AI Data Centers Are Not the Railroads of Today

The AI boom shares all the risk profiles of previous speculative manias but lacks society-wide benefits while generating fast-metastasizing negative consequences and costs.

The idea that the current bubble in AI data centers is an echo of the railroad-construction bubble of the 1870s is appealing--but only half-right. The completion of the first transcontinental railroad in late 1869 sparked a speculative mania of raising capital to build railroads, which were seen as “can’t lose” investments in a technology that lowered transport costs from $1 to ten cents.

But not all routes had the potential to become profitable, and the resulting collapse of the railroad bubble devastated the developed-world economies, triggering a deep economic downturn from 1873 to 1879 that was called “The Great Depression” at the time (or “The Long Depression”).

The term for speculative frenzies channeling vast sums into investments with difficult-to-assess risk profiles is mal-investment, and mal-investment on a large scale triggers financial panics and economic depressions in a well-understood feedback loop.

Money invested in digging a mine that doesn’t yield any gold can’t be recovered. That capital is gone. There is an opportunity cost to every investment: that capital could have been invested in something else that was more productive than the speculative bet on something with unclear risks and payback.

As the scale of losses become apparent, credit tightens and the pool of capital available shrinks. Short-term loans that can’t be rolled over into longer duration loans trigger bankruptcies which quickly lead to bank runs (financial panics) and layoffs as businesses close. This decline in wages, revenues and the velocity of money is self-reinforcing, and the recovery process--being both financial and psychological--takes years.

The parallels with the AI speculative investment mania are obvious. Just as any railroad was viewed as guaranteed to be immensely profitable because railroads generated enormous efficiencies that reduced costs, all AI is guaranteed to be immensely profitable because AI generates enormous efficiencies that reduced costs. But in the real world, use cases for specific railroads and AI applications are stretched along a spectrum which isn’t visible in the early stages of a speculative boom.

Individual use cases don’t automatically guarantee an entire class of use cases will be successful. That one railroad--or application of AI--profitably reduced costs does not necessarily extend to all railroads or AI applications.

Nobody wants to wait around for the long process of sorting which use cases are actually beneficial and which are mal-investments, as the big money is made by making big bets in the early days. Human greed is a remarkable force, especially when combined with self-serving hype and the euphoria of the herd running.

In the current confluence of greed, hype and euphoria, the possibility that the inevitable aftermath of vast mal-investment is a Great Depression doesn’t exactly resonate. AI isn’t a railroad, it’s the most amazing force in the Universe, etc. This is Wetware 1.0 in action: the psychology of speculative frenzies doesn’t change, and so here we are--again.

Those are the parallels of the railroad mania of the 1870s and the current AI mania. But that’s only half the story. Railroads did dramatically lower costs, turning unprofitable ventures into profitable ventures not by reducing production costs but by reducing transport costs, which prior to railroads might equal production costs.

The differences between railroads and LLM / generative AI are significant. While many railroads went bankrupt when the bubble burst, those that actually served expanding markets were eventually put to use as the tracks were still useful many years after being laid. A new locomotive type might enter service decades later, but the tracks remained useful and valuable for decades--with proper maintenance. The rails were not obsoleted every few years, nor did the the entire rail lines have to be replaced every few years.

AI is not permanent. It is constantly being obsoleted. A new class of lower-power consumption chips could obsolete the current class of AI chips, requiring a mass replacement of the entire processing foundation of AI. Innovations in software could reduce the processing demands, turning existing data centers into expenses rather than profit generators. AI software that users download onto their own computers negates the need for “renting” data centers (i.e. buying processing power with tokens) by generating models from the user’s own data. These are just a few potential forces undermining the utility, lifespan and profitability of the current build-out of data centers.

While the cost structure of railroads were relatively straightforward, the costs of AI are complex and difficult to assess as initial costs are not total ownership costs, as maintenance expenses are still unfolding and future costs of resources and energy are trending higher.

While the cost reduction and efficiency benefits of depending on AI are as yet unclear, the costs of sorting “good AI” from “bad AI” are already mounting as real-world expenses. The market continues to underestimate the AI slop problem and what it means for enterprise adoption and spending.

Create enough hallucinated legal arguments, flawed engineering calculations and backdoor-ridden code, and the slop vats fill faster than our capacity to tell good work from bad, writes Tim Harford. How can we tell good AI from bad? (Financial Times)

Cedar Owl recently published a comprehensive overview of the Total Costs of Ownership of AI / Robotics and concluded they may exceed the costs of human employees. Will the cost of an AI Robot be higher than the salary of a Human Employee? AI Robot vs. Human Worker Total Cost of Ownership (cedarowl.substack.com)
“AI didn’t remove cost--it changed where the cost lives.”

As for profitable use cases, it’s too soon to tell. Individual cases don’t necessarily scale to the entire sector or economy. The hype is AI is scalable and applicable everywhere, but this isn’t what real-world experience is finding.

Unlike railroads, whose cost-reduction benefits were immediate and measurable, the sum total of AI benefits is not just unclear but potentially negative. The negative effects of AI slop and malicious applications are already visible but the full consequences of their expansion cannot yet be determined.

Recent polls reveal a profound skepticism in the younger generations whose lives will be most impacted by AI. Gen Z Is Using A.I., but Doesn’t Feel Great About It. Only 15 percent said they saw A.I. as a net benefit.

The structural limits of AI are equally visible but the full consequences of these multi-factor limitations cannot yet be determined. A recent article in Scientific American summarized one key limitation: the illusion that AI is “thinking,” “understanding” and “reasoning”: AI and human intelligence are drastically different--here’s how:
“They are extraordinarily powerful tools when used as what they are: engines of linguistic automation, not engines of understanding. They excel at drafting, summarizing, recombining and exploring ideas. But when we ask them to judge, we unintentionally redefine judgment--shifting it from a relation between a mind and the world to one between a prompt and a probability distribution.”

There are many other structural limitations whose nature limits “quick fixes.” “To grow skills, people need to go through hardship. They need to develop the muscle to think through problems,” he said. “How would someone question if AI is accurate if they don’t have critical thinking?”

“This is the contradiction that has many AI boosters talking out of both sides of their mouths: The use of coding agents is actively diminishing the very skills needed to effectively manage the coding agents.” (via Manoj S.)

CEOs are quietly realizing the AI replacement plan has a problem. Two problems, actually.
“One: the token costs for running AI agents are now exceeding what they were paying the employees they fired.

Two: when the tokens run out, the AI stops. Just stops. No continuity. No workaround. Just a spinning wheel where your workforce used to be.”

AI coding frontloads one form of productivity by backloading the entire system with higher maintenance costs down the line. These costs are not visible in the initial phase, and by the time they’re piling up, it’s too late to reverse these structural costs.

The sums invested in AI data centers--and committed to planned data centers--are on a large enough scale that even the most robust economy is vulnerable to disruption when the revenues needed to justify these extraordinary sums fail to materialize and the total operational costs and costs of ownership become measurable.

Matt Stoller offered an apt analogy of AI data center capital investments:

But in a sense, the entire AI narrative is a bit like selling huge amounts of picks and shovels as everyone rushes to the mines, and then betting there will be gold when they all start digging. Much of the stock market is made up of investor speculation that pick and shovel companies are about to hit the motherlode. But we don’t actually know how much gold there is, or even if there is any gold at all. So far, every powerful and rich person has insisted that there’s so much gold we can’t imagine it all, and anyone who thinks otherwise is a Luddite Marxist loser.”

Perhaps most importantly, once we subtract the hype, there is no evidence-based answer to the question: will our society / the public benefit from AI? Or are all the proposed benefits of reducing costs and generating innovations concentrated in the hands of AI’s owners and corporate users?

Cui bono--to whose benefit? What’s being touted as beneficial to all--equivalent to railroads--is at this point only beneficial to owners and monopolistic-cartel corporations, the very asymmetry that is fast undermining the foundations of our social and economic systems.

Put another way: is AI actually solving the core problems undermining our society and economy--systemic asymmetries of costs, wealth, power, agency and opportunity--or is AI adding new problems--brain rot, dependence on black box systems owned by a handful of tech corporations, AI slop, deepfakes, and a tsunami of malicious AI?

For all these structural reasons, AI data centers are not the railroads of today. The AI boom shares all the risk profiles of previous speculative manias but lacks society-wide benefits while generating fast-metastasizing negative consequences and costs.