Four companies will spend roughly $725 billion on capital projects this year. Whether that spending pays off turns, to a surprising degree, on one estimate buried in the footnotes: how long a graphics processor stays economically useful. Here is the bear case on the AI buildout, the strongest case against it, and why most investors already own the outcome whether they chose it or not.
Amazon, Alphabet, Microsoft and Meta have guided to roughly $725 billion of capital spending in 2026. That is not a forecast. It is what they told their own shareholders, on their own earnings calls, with their names on it.
It is up about 77% from roughly $410 billion in 2025. Goldman Sachs now models $5.3 trillion of hyperscaler capital spending between 2025 and 2030. Technology equipment and software investment reached something near 4.4% of GDP last year — a level last seen at the peak of the dot-com build.
Somebody has to be right about this. Either the buildout produces enough revenue to justify it, or it does not. There is no third outcome in which a trillion dollars of concrete, copper and silicon quietly turns out not to matter.
Ed Zitron argues it does not. Michael Burry argues the accounting is already obscuring how thin the returns are. I do not agree with everything either man says, and I have given their critics a full section further down — a long one, because the counterarguments are strong and nobody should hold this view casually. But the arguments have not been answered. And a great many people who own Nvidia and AMD today, including people who never chose to own them, have never heard the arguments at all.
This page is about that second group.
Here is the part that gets the least attention and carries the most weight.
A graphics processor is depreciated over a useful life the company itself estimates. The hyperscalers generally use five to six years. Burry's argument is that the real economic life is closer to two or three, because each new generation is materially more efficient per watt and per unit of compute — which makes the prior generation uneconomic for frontier work long before it stops turning on.
By his arithmetic, that gap understates depreciation by roughly $176 billion across the industry between 2026 and 2028, and leaves reported operating income at some of these companies more than 20% above economic reality.
Nvidia disputes it, and says customers land on four-to-six-year lives based on observed utilization and longevity. That is a real argument, not a dodge. Old chips do keep working, and serving a model is far less demanding than training one.
But notice what happened inside the buildout itself.
Reduced the useful life on a subset of its servers — accepting lower reported earnings today in exchange for an assumption it evidently believed it could defend.
Lengthened its useful-life assumption further, spreading the same kind of hardware cost across more years and supporting reported earnings in the process.
Same hardware. Same physics. Opposite accounting. Satya Nadella has said publicly that he does not want to be stuck carrying years of depreciation on a single chip generation. When the operators closest to the machines cannot agree on the one variable that decides whether the spending works, that variable is not settled. It is contested from the inside.
Zitron's case is narrower, and in some ways harder to wave off. He argues that the whole structure rests on one company that does not make money.
By the audited figures he has published, OpenAI generated about $13.07 billion of revenue in 2025 against roughly $34 billion of expenses. In the first quarter of 2026, revenue reached about $5.7 billion with an operating loss near $9.3 billion. The company projects cash burn of roughly $25 billion in 2026, rising toward $57 billion in 2027.
Against that, its commitments to cloud providers for computing capacity run into the hundreds of billions of dollars, with contracts extending through 2030 — and most of them do not sit on the balance sheet.
Oracle is the clearest expression of the exposure. Its remaining performance obligations have ballooned past $600 billion while free cash flow ran deeply negative, and S&P cut its rating to the lowest rung of investment grade with the OpenAI relationship named in the agency's own reasoning.
Read that last clause again. A rating agency has said out loud that a single customer is now a credit factor for a company of Oracle's size.
If this were all funded out of operating cash flow, a bad outcome would be an earnings problem and nothing worse. Increasingly, it is not.
AI-related borrowers tapped the debt markets for at least $200 billion in 2025 — likely an undercount, since a great deal of it is private. Hyperscalers alone issued roughly $121 billion of bonds that year, more than four times their five-year average. GPU-collateralized lending has become its own asset class: CoreWeave's $7.5 billion facility carried an average rate near 11%, with repayments beginning in January 2026. JPMorgan projects $30 to $40 billion a year of data center securitization in 2026 and 2027.
Now set that beside one more number. Rental rates on the prior generation of Nvidia chips have fallen by something on the order of 70% to 90% since 2023.
The collateral is depreciating faster than the loan amortizes.
The comforting story about Nvidia and AMD is that they sell the shovels, so they get paid regardless of who strikes gold. That story has a hole in it.
Roughly half of Nvidia's data center revenue flows from a handful of hyperscalers. Its recent filings disclose about $119 billion of supply-related commitments and $30 billion of multi-year cloud service commitments. This is not a business diversified across a thousand customers. It is a business levered to the capital budgets of five.
Then there is the circularity. Nvidia announced an intent to invest up to $100 billion alongside OpenAI's buildout; a quarter later, the Wall Street Journal reported the arrangement had gone cold. AMD's October 2025 agreement with OpenAI paired six gigawatts of deployment with warrants that could convey up to roughly 10% of AMD's shares.
When a chipmaker helps finance the customer who buys the chips, some portion of the reported revenue is that chipmaker's own capital making a round trip. That is not fraud, and I am not alleging any. It was also standard practice in telecom in 1999.
Valuation leaves very little room to be wrong about any of it. AMD has recently traded north of 150 times trailing earnings. Nvidia's forward multiple looks far more reasonable — but every forward multiple in this complex is a function of hyperscaler capital spending staying voracious. The denominator is the assumption.
None of these requires a collapse. That is the point of listing them. Each one is an ordinary business event that would nonetheless reprice a great deal of capital.
Useful lives get shortened. Earnings fall across the buyers with no change in demand whatsoever, and the multiple on the sellers compresses in sympathy.
Nobody has to cancel anything. One hyperscaler saying “we are pacing ourselves” on one earnings call is enough to reprice a sector built on 77% growth.
If the cost of running a model keeps falling, you need fewer chips per unit of intelligence. Excellent for AI as a technology. Poor for companies whose revenue assumes chip scarcity.
Revenue owed by a customer who is financed is not the same asset as revenue owed by a customer who is funded. If the funding rounds stall, it lands on the cloud providers first and the order book second.
Under Chair Warsh, roughly half the Fed's policymakers now project a 2026 rate increase, up from none in March. Long-duration growth assets are exactly where a higher discount rate does the most arithmetic damage.
The passive flows that concentrated the market on the way up do not politely stop on the way down. They reverse, mechanically, into the same names.
I would not publish the preceding sections without this one. These objections are strong, several of them are held by people who have been right for a decade, and anyone who reads only the bear case has been given half an argument.
Microsoft's commercial remaining performance obligations sit north of $600 billion. Google Cloud's backlog roughly doubled to about $460 billion. Microsoft's AI business passed a $37 billion annual run rate, more than doubling year over year. These are signed dollars, not projections, and one Jefferies analyst was blunt enough to call the bear thesis “garbage.”
Unlike the telecom build of the late 1990s, the largest spenders here generate enormous operating cash flow. Free cash flow compresses; solvency does not come into question. That is a materially different animal from a company borrowing to build something it hopes someone will use.
High-bandwidth memory is reported sold out through most of 2027. Shortage is an odd signature for a demand mirage. If this were a bubble in the classic sense, we would expect gluts, discounting and idle capacity — and we are not seeing them yet.
Cisco traded past 100 times earnings at the dot-com peak. Nvidia's forward multiple is a fraction of its own five-year average, and FactSet has second-quarter 2026 semiconductor earnings growing triple digits. Expensive is not the same as absurd, and this cohort earns real money.
Overbuilt railroads and dark fiber both ruined their first owners and then served the economy for decades at the second owner's cost basis. Malinvestment and worthlessness are different words. A capital cycle that punishes today's shareholders can still leave the country better equipped.
Burry's positions are trades with expiration dates, not forecasts with timelines. Zitron has been directionally consistent and repeatedly early. For an investor, being right about the arithmetic and wrong about the year is indistinguishable from being wrong — and I include myself in that warning.
Here is why I am writing this for families rather than for traders.
The ten largest holdings of the S&P 500 now account for roughly 41% to 43% of the index — a record, and close to double the 18% to 23% that prevailed from 1990 through 2015. The Magnificent Seven alone are around a third of it. Nvidia by itself has run near 8%.
Which means the index fund in your 401(k) is not the diversified cross-section of American business it was when you started contributing to it. More than $40 of every $100 you add goes to ten companies, and those ten are not spread across unrelated industries the way the top ten were in 1990. They are linked by one theme.
For the households I work with, it compounds twice.
If you carry a large position in your employer's stock and hold a broad index fund alongside it, you have a concentrated single-name risk sitting on top of a concentrated single-theme risk. And if that employer is a utility, notice that a good portion of the load-growth story now supporting utility capital plans — and utility equity valuations — is the same data center buildout.
A slower buildout does not only reprice the chipmakers. It reaches the demand forecasts underneath the regulated business too.
That is what a single point of failure looks like when it is spread across three account statements and therefore feels like diversification.
I have written for years that we are living through a historic period, and that the reset ahead will be the largest of my lifetime. I hold that view because of debt, deficits, and a monetary posture that has been wrong repeatedly and expensively — not because of anything happening in a data center in Ohio.
So let me be precise about what I am and am not saying.
I am not saying AI is a hoax. It is not. I am not saying these companies fail. Most of them will not. I am not predicting a date, because I do not have one, and anyone who hands you one is selling something.
What I am saying is narrower, and harder to argue with: an enormous share of American retirement capital is now positioned on the assumption that hardware bought in 2026 will still be economically productive in 2031 — and almost nobody holding that position knows they are holding it.
That is not a market call. That is a description of a portfolio. And a portfolio is something you can act on without predicting anything at all.
The response is not to sell everything, and it is not to time a top. It is to know what you own, size it deliberately, rebalance on a schedule rather than on a headline, and stop mistaking a market-cap-weighted index for diversification.
Not one of these requires a market forecast. All five are answerable this week, with statements you already have.
What percentage of my total portfolio sits in the ten largest companies in the S&P 500 — counting every account, not just the statement I look at most?
If I hold company stock, how correlated is my employer's business to the same buildout my index fund is exposed to?
What would a 40% drawdown in that top-ten cohort do to the withdrawals I am counting on for the first five years of retirement?
Do I have a rebalancing rule written down, or do I have an intention?
Which of my holdings did I choose, and which did I simply inherit from an index committee's weighting methodology?
If you would like those questions answered with actual numbers instead of impressions — across every account, including the company stock — that is a large part of what I do. Start wherever fits.