AI infrastructure spending has outrun its own revenue by $230 billion and the debt is now repricing everything it touches.

CoreWeave co-founders Michael Intrator, Brian Venturo, and Brannin McBee have sold $3.3 billion in shares since the March 2025 IPO, with $2.4 billion of that coming in 2026 alone. The stock is down more than 60% from its June 2025 peak. Venturo sold over 35% of his original stake. Intrator told analysts in August that conviction in the strategy had never been stronger. The bond market, priced at a 60%-plus default probability, disagrees.

If you bought CoreWeave shares near the June 2025 peak, you're sitting on a 60%-plus loss. The 2031 bonds now yield above 15%, up from 12% two months ago. That means every $1,000 in CoreWeave debt you hold has lost roughly $180-$200 in market value since August. The founders' pre-IPO tender sales in 2023-2024 added another $488 million on top of the $3.3 billion. Their total exit from paper to cash: north of $3.7 billion.

Law of Projection: what someone says about their conviction is often a precise mirror of what they're doing with their money. The founders filed every sale legally, under pre-scheduled 10b5-1 plans. The company called it routine financial planning. That's not a lie. It's also not the whole truth. What it does to your head is make you feel the situation is controlled, normal, disclosed. The thing to understand is that disclosure and alignment are different things.

Volta Infrastructure Holdings, an AI neocloud startup founded this year, needed $5 billion to build a data centre in Norway and buy 36,000 Nvidia GPUs. JPMorgan and Goldman arranged the loan but only underwrote $1.3 billion of it. The remaining $3.7 billion GPU loan was sold on a best-efforts basis, meaning the banks made no promise to absorb unsold portions. The first attempt failed to clear. The second attempt priced at yields above 11%, among the highest in the AI financing market. Anthropic has a six-year compute deal at the site.

The Volta loan cleared at 6.5 percentage points above the benchmark rate, at 95 cents on the dollar. That 5-cent discount is immediate loss for buyers. The all-in yield: above 11%. Compare that to US investment-grade bonds at roughly 5-6%. The extra yield compensates for the risk that Volta, founded in 2026 with $300 million in equity against $5 billion in debt, can't service what it owes. If Anthropic ever cancels that compute contract, the chips are the collateral. Nvidia GPUs depreciate fast.

Law of the Trap: the structure looks like opportunity because the yield is high. Double-digit returns on infrastructure debt backed by Anthropic contracts sounds like a safe bet. The trap is that 'best efforts' means the bank won't catch you if the market doesn't want it. The banks positioned themselves as arrangers, not guarantors. Private credit funds and retail investors in credit vehicles are the ones who absorb the unsold risk. The yield is the price of being the last buyer in a thinning market.

Goldman Sachs published research in September 2026 showing the five largest US hyperscalers need $300 billion in annual AI revenue to break even on their infrastructure spending. They're currently generating about $70 billion above pre-AI trend. The gap is $230 billion a year. Hyperscalers are on track to issue $250 billion in bonds in 2026 to bridge it. Goldman notes this spending is now large enough to push up the global cost of capital. The Bank of England flagged in July 2026 that a reassessment of AI earnings could tighten financial conditions for UK borrowers.

UK defined-benefit pension schemes held over £1 trillion in assets at September 2025, with AI companies now accounting for roughly half the S&P 500 by value. Passive pension funds tracking global indices are automatically increasing AI debt exposure as hyperscalers issue more bonds. Morgan Stanley estimates private credit markets could supply over half the $1.5 trillion needed for the data centre buildout to 2028. The debt ends up in pension portfolios and insurance products through securitised vehicles that most savers can't see. If hyperscaler leases cancel early, the losses transmit directly into retirement portfolios.

Law of Entropy: systems that expand through borrowed energy eventually demand repayment in full. The AI infrastructure boom was funded with cash first. Then free cash flow ran short and bonds took over. Now Goldman says the US dollar bond market alone can't absorb what's needed. The next stop is private credit, structured vehicles, and ultimately pension money. Each step away from the source is one step harder to trace, and one step harder to exit when the direction changes.