170 years later, the same mechanism is playing out with artificial intelligence. The question is who, this time, is selling the shovels.
A race to invest
Since 2023, Amazon, Microsoft, Alphabet and Meta, the hyperscalers, the cloud giants that operate the largest data centers on the planet, have been locked in a capital-spending (capex) race unlike anything seen in recent years, almost entirely devoted to AI. The combined total for these four groups has risen from around 136 billion dollars in 2023 to 354 in 2025, and estimates put it at 640 billion in 2026, then close to 780 in 2027. In four years, their investment effort will have been multiplied by nearly five, and it doubles in 2026 alone. Such a sum is hard to grasp. Spent at a rate of one million dollars a day, it would take more than two thousand years to run through, and it amounts to nearly 80% of Switzerland's annual GDP.
These amounts end up weighing on their cash. The combined free cash flow of the four companies peaked at 237 billion in 2024. It fell back to 200 in 2025, then to an expected 77 billion in 2026, two thirds gone in two years, even as their revenue keeps growing. Investment now absorbs almost all the cash they generate. At Amazon, investment even exceeds it, with the group expected to post negative free cash flow in 2026. A sign that the cash generated no longer covers the capex, Alphabet raised nearly 85 billion dollars in equity in June, its first equity raise since its 2004 IPO.
Nothing suggests financial fragility, though. These companies generate far too much cash to be at risk, and this race is largely financed by their other core businesses, highly profitable ones, such as advertising for Alphabet and Meta or software for Microsoft. Their own forecasts even point to a free-cash-flow recovery as soon as 2027. But spending on this scale only makes sense on one condition, that the sums invested are eventually recouped.
Who captures the value
Where does all this money go? As in Brannan's day, the real winners aren't those digging for gold, they're the ones selling the shovels. Today, those shovels are semiconductors. The chips that run the computations, the components that feed them, the machines that etch them. Of every dollar of capex the hyperscalers commit, a large share ends up in this hardware chain. And as long as the infrastructure race lasts, it's these suppliers that collect the most, and with the best margins. The value goes to those who equip AI, far more than to those who sell its use.
Are we paying the true price of AI?
Let's recap who does what. The chipmakers sell the hardware, and they're the ones collecting the most today. The hyperscalers, for their part, buy those chips, build the data centers and rent out computing power. Their business is profitable today. But the whole question, for them, is whether these hundreds of billions invested will pay off tomorrow. That leaves the last link in the chain, the one that puts AI in the hands of the end customer. And that one is already losing money.
Today, you can use ChatGPT, Gemini or Claude without paying anything, and even a paid subscription doesn't always cover the real cost of everything a user consumes. Those who build these models therefore lose money, sometimes enormous amounts. OpenAI, the most visible case, ran up nearly 21 billion dollars in operating losses in 2025 and doesn't expect a profit before the end of the decade. But these losses don't reflect a lack of interest, quite the opposite. Demand is exploding: OpenAI claims more than 900 million weekly users and an annualized revenue (ARR) above 20 billion dollars, while Anthropic's went from around 1 to 19 billion in fifteen months. Their problem isn't attracting customers, they have hundreds of millions of them. It's that every use costs them more than it brings in.
To keep going, these labs continuously raise money. The hyperscalers are among their main backers, but not the only ones, and capital keeps flowing in as long as the AI bet stays credible.
This is the crux of it. For an economic chain to hold over time, every link must eventually make money. In AI, that isn't yet the case. The models are sold at a loss, the end user pays only a fraction of the real cost, and what keeps the whole thing standing isn't revenue from usage, but capital betting on what comes next.
So the market hasn't found its balance. For it to do so, one or more of these variables will have to change. Either the user pays more, or the cost of computing falls, or AI ends up creating enough value to justify these hundreds of billions invested. Until that moment comes, the profitability of the entire chain remains a promise, funded by those who believe it will come true.
What happens when the market has to rebalance?
Three levers can restore the balance, and they can work together.
First, the price rises. Free models become paid, subscriptions go up, businesses charge for AI at its real cost. Each user then brings in more. The risk is that demand pulls back as soon as the service is no longer subsidized.
Second, the cost falls. More efficient chips and lighter models reduce what each query costs. The equation is solved not by charging more, but by spending less to deliver the same service.
Third, volume grows massively. Part of this spending is fixed costs, developing and training the models, building the data centers. The more widely AI is adopted, the more these initial costs are spread across a large number of users and weigh little per head. The nuance is that each additional query still consumes computing power, so part of the cost keeps rising with usage. Volume alone isn't enough, but it dilutes the heavy upfront bill, all the more so as AI settles in everywhere.
And now?
One thing is certain. AI is here, demand is massive and very real, and the infrastructure is being built at an unprecedented pace. What isn't certain is who, in the end, will make money.
Today the value is concentrated in those who sell the tools, while those who sell its use operate at a loss, carried by capital betting on what comes next. This imbalance is nothing unusual at the start of an investment cycle. The whole question is whether it resolves upward, when revenue catches up with spending, or downward, when investors stop waiting.
The great technological waves offer both precedents. The internet delivered on its promises, but along the way it ruined a good many of those who had financed its pipes. The AI bet may well be right in substance and painful for many of those funding it. It's up to each person to decide which side of the chain they would rather stand on.
Sources: Zonebourse (capex and free cash flow, reported figures and consensus estimates); Fortune, CNBC and company press releases (AI labs' losses and revenue, Alphabet's capital increase).