There are moments in this industry when someone with real authority stands up, clears their throat, and says the quiet part out loud. Last week, IBM CEO Arvind Krishna did exactly that. In a world where every hyperscaler, startup, and sovereign cloud ministry is tripping over itself to announce new AI datacenters, Krishna basically walked onto the parade route and pointed out that the emperor might be… well… a bit underdressed.

His message? The math doesn’t work.

Now, if you’ve been reading my stuff for the past few months, you know I’ve been shouting into the void about this very thing. The capex required to build the AI future everyone is promising simply doesn’t align with the economics that would be needed to sustain it. But Krishna isn’t just humming this tune—he’s bringing a brass section, a marching band, and the sheet music.

Let’s start with the big number: $8 trillion.

According to Krishna, that’s the real ballpark for what it would take to build out the projected AI datacenter capacity the industry is currently salivating over. Eight trillion. With a “T.” Even by today’s monopoly-money standards, that’s not pocket change.

And that’s just the beginning. He goes on to note a few inconvenient truths. First, he pegs the chances of achieving AGI anytime soon at something like 1%. That’s not me rounding up or down—that’s his number. If you’re making bets on trillion-dollar infrastructure with a 1% shot at delivering the return story being sold, well… you might want to check your risk profile. Or get a new hobby.

Second, these enormous facilities—the very temples of the AI religion—are on a five-year refit cycle. Every five years. Think about that. We’re not talking about repainting your garage or swapping a router. We’re talking about ripping and replacing some of the most complex, power-hungry, GPU-thirsty systems humanity has ever built.

And then comes the third punch: Krishna says outright that datacenter economics cannot deliver enough revenue to service the debt required to build them—much less generate the kind of profits Wall Street expects. This is the part where his résumé matters. Krishna isn’t a guy yelling in a parking lot. He’s the CEO of IBM, former head of cloud and cognitive software, a technologist who actually understands the plumbing behind this stuff.

In other words:

If Arvind Krishna says the math doesn’t work… It’s time to run the numbers again.

To back this up, you don’t need to take my word for it. I’ll point you to reporting that lays out the details far better than a single column can:

• Data Centre Magazine’s coverage of Krishna questioning the trillions being poured into AI buildouts.
• TechRadar Pro’s piece titled—rather bluntly—that current AI datacenter trends are unsustainable.
• EE News Europe summarizes his skepticism around AGI and the trillion-dollar capex journey ahead.
• Futurism highlights his view that the economics simply don’t reconcile. https://datacentremagazine.com/news/ibm-ceo-questions-trillions-in-data-centre-spending

That’s not fringe commentary. That’s not a contrarian blogger on LinkedIn firing off a spicy Monday take. That’s a growing chorus of industry analysts, journalists, and executives saying: “Maybe we should check the receipt here.”

What If He’s Wrong?

Now, before we all start panic-selling GPU futures, let’s be fair and ask the question any reasonable engineer would: What if Krishna is wrong?

If you disagree with him, what exactly is the flaw you’re calling out?

Does he underestimate the demand curve?

Possible, but unlikely—this is a man who has been analyzing compute economics for decades.

Is he ignoring future technical breakthroughs?

Maybe, but that would require breakthroughs of a sort that change the laws of thermodynamics, fabrication costs, energy density, and interconnect speeds. Those don’t exactly show up overnight.

Is he too conservative about AGI timelines?

Perhaps—but even if you’re wildly optimistic and say AGI is 20% likely, that’s still not a business case for raising trillions in debt at today’s interest rates to build out an infrastructure layer that needs replacing every presidential cycle.

Is a new revenue model coming that we can’t see yet?

Sure. Maybe. Metaverse credits? AI-assisted astrology subscriptions? But if you’re hanging your trillions on undiscovered revenue models… well… Godspeed.

So again: If you can’t poke real holes in Krishna’s assumptions, what are we actually doing here?

We’ve Seen This Before — But This One Is Louder

The closest historical analog is the late 1990s/early 2000s fiber and datacenter buildout. I remember it well. Back then, analysts claimed we were building way more capacity than anyone could ever use. Some even joked we could run the entire internet over a single strand of dark fiber in Kansas.

And for a while, they seemed right. Companies collapsed. Billions evaporated. Capacity sat idle.

But eventually—eventually—the usage caught up. Video streaming, mobile computing, SaaS, cloud… all stood on that “overbuilt” foundation.

The difference?

That bet, enormous as it felt at the time, looks quaint compared to this AI datacenter boom. Back then, we were dealing in billions. Today’s planning decks start with trillions and end with terawatts. And unlike fiber, which you bury once and light up for decades, these AI datacenters are in a constant state of depreciation and replacement.

This isn’t “build once, harvest forever.”

It’s “build, pay, rip out, rebuild, pay again.”

Shimmy’s Take: Gravity Always Wins

Here’s where I land on all this:

Economics always wins. In fact, it’s undefeated.

You can hype your way through an earnings call. You can announce a new GPU cluster every three weeks. You can show shiny renderings of future AI campuses that look like Apple Stores crossed with NASA habitats.

But eventually the bill comes due.

And it’s a really, really big bill.

Either we will:

1. See meaningful technical innovation—something that dramatically lowers the cost of compute, energy, cooling, or networking;
2. Slow-walk the trillion-dollar AI buildout as the market digests reality;
3. Or some hybrid scenario where the industry quietly turns the dial back while inventors and researchers race to make the numbers less terrifying.

But it’s going to be one of those. Because the alternative—charging headfirst into an $8 trillion capex wall with a 1% AGI lotto ticket—doesn’t pencil out, no matter how many PowerPoints you throw at it.

And here’s the part everyone in our industry should take seriously:

When the CEO of IBM—one of the few people on the planet with an unobstructed view of the technological, economic, and operational realities of global compute—steps forward and says, “Folks, this isn’t sustainable,” you don’t shrug.

You listen.

You think.

You run your own numbers.

And maybe you stop assuming the future is guaranteed just because NVIDIA’s stock chart looks like a SpaceX launch trajectory.

We’re in extraordinary times. AI is real. The breakthroughs are astonishing. The potential is enormous.

But potential doesn’t pay interest.

Math does.

And right now, as Krishna bluntly pointed out, the math doesn’t work.

If it ever does, it’ll be because we changed it—through innovation, discipline, or cold, economic necessity. Until then, keep your eyes open. Because when the hype clears, the gravity of real-world economics will be waiting.