Is the AI buildout a bubble or a bet?

Every debate about the AI buildout starts from one staggering fact: the amount of money being spent. The largest cloud and AI companies, the hyperscalers, have collectively raised their capital spending into the range of hundreds of billions of dollars per year, much of it aimed at AI data centers, and their own guidance points higher still. Individual firms now cite annual capital budgets that a few years ago would have described a whole industry.

Numbers this large invite a single question: is this wise investment or a bubble? That question has no glib answer, and this lesson will not hand you one. Instead it gives you the concepts to reason about it yourself.

To do that honestly you need four ideas, one per step: what this spending is (capex), why the chips being bought lose value fast (depreciation), how the money flows in loops that can flatter the picture (circular financing), and how to weigh the genuine case for and against. Get those, and you can read the AI-economy debate without being at the mercy of whoever is talking.

The spending has a name: capital expenditure, or capex, money spent to buy long-lived assets, chips, buildings, power gear, that will be used for years. It is different in kind from operating expense, the day-to-day cost of running the business, and the difference matters for how you judge it.

The logic of capex is simple: spend a large sum now to earn a return later. A railroad lays track for decades of freight. A factory buys machines to make years of product. The bet is always that future earnings from the asset exceed today's outlay.

So capex is not reckless by nature. Spending 100 billion dollars is completely rational if the assets reliably generate more than 100 billion dollars over their life. The entire AI-buildout question reduces to that conditional: will these data centers earn back their enormous cost?

Two things make that hard to answer here. The sums are unprecedented for such a young revenue base, and, as the next step shows, the main asset being bought ages unusually fast. Those two features are exactly what separate the AI buildout from a boring, obviously-sensible railroad.

Here is what makes AI capex riskier than a railroad. Track lasts decades. The most valuable asset in an AI data center, the GPUs, is superseded within a few years as faster chips arrive. In accounting terms, they depreciate fast, they lose value quickly, so the window to earn a return is short.

This creates a punishing treadmill. To keep a frontier position you must keep replacing chips every few years, which means the capex is not one big spend followed by a long harvest, but a recurring spend that repeats before the last batch has paid off. The clock is short and it resets constantly.

Even the accounting is contested. Companies choose an assumed useful life for their chips, and a longer assumption makes reported profits look better today by spreading the cost over more years. Critics argue some estimates are optimistic, which would mean profits are flattered now and a reckoning is deferred. The point is not to referee it, but to see that a boring assumption, how long a chip counts as useful, materially changes how healthy the whole enterprise looks.

Depreciation is the single biggest reason a reasonable person can doubt the returns.

A feature of this boom that draws heavy scrutiny is circular financing, arrangements where the same dollars appear to loop between companies that are simultaneously each other's investors, suppliers, and customers.

The pattern, in the abstract: Company A invests in Company B; Company B uses that money to buy products from Company A; Company A books the sale as revenue. Value may be real, but some of the reported revenue is, in effect, A's own money returning to it. When chip makers, cloud providers, and AI labs take stakes in one another and buy one another's products, the web of who-funds-whom gets genuinely hard to untangle.

Why it matters: circular flows can make an industry look more self-sustaining than it is. Revenue that ultimately traces back to your own investment is not the same as revenue from an independent customer who freely chose to pay. If a chain like that unwinds, the losses can cascade through every linked party at once.

The honest reading is not that all such deals are shams, many fund real capacity, but that these structures make headline growth harder to interpret, so treat revenue figures inside a tightly interlinked ecosystem with extra care.

Put capex next to revenue and you see why the debate is live. AI is generating real and fast-growing income, subscriptions, enterprise services, cloud usage, but the annual capital spending currently runs well ahead of the direct AI revenue it produces. The infrastructure is being built for demand expected in the future, not demand fully arrived today.

That gap is not automatically damning. Almost every capital-intensive buildout in history spent ahead of revenue: railroads laid track before the freight existed, telecoms built fiber before the traffic came. Building capacity for anticipated demand is the normal logic of infrastructure, and getting there first can be decisive.

But the gap is the crux. If demand grows into the capacity, today's spending looks visionary. If demand disappoints, a mountain of fast-depreciating assets earns too little, and the losses are enormous. The same facts support both stories, which is exactly why smart, honest observers disagree.

So the whole argument turns on one forecast: how fast and how large will real, paying AI demand become? The next two steps lay out the strongest case each way, using only the mechanics you now have.

Take the optimistic case seriously, on its mechanics, not its enthusiasm.

Demand is real and compounding. Paying users and enterprises are adopting AI quickly, and usage per user tends to rise as models improve and reach into more tasks. If that curve continues, today's capacity gets filled and then some.

The cost of being late is severe. If AI is as important as the builders believe, a company that under-invests risks being permanently locked out of the defining platform of its era. Faced with that asymmetry, spending too much is a smaller mistake than spending too little, so heavy investment is the rational hedge.

The builders can afford it. Unlike the debt-fueled manias of history, much of this capex comes from a few extraordinarily profitable companies funding it largely out of their own cash flows, which makes the bet more survivable if it takes time to pay off.

Infrastructure outlives the hype. Even if specific chips age out, the power, land, cooling, and grid connections are durable and hard to replicate, and could underpin decades of computing demand beyond today's models.

On this view the spending is a rational, if aggressive, bet on a genuinely transformative technology by players who can absorb the risk.

Now the skeptical case, equally on mechanics.

Spending far outruns revenue. Capex is scaling faster than paying demand, and much rests on a forecast that AI usage and prices will grow enough to justify it. If that forecast is even moderately wrong, the assets earn too little.

The assets age fast. Because GPUs depreciate in a few years, there is no long harvest to bail out an over-build. Unlike railroad track that earns for decades, mistimed AI capacity becomes near-worthless quickly, so errors are punished hard.

Circular financing hides fragility. The interlinked web of investment and revenue can make the ecosystem look more self-sustaining than it is, and tightly coupled systems can unwind together if one link breaks.

History rhymes. Railroads, telecom fiber, and the dot-com boom all built real, lasting infrastructure and still destroyed enormous amounts of investor capital along the way, because supply arrived faster than profitable demand. Useful technology and a financial bubble are not mutually exclusive; they often coincide.

On this view the technology can be revolutionary and the current pace of spending still be a bubble that corrects painfully before the long-run value is realized.

You now have the tools, so here is how to think, not what to conclude.

First, separate the technology from the trade. "Is AI valuable?" and "Is this level of spending, at this moment, going to pay off?" are different questions. History shows a technology can be transformative while the investment that built it is still a bubble. Holding both at once is the mature position.

Second, watch the gap between capex and paying revenue. That single relationship is the heart of it. Revenue rising to meet the spending supports the rational-bet case; a widening gap supports the bubble case. It is the number to track over time.

Third, discount tightly interlinked revenue. Growth inside a web of mutual investment and buying deserves more scrutiny than sales to independent customers. Ask where the money ultimately originates.

Fourth, respect the depreciation clock. Because the core assets age in a few years, this bet has less margin for being early or wrong than slower infrastructure did.

The goal is not a verdict. It is to reason from mechanisms, capex, depreciation, circular flows, and the revenue gap, instead of headlines, so that whatever happens, you understand why, rather than being told what to think.

Enormous capex buys fast-depreciating chips; whether that pays off hinges on the gap between spending and real paying revenue, read through the lens of circular financing, with a coherent case on each side.