Traffic equated to revenue. Adding .com to the end of something made investors throw money at aspiring entrepreneurs. Startups with no business model bought Super Bowl ads. People became paper millionaires overnight.

On a book tour in ’99, I’d drive down 101 past office buildings wrapped in billboards for AltaVista, Excite, and other names destined for extinction. By 2001, those buildings stood empty. I went to a launch party where a startup dropped what must have been half a million dollars just to announce they were going to start charging for something they’d been giving away for free. The room was filled with VCs — and no one blinked.

By August 2000, the party was over. From the MoneyHunt studio, I had the perfect view — just uptown from the collapse.

And now, 25 years later, we’re back.

The New .com: Intelligence for Rent

“AI-powered” is today’s “.com.”

Only this time, many startups aren’t even pretending to own the tech they’re built on. Most so-called “AI products” are just wrappers—UIs stapled to OpenAI’s API. No IP. No infrastructure. Just a prompt pipeline and a Stripe integration.

Upload a transcript. Prompt it. Parse the response. Drop it into formatted boxes.

They’re charging $60/month for what costs $4 to do manually.

It’s not a product. It’s a disguise.

Fragile Layers, False Leverage

Wrappers rely on OpenAI. OpenAI relies on Microsoft. Microsoft needs NVIDIA. NVIDIA owns the chips that power it all.

No one’s in control. Everyone’s exposed.

And yet, each party pretends they hold the leverage.

But here’s the real problem: OpenAI doesn’t own the user—wrappers do. And wrappers are brittle. They’re margin-negative, burn-heavy, and one pivot away from irrelevance.

If they go down, OpenAI’s growth engine stalls. That’s the loop: OpenAI has the model, wrappers have distribution, but neither is insulated. And the whole stack is built on chips from one company in one geography.

The Silent Risk: NVIDIA

NVIDIA isn’t just a GPU maker. It’s the silent kingmaker.

GPT-4? Trained on NVIDIA chips. OpenAI’s inference engine? NVIDIA clusters inside Azure. Most AI you use today touches NVIDIA in some way.

If their supply chain cracks—whether by regulation, war, or export ban—AI progress halts. That’s not a doomsday scenario. It’s already happening.

Who Survives?

Not the prompt wrappers. Not the pitch decks that look good on Product Hunt. Not the startups charging $100/month for markup on a free API.

The ones who survive are the ones who build infrastructure:

• Real memory layers
• Workflow engines
• Durable moats

Survivors own experience, not just endpoints. They build rails, not rides.

Ask These Questions

Want to know if you’re building something that lasts? Ask yourself:

• Would anything break if you vanished?
• Would your customers rebuild you?
• Do you own your stack—or just lease the smarts?

If the answer’s no, you’re not infrastructure. You’re noise.

The real builders don’t just chase the wave. They shape it. They don’t just appear in the story. They become the setting.

Be the thing they can’t delete.

About the Author

Miles Spencer is a multi-exit founder, investor, and storyteller. His work at the intersection of memory and technology is rooted in personal experience and a deep belief in legacy.

FAQ

Q: What is an LLM wrapper?
A: It’s a UI or product built on top of a large language model like GPT-4, without owning the model itself.

Q: Why do so many AI startups look the same?
A: Because many are using the same underlying model, with only slight prompt and design tweaks differentiating them.

Q: What makes infrastructure companies different?
A: They provide foundational layers others depend on — systems that would break if they disappeared.