Eventually, railroad projects began to outpace the demand for their capacity, reducing returns and contributing to widespread defaults. 

Fiber builders extrapolated the extraordinary early growth of internet traffic. They assumed that new applications and fiber connections to homes and businesses would arrive rapidly enough to absorb the new long-distance capacity. Those assumptions proved too optimistic, at least in the short run.

AI companies are making similarly difficult forecasts. No one knows with much confidence:

• how much computing future models will require;
• how quickly businesses will reorganize around AI;
• how much customers will pay for AI services;
• whether most AI revenue will accrue to model developers, cloud companies, application providers, or users; or
• how much improved chips and algorithms will reduce the computing required for a given task.

The uncertainty does not mean that the investment is irrational. It means that large forecast errors—in both directions—are likely.

4. Rapid technological progress can itself produce overcapacity

This was especially important during the fiber boom. In 1996, one fiber strand could carry approximately 2.5 gigabits per second. By 2000, improved multiplexing equipment could raise the capacity of the same strand to approximately 100 gigabits per second. Thus, companies were not merely laying more fiber; technology was multiplying the capacity of fiber already in the ground. Communications-equipment investment rose from about $62 billion at an annual rate in early 1996 to more than $135 billion in late 2000, before falling sharply. 

AI faces an analogous risk. The cost of running a model with roughly GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024, according to Stanford’s AI Index. Hardware costs were declining by about 30 percent annually and energy efficiency was improving by approximately 40 percent annually. 

Consequently, today’s estimates of the number of data centers needed in 2030 may prove too high if chips, models, and software become much more efficient. That is the closest parallel to the fiber glut.

But there is an important countervailing effect: cheaper computation can generate much more use. The IEA estimates that energy consumption per basic AI task has recently fallen by at least an order of magnitude annually, yet more demanding reasoning, video, and agentic applications are expanding so quickly that total data-center electricity consumption is still projected to roughly double between 2025 and 2030. Efficiency may therefore lower the cost of AI without lowering total demand for computing or electricity.

5. A private investment failure can leave a valuable social legacy

Many railroad investors lost money, and many railroad companies entered receivership, even though the resulting network transformed the American economy. The Panic of 1873 began in substantial part because railroad construction had outrun demand and heavily indebted companies could not refinance, but the tracks themselves did not cease to be useful. 

The fiber experience was similar. Long-distance fiber companies failed, telecommunications shares collapsed, and equipment investment fell, but the underlying fiber remained available to carry later increases in internet traffic. The glut helped drive down the price of communications capacity and made subsequent internet businesses less expensive to operate. The Richmond Fed concluded that excessive investment and rational responses to genuinely revolutionary—but highly uncertain—technology were not mutually exclusive explanations. 

AI could follow the same pattern. A future decline in data-center rents or computing prices might be bad for data-center owners and chip suppliers but good for application developers and users. AI could be economically transformative while a substantial fraction of current AI investment earns a disappointing return.

Important differences

1. A data center is not a network in quite the same sense

A railroad line moves something between two locations. A fiber-optic cable transmits information between locations. Their usefulness depends strongly on connections to the rest of the network.

A data center primarily produces computation. It resembles an electricity-generating plant, steel mill, or semiconductor fabrication plant more than a railroad line. It needs fiber connections to communicate with users and other data centers, but its principal economic output is processing rather than transportation.

This matters because railroad and fiber networks have especially strong physical network effects: a route becomes more valuable when it connects to additional routes. AI has powerful scale economies and ecosystem effects, but much of that advantage resides in models, data, software, cloud platforms, and proprietary chip architectures rather than merely in connecting one data-center building to another.

2. Technological progress affects the installed assets differently

With fiber, new electronics could greatly increase the capacity of fiber that was already buried. Technological improvement therefore often enhanced the usefulness of the old physical asset.

With AI, a new generation of accelerators may make the previous generation significantly less economical. The site, building, electrical connection, cooling system, and fiber links can remain useful, but the chips and servers may have to be replaced.

Meta’s 2025 annual report estimated useful lives of five to five-and-a-half years for servers and network assets, compared with 25–30 years for its buildings. Even five years may overstate the competitive life of some frontier AI equipment if newer chips provide much better performance per dollar or per watt. 

This makes unused AI capacity more perishable than “dark fiber.” A buried fiber strand can sit unused for years and later be equipped with improved electronics. A warehouse full of unused accelerators loses value simply as better chips arrive.

3. AI data centers have unusually large continuing resource costs

After fiber has been installed, transmitting another unit of information has a very low marginal cost. Railroads have significant continuing costs—labor, fuel, maintenance, and rolling stock—but the track itself does not require a massive continuous energy input merely to remain economically relevant.

AI computation consumes large amounts of electricity whenever it is used and requires cooling, maintenance, networking, and periodic hardware replacement. The IEA expects global data-center electricity consumption to rise from approximately 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030. AI-related servers are expected to account for much of the increase. 

Thus, an overbuilt AI center can continue generating costs even after the initial construction bill has been paid. Its economics depend not merely on filling the building but on earning enough revenue to cover electricity, maintenance, hardware depreciation, and financing.

4. Today’s leading builders are financially stronger than many railroad and fiber promoters were

The railroad boom was heavily financed through bonds. When railroad returns declined and European investors withdrew, defaults spread to banks and securities firms and helped produce the Panic of 1873. 

The 1990s fiber boom included numerous new or rapidly expanding companies—Global Crossing, Level 3, Qwest and others—whose survival depended on continued access to capital and rapid growth in wholesale bandwidth demand. When demand disappointed, many failed or were reorganized. 

The principal AI builders—Amazon, Microsoft, Alphabet and Meta—began with enormous revenues, cash flows, and valuable existing businesses. That makes a wave of immediate hyperscaler bankruptcies considerably less likely than nineteenth-century railroad failures or the failures of many fiber entrants.

The distinction is narrowing, however. Amazon, Alphabet, Meta and Oracle had issued about $194 billion of bonds during 2026 through July 7, 79 percent more than they issued in all of 2025. Goldman Sachs estimated that debt would finance roughly one-third of 2026 hyperscaler capital spending. 

Companies are also shifting projects into leases and joint ventures. A $14 billion Meta–BlackRock data-center project in El Paso, for example, will be 80 percent owned by BlackRock-managed funds and partly financed with $12.5 billion of debt, while Meta leases the capacity.

My inference is that a downturn would be more likely to produce project cancellations, asset write-downs, lower returns, and losses for developers, lenders, lessors, utilities, and equipment suppliers than the wholesale disappearance of the leading technology companies. But increasing reliance on debt and outside capital means the financial consequences would no longer be confined to their own shareholders.

5. The return on AI investment is harder to observe

A railroad could measure freight and passenger revenue against the cost of operating a route. A wholesale fiber company could measure the bandwidth it leased and the price per unit of capacity.

Much AI infrastructure is used internally. AI may improve advertising, search, recommendation systems, programming, logistics, customer service, or the defensive position of an existing business. The return may not appear as separately identified “AI revenue.”

That gives the builders greater flexibility, but it also makes investment discipline more difficult. A company may continue spending because it fears falling behind, even when it cannot demonstrate that the next data center will earn an adequate direct return. Meta, for example, derives most of its revenue from advertising while making enormous infrastructure investments intended both to support that business and to develop new AI products.

6. The principal bottleneck is electricity rather than merely construction capital

Railroads competed for land and rights-of-way. Fiber companies needed trenches, conduits, poles, and access to local networks.

AI data centers compete above all for dependable electricity and rapid grid connections. A data center can be completed in two or three years, while new transmission, generation, and other energy infrastructure can require much longer. In some regions, expected data-center loads already exceed available capacity, causing projects to be delayed or paired with on-site natural-gas generation. 

This creates a distinctive public-policy problem. Utilities may build generation and transmission for projected data-center demand, only to be left with underused assets if projects are cancelled or use less power than anticipated. Depending on rate design, households and other businesses could bear some of those costs.

7. The form of government assistance is different

Four of the five transcontinental railroads received federal land grants, giving them both rights-of-way and land they could sell to finance construction. 

AI data centers have not received an equivalent nationwide transfer of federal land. Assistance has instead generally taken the form of state and local tax concessions, discounted or specially designed electricity tariffs, accelerated permitting, semiconductor subsidies, and publicly regulated grid investment. That creates similar questions about who receives the gains and who bears the costs, but through less visible mechanisms

Which historical analogy is the more useful one?

For the next several years, the fiber analogy is probably more informative. Both involve rapidly improving technology, uncertain demand for applications that do not yet fully exist, extrapolation of very high early growth rates, and the possibility that efficiency improvements will create more usable capacity than planners expected.

For the longer-run physical and economic consequences, the railroad analogy may be more informative. AI is driving investment not only in servers but also in power plants, transmission lines, substations, gas pipelines, batteries, cooling technology, chip factories, and fiber connections. Those investments may alter the economic geography of the United States and remain useful long after the first generation of AI chips has been replaced.

My tentative expectation is therefore a fiber-style shakeout within a railroad-scale infrastructure transformation:

• Some announced data centers will never be completed.
• Older accelerators will depreciate faster than expected.
• Computing prices and profit margins may fall.
• The strongest cloud and technology companies will absorb or outlast weaker developers.
• Much of the underlying electrical, building, and communications infrastructure will survive and support applications that are not yet apparent.

That is an inference rather than a confident forecast. AI demand might grow quickly enough to absorb most of the planned capacity, especially as lower costs generate more intensive applications. But the historical comparisons show why “AI will transform the economy” and “too much AI capacity is currently being built” are entirely compatible propositions.

Bottom line

The three buildouts share a common economic pattern:

A genuinely revolutionary technology creates a race to build costly infrastructure ahead of uncertain demand. Competition and optimism produce duplication and sometimes overinvestment. Investors may suffer when expectations are revised, yet the surplus infrastructure can lower costs and accelerate the technology’s eventual diffusion.

The largest difference is that railroad rights-of-way and buried fiber are unusually durable, while the most expensive part of an AI data center—the computing equipment—is rapidly depreciating and must be continuously powered. The crucial question is therefore not simply whether demand for AI will grow. It almost certainly will. The question is whether monetizable demand will grow quickly enough to cover the extraordinary depreciation, electricity, and financing costs of the capacity now being installed.