Trillion-Dollar AI Gamble

OpenAI had reportedly lined up more than $1 trillion in infrastructure commitments in late 2025. Nvidia committed up to $100 billion. AMD offered warrants to purchase 160 million shares, nearly 10% of the company, at one cent per share. In exchange, OpenAI committed to purchase six gigawatts of AMD GPUs, representing about $90 billion in spend. Those $160 million in AMD shares are worth nearly $40 billion. At $90 billion of revenue and 20% margins, AMD gets $18 billion of EBITDA back, but at a cost of $40 billion in stock.

Then came Broadcom, with a plan to deploy ten gigawatts of OpenAI-designed AI accelerators, an estimated $350 billion deal. Oracle signed for $300 billion, Microsoft for $250 billion, Amazon for $38 billion over seven years, CoreWeave for $22.4 billion. Put together, the announced deals total roughly $1.15 trillion in infrastructure commitments over five years.

That number is hard to process on its own, so use a simpler frame. A company doing roughly $20 billion in annualized revenue was acting like it had the balance sheet of the largest business in history. How does a company with $13 billion in revenue sign $1.4 trillion in spending commitments?

That does not automatically mean fraud, and it does not mean AI is fake. It does mean the economics deserve far more scrutiny than the market is giving them.

The Math Is Absurd

The easiest way to get lost in AI finance is to stare at single deals. One vendor offers equity sweeteners, another provides cloud credits, another signs a massive chip partnership. Each announcement sounds strategic in isolation, but put them together and the scale starts looking detached from reality.

OpenAI's reported commitments add up to a sum that sits uncomfortably close to total annual corporate capital expenditure in the United States. Annually, US companies spend around $1.2 trillion on CapEx combined. OpenAI is committing to spend nearly 100% of that amount over five years. Even if the headline numbers overstate what will actually be spent, the direction of travel is obvious. This is not normal venture-backed expansion. It is a wager on a future so large that almost no existing company offers a useful comparison.

The usual defense is that OpenAI will simply grow into the spending. That argument falls apart as soon as the revenue required is penciled out. If capital needs ramp from $6 billion in 2025 to $295 billion in 2030, then revenue has to scale from $12 billion to $983 billion in the same period. That is an 85X increase in five years. OpenAI essentially needs to become the most profitable and biggest company on earth to meet its obligations.

And that is before operating expenses, energy, maintenance, debt service, leases, or the cost of keeping the whole machine running. These numbers do not contemplate any of those costs. Although this is not impossible, the setup is insane.

The optimistic case still asks OpenAI to become one of the biggest and most profitable companies ever assembled, on a timeline that borders on fantasy.

The Circular Money Machine

The scale is only half the problem, and the structure of the deals is the other half. Much of the AI boom is built on circular flows of money. Look at the Microsoft relationship. Microsoft invested over $13 billion into OpenAI for equity and revenue share, but much of that investment was not cash. It was Azure cloud credits. Microsoft invests in OpenAI so that OpenAI can buy Microsoft products. OpenAI consumes that Azure capacity, which justifies Microsoft ordering more hardware from Nvidia. Microsoft takes the booked revenue from OpenAI and places massive orders with Nvidia for millions of GPUs.

Now Nvidia has cash coming in from Microsoft and others, and it starts its own loop. Nvidia commits up to $100 billion to invest in OpenAI. OpenAI then commits to using that investment to purchase a huge volume of Nvidia chips. With that leftover money, OpenAI can pay off others like Microsoft. So Nvidia pays OpenAI so OpenAI can pay Nvidia and also pay Microsoft, so that Microsoft can invest in OpenAI, so that OpenAI can buy from Microsoft, so that Microsoft can buy from Nvidia, so that Nvidia can invest in OpenAI.

The same pattern appears elsewhere. Venture investments with companies like Ambience, Harvey, and Anysphere follow the same script. OpenAI invests in startups so those startups can use the money to spend more with OpenAI. The revenue of one company becomes the expense of another, which becomes the investment case for a third.

These ouroboros-like dealings are impossible to track and measure because money flows from entity to entity. The revenue of one company is the cost of another is the investment of another. Dollars get counted as revenue, investment, demand, and partnership all at once. Layer on complex contingencies and financial instruments, and it becomes impossible to assess the quality of cash flows.

When the revenue of one company is the expense of another, and the investment case for a third, the line between real demand and self-reinforcing financial theater gets blurry. The market sees growth, suppliers see backlog, and investors see momentum, but momentum is not the same thing as durable cash generation.

None of this is illegal, but it makes the system much harder to read.

This is what makes the bubble more dangerous than a simple valuation story. It is not just that prices may be high. It is that the underlying cash flows are increasingly entangled, contingent, and hard to separate from the financing arrangements producing them.

Not 2008, Not 1999, Still A Problem

The most common comparison is 2008, which is too neat. The AI boom does not look like the global financial crisis in its pure form. There is not yet the same degree of securitized garbage, derivative layering, and extreme leverage sitting on top of household balance sheets. The biggest companies involved also have real cash flow and stronger balance sheets than the institutions that blew up in 2008.

That lowers the odds of an immediate systemic blowup. The better comparison is the dot-com era, with one important difference. Back then, many of the companies at the center of the mania had weak businesses and little profit. Today, the largest firms funding the AI arms race are real machines. Microsoft, Nvidia, Amazon, and Meta are not Pets.com, they print cash.

That makes the bubble sturdier, but also stranger. A profitable set of giant companies can keep bad incentives alive much longer than weak companies can. They can absorb losses, subsidize experiments, and keep bidding up the story long after the economics stop making sense for smaller players.

So no, this is not a replay of 2008. No, it is not a clean repeat of 1999 either. It is its own thing, a bubble supported by real profits, circular deal-making, and a belief that future AI demand will justify almost any present commitment.

Too Big to Fail Already?

Once a company starts floating the need for a public backstop, even indirectly, the frame changes. This stops being a simple growth story and starts becoming a political one. If the infrastructure build-out is treated as strategically important, then private upside and public downside begin creeping toward each other.

In a Wall Street Journal interview, OpenAI CFO Sarah Friar suggested the company may eventually need a financing ecosystem that includes banks, private equity, and potentially government support. A federal backstop. That way, if OpenAI can't meet those obligations, the taxpayers will be on the hook.

A private company making gigantic speculative commitments is one thing. A private company making gigantic speculative commitments while hinting that the state may eventually need to stabilize the system is something else entirely. That is the point where "ambitious" starts sliding toward too big to fail.

And OpenAI is only the center of the story because it is the loudest. Anthropic, Meta, and a long tail of AI startups are all participating in the same logic. Anthropic raised another $13 billion in its Series F and hit a $183 billion post-money valuation, with commitments to purchase $30 billion of Azure compute and possibly $14 billion with Amazon. Meta agreed to a $14.8 billion Scale AI deal and is projecting $70 billion in data center spending, plus $1 billion in wages for top AI talent. There are already reports that people either turned down those pay packages or are leaving to go back to their previous jobs or start their own companies.

Huge valuations, huge compute promises, huge assumptions about future demand, too much capital chasing too few durable business models.

There will be real winners, and there already are some, but there's also a lot of AI slop wrapped in premium narratives and sold as inevitability.

Raisins and Turds

Charlie Munger once said that if you mix raisins with turds, you still have turds. That is a crude line, but it fits.

There are real businesses in this market. There are real products people want. There are companies with genuine moats, real margins, and actual cash flow. The problem is that these are now mixed in with circular transactions, inflated expectations, and spending commitments that require absurdly favorable futures to look sane.

That is why the AI bubble is worse than it first appears. Not because every company involved is fake, and not because AI itself is useless, but because the system is becoming harder to value honestly. The same dollars are being counted as revenue, investment, demand, strategic partnership, and future growth story all at once.

When that happens, price discovery degrades. Risk gets hidden. Everyone looks stronger than they are, and when the story finally breaks, the question will not be whether AI mattered. It will be who was left holding the circular money.