By Tyler Kelley
In the first quarter of this year, the U.S. economy grew at an annual rate of just over 2 percent. Break that number apart and nearly half of it came from the AI buildout, the chips, servers, and software being poured into new data centers. Take away the AI buildout and growth falls to about 1 percent. Half of American economic momentum, riding on one bet. And however that bet turns out, the result is going to land on your side of the counter.
You've probably heard the word "bubble" attached to that bet. There's a reason it keeps coming up. The five biggest technology companies plan to spend over $600 billion on AI infrastructure this year, capital spending equal to 34 percent of their revenue. That's more than double the peak internet companies hit in the late 1990s. Alphabet just posted its first quarter of negative free cash flow in its two decades as a public company. Researchers at Allianz measured the gap between what's being spent on AI and what AI actually earns, and found it's already wider than the gap that preceded the telecom crash of 2001.
Then there's the financing. Nvidia sells the chips, and Nvidia invested $30 billion in OpenAI, the company buying the chips. In the late 1990s, telecom equipment makers lent money to their own customers so those customers could keep buying equipment. Demand looked strong right up until it wasn't.
So the conventional wisdom has hardened into simple advice. AI is a bubble, bubbles pop, and a careful business owner should stand back and wait for the smoke to clear.
I wanted to test that advice. So I looked at what happened to ordinary businesses after the last two great infrastructure bubbles actually popped. What I found flips the advice on its head, because the people who fund a bubble and the people who benefit from one are almost never the same people.
Start with the railroads. In the 1840s, Britain poured roughly 7 percent of its entire economy into laying track. The rail network roughly tripled in less than a decade. Promoters promised investors dividends of 10 percent or more. By 1850, the average railway was paying under 2. Thousands were ruined. But the track didn't disappear when the shareholders did. Every merchant and farmer in Britain woke up connected to markets they could never reach before, at freight rates the bubble's victims had unintentionally paid for.
The dot-com version is even cleaner. Companies like Global Crossing laid over 100,000 miles of fiber optic cable and collapsed doing it. After the crash, roughly 85 percent of that fiber sat dark and unused, and the price of bandwidth fell by as much as 90 percent. Netflix launched streaming in 2007 without building a single mile of network. Amazon Web Services showed up in 2006, renting capacity priced by desperation. The crash made the internet cheap enough for everybody else. The companies that defined the next twenty years were customers of the wreckage, not investors in the boom.
When giants overbuild, overcapacity becomes a discount. Excess supply drives prices down, which works out to a subsidy flowing from the builders to their customers. The railway shareholders subsidized Victorian shopkeepers. The fiber investors subsidized Netflix. If today's AI spenders have overbuilt, they will end up subsidizing whoever is positioned to use cheap computing power. That could be you.
The subsidy has already started, no crash required. Stanford's AI Index tracked what it costs to get a fixed level of AI performance and found the price fell more than 280-fold in about two years. The capability that cost a dollar at the end of 2022 cost a fraction of a penny two years later, and it has kept sliding since. Every billion the giants pour into competing with each other pushes your price down a little more.
Cheap tools only matter if they work, and the record on that is messier than the sales pitch. A team at MIT reviewed more than 300 corporate AI initiatives and reported that 95 percent of enterprise AI pilots produced no measurable financial return. Zero impact on profit and loss. The failures were concentrated in big companies buying big, complicated custom systems. That statistic gets cited everywhere as proof the whole thing is smoke.
But buried in the same report is the opposite story. Inside more than 90 percent of those same companies, employees were quietly using cheap consumer AI tools on their own, off the books, because those tools actually worked. MIT called it the shadow AI economy. The $20-a-month subscription was delivering what the seven-figure pilot couldn't.
Put those two findings side by side. The losers in this cycle are the ones spending like investors, making massive commitments to custom systems with returns that may never arrive. The winners are behaving like customers, making small bets on cheap tools they can use today. In practice that looks unglamorous. It's the shop owner whose leads get answered within five minutes, or the contractor who drafts estimates in a quarter of the time. No pilot program. No consultant. A small business can't build a data center, and that turns out to be an advantage. You're not the investor here. You're the customer. You hold none of Nvidia's risk and all of the benefit of Nvidia's spending.
Which means the standard question, "is AI a bubble?", is the wrong one for a business owner to ask. If it's not a bubble, the tools keep improving and you want to be fluent in them. If it is a bubble, history says the pop makes those tools dramatically cheaper, and you still want to be fluent in them. The better question is which side of the bubble you're standing on.
Somewhere right now, a trillion dollars of other people's money is being poured into concrete and silicon on your behalf. The Victorians got railroads out of their mania. We got the internet out of ours. The wreckage, if it comes, has always belonged to whoever showed up ready to use it.
Tyler Kelley is Co-founder and Chief Strategist of SLAM Agency, where he helps organizations put AI to work without betting the company on it.