The AI Bubble Is a Lie. The Real Threat Is Compute Scarcity.

You’ve heard the whispers. You’ve seen the stock charts. Every smart investor you know is repeating the same mantra: The AI bubble is about to pop. They pull out their fancy valuation models, point to the dot-com crash, the fiber optic glut, the railroad mania, and confidently declare that AI infrastructure is massively overbuilt.

But what if you’re being conned by a comfortable lie? What if the real danger isn’t that we’ve built too much AI infrastructure—but that we’ve built almost nothing at all?

Recently, David George and Gavin Baker from a16z sat down to dissect the current state of AI. They didn’t use macroeconomic jargon. They used hard numbers. And the picture they painted is completely upside-down from what the mainstream media is selling you.

Pessimism isn’t always wisdom. Sometimes, it’s just cognitive laziness—a way to avoid facing a reality that breaks your entire framework.

Over the summer, Gavin Baker met with countless founders, investors, and executives. He asked them all the same question: Can you give me just one quantifiable data point from your business that is getting worse? Just one. From July to August, not a single person could answer him. OpenAI is accelerating. Open-source models are accelerating. The entire industry is shifting into a higher gear.

Yet, AI stocks on the secondary market got crushed. You might drown in a river with an average depth of two feet. The market is currently pricing a “bubble narrative,” systematically misreading reality. It’s a classic split: fundamentals are accelerating, and prices are falling. This is the exact phase where the most critical wealth is transferred to those who actually understand the data.

Let’s talk about the data centers. The mainstream narrative says we are overbuilding them. The hard math says otherwise. Building a Nvidia data center today has a payback period of nine to ten months. You read that right. Top-tier asset managers like Blackstone, KKR, and Apollo are throwing cheap debt at these projects because the residual value of GPUs keeps going up, and Nvidia even provides residual value guarantees. This isn’t a heavy, risky capital expenditure. It’s a cash-printing machine wearing a heavy-asset disguise.

But here is the mind-bending part: the actual global market of heavy, paying AI users is less than 10 million people. Even in the most aggressive AI-native companies, token spend is barely 10% of human capital costs. In traditional companies, it’s 1%.

We are at less than one percent of the AI adoption curve, and we’re already panicking about a surplus. We haven’t even started the race, and we’re worried about running out of track.

The total addressable market isn’t 10 million tech bros in San Francisco. It’s 1.5 billion knowledge workers globally. When adoption scales from a few million early adopters to hundreds of millions of daily users, demand won’t double or tenfold. It will experience an exponential, violent jump. And this doesn’t even require a technological breakthrough. It just requires time and habit formation. A 23-year-old today uses AI like oxygen; the next generation will consume tokens at an order of magnitude higher than we do.

So, what happens when severe supply shortages meet this incoming tsunami of demand? The mainstream assumes AI costs will forever trend downward, just like chips and storage did. But that historical precedent relies on supply keeping pace with demand.

You assume AI will get cheaper forever. But in a world of severe compute scarcity, tokens won’t just get expensive—they’ll become a luxury good.

If supply is systematically suppressed, token prices won’t drop. They will spike. And when compute becomes scarce and expensive, who gets squeezed out? Small businesses. Independent developers. The middle class. Large corporations and the ultra-wealthy will hoard the available compute. This creates a terrifying new dynamic: compute inequality.

And here is the darkest twist of all. Who is helping create this digital serfdom? The people blocking data center construction under the banner of protecting their local communities. They think they are saving their neighborhoods, but they are actually building the walls of a new resource monopoly.

The people blocking data centers to protect their towns aren’t saving their communities. They are accidentally building the walls of a new digital serfdom.

They don’t realize that data centers are the best thing to happen to the working class in decades. Look at Loudoun County, Virginia—the highest-income county in America, powered entirely by data center density. These facilities bring ten times the tax revenue, create high-paying local jobs for electricians and HVAC technicians that can’t be outsourced, and reindustrialize forgotten towns. But the tech industry is terrible at telling this story. They need to stop talking in abstract policy and start pointing to the local bakeries and plumbers whose lives are changed by compute infrastructure.

And if you think the ground is the only place for compute, you’re already behind. SpaceX and Nvidia are designing orbital data centers to launch by 2027. You might think it’s physically impossible, but 10,000 elite engineers have spent thousands of hours solving the physics. In space, power is solar, cooling is the infinite shadow of the void, and land is free. Once Starship slashes launch costs, the economics of orbital compute flip instantly.

Ultimately, the battle isn’t just about who has the smartest model. It’s about who becomes the abstraction layer—the default router that enterprises trust to manage their data, route tasks, and deliver intelligence. Whoever wins that trust wins a historic ticket.

But hovering above all of this is one undeniable truth. Nvidia is the central bank of the AI economy. Jensen Huang locked up global wafer, DRAM, and laser capacity years ago. He supports open-source not out of charity, but because open-source lowers token costs, which drives consumption, which requires more GPUs. His incentives are perfectly aligned with total compute democratization. You don’t fight the Fed; you find a niche and plug into his ecosystem.

The market is currently pricing a bursting bubble. It is systematically misreading a famine. The real risk isn’t that AI fails to deliver. The real risk is that it succeeds, and we don’t have the compute to share it.