The future is already here – it’s just not evenly distributed.

– William Gibson, author of Neuromancer

 

Those who cannot remember the past are condemned to repeat it.

– George Santayana, The Life of Reason

 

We are in the midst of an unprecedented boom. A technological revolution that will change everything. As a certain former and current president might say, like nothing anyone has ever seen before. AI has certainly changed things since ChatGPT was released in November 2022. By the end of this year, we will have spent roughly $2 trillion on AI infrastructure in four years. AI spending is tracking at over 1% of global GDP and roughly 2–2.5% of US GDP. Large numbers, no question.

But unprecedented?

Unequivocally no. The US railroad boom in the second half of the 19th century was actually quite a bit larger as a percentage of GDP. Railroad investment averaged 2–3% of GDP throughout the boom, with the peak from about 1868 to 1873 approaching 6% of GDP — at one point representing roughly 20% of all capital formation in the country. Total AI spending today, even at 14% of non-residential investment, isn’t close. So no, this isn’t unprecedented. The only difference is the nature of the technology. In both cases, the expectation was a large increase in productivity and the railroad boom did transform the US economy — just not fast enough to save the early investors. Will AI turn out the same?

The railroad boom was fed by several independent developments: the steam engine, the canal boom, and eventually the Bessemer process, patented in 1856, which dropped the price of steel by roughly 80% and allowed railroads to replace brittle iron track with something that could actually support large locomotives and freight cars. The financial infrastructure to support the rail boom was a byproduct of the canal boom that preceded it: largely debt financing, funded globally.

That boom also produced two busts — 1873 and 1893. In both cases, roughly a quarter of all railroads ended up in bankruptcy. The first was triggered by the collapse of Jay Cooke & Co., one of the earliest and most prolific financiers of the expansion. The NYSE closed for ten days and the ensuring depression – known as the Great Depression until the one in the 1930s took that crown – lasted until 1879. The Panic of 1893 was a different animal — a credit crunch triggered by a shortage of gold, ultimately resolved only by a $65 million gold loan from J.P. Morgan to the US government. The railroad bust was collateral damage; the credit crunch was deadly for highly indebted railroads.

The technology of AI can seem like magic, which makes comparing it to 19th century railroads feel almost silly. But the parallels are numerous, and this pattern has repeated itself many times — from the South Sea Bubble to canal mania to the panics of 1825 and 1837, right through to modern times. Each transformative technology or financial innovation follows a similar arc*:

The Installation Period, driven by financial capital and speculation. This divides into the Irruption phase (the technology breaks out) and the Frenzy phase (everyone throws money at it, afraid of missing out).

The Turning Point, where a massive structural mismatch occurs: the infrastructure has been overbuilt, but actual applications haven’t caught up to generate a clear return on investment. This almost always results in a crash or significant correction — see the dot-com boom/bust.

The Deployment Period, where the technology enters its Synergy and Maturity phases. Production capital takes over. The technology becomes cheap, ubiquitous, and deeply integrated into the institutional fabric of society, finally delivering on its productivity promises.

We’ve seen this with canals, railroads, telegraphs, automobiles, electricity, semiconductors, PCs, biotech, the internet, shale oil, and crypto.

And now AI.

Where Is AI In This Cycle?

I think we are nearing the end of the Frenzy phase and approaching the Turning Point.  The Big Tech hyperscalers allocated roughly $342 billion to capital expenditure in 2025, and projections for AI-related CapEx put annual spending in 2026 well north of $500 billion—with some models tracking closer to $700 billion+ as companies race to build next-generation data centers, secure power grids, and buy up advanced silicon.

However, a structural divergence has now emerged between the retail consumer of AI and the corporate, with cost driving the difference. While consumers have embraced AI tools at a spectacular rate, most of it remains heavily subsidized or free. On the enterprise side, the vast majority of companies are still in the pilot phase. They are trying to figure out how to securely deploy AI without hallucinations, data leaks, or legal liabilities, and at a reasonable price. So far, the result has been sticker shock and return on investment is elusive. 

In the railroad boom, the irruption phase was the development of the Bessemer process. The frenzy phase was the Pacific Railway Acts that authorized the transcontinental lines, backed by land grants and government financing. That speculative frenzy phase ran from the late 1860s to the crisis in 1873 and was marked by the Credit Mobilier fraud (direct government involvement in the private economy always ends up in corruption). Today, the irruption phase was the initial buildout of AI infrastructure for the initial “chatbot” period which broadened into Generative AI, the foundational models and the “co-pilot” (AI as assistant) stages. We’re now in the agentic era of AI where AI agents plan and execute multi-step tasks. The speculative frenzy can be seen in the massive jump in capital spending planned for this year where multiple companies are building data centers with little idea how or if they will ultimately be profitable. In the railroad era, multiple rail lines were built for the same route and while the early ones made money, in the end none did.

The financing of the AI frenzy also has a familiar feel to it. In the railroad boom, early rail lines were profitable because they carried building materials for the next line. Most of the rail expansion was financed with debt, railroad bonds and various government financing schemes (land grants and government bonds). In the dot com era, public underwriting standards were essentially non-existent and companies used the capital raised to buy advertising and other services from the companies that were already operating, all in a race to get big fast. Today, one of the hyperscalers (Google, Microsoft, Amazon, Oracle) invests in the AI labs (OpenAI, Anthropic). The hyperscaler then buys equipment from the AI infrastructure companies (Nvidia, Micron) for the AI labs to lease. Nivida then takes their cash pile and invests it back into the AI companies (OpenAI, Anthropic) so they can buy more compute from the hyperscalers. In a sense the hyperscalers are financing the purchase of their own services by the AI Labs.

At one point it was common knowledge that the hyperscalers were all financing this out of their cash flow but that narrative is dying a quick death. While Microsoft and Alphabet may be able to still make that claim with a straight face, Meta, Oracle, and Amazon have issued a combined total of $95 billion in new bonds over the last year. The newer cloud providers like Coreweave are financing themselves by offering chips as collateral. Chips that may be obsolete in a few short years I might add.

The Turning Point

The reality check is coming due.

What companies are discovering as they try to deploy these agentic systems is that all of this is expensive in ways they didn’t expect. This is not me chatting with Claude for a low monthly fee. When an AI agent enters an autonomous reasoning loop, a single user request can trigger dozens of background calls consuming millions of tokens (dollars). Companies are watching project budgets evaporate before the problem is solved — and that assumes the agent actually can solve the problem, which may be just wishful thinking. Many of these agents don’t make it out of testing, much less into production.

There are plenty of reports of projects abandoned when the cost turns out to exceed what the humans they were meant to replace would have cost. Sometimes it’s about data quality or security, but in the end it’s about accountability. These systems require human oversight (HITL: humans in the loop) because, ultimately, the bots can’t yet be trusted.

Uber recently disclosed it burned through its entire annual budget for an AI coding tool in a single month. The COO’s take: the expected link between spending and useful output “is not there yet.” Microsoft is reportedly scaling back some AI coding licenses and redirecting to internal resources. The CEO of Duolingo, an early AI advocate, said recently he doesn’t see it replacing what his employees do. An AI consultant reported that a client burned through $500 million in one month by failing to put usage limits on employees. Some companies have had to lay off workers just to cover their AI bills.

Color me skeptical that this is the productivity revolution everyone ordered.

The reality is that AI works well for certain coding tasks and some narrow applications. But someone with no prior programming knowledge using AI to write a program will end up with something they don’t understand — something that might look functional but won’t actually do the job, or worse, will carry security flaws that create liabilities well beyond its usefulness.

For a lot of workplaces, AI is creating more work or providing a convenient way for employees to look productive based on their token usage. That’s what is currently driving revenue at the AI labs. But how long does that last if companies aren’t getting meaningful returns on the spending?

This isn’t unusual for a new technology. AI may well prove to be transformative. But it was always going to take time for people to learn how to actually use it. The internet boom — which blew up any stock with dot-com in its name — eventually delivered on its promise. All that dark fiber sitting dormant in 2002 has been put to use in ways nobody (except maybe Mark Cuban) anticipated in 1999. The beneficiaries were the surviving telecoms, private network builders like Google, the cloud computing companies, and the video streamers. The fiber was useful but in 2000 nobody had figured out how to use it efficiently and profitably. That got solved in the Deployment Period, when synergies emerged and capital was finally put to efficient use.

The Crash

If this emerging awareness, that AI isn’t quite ready for primetime, continues to spread — if companies keep finding that the HITL aren’t the bottleneck they assumed — it will ripple through the entire ecosystem. If hyperscalers have capacity they can’t sell at a profitable price, they stop building, they stop buying chips and servers and networking gear and the fiber connecting it all. That’s the crash phase. We’re not there yet. But it sure seems like it’s coming.

I am often asked why I haven’t been more aggressive about investing in AI-related companies. This is why.

I was there for the PC boom in the 1980s, when hundreds of companies were making IBM clones. Today five companies control roughly 90% of computer manufacturing globally. I was there for the first biotech boom in the early 1990s, when over 100 companies with “bio” in their names went public in three years. About 85% were complete failures. The early pioneers — the OpenAIs of biotech — were Amgen and Genentech. Amgen still churns out profits. Genentech was acquired by Roche. There were other successes — Gilead, Vertex, Regeneron — but the vast majority either failed completely or never recovered their IPO price. The turning point came when high-profile drugs from Centocor, Xoma, and Synergen couldn’t get through Phase III trials. The beneficiaries of the bust were companies like Amgen, which bought Synergen cheaply and turned its failed sepsis drug into an arthritis treatment, or Johnson & Johnson, which acquired Centocor.

I was also there for the dot-com boom. From 1995 to 2000, over 800 dot-com companies went public and fewer than half survived — most as zombies. Of those that thrived: Amazon, eBay, Booking Holdings (née Priceline), and, believe it or not, Nvidia. Who benefited most from the bust? Anyone who wanted to advertise at prices not inflated by dot-com silliness, the bricks-and-mortar companies mocked for not having an internet strategy, and ultimately everyday consumers who got the on-demand economy that followed. The bust also left behind a highly capable tech workforce, miles of dark fiber, and cheaply available server farms. That set up everything that came after.

The pioneers who build the infrastructure for a new technology are often not the long-term beneficiaries. It was J.P. Morgan who took over the failed railroads and made them profitable. It was Rockefeller who built Standard Oil on access to cheap rail rates, wrung from desperate railroads in exchange for load guarantees.

Who will be the winners this time? If history holds, it will be the companies that buy excess AI compute capacity cheaply and find new profitable uses for it. It will be the companies that figure out how to deploy AI in ways that genuinely improve their efficiency. It will be industries that process large volumes of data — pharma, climate modeling, finance — and some SaaS platforms like Salesforce, Adobe, or ServiceNow. It will be the private equity firms that acquire failed assets. And it will be the workers who discover they’re more valuable than expected — the ones with the judgment to guide AI, not be replaced by it**.

The question for investors today is not whether AI works. The question is whether it can work profitably, at scale, soon enough to service the debt being used to build it.

I don’t know. Neither does anyone else.

But based on history, I’d put the odds roughly like this (just my guess):

60% probability of a crash and reset.

30% chance today’s leading AI companies make enough to survive if not thrive.

10% chance AI just continues to boom and never looks back.

I prefer to wait out most of the installation and buildout phase and invest heavily after the bust. There will be long-term winners from AI; we just don’t know who they are yet. The implementation phase can produce enormous short-term profits, but they rarely last. The long-term profits usually go to the patient investors who ignore the initial hype and buy when things are cheap.

Find your inner J.P. Morgan. Be patient.

Joe Calhoun

 

*This sequence of events is outlined in Carlota Perez’s 2002 book Technological Revolutions and Financial Capital. I read it about 15 years ago so I didn’t remember the name of the book or the author but I described it to Claude and it immediately found it. I then used AI to help me remember some of these details. That is one thing AI is really good at; give AI one source of trusted material and it cuts down on the hallucinations.

**The winners after the bust may well be some of the people who are financing or leading the current AI boom. Jay Cooke, whose eponymous company went bankrupt in the Railroad crash of 1873, ended up in bankruptcy and penury after the bust. But he managed to scrape together enough capital to buy the Horn Silver Mine in Utah in the late 1870s. It turned out to be one of the richest silver strikes in US history. His advantage? He built a pipeline (dedicated) railroad directly to the mine (the Utah Southern Railroad extension) that made the mine more efficient. Ore went from the mine directly onto the rails and straight to the smelter. After making a new fortune, he hunted down all of his old creditors from his bankruptcy, paid them off and restored his name and reputation.