AI Capex Warnings: Separating Signal from Noise in Tech's Biggest Bet

How to Tell Which Companies Are Actually Burning Cash

On the Value of Information

One of my core investing principles is obsessing over the value of information. I’d estimate 95%—probably more—of what I read is pure noise. Most of it won’t matter in one moth let alone six months. Most of it exists solely to make me do something stupid.

So whenever I read something, I ask: Does this actually matter? Or is this the financial media equivalent of empty calories—filling, forgettable, and potentially harmful?

More information isn’t better. The right information is crucial and the timing matters. I actively label things as wrong or useless so they don’t accumulate in my brain like some kind of intellectual junk drawer. It takes effort. But I’d rather remember three things that matter than three hundred things that don’t.

My favorite quote from 100 to 1 in the Stock Market by Thomas Phelps is:

When you read a bearish story on a company whose stock has declined to a third of what it was two or three years ago, ask yourself not only whether the story rings true but also why it was published at this late date. It may be factual but still highly misleading to investors because of its timing.

Factual doesn’t mean valuable. Timing changes everything.

When Everyone Agrees, Ask Why Now

This brings me to the sudden flood of AI capex warnings—articles arguing that tech companies may be overspending on data centers and infrastructure to power their AI ambitions. The concerns aren’t new; I saw early articles on this more than a year ago. But lately? They’re everywhere.

Some are genuinely well-researched pieces, I very much like the two recent pieces by Harris Kupperman from Praetorian Capital. But the sheer volume, and the timing? Suspicious.

Even Jeff Bezos recently said at a talk in Italy that artificial intelligence is currently in an “industrial bubble” where both good and bad ideas are being funded. However, he added that the technology is real and will ultimately bring significant benefits to society.

The question for an investor isn’t whether the warnings are factual, but whether they are valuable information or just well-timed noise. To answer that, you need a framework.

AI Will Change Everything. That Doesn’t Make It a Good Investment.

I personally think AI will have a transformative impact on society—probably as significant as, if not greater than, the internet. There’s no question about it. The technology is real.¹

But that doesn’t mean you should be the one funding the buildout. The railroads changed America. The fiber-optic networks power everything we do online today. Both were essential infrastructure. Most of the companies that built them went bankrupt anyway.

The question isn’t whether AI matters. It’s whether the companies burning billions on capex (and opex) today will be the ones capturing the value tomorrow—or whether they’re just laying expensive groundwork for someone else’s profit.

From Moat to Treadmill: Why AI Is Different

Everyone uses the same analogies—railroads, fiber-optic cables, shale drilling. I just used them myself. But the details matter, and they are exactly why I don’t like the comparison.

Railroads and fiber-optic cables, in theory at least, build a moat. Once you lay down tracks or bury fiber, nobody else is doing it. It would be economic suicide. This infrastructure lasts. In Germany, we still use train tracks and signaling from the pre-WW1 era. Fiber cables are the same. Once they’re in the ground, nobody’s ripping them out to build “better” fiber.

AI is different. What you build today is obsolete in 3 to 5 years. New chips, new architectures, faster models—they make last generation’s infrastructure worthless.

This creates a trap. If you stop investing and your competitors don’t, everything you spent becomes useless. You’re not building a durable moat. You’re running on a treadmill that speeds up every year.

Worse: rapid depreciation means any latecomer with fresh capital can deploy newer architectures and close the gap without first wasting billions on infrastructure that’s already obsolete.

This is structurally identical to the prisoner’s dilemma: mutual restraint would benefit everyone, but individual incentives guarantee an arms race. If you keep spending, you burn capital on depreciating assets. If you stop and others don’t, you fall behind. The only winning move—coordinated restraint—is unavailable because no one can trust competitors to hold back. This makes the AI buildout even more brutal than historical infrastructure booms.

The Hyperscaler’s Defense: Ecosystems and Cash Flow

On the other hand, some of the companies building AI infrastructure today aren’t like railroad companies. Microsoft, Google, Amazon, Tencent—they aren’t pure infrastructure plays. They have massive cash-flow engines elsewhere. They can subsidize AI to reinforce broader ecosystems, even if AI itself runs at break-even or a loss.

I’ve written about this before with bike rental businesses in China. Meituan, Alipay, Didi—they all run bike rentals that barely break even, maybe lose money. But the point is bringing people into the ecosystem.

The difference? Bike rentals are tiny compared to the overall scale of these businesses. The losses don’t matter. AI capex? That’s serious business. We’re talking hundreds of billions. That’s not a rounding error.

For AI-first startups like OpenAI, with minimal revenue and massive burn, the bearish case writes itself. They’re spending far ahead of any proven business model.

But the picture is more nuanced than the blanket warnings suggest. The noise of capex concerns fails to make this critical distinction. Details matter, and the details vary wildly. A blanket warning about AI capex is about as useful as a blanket warning about tech stocks—which is to say, useless.

I went through all of my AI-related holdings and analyzed other players in the field, comparing their capex, opex, revenue trajectories, and whatever AI-related financial data I could extract. The picture is anything but uniform. For some I’m deeply pessimistic. The math simply doesn’t work. Pessimistic doesn’t mean doomed—I just can’t see the path yet. My strategy is to track them closely and wait for the fog to clear, then identify who survives and wins big, because there will be winners in that group. For others, it looks less like reckless spending and more like a well-calculated gamble with asymmetric upside.

To see if the noise has merit, let’s take Tencent as a case study. I went through their filings, management statements, and reliable online information.

A Case Study in Cutting Through the Noise: Tencent