The AI Economy Is Cracking. Here's What Nobody's Saying Out Loud.

---

Five people who literally touch every layer of the AI supply chain—from chips to data centers to model training—sat down at the Milken Global Conference last week and basically admitted something the industry has been avoiding: the system is breaking.

Not slowly. Not metaphorically. The actual wheels are coming off.

This matters because these aren’t doomers. These are the architects. The people who built this thing. And they’re talking chip shortages, orbital data centers, and the possibility that “the whole architecture that undergirds the tech is wrong.” That’s not pessimism. That’s diagnosis.

But here’s what’s wild: while they’re saying this in Beverly Hills, the supply chain is simultaneously collapsing in real time. Ubuntu’s infrastructure went down for more than a day. A widely used disk app called Daemon Tools got backdoored in a monthlong supply-chain attack that nobody caught until it was too late. Reddit is now actively blocking mobile web access to force users into their app. And GameStop just offered $56 billion for eBay—a company that barely moves the needle for them, which tells you they’re swinging for the fences because their core business is done.

These aren’t unrelated stories. They’re the same story told five different ways.

Photo by Chris F / Pexels

The Cracks Are Structural

Let me be direct: the AI economy as currently architected is unsustainable.

The math doesn’t work. You need massive compute to train models. That compute costs billions. The models generate useful outputs, but not nearly valuable enough to justify the infrastructure spend. So you need either (a) venture capital, (b) sovereign wealth, or (c) a completely different business model. We’re running on fumes of (a), starting to see (b), and nobody’s found (c) yet.

Barry Diller—who’s been in tech and media long enough to smell when something’s fundamentally wrong—said something that stuck with me: “Trust is irrelevant as AGI nears.” He was defending Sam Altman, but his actual point was darker. In a system approaching AGI, human judgment stops mattering. The machine becomes the variable. And if you can’t predict what AGI will do, then the entire financial and governance structure built around “trusting the CEO” collapses. You’re betting on a black box.

That’s not a governance problem. That’s an architecture problem.

Now look at the supply chain failures. Daemon Tools. Ubuntu down for 24+ hours. These are mission-critical tools that billions of people depend on. They went down because security at scale is exponentially harder than anyone admits. One compromised supplier, one overlooked vulnerability, one monthlong gap before detection—and suddenly the whole infrastructure is suspect.

And we’re building AI on top of this same infrastructure.

The Deals Are Dying

Snap just killed its $400 million deal with Perplexity. They announced it last November with genuine excitement—Perplexity’s AI search engine integrated directly into Snapchat, supposedly the future. Seven months later: “amicably ended.” That’s corporate-speak for “we looked at the actual metrics and realized this doesn’t move the needle.”

This is what happens when you tie your business model to unproven AI applications. Snap bet that AI search was a Snapchat problem worth solving. Turns out it wasn’t. Or more likely: Perplexity needed the distribution more than Snap needed the technology, and once that asymmetry became obvious, the deal became ballast.

GameStop offering $56 billion for eBay is the same signal, just louder. GameStop is a company that survives on nostalgia and Reddit hype. They don’t have $56 billion in cash. They don’t have a credible path to integrate eBay’s logistics. This isn’t a real offer. It’s a desperate flag. It says: “Our current business is dead, we’re throwing everything at the wall.”

These deals dying or imploding—that’s not noise. That’s the market signaling that the easy money is gone.

Photo by UMA media / Pexels

Where’s the Real Money?

Here’s what I find interesting: xAI might actually be telling the truth about what matters right now.

xAI’s stated mission is AI. But as one headline asks—“Is xAI a neocloud now?”—the real business might be data centers. Not models. Not inference. Infrastructure. Building the pipes that everyone else will depend on. That’s actually harder to disrupt, more defensible, and generates recurring revenue. It’s boring. It’s not as sexy as “we’re training AGI.” But it’s real.

Compare that to Pronto, an Indian startup that just got backed by Lachy Groom after a 20-minute pitch. Pronto’s scaling to 26,000 daily bookings in a market heading toward $18 billion. This is a logistics play. Not a model. Not a chip. A real service that solves a real problem at scale. It’s not going to change the world. It’s going to make money.

I think the AI economy is bifurcating right now. One branch is the hype cycle—models, APIs, “foundation” everything. That’s collapsing because the economics don’t work and the applications are still vaporware. The other branch is infrastructure and services—data centers, logistics, supply chain optimization. That’s where the actual value is being created.

The first branch gets the press. The second branch gets the revenue.

The Real Problem

There’s something deeper here that the five architects at Milken were hinting at but not saying explicitly: the whole stack might be built on the wrong foundation.

We’ve architected AI around the assumption that scale = capability. Bigger models. More compute. More data. This works up to a point, and then it stops working. We’re hitting that point now. The models aren’t getting proportionally smarter. The inference costs aren’t coming down. The real-world applications aren’t appearing.

So you’ve got three options:

One: Double down. Build even bigger. Spend even more. Hope the exponential curve resumes.

Two: Pivot. Stop chasing scale. Build smaller, more specialized models for specific problems. Accept that general AI might not be the right architecture.

Three: Collapse. The whole thing turns out to be a dead end, and we spend a decade building something different.

Right now we’re oscillating between one and two. Snap killed the Perplexity deal because they were betting on one. GameStop is swinging at eBay because they’re hedging against two. Ubuntu went down and Daemon Tools got backdoored because we’ve been so focused on one that we ignored the supply chain risks.

I genuinely don’t know which outcome is most likely. That’s the uncomfortable part.

What I’m Watching

1. Data center capacity announcements. If xAI’s real business is infrastructure, watch who’s actually building compute centers vs. who’s just talking about it. Announcements from Crusoe, Lambda, CoreWeave on actual megawatt commitments by Q2 2025. The real money follows the real pipes.

2. The next major supply chain incident. Daemon Tools taught us that security at scale fails invisibly for months. Watch for the next compromised tool everyone depends on. When it happens (not if), the supply chain gets rearchitected overnight. That’s when you see real innovation.

3. Enterprise AI adoption rates. All the hype talks about consumer AI. But enterprise is where money actually moves. Track: Are Fortune 500 companies building internal AI teams or licensing from OpenAI/Anthropic? If they’re building in-house by late 2025, that means they don’t trust the external APIs. That’s a massive signal that the third-party model is failing.

4. Venture capital allocation shift. Watch where Sequoia, a16z, and Benchmark are deploying capital over the next two quarters. If it’s still 70%+ to foundation models and applications, the industry hasn’t woken up yet. If it drops below 50%, they’re hedging. Below 30%, they’ve lost faith.

The AI economy isn’t dead. But it’s reorganizing, and the people at Milken last week were the first ones honest enough to say it out loud.