The AI Gilded Age: A Good Idea Can Still Be a Bad Investment
Every speculative mania begins with a good idea (and that’s what makes them dangerous).
Railroads transformed the United States, canals opened new markets, and the internet changed nearly every aspect of modern life. Early successes captured the public’s imagination, and capital rushed in to fund the possibilities.
Then, in each case, participants borrowed too heavily to build too much, too quickly, at prices that made little economic sense.
It appears that Artificial intelligence may be the next world-changing technology to follow the same familiar pattern, and we’re watching it happen in real time.
A Familiar Kind of Progress
The Gilded Age was a period of tremendous innovation. Railroads connected communities, reduced transportation costs, and opened vast new markets while families accumulated massive amounts of wealth. That rush of cash also attracted speculation, aggressive financing, redundant construction, fraud, and eventually a wave of failures.
As railroad investment accelerated, new projects eventually outpaced the demand for additional capacity. Returns declined, heavily financed companies failed, and the resulting instability contributed to the Panic of 1873. The technology was real, useful, and transformative, and investors simply financed far more capacity than the economy could profitably absorb.
Artificial intelligence appears to be following a similar script.
Companies are committing extraordinary amounts of capital to chips, data centers, power generation, cooling systems, networking equipment, and other infrastructure. The current focus seems to be on how much capacity can be built rather than how much profit each additional dollar of spending is likely to produce.
If several companies spend hundreds of billions of dollars training models that perform largely the same functions, is that meaningfully different from building two rail lines between the same cities when the market only needs one? After all, Microsoft CEO Satya Nadella recently stated that “every model is substitutable.”
The passengers may benefit from the competition. The people financing the redundant railroad do not.
It’s Not Just a Valuation Bubble
Most conversations about an AI bubble focus on stock prices. Investors recognize (or at least should recognize) that many AI-related companies trade at unusually high valuations.
But the larger concern may be that we are also experiencing an earnings bubble.
When a technology company buys chips and builds a data center, it does not record the entire investment as an expense on the day the money is spent. The asset is placed on the balance sheet, and the expense is recognized gradually through depreciation.
In other words, the cash goes out today, but much of the accounting expense appears tomorrow.
Meanwhile, the companies selling the chips, memory, electrical equipment, and construction services recognize revenue from that spending today. Bottlenecks can make those profits especially large because scarce components command premium prices.
The result is a temporary but powerful boost to corporate earnings throughout the AI supply chain. Earnings are also boosted by paper gains. Hyperscalers buy stakes in their AI model building customers at ever higher valuations. They can then claim a one off profit because their existing investments traded at higher prices despite the fact that they were the ones paying it. Those earnings make the market appear less expensive than it might be under more normal conditions. For fans of Whose Line is it Anyway?, everything is made up and the points don’t matter.
How can they afford this? Debt. Tremendous amounts of debt. In addition to the equity raises and direct debt, hyperscalers now have roughly $3T in disclosed off balance sheet liabilities. Private credit funds are stepping up with capital (partially funded by life insurance premiums). And Wall Street has begun crafting a plan to securitize debt backed by semiconductors much the same way they packaged risky mortgages in the mid 2000s. They even got the SEC to provide them a waiver for it on disclosure and risk retention rules designed after the financial crisis to protect investors from this very behavior. This is all necessitated by the fact that OpenAI and Anthropic burn more cash the larger they get. Much like the NINJA (no income, no job) homebuyers in the mid 2000s, the plan is to hope they can afford the payment later. The card holding up this leaning tower of a stock market mania is a hope/prayer.
Eventually, however, tomorrow arrives. The points will eventually matter, and only reality will be counted.
The buyers must record depreciation. Older equipment may become obsolete. Suppliers may add capacity. Shortages may turn into gluts. The bill comes due and we hope the same people ordering the cowboy ribeye and top shelf whiskey can afford to pay. Otherwise, those losses will likely be socialized through life insurance policyholders, pension funds, taxpayers, etc. If the debt spigot runs dry before AI revenue catches up, AI supply chain profits will drop faster than Wiley Coyote can run off a cliff.
Meta, for example, disclosed in its 2025 annual report that it expects to spend approximately $115 billion to $135 billion in 2026, largely in support of AI and its core business. The company also extended the estimated useful life of most servers and network assets to 5.5 years. A longer assumed life reduces annual depreciation and increases reported profit in the near term.
That accounting may ultimately prove reasonable. But investors should ask a basic question: When new and better chips arrive every year, will today’s equipment remain economically productive for as long as the accounting assumes?
Training a Model Is Kind of Like Drilling an Oil Well
An oil producer spends heavily upfront to drill a well and then tries to earn a return from the oil it produces. The company generally has a reasonable estimate of the well’s productive life and how quickly production will decline.
Training an AI model involves a similar upfront investment. The difference is that its economic life is far more difficult to estimate.
In theory, a trained model can continue operating for years. In practice, it may become commercially obsolete shortly after a competitor releases a better or cheaper one.
That makes an AI model resemble a fast-declining shale well. The owner must keep drilling, or in this case, training, just to maintain its competitive position.
The challenge is not whether AI can generate revenue. It already does. The question is whether that revenue will be large and durable enough to justify the continuing investment required to remain competitive.
Where Will the Profits Go?
Transformative technology doesn’t automatically create a durable competitive advantage.
As the technology becomes more widely available, the model itself may become a commodity. If several companies offer similar capabilities, customers will naturally gravitate toward whichever option is cheapest, easiest to access, or best suited to their particular needs.
That suggests the long-term winners may have one or more of three advantages:
Superior distribution. A company with an established customer base can place AI tools directly into products people already use.
A lower cost structure. If the underlying technology becomes commoditized, the low-cost provider has an obvious advantage.
Unique and valuable data. Companies with proprietary information may be able to use AI to unlock value that competitors cannot reproduce.
Think about a farmer living in the interior of the United States before the arrival of a railroad. The farmer may have produced plenty of grain, but transportation costs consumed most of its economic value. When the railroad arrived, the grain suddenly became far more valuable because it could reach distant customers cheaply.
AI may do something similar for companies sitting on enormous quantities of specialized data. The model is the railroad. The proprietary data is the grain.
The railroad itself may not be where most of the lasting profits reside. The value often accrued to the land it cut through.
When Speculation Starts to Look Like Investing
In 2021, many people buying meme stocks, speculative technology companies, SPACs, or questionable cryptocurrencies knew they were gambling. They may have believed the gamble would pay off, but they generally understood what they were doing.
Today, the risk may be a bit more difficult to recognize.
AI companies have real products. Semiconductor companies have real revenue. Data centers are real buildings filled with real equipment. Because the infrastructure is tangible and the revenue is growing, investors can convince themselves that any price or financing structure is reasonable.
That does not make the assumptions embedded in those prices economically sound.
Speculative manias tend to move through a predictable progression. A breakthrough produces real success. That success attracts capital. The capital attracts competitors. Eventually, it also attracts opportunists who recognize that investors want exposure to the theme more than they want a careful explanation of the business.
If the potential cannot be easily quantified, the imagination supplies the number. And the imagination can justify any price it wishes.
Risk feels especially distant after a long period of strong returns. When I began working in the investment industry in 2008, investors had endured an almost nine-year period in which the S&P 500 lost more than 7% per year on a compounded basis. At that time, people struggled to believe the stock market could produce attractive returns.
Today, after more than a decade of exceptional performance, investors struggle to believe that it might not.
Neither response is rational, but both are human.
What Should Investors Do?
I do not know when enthusiasm for AI will peak. I do not know which model will win, how much computing capacity the economy will ultimately need, or how quickly current equipment will become obsolete.
“I don’t know” is not a particularly exciting investment thesis, but it is often a useful starting point.
Instead of trying to predict the exact ending, investors can ask more practical questions:
Where is the cash actually coming from?
Does the customer generate enough cash to pay, or does it depend on additional financing?
How long will the equipment or model remain economically useful?
What competitive advantage will survive when the technology becomes cheaper and more widely available?
Does the price leave room for anything to go wrong?
Most importantly, investors should remember that they are not required to participate in every opportunity.
While capital crowds into the shiny object, useful and profitable businesses elsewhere in the market can be ignored. We continue to find productive assets, durable businesses, and reasonable valuations in areas where expectations are considerably lower.
That does not produce the same excitement as trying to identify the next AI winner. Fortunately, excitement isn’t the goal (sorry to disappoint).
The goal is to earn an attractive return without accepting risks that defy economic logic.
AI will almost certainly change the world. Railroads did, too.
Just remember that a great technology and a great investment are not always the same thing.
Invest Curiously,
Austin
Austin Crites, CFA
Chief Investment Officer
Aurora Asset Management/Aurora Financial Strategies
Austin Crites is the Chief Investment Officer of Aurora Asset Management, an Indianapolis-based subsidiary of Aurora Financial Strategies, which is located in Kokomo, IN. He can be reached via email at austin@auroramgt.com. Investment Advisory Services are offered through BCGM Wealth Management, LLC, a SEC-registered investment adviser. Registration with the United States Securities and Exchange Commission does not imply that BCGM or any of its principals or employees possesses a particular level of skill or training in the investment advisory business or any other business. This blog does not constitute advice. This is not an offer to buy or sell securities. Advisor is not licensed in all states. Any mention of a particular security and related performance data is not a recommendation to buy or sell that security. BCGM Wealth Management, LLC manages its clients’ accounts using a variety of investment techniques and strategies, which are not necessarily discussed in the commentary. Investments in securities involve the risk of loss. Past performance is no guarantee of future results. Clients may own positions in the securities discussed.
