How The $3 Trillion AI Pyramid Could Crack

I remember sitting in a Midtown coffee shop last spring, listening to two analysts argue about whether Nvidia’s stock was a bargain or a bomb. One of them pulled out his phone, scrolled to a chart, and said something that stuck with me: “The math either works in 2027 or it never works at all.”

He wasn’t wrong.

You’ve probably seen the headlines by now. AI is the biggest capital deployment in modern corporate history, somewhere between $2.9 trillion and $3 trillion in data center construction alone projected through 2028, according to Morgan Stanley Research. Moody’s pegs the same figure for global data center investment over the next five years. The number is so large it’s almost meaningless. Three trillion dollars. That’s roughly the GDP of France.

But here’s the thing about pyramids: the view from the top is spectacular. It’s the structural integrity underneath that keeps engineers up at night.

The Revenue Gap Nobody Wants to Talk About

Let’s start with the core arithmetic, because everything else flows from it.

Sequoia’s David Cahn first did this math in 2023, when he looked at Nvidia’s $50 billion in GPU revenue and worked backward. Add operating costs, add the margins operators need, and the industry would need to generate $200 billion in revenue just to break even. That was three years ago. Now? Cahn’s updated 2026 infrastructure spending figure is $1.5 trillion, and he estimates the AI industry needs to earn roughly $3 trillion to justify all those chips and data centers.

Three trillion dollars in eventual revenue. Against what?

Anthropic has hit roughly $60 billion in annualized recurring revenue. OpenAI, depending on which number you trust on any given week, is somewhere between $36 billion and $70 billion annualized. Even at the generous end, that’s somewhere in the low hundreds of billions combined against a $3 trillion bill. The gap doesn’t just exist, it yawns.

And OpenAI’s audited 2025 statements showed a $38.5 billion net loss on $13 billion in revenue. Revenue, it turns out, is not the same thing as profit. In this industry, the distance between them is measured in compute.

Now, I want to pause here because I can already hear the pushback. “The internet also looked unprofitable in 1999.” True. But the internet didn’t require $60 million per megawatt of data center capacity, that’s the all-in figure Crusoe puts on a modern AI facility, with roughly half the cost sitting in the compute itself. The infrastructure intensity here is unprecedented. And unlike railroads or fiber, GPUs have a depreciation problem that nobody has fully solved.

We’ll get to that.

When Your Biggest Customer Is Also Your Biggest Investor

Here’s where the story takes a turn that would make a securities lawyer reach for the aspirin.

There’s a financing structure that’s been quietly building underneath the AI ecosystem, and it has a name that sounds like a Greek myth: Ouroboros financing. The serpent that devours its own tail.

The mechanics work like this. Six companies, OpenAI, Nvidia, Microsoft, Oracle, AMD, and CoreWeave, have become simultaneously investors, customers, and suppliers to one another. Morgan Stanley mapped the web: OpenAI agreed to buy $300 billion in computing power from Oracle over five years. AMD granted OpenAI the right to purchase 10% of its shares at a penny apiece in exchange for a long-term customer commitment. Nvidia announced investments of up to $100 billion in OpenAI and $10 billion in Anthropic, with critics noting that a significant portion of that money ultimately flows back to Nvidia as chip purchase payments.

Follow the cash around the circle. Nvidia money goes into OpenAI. OpenAI spends it on Nvidia chips. Oracle builds data centers to serve OpenAI’s contracts. Oracle buys Nvidia hardware. Nvidia books the revenue.

“It’s the same dollars doing laps,” as Nigel Green of deVere Group put it, describing the Anthropic-SpaceX-Nvidia arrangement that’s part of a proposed $2 trillion IPO.

This isn’t illegal. And it’s not even necessarily irrational, in the early stages of building an ecosystem, vendor financing can accelerate adoption. It’s essentially what Lucent and Nortel did for telecom startups in the late 1990s, right before that bubble popped. The difference is scale, and the fact that the same handful of hyperscalers now dominate both cloud infrastructure and AI semiconductor markets.

The Bank of England has flagged the “circular arrangements” as a risk factor. The IMF has included them among the biggest threats to financial stability. Michael Burry, the man who saw the subprime crisis coming, has swapped his short positions for put options, betting the bubble may burst sooner than later.

Burry’s specific concern cuts deeper than circularity itself. He’s estimated that hyperscalers could be understating depreciation by more than $175 billion between 2026 and 2028.

The Depreciation Time Bomb

This is the part that keeps credit analysts awake.

Every major GPU operator books a useful life of five to six years for their AI accelerators. That assumption determines how much they charge to depreciation each quarter, and therefore how profitable they appear to be.

But here’s what a 1,397-day panel of actual GPU rental transactions shows: the rental contribution per GPU decays at 17% to 32% annually, against the 16.7% to 20% implied by those five- and six-year useful lives.

In other words, GPUs lose their earning power faster than the accountants say they should.

The A100, Nvidia’s workhorse from 2020, retains about 25% of its original value after six years. The H100, the current standard, holds around 58% after three to four years. That sounds reassuring until you realize that newer chips are becoming dramatically more efficient, which could shorten the economic life of older hardware even if resale values stay high.

Jensen Huang insists that high-end GPUs can generate revenue for up to a decade. Bankers counter with three to four years. That gap, between what the CEO says and what the lenders believe, is where the real risk lives. Because if the actual economic life is closer to four years than ten, then every depreciation schedule in the industry is wrong, and the profits being reported are considerably more optimistic than reality.

Bankers pricing a covenant-levered GPU-backed facility must amortize the hardware before contract expiry. If the GPU doesn’t earn enough in that window, the debt doesn’t get serviced. And you can’t repo a GPU the way you can repossess a car, it depreciates faster than the loan amortizes.

The Concentration Problem Hiding in Plain Sight

Let me tell you something about my own portfolio that I’m not proud of. I thought I was diversified. Index funds, a few sector ETFs, some international exposure. Turns out I own the same bet in seven different wrappers.

The S&P 500’s top ten companies now hold 39% of the index. Nvidia, Apple, and Microsoft alone account for over 21%. Goldman’s Prime Book data shows hedge fund net exposure to the Magnificent Seven at roughly 22% of total US equity exposure, the highest on record since tracking began in 2022.

This isn’t just a portfolio problem. It’s a structural fragility. When everyone owns the same thing, exits get crowded. And the Magnificent Seven now trade at forward price-to-earnings ratios roughly 40% above the rest of the US large-cap market, pricing in not just continued dominance but flawless execution.

Flawless execution is not a thing that happens in complex systems.

The AI buildout is now large enough to contribute roughly 25% of US GDP growth this year, according to Morgan Stanley. When a single investment theme becomes a macro variable, the feedback loops get dangerous. A valuation reset doesn’t stay contained in tech. It ripples through credit markets, consumer spending, and, eventually, employment.

Torsten Slok at Apollo put it plainly: if the hyperscalers don’t meet their free cash flow targets by 2028, “a slower payoff wouldn’t just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction”.

The Debt That Isn’t on the Balance Sheet

Now we get to the really interesting part. The part that rhymes with 2008.

AI companies raised $466 billion in AI-linked debt in 2026 so far, up from $101 billion for the same period last year. But a lot of it doesn’t show up where you’d expect.

In June, Apollo and Blackstone led a $35 billion private credit deal, the largest on record, that technically never reached Anthropic, the company the debt ultimately rests on. Instead, the money funded a special-purpose vehicle that bought roughly a million of Google’s custom AI chips and leased them to Anthropic. Anthropic’s rental payments service the debt. Broadcom supports the senior notes. Google backstops the site leases.

Anthropic added roughly a gigawatt of compute without adding a dollar of debt to its own balance sheet.

This is the financial equivalent of building a house on stilts and calling it a bungalow. The debt exists. The risk exists. It’s just been moved to a different address. And the same pattern is repeating across the industry, datacenter-specific issuance has topped $169 billion, with data-center ABS issuance climbing to $15.5 billion last year from $2.4 billion in 2020.

The credit market is starting to notice. AI-related debt issuance dropped to $23 billion in September, less than half the prior month. New financing has been falling since June’s record $113 billion peak. In the US investment-grade market, where blue-chip tech groups borrowed around $306 billion between January and August, AI-related bond issuance ground to a halt last month.

“Everybody’s being tapped and everybody’s getting a bit tighter,” said John Aylward of Sona Asset Management at a recent industry summit.

When the credit window starts to close, the refinancing math changes. And a lot of these projects depend on continuous refinancing.

What Would Actually Crack the Pyramid

I’m not here to tell you the AI boom is a fraud. The technology is real, the adoption is real, and 21% of S&P 500 companies are now citing AI benefits in their earnings, up from 10% in 2024. Morgan Stanley sees paths to 25% to 50% returns on invested capital across emerging AI business models.

But pyramids don’t crack from the top. They crack from the bottom, where the foundation meets the ground.

Scenario one: the revenue timeline slips. The hyperscalers are forecasting a massive acceleration in free cash flow starting in 2028. That’s a promise, not a fact. If enterprise adoption slows, or if cheaper open-weight models eat into the frontier labs’ pricing power, the gap between investment and return widens. And credit markets are already showing they won’t fund that gap indefinitely.

Scenario two: the depreciation bill comes due. If GPU economic lives are closer to four years than six, the write-downs would cascade through every balance sheet in the ecosystem. Burry’s $175 billion estimate could prove conservative. The equity investments become impaired. The debt gets restructured. The circular deals unwind in a hurry.

Scenario three: a single node fails. The system is a network of interdependencies. OpenAI depends on Oracle for compute, on Nvidia for chips, on Microsoft for cloud. Oracle depends on OpenAI’s payments to service the debt it raised to build the data centers. If any one of these relationships breaks, a missed payment, a contract renegotiation, a bankruptcy filing, the whole loop starts spinning backwards.

The Thing About Circles

Circles are beautiful. They’re also inherently unstable. A circle has no foundation, no base, no load-bearing wall. It holds together only as long as every point on the circumference maintains its relationship to every other point.

Nigel Green said it best: “At $2 trillion, the circle has to spin faster forever. Circles rarely manage it”.

The AI industry has built something genuinely remarkable. The compute infrastructure being assembled right now will power breakthroughs we can’t yet imagine. But the financial architecture supporting it looks increasingly like a Rube Goldberg machine, elaborate, interconnected, and dependent on every component working exactly as designed.

The $3 trillion question isn’t whether AI will change the world. It will. The question is whether the companies building it can survive long enough to see the payoff.

That coffee shop conversation from last spring keeps replaying in my head. “The math either works in 2027 or it never works at all.”

We’re about to find out which it is.