🧳 Dumb Money, Smarter Models and the $2 Trillion Briefcase

There’s a beautiful moment in Dumb and Dumber when a briefcase that was supposed to contain a fortune turns out to contain something much less liquid: a magnificent collection of IOUs.

AI infrastructure, circa 2026, occasionally gives me the same feeling.

NVIDIA invests in an AI company. The AI company rents NVIDIA GPUs through a hyperscaler. The hyperscaler orders more NVIDIA GPUs. NVIDIA reports booming demand. Investors see booming demand and provide more capital to the AI company.

Everybody has money. Everybody owes money. Everybody is buying something from everybody else.

The models really are getting better

The cleanest bubble argument would require the technology to have stopped improving. That argument is difficult to defend. Frontier systems continue to improve on difficult evaluations, and software agents can complete tasks that were previously out of reach.

But model quality is not the same thing as product quality. Routing, context management, rate limits, available compute, latency, tools and cost controls can make a product feel worse even while the underlying model improves.

The economics are the uncomfortable part

The frontier labs are raising extraordinary amounts of capital while building extraordinary infrastructure commitments. OpenAI and Anthropic provide different evidence for the bullish case, but neither removes the central question: how much of the infrastructure boom reflects profitable demand today, and how much reflects capital deployed in anticipation of profitable demand tomorrow?

NVIDIA, Microsoft, Amazon and the hyperscalers sit inside an increasingly interconnected system of customers, suppliers, investors and infrastructure commitments. That does not make the revenue fake. It does make the financing structure worth inspecting carefully.

The boring apocalypse

The most plausible failure mode may not be that AI is fake or that the technology disappears. It may be that the financial structure of the race proves unsustainable while the useful technology survives.

If the bubble bursts, GPUs can be written down, data centres can change hands, contracts can be renegotiated, training can slow and inference can become cheaper. The financial revolution could die while the infrastructure becomes more useful.

I don’t think AI is fake. I also don’t believe that “AI is transformational” automatically makes every valuation, data-centre project, financing structure and compute commitment rational.

Someone has to pay the bill.

This is the local Angelic Articles edition of the article published on Medium.