The AI Bubble Is a Risk Management Problem Before It Is a Market One

Here is what keeps me up at night, and I say this as someone who leans on AI every day and believes it earns its keep. 

The four largest hyperscalers plan to spend roughly 725 billion dollars on AI infrastructure in 2026, up 77 percent from last year, with capital intensity now running 45 to 57 percent of revenue, a level that used to be unthinkable, and the debt issued to fund it could top 1.5 trillion dollars over the next few years. 

Then let’s look at how the revenue gets manufactured. Nvidia committed up to 100 billion dollars to OpenAI, AMD struck deals near 200 billion with equity warrants attached, and Oracle pledged 300 billion in cloud capacity, all to firms that turn around and buy the vendors right back. Roughly 800 billion dollars of this circular financing now props up the boom while OpenAI is still projected to lose 14 billion this year. 

We watched this exact movie during the dot com fiber glut. Money moving in a circle is not demand. Now the part nobody wants to reconcile. Oracle cut about 30,000 jobs, eighteen percent of its workforce, while free cash flow swung to negative 24.7 billion and debt climbed past 134 billion, and leadership framed it as a generational reallocation of capital from people to compute. Meta cut 8,000, Microsoft 9,000, and Goldman estimates AI is removing 16,000 American jobs a month. Yet that same stretch Sam Altman and Dario Amodei both walked back their white collar bloodbath warnings, conveniently as OpenAI filed IPO paperwork. 

You cannot fire to fund the buildout and tell workers the buildout will not touch them. My guesstimate on timing is a revenue reckoning through late 2026 into 2027, a sharp repricing, and writedowns where the loops unwind, not a clean collapse but there will be a great deal of collateral damage. 

Here is the deeper tell. Generative AI is firms refusing to crawl before they walk, let alone run. Everyone wants agentic AI now, autonomous systems acting on their own, but most companies have neither the in house expertise nor the data discipline to deploy agents without burning their own house down. You do not give a new intern root access on day one. Reliable agentic AI is at least one full AI generation out, which in my book is two to five years. 

And here is the upside, because there is one. The dot com bust left behind the fiber that built the modern internet. This cycle is laying down data centers, power, and chips that will outlast whoever overbuilt them, the productivity gains are real for anyone who learns the tools. 

But as of today the tech bros are writing checks with infrastructure that their revenue cannot cash. This will end badly so make sure to limit your firms exposure. 

Reach out to discuss.