AI & Money

Two of the most important chip CEOs say the biggest threat to the AI boom is a shortage of money. Read their words closely and they're describing something else.

Speaking at DBS's Asia Future Forward conference in Singapore this week, Broadcom CEO Hock Tan said something that should have gotten more attention than it did. Capital, he suggested, might become a bigger obstacle to AI's expansion than shortages of chips, memory, or electricity. Intel CEO Lip-Bu Tan agreed, pointing to the rising cost of borrowing and to open questions about whether the companies spending so heavily on AI will ever earn enough back.

On the surface it sounds like a supply-chain story. The industry has been bottlenecked for two years on one thing after another: GPUs, high-bandwidth memory, wafers, power hookups. Hock Tan expects many of those to ease over the next few years. Now the next bottleneck is money.

But stop and think about what it means for money to be the bottleneck on a technology that is supposedly the most valuable ever built. Lenders are getting cautious. Equity investors are getting cautious. Private backers are getting cautious. When every kind of investor hesitates at once, the problem usually isn't the supply of investors.

Demand that doesn't pay

The pitch for all this spending rests on one claim: demand for compute outstrips supply. Both executives repeated it. There is more demand for computing infrastructure, they said, than the industry can provide.

That's true in a narrow sense. People use AI constantly. But a large share of that usage is free, and a large share of the paid usage, by most outside estimates, doesn't cover the cost of serving it. A high-end monthly subscription that loses money on heavy users is not demand in the way a business normally means the word. It's consumption. Consumption you subsidize will always outstrip supply, because you've removed the price that would normally keep it in check.

The headline figures make this hard to see. AI companies love to talk about annualized revenue, and those numbers are genuinely huge and growing fast. What they talk about much less is profit, the real kind, after the cost of running the models. Revenue tells you people will take something. Profit tells you they'll pay what it costs.

A story about chicken wings

There's an old story from the pizza business that explains this better than any spreadsheet.

Back in the '90s, Little Caesars locations sold chicken wings, and the wings sold well. Customers loved them. Then word came down from corporate: stop selling wings. Staff were baffled. Why kill the thing everyone was buying?

The answer was in the ingredients. Pizza is cheap to make: flour, water, yeast, a little cheese, some sauce. Chicken is not. A bag of wings was a real expense, and the margin on each order was thin. Add in the wings that quietly disappeared into employees' stomachs during a long shift, and the store wasn't making a thin margin on wings. It was losing money on every one it sold. The more wings it moved, the worse things got.

So corporate did the obvious thing. It stopped.

Imagine instead that Little Caesars had gone to the bond market and said its real problem was capital. It just needed more money to buy more wings.

Now imagine it had done the opposite. Imagine it had looked at the long line of wing customers and decided the issue was capacity. It just didn't have enough money to buy enough wings. So it went to lenders, to private equity, to anyone who would listen, and raised funds to buy more wings, so it could sell more wings, so it could lose more money on more wings.

Nobody would call that a capital problem. They'd call it a business model problem. And that is roughly where the AI industry is.

"Sufficient" is doing a lot of work

Look at the language the two CEOs actually used. Lip-Bu Tan said people want to see the return on investment, and that companies will need to show tangible returns through efficiency gains and new business models. Hock Tan warned that demand could slow if new models disappoint or if their improvements start to level off. Both framed the long-term outlook as depending on whether AI delivers meaningful economic benefits.

These are not fringe skeptics. These are the people selling the picks and shovels. And in late 2026, four years after ChatGPT, they are publicly saying the question of whether AI will produce sufficient returns is still open. Not outsized returns. Not world-changing returns. Sufficient ones.

That word should stop you cold.

Good technology doesn't need a sales pitch

Anyone who's worked in corporate IT knows how most technology gets adopted. Nobody had to spend an hour convincing a CEO that a networked print server was a good idea. You explained that the whole office could share one printer instead of plugging a printer into every desk, and the answer was "buy one." Email, websites, mobile apps: the case made itself. The hard part was never the pitch. The hard part was getting the budget, because IT has always been treated as a cost center, and executives fight over every dollar.

That's what makes AI so strange. Money is pouring in, yet many of the companies spending it can't clearly say what they're getting back. That can't last. At some point shareholders ask what all the spending did, and at some point a CFO starts looking hard at the line items.

When CFOs do that, they don't care how impressive the technology is. A common story from the early cloud era goes like this: a CIO walks into the CFO's office asking for around $50,000 for new Exchange servers, expecting a quick yes. The CFO says no. They're moving to Office 365. The CIO protests that this is a technical decision. The CFO walks through depreciation schedules, tax treatment, and the ongoing cost of owning hardware, and wins the argument.

The same thing will happen to AI. If a company can get 99% of the usefulness from a product that costs 10% as much, a smaller model, an open-weight model, something running locally, that's what it will buy. And the local option is arriving fast. Google is building language models straight into Android and Chrome. New PCs ship with hardware that handles AI tasks on the device and only reaches out to the cloud when it has to. Every task that stays on the device is a task nobody pays a frontier lab for.

Four years in

The best comparison is the web. For most people, the web arrived with Windows 95, which put a browser in front of ordinary users the same way ChatGPT put a chatbot in front of them in 2022. Compare what happened in the four years after each.

From 1995 to 1999, the world changed in ways nobody had to explain. Businesses moved online, people started emailing instead of calling, and entire industries reorganized. No one had to be told the web mattered. They could see it.

From 2022 to 2026, the picture is more mixed. AI has genuinely changed some fields, software development most obviously. But outside those pockets, the transformation is mostly something we're told about rather than something we see. That gap between the story and the lived experience is exactly what the two CEOs are hinting at when they talk about proving returns.

And even if AI does turn out to be as essential as the web, that doesn't mean the companies selling it will capture the value. Web browsers are essential. You can't do much of anything in 2026 without one. Netscape used to sell its browser for around $40. Then Microsoft bundled Internet Explorer into Windows for free, and Netscape's business collapsed. Today nobody pays for a browser. Essential technology and valuable products are not the same thing.

The fair counterargument

The bulls have real points. Amazon lost money for years before its infrastructure became one of the most profitable businesses in history. The cost of running a given level of AI capability has dropped sharply and keeps dropping, so today's money-losing usage might be tomorrow's profitable usage. And in coding, customer support, and some research work, there are companies reporting measurable productivity gains, not just hype.

But those arguments all ask for the same thing: patience, plus more money, on the promise that the economics will work out later. That's a reasonable bet to make with some of your capital. It's a strange bet to make with trillions of it, at rising interest rates, before the business model has been shown to work.

The technology itself is fine. The models are useful. The problem is the assumption that the only thing standing between AI companies and success is more cash. Little Caesars didn't need more money for wings. It needed to stop selling wings at a loss. The AI industry is going to face the same choice, and the people building its hardware are starting to say so out loud.

• • •