Is This an AI Bubble? A Computer Engineer Reads the Financials

Every few months someone asks me, straight out, whether AI is a bubble. And I get why. The chart looks like one. Nvidia went from a gaming-chip company to the most valuable business on Earth in about three years. The word “bubble” is doing a lot of work in those conversations, and almost none of it is precise.

Here is the thing that bugs me as someone who reads financial statements for fun. The default comparison is always the dot-com crash, and that comparison is lazy. Dot-com was a valuation event built on companies that had no earnings. This is not that. Whatever is happening in 2026, it is not 1999, and pretending it is means you will be watching the wrong risk. So let me actually read the numbers, because the real question is more interesting than the framing.

The dot-com comparison falls apart on the first line item

In 2000, the poster child was Cisco. It made the routers and switches that the internet ran on, so the story wasn't crazy, but the stock was. Cisco peaked at a P/E ratio north of 200 while its margins were actually shrinking.^[6]Think about what a 200 P/E means. You are paying two hundred years of current earnings for the stock. The math only works if growth stays violent essentially forever. It didn't, and the stock lost about 80% of its value and never came back to that peak.

Now look at Nvidia. In fiscal 2026 it reported revenue of $215.9 billion, up 65% year over year, and GAAP net income of $120.1 billion.^[1] That is a net margin around 56%. Software companies dream about margins like that. A chip company posting them is close to unheard of. Data center revenue alone was $193.7 billion. Then the next quarter got louder. In Q1 of fiscal 2027, ended late April 2026, Nvidia did $81.6 billion of revenue in a single quarter, up 85%, with a 74.9% gross margin and $58.3 billion of net income.^[2]One quarter of profit that dwarfs most companies' entire market cap.

And here is the part that should stop the lazy comparison cold. At mid-July 2026, Nvidia traded near 20 times forward earnings, and during a selloff it briefly dipped to 18x, which is belowthe S&P 500 average.^[6] Cisco at 200-plus with shrinking margins. Nvidia at 18 to 20 with record margins. These are not the same animal. They are barely the same phylum.

So if you define a bubble as “price wildly disconnected from earnings,” the market leader flunks the test. The earnings are there. The margin is there. The multiple is sane. Whatever the AI risk is, it is not sitting in Nvidia's P/E ratio. That is the whole reason the dot-com frame misleads people. They go looking for a 200 multiple, don't find one, and conclude everything is fine. That conclusion is also wrong.

Where the risk actually lives: the spending, not the price

The scary number in 2026 isn't a valuation. It's a capex line. Google, Amazon, Microsoft, and Meta collectively guided to about $725 billion of capital expenditure in 2026.^[3]That is up roughly 77% from the prior year's record of about $410 billion. Amazon near $200 billion. Microsoft near $190 billion. Google in the $175 to $185 billion range. Meta $125 to $145 billion. These are annual figures, from four companies, and they are still climbing.

Goldman Sachs projects the four largest hyperscalers will spend a combined $5.3 trillion on capex from fiscal 2025 through fiscal 2030, and expects total big-tech capex to top $1 trillion in a single year by 2027.^[3] A trillion dollars a year. For context, that is a meaningful fraction of what the entire United States spends on defense. It is being deployed on the bet that demand for AI compute keeps compounding.

Now put that spending next to what it's buying. An MIT study found that 95% of companies saw zero return on their generative-AI investments, despite $30 to $40 billion poured in.^[11] Read that twice. The infrastructure is being built at a trillion-dollar annual clip, and the enterprises meant to use it are, so far, mostly not making money on it. That is the gap. Not price versus earnings. Spending versus realized return.

This is the actual bear case, and it's a good one. Nvidia's margins are real precisely becausethe hyperscalers are spending like this. Nvidia's revenue is their capex. If even a couple of those four decide the returns aren't materializing and pull back, the most profitable company in the world loses its biggest customers at the same time. The 56% margin is a function of demand that is, right now, driven by faith more than by proven downstream profit.

Concentration: the whole index is riding on seven names

Here's a different kind of risk, and it's structural. As of July 2026 the Magnificent Seven, which is Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla, carried a combined market cap of about $22 trillion and made up roughly 32.5% of the S&P 500.^[7] Seven companies. Nearly a third of an index of five hundred. That is one of the highest concentrations since the late-1990s tech era.

This matters even if you never buy a tech stock directly. If you own an S&P 500 index fund, and most people's retirement money does, then a third of your exposure is seven correlated bets on the same theme. The diversification you think you own is thinner than it looks. When people say “just buy the index,” in 2026 that is quietly a large AI position.

The interesting wrinkle is that the market has already started to pull these seven apart. As a group the Magnificent Seven rose only about 2.6% over the trailing year through mid-2026 and actually underperformed the S&P 500.^[13] But the dispersion was enormous, from Tesla down about 17% to Nvidia up about 51%. That is not the behavior of a mania where everything with a story goes vertical. That is a market doing discrimination, rewarding the names with real AI earnings and punishing the ones coasting on the label. Healthy, actually.

The financing is starting to look circular

This is the part that would keep me up at night if I had money in the whole complex. In September 2025, Nvidia announced an investment of up to $100 billion into OpenAI. OpenAI then uses that money to buy, among other things, Nvidia GPUs.^[5]So Nvidia funds a customer, the customer buys Nvidia's product, and that shows up as Nvidia revenue. Critics call this circular financing, and they are not wrong to be nervous. By 2026, more than $800 billion in such interlocking AI-supply-chain arrangements had been identified.^[5]

Circular deals aren't automatically fraud. Vendor financing is a normal thing. But it does inflate the appearance of demand, because the same dollar can be counted as revenue at more than one link in the chain. When a big chunk of your customer's ability to pay comes from you in the first place, the “demand” is partly manufactured. And that is exactly the kind of structure that looks brilliant on the way up and unwinds fast on the way down.

Look at OpenAI specifically, because it's the load-bearing pillar of the whole AI demand story. OpenAI generated about $13.07 billion of revenue in 2025, up from $3.7 billion in 2024, and reached roughly $25 billion in annualized revenue by mid-2026.^[8] Explosive growth, no question. But it is projected to lose about $14 billion in 2026, and it has committed to roughly $1.4 trillion of data center infrastructure spending over about eight years.^[8] Sit with those three numbers. Twenty-five billion of revenue. Fourteen billion of losses. A 1.4 trillion dollar spending commitment. The gap between the last number and the first two is the bet.

The accounting question nobody wants to answer

Michael Burry, the investor who shorted the housing bubble, resurfaced with a specific and technical bet. He took notional short positions of about $1.1 billion against Nvidia and Palantir.^[4]His argument is not “the stocks are too high.” It is sharper than that. He claims the hyperscalers are understating depreciation by extending the assumed useful life of their GPUs.

The engineering angle here is the whole point, so let me unpack it. A GPU is a depreciating asset. You buy it, you write down its value over its useful life, and that depreciation is a real expense that reduces reported profit. If a company assumes its chips last six years instead of, say, three or four, it spreads that expense thinner each year and reported earnings look better. But anyone who has watched this hardware cycle knows a cutting-edge accelerator does not stay cutting-edge for six years. Nvidia ships a materially better generation faster than that. Burry estimates the extended-life assumptions could understate industry depreciation by roughly $176 billion between 2026 and 2028.^[4]

There's a related off-balance-sheet issue that got less attention. A February 2026 Moody's report found the five major hyperscalers carry a combined $662 billion in signed-but-not-yet-started data center lease commitments that stay off their balance sheets until the leases activate.^[10]That is $662 billion of obligation that doesn't show up in the leverage numbers people are looking at today. It is coming, it is contractually real, and it's not in the tidy debt figure on the current statements.

So, is it a bubble? My honest read

I'm going to give you an actual answer instead of hedging it into paste. Here is how I hold it.

The chip layer is not a valuation bubble.Nvidia at 18 to 20 times earnings with a 56% net margin and $120 billion of profit is not Cisco at 200 times with shrinking margins. If you are waiting for a dot-com-style valuation blowup to signal the top, you will miss the actual risk, because it isn't priced like a mania. At mid-July 2026 Nvidia sat near a $4.9 to $5.1 trillion market cap, the world's most valuable company, and it had already shed close to $1 trillion of value from a mid-May peak around $5.5 trillion.^[12]The market is repricing it in real time, and it's doing it on fundamentals, not fear alone.

The spending layer is where a bubble could actually be.A trillion dollars a year of capex, against enterprise AI returns that 95% of companies say they haven't seen yet, financed in part by circular deals, propped up by a customer losing $14 billion a year, with depreciation schedules that flatter the earnings and $662 billion of leases hiding off the books. That is not a valuation problem. It is a capital-cycle problem, and capital cycles are exactly how real bubbles form. Overbuild first, discover the demand second.

The honest version is that both things are true at once, and people keep collapsing them into a single yes-or-no because that's easier to tweet. The technology is real and the economics of the chip maker are real. The financing structure and the return-on-investment gap are genuinely fragile. If this ends badly, it won't look like 2000, where the leaders had no earnings. It will look like a classic infrastructure overbuild, railroads and fiber optics, where the technology was real, the demand eventually showed up, and the people who financed the first wave still got wiped out because they built too much, too fast, on borrowed conviction.

So when someone asks me if it's a bubble, my answer is that they're asking the wrong question. Stop staring at Nvidia's P/E, because it's the least alarming number in the whole picture. Watch the capex guidance. Watch whether enterprise AI returns start showing up in the next few earnings seasons. Watch whether the circular deals keep growing. Watch the depreciation footnotes, boring as they are. That is where the story actually gets decided. The valuation is the part that looks scary and isn't. The financing is the part that looks fine and might not be.