AI’s trillion-dollar bet has no proven path to profit

The U.S. economy is staking its growth on an industry where compute costs are soaring and the numbers still don’t close.

A bet that can’t cash out yet

Let me paint you a picture. It’s a Tuesday afternoon in May 2026. A crypto account named @Exeswap posts a thread that sounds like it was written by a jittery grad student who just discovered that the “AI revolution” might require, you know, actual revenue. The thread is full of jargon, but the core point is brutally simple: The U.S. economy has placed an enormous bet on an AI industry that still has no fully proven path to profitability at scale.

This isn’t some fringe take. It’s the quiet anxiety gnawing at everyone from Silicon Valley venture capitalists to Treasury bond traders. We’ve spent the last two years watching a handful of companies — OpenAI, Anthropic, Meta, Google — pour hundreds of billions into data centers, chips, and electricity. The numbers are staggering. In 2024 alone, capital expenditures on AI infrastructure hit roughly $150 billion globally, according to Goldman Sachs. By early 2026, that figure has climbed past $250 billion. And what’s the return? Mostly promises.

I sat in a meeting last month with a mid-sized SaaS company that had just dropped $2 million on an AI “co-pilot” tool. The CEO told me, with a straight face, that it had “reduced support ticket response time by 12 percent.” That’s nice. But it doesn’t pay for the $2 million license fee plus the $400,000 annual compute bill. The math doesn’t close. And Wall Street is starting to notice.

The compute trap

Here’s the dirty secret that the hype machine doesn’t want you to dwell on: AI is incredibly expensive to run. Every time you ask a large language model a question, you’re burning through compute power that costs real money. OpenAI reportedly spends something like $700,000 per day just to keep ChatGPT running. That’s before any profits. And the costs scale with usage — the more people use these tools, the more the bills pile up.

This isn’t like software, where once you write the code, the marginal cost of each new user is near zero. This is more like building a factory that consumes diamonds as fuel. The more you produce, the more you burn. Jensen Huang, CEO of Nvidia, has been selling the shovels, and he’s made a fortune. But the gold miners? They’re still digging.

The U.S. economy has effectively become a giant venture capital fund for this experiment. The stock market’s valuation of the “Magnificent Seven” — Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, Tesla — now accounts for something like 30 percent of the S&P 500’s total market cap. That’s an insane concentration. And it’s built on the assumption that AI will eventually deliver profits that justify those valuations.

But here’s the thing: “eventually” is not a business plan. The Federal Reserve Bank of San Francisco published a paper in late 2025 estimating that AI adoption could boost U.S. productivity growth by 0.5 to 1.5 percentage points over the next decade. That’s a plausible range. It’s also not the kind of number that justifies a $3 trillion market cap for Nvidia.

The Iran distraction

Meanwhile, we’re fighting a conflict with Iran. I know, I know — you’re wondering what that has to do with the economics of AI. But look at the Twitter feed. A user named @DukeORiordan asked bluntly: “In what way is the current Iran conflict helping the US economy?” The answer, of course, is that it isn’t. Wars are expensive. They divert resources, increase uncertainty, and drive up energy prices — which, in turn, makes compute even more expensive.

The U.S. military budget is already pushing $900 billion. Add a prolonged engagement in the Middle East — or even just the threat of one — and you’re looking at hundreds of billions more in spending. That’s money that could have gone into, say, infrastructure, or education, or even subsidizing AI research. Instead, it’s going into bombs and fuel.

And here’s where it gets ironic: the Iranian community that’s being harmed by the visa pauses — the highly educated, law-abiding Iranian immigrants that @behthr mentioned — are exactly the kind of people we need for the AI industry. Iran has a strong tradition of engineering and math education. Many of the best AI researchers are Iranian or of Iranian descent. But we’re making it harder for them to come here. That’s not just cruel; it’s economically stupid.

The self-sustaining loop that isn’t

The core argument from the AI boosters is that “if intelligence becomes powerful enough, the economics will eventually solve themselves.” That’s the phrase from the Exeswap thread, and it’s maddeningly vague. What does “powerful enough” mean? When do we know we’ve reached it? And what happens if we don’t?

I remember a conversation I had with a former student who now works at a major AI lab. He told me, “Paul, we don’t even know if the current approach will scale to AGI. It might hit a wall. We’re just throwing more compute at it and hoping.” That’s not a strategy. That’s a prayer.

The numbers aren’t closed. Compute costs are still high. Inference costs — the cost of actually using the models — are still high. Scaling, which means building even bigger models, is even more expensive. And the revenue streams — subscriptions, enterprise licenses, advertising — are real, but they’re not closing the gap. OpenAI is reportedly losing money on every ChatGPT subscription. Anthropic is burning through cash. Even Google, which has the deepest pockets, is watching its cloud margins shrink as it invests in AI.

What happens when the music stops?

I’ve been writing about economics long enough to recognize a bubble when I see one. The dot-com bubble, the housing bubble, the crypto bubble — they all had the same pattern: genuine technological promise, excessive optimism, and a belief that “this time is different.” It’s never different. The laws of economics don’t suspend themselves for Silicon Valley.

The U.S. economy can absorb a correction. It’s done it before. But the concentration of risk is alarming. If the AI bubble bursts, it won’t just be a few tech stocks that get hit. It’ll be the whole market. The pension funds that bought into the hype. The municipalities that invested in AI startups. The workers who quit stable jobs to join the gold rush.

We need to stop pretending that the AI industry has already proven itself. It hasn’t. It’s still an experiment. A big, expensive, fascinating experiment. But an experiment nonetheless. And until the numbers close — until compute costs come down and revenues go up — we should be a little more skeptical and a lot less willing to bet the whole economy on it.