Rosenberg Research: Big Tech’s capex race and the looming earnings-quality problem

Open this photo in gallery:

For the better part of last three decades, Big Tech had the dream business model. Build the software once, distribute it globally, add users at very low incremental cost, and watch margins expand. That was the magic of the internet era: asset-light scale, huge network effects, and very little need to keep pouring capital into physical infrastructure to scale up operations.

That model is changing. The artificial intelligence boom is dragging Big Tech into a far more capital-intensive world. This is no longer just code, cloud subscriptions and digital advertising. AI models require land, power, cooling systems, data centres, servers, GPUs, memory chips, and the entire physical supply chain behind them, before they can even be introduced to the market. The data centres are, in effect, large factories without workers. And factories come with very different economics: big upfront costs, ongoing maintenance, replacement cycles, and sunk capital that cannot easily be walked back.

This is where the earnings-quality question enters the picture.

The issue is not that the earnings are fake. That is not the point. The point is that reported earnings now depend much more heavily on accounting assumptions related to depreciation. How long should a company assume that a cutting-edge AI chip or server remains economically useful? Four years? Five years? Six years? That may sound technical, but it is not trivial. A one-year change in useful-life assumptions can move reported earnings by billions of dollars when the sheer size of these data centres are considered.

When will the AI boom end? Nobody knows, and the market doesn’t want to think about it

This depreciation issue is no longer buried in the footnotes. It came up in the latest earnings calls last week. Alphabet, Amazon, Meta and Microsoft spent a combined $130-billion (with a “b”) on capex, or capital expenditures, in the first quarter alone. Their combined capex plans are tracking toward roughly $750-billion just this year. That is not normal corporate spending. That is major industrial build-out dressed up as a technology story.

Alphabet made clear that higher infrastructure spending is flowing through the income statement in the form of higher depreciation and data-centre operating costs. Meta said much the same thing.

Microsoft said roughly two-thirds of its capex is going into shorter-lived assets such as GPUs and CPUs. That is the key point. The building may last for decades. The chip inside the building may not.

And that is where the market may be too relaxed. A GPU can still turn on after five years. But if the next generation of chips is faster, cheaper and more energy-efficient, the old hardware may no longer be competitive for the highest-value AI workloads. So the relevant question is not whether the chip still works. The question is whether it still earns an adequate economic return, as new AI models and new chips are released almost each month.

Is our current form of politics capable of governing AI?

There is also a deeper strategic problem. These companies are spending enormous sums on a technology that may disrupt the very software and internet profit pools that made them so profitable in the first place. That makes the return-on-capital question far more complicated than it was in the old asset-light era.

And we have not even touched the Chinese competition angle. Nor do we know who ultimately wins the AI race. One month it is OpenAI. Then Anthropic. Then DeepSeek. Then Google. The pace of change is relentless, and yet the accounting assumptions often look far more stable than the technology cycle itself. That is why current tech valuations might be carrying a higher premium than the legacy assumptions imply.

So here is the bottom line. Big Tech is moving away from the old asset-light model and toward something that looks much more like industrial infrastructure. That does not make these companies bad businesses. But it does mean investors should be more careful with earnings quality, depreciation schedules, replacement costs and valuation multiples that are still anchored to the old world.

The investment conclusion is straightforward in our view. The cleanest part of the AI story is still upstream. The companies selling the chips, equipment, power systems, cooling technology and critical materials are already getting paid. The platforms still have to prove that this extraordinary spending will generate returns even as the competition and disruption rate intensify.

Mehmet Beceren is senior market strategist for Rosenberg Research