“Is AI a bubble?” is probably the question the most people are asking right now, and the one most likely to start a fight. The interesting part is that the bulls and bears often hold the same set of numbers yet argue their way to opposite conclusions.
This piece isn’t out to give you a “yes” or “no,” because nobody can reliably predict whether or when a bubble will burst. What’s more useful is to hand you a ruler: four metrics to read together, plus the strongest arguments from each side, so you can judge for yourself instead of listening to whoever shouts loudest. This is one entry in the “AI Industry Watch” series; to see who’s actually making money, pair it with the AI Stocks Money-Flow Map, and to see the web of capital where chipmakers and clouds invest in each other, pair it with the AI Circular Financing Map.
Core-Data Snapshot
The numbers below get cited by both sides; only the interpretation differs.
| Metric | Data | Nature |
|---|---|---|
| Top-five cloud capex, 2026 | About $745B–$775B combined | Official guidance |
| NVIDIA data-center quarterly revenue | About $75.2B, up about 92% year over year (Q1 FY27) | Official filings |
| Sequoia’s “AI’s $600 billion question” | Estimates AI needs ~$600B in revenue a year to pay back | Venture view |
| Bain estimate | About $2T in annual revenue needed by 2030 to support AI’s scale | Consulting report |
| Oracle remaining performance obligations (RPO) | Over $523B | Official filings |
| Model companies’ annualized revenue | OpenAI and Anthropic estimated by the press at ~$25B and ~$30B respectively, still burning cash | Press estimate, unaudited |
Why Some People Cry Bubble
The bears’ core worry is one line: spending too fast, earning it back too slowly.
The most-cited point is the “$600 billion question” raised by the venture firm Sequoia: at the current scale of infrastructure spending, the AI industry needs to generate roughly $600 billion in revenue a year to hold up, and actual AI revenue is still a fair distance from that figure (estimates vary). The consultancy Bain also estimates that by 2030, about $2 trillion in annual revenue will be needed to support this round of compute expansion.
Beyond payback, a few other worries get named regularly. One is corporate adoption results: an MIT study found that most companies’ AI pilots show no clear bottom-line contribution yet. Two is that the model companies are still investing heavily: OpenAI and Anthropic are growing revenue fast, but at the same time they’ve signed astronomical compute contracts, and by press estimates they’re still burning cash (private-company financials are unaudited); even CoreWeave, the listed company that rents GPUs specifically to them, is still posting net losses. Three is chip depreciation (more on that in the next section). Four is the “circular financing” in which chipmakers and clouds invest in each other, raising the worry that demand is being propped up by insiders’ own money; for that part, see the AI Circular Financing Map.
Why Some People Say It’s Not a Bubble
The bulls’ core rebuttal is also one line: demand is real, and this time it’s people who can afford it doing the spending.
On the evidence, NVIDIA’s data-center revenue tops $75 billion in a single quarter, up around 90% year over year, which bulls read as demand still being strong. The cloud order book (backlog, future revenue contracted but not yet recognized) is also thick; Oracle’s remaining performance obligations, for instance, exceed $523 billion, which bulls take to mean demand has already been booked.
More important, the people paying have changed. The ones spending big this round are giants with strong cash flow, like Alphabet, Amazon, Microsoft, and Meta, not startups living off borrowed money. Bulls therefore liken it to the “industrial-grade infrastructure” of electrification or the railways: the upfront investment looks scary, but the foundation, once laid, can serve for a long time and underpin decades of later growth.
A Ruler: Four Metrics to Read Together
Rather than agonizing over “bubble or not,” it’s more useful to keep an eye on four metrics yourself. The key is that they’re only truly dangerous when they worsen “at once”; when only one or two light up, it’s usually just localized overheating.
- Capex vs. payback. Watch whether capex as a share of revenue (capex intensity) is getting out of hand. Morgan Stanley estimates that several cloud giants’ capex-to-revenue ratio could climb to about 38%, 44%, and 45% in 2026 through 2028, far above pre-AI levels. A high ratio isn’t necessarily bad, but if it runs far ahead of revenue for several years running, it’s worth watching.
- Revenue quality. Especially for the private model companies, their “annualized revenue” is mostly a press estimate, not an accountant-audited figure, and there’s little disclosure of retention rates or customer concentration. The same growth number can mean very different things depending on whether it’s backed by solid renewals or one-time curiosity trials.
- The degree of circular financing. When chipmakers and clouds invest in model companies on one hand and collect money from them on the other, demand can get propped up by insiders’ own funds. Look at how much of that chain of deals is real external demand and how much is money going in circles among insiders.
- Chip-depreciation life. This is the most technical and most easily overlooked one (more on that in the next section).
Putting these four side by side gives you a far clearer picture than watching the stock price alone or listening to a single bull or bear take. To be clear: this is a ruler to help you check things for yourself, not a conclusion I’m drawing for you.
Why Depreciation Life Is So Pivotal
The fourth metric deserves its own section, because it’s the least intuitive.
The GPUs cloud companies buy get their cost depreciated gradually over several years, currently mostly 5 to 6 years. The problem is that AI chips turn over fast, and some question whether their real “economic life” might be only 2 to 3 years (once a new generation ships, the old chip stops being worth running). If the depreciation life is assumed to be too long, the annual cost recognized comes out low and book profits get padded up; shorten that life, and profits shrink visibly.
This isn’t idle talk. In 2025 Amazon cut the depreciation life of some servers and networking gear from 6 years to 5; that single adjustment added about $1.4 billion in depreciation and cut about $1 billion from net income. Meta, going the other way, extended its server life to about 5.5 years, saving about $2.9 billion in depreciation for 2025. With the same batch of hardware, change the life assumption and the profit figure jumps along with it. Bears therefore argue that some current book profits may be flattered by depreciation assumptions that run too long.
Can History Offer a Lesson?
The most common analogy drawn is the dot-com bubble of the late 1990s.
Back then, telecom operators were convinced that “internet traffic would grow without limit” and took on heavy debt to lay fiber, only for demand to fall short, leaving vast amounts of fiber idle and operators collapsing in droves. Sound familiar? Both poured big money into building infrastructure before demand had fully arrived. But there are key differences too: back then, most of those spending were debt-laden startups, while this round’s spenders are giants with strong cash flow; and fiber was an asset good for decades, whereas AI chips turn over far faster. Another common analogy is the 19th-century railway mania: many investors lost money, but the rails stayed behind and underpinned the industrial age that followed.
The lesson of history is usually this: a bubble can be a disaster for “investors” yet may leave “society” with useful infrastructure. Which script AI follows hinges on whether what’s being built today can last long enough and be used fully enough to earn the cost back.
The Key Points of This Piece
Nobody can reliably predict whether or when a bubble will burst. Rather than listening to bulls and bears shout, it’s more grounded to keep a ruler of your own handy.
This ruler has four slots: the gap between capex and payback, revenue quality, the degree of circular financing, and chip-depreciation-life assumptions. When all four worsen at once, the market often reads it as elevated risk; when only one or two light up, it’s usually just localized overheating. Bulls and bears actually tend to hold the same set of numbers, differing only in how they read them, and once you see that, you’re less likely to get swept along by either side’s emotions.
The judgment is left to you. This piece lays out a framework and both sides’ arguments; it makes no call on the direction of any stock or the broader market, nor does it constitute investment advice.
To see who’s actually making money along this chain, read the AI Stocks Money-Flow Map; to see the full web of capital behind circular financing, read the AI Circular Financing Map; to understand the new cloud players that rent out GPUs, read Neocloud Compute Rental; to see the physical hardware supply chain, head back to The AI Hardware Supply Chain, End to End.