Gary Marcus just dropped two estimates of AI capital expenditure that should make everyone pause. We’re not talking about normal tech spending anymore. We’re talking about something closer to a infrastructure arms race, and it’s happening at exactly the moment when public sentiment on AI is turning sharply negative.

Here’s the thing that nobody wants to say out loud: the data center buildout we’re seeing right now might be the biggest bet in tech history, and there’s a real chance it doesn’t pay off.

The Numbers Are Absurd

The capex figures floating around aren’t just big. They’re historically unprecedented. And they’re coming at a weird time. Public opinion on AI has, according to Marcus’s analysis, “totally soured.” That’s not just one survey or one demographic. That’s a pattern.

So you have this contradiction. The money people are building like there’s infinite demand ahead. The actual people are increasingly skeptical about whether they want what’s being built.

Nvidia’s announcement that it’s partnering with data center developer Cloverleaf is just the latest example. Nvidia continues to make money hand over fist selling the pickaxes for this gold rush, so of course they’re investing in more mines. But pickaxe sellers get rich in both real gold rushes and fake ones. The question is which kind this is.

The Infrastructure Trap

There’s a classic mistake in tech: building massive infrastructure for demand that never materializes. Remember how many fiber optic cables got laid in the late 90s? Remember how many of those companies went bankrupt?

The current AI buildout has that same energy. Lots of smart people convinced that if you build it, they will come. Lots of capital flowing to infrastructure before anyone’s really figured out the sustainable business models.

And here’s what makes it weirder: the research coming out suggests that maybe the infrastructure isn’t even the bottleneck we think it is. Nvidia just published research showing that AI agents can perform well through fine-tuning and good harnesses, even when the underlying model isn’t great at the task.

Read that again. The harness matters more than the model.

If that’s true, then what exactly are we building all these data centers for? To train slightly better models that might not actually matter as much as the software wrapped around them?

The Uncomfortable Questions

I’m not saying AI is fake or that LLMs aren’t useful. I use them every day. But I am saying that the current spending trajectory doesn’t match the current revenue trajectory, and it doesn’t match public sentiment, and it increasingly doesn’t match what the research is telling us about where the real value lives.

When you see companies cutting staff on high-profile products like Vision Pro while simultaneously reading about billion-dollar data center partnerships, you start to wonder who’s paying attention to the fundamentals.

Apple just laid off over 200 people, including largely shutting down a Vision Pro gaming team. That’s a company reading the room on a product that isn’t getting traction. Meanwhile, the AI data center spending continues to accelerate.

One of these groups is looking at actual demand. The other is looking at projected demand. History suggests you should bet on the people responding to reality.

What Happens Next

There are two ways this plays out. Either the AI applications break through in a major way and justify all this infrastructure spending, or we’re looking at a correction that makes the dot-com bust look quaint.

I don’t know which it will be. But I know that when public opinion is souring, when the research suggests infrastructure might not be the constraint, and when the spending shows no signs of slowing down, someone is going to end up holding a very expensive bag.

The smart money isn’t just building data centers. It’s figuring out what people actually want to use AI for and building that. The harness research from Nvidia should be a wake-up call. Maybe we don’t need another generation of slightly better models trained on slightly bigger clusters. Maybe we need better software, better interfaces, better actual products.

The cathedrals are impressive. But if nobody’s coming to services, they’re just very expensive buildings in the desert.