Tech • AI • Robotics • Game
The revolt against AI has begun… and it is enriching those it targeted
7/10TL;DR
The recent selloff in AI infrastructure stocks reflects a funding squeeze on debt-heavy data-center subcontractors rather than weakening demand for computing.
KEY POINTS
Two-tier AI infrastructure market
The market is splitting between hyperscalers such as Microsoft, Amazon, Meta and Alphabet, which can fund much of their infrastructure from operating cash flow, and smaller “neo-clouds” such as CoreWeave, Nebius and IREN, which rely heavily on debt and equity issuance. Both groups serve the same AI boom, but the financial risk sits in different places.
Stocks diverged sharply
Between October 8 and November 6, shares of CoreWeave, Nebius and IREN fell by roughly 40% to 50%, while Microsoft and Amazon were broadly flat and Meta and Alphabet declined far less. That gap suggests investors are punishing financing structures, not the underlying demand for AI services.
Bond markets flashed distress first
Pressure first appeared in credit. A CoreWeave bond issued in June at a yield a little above 10% was yielding about 15% by October 13, a classic sign that the bond price had dropped and buyers now saw elevated default risk. For leveraged infrastructure builders, losing access to refinancing can be existential.
Heavy leverage is now the core vulnerability
By the end of June, CoreWeave reportedly carried about $35 billion in debt, with some borrowing costs reaching 15%. That creates fixed repayment obligations regardless of whether a facility is finished, energized and generating rent. Oracle, despite its scale, also faced pressure after funding new data-center expansion with debt, prompting a credit downgrade and a need to raise nearly $20 billion more.
Political resistance is slowing projects
In 10 U.S. states, localities held 41 votes related to data centers, and 40 of them favored slowing or restricting projects. The main tools are administrative: delayed permits, local referendums and extra guarantees. These barriers raise the cost of building and, more importantly for leveraged firms, extend the time before rental revenue begins.
Power costs are driving the backlash
Opposition is tied to electricity bills and grid stress. In the largest U.S. power market, a grid monitor estimated $6.2 billion in added costs in the December 2025 auction tied to planned data centers, including projects not yet built. Yet by September 21, about 42% of that announced demand had already been canceled or delayed, leaving households effectively paying in advance for capacity that may never arrive on time.
Hyperscalers gain from tighter barriers
What hurts subcontractors can strengthen incumbents. If permits become scarce and power hookups harder to secure, infrastructure that is already built and authorized becomes more valuable. That helps large platforms with existing footprints and deep cash reserves, turning local resistance into a form of protection for dominant players rather than a broad brake on AI adoption.
Contracts push risk downward
Large customers often sign long leases for land and buildings, giving financiers comfort based on the credit quality of names like Microsoft or Meta. But neo-cloud lenders are frequently secured by the value of chips and short-lived hardware. Those assets depreciate quickly, often over 4 to 6 years, and possibly faster in practice, making the collateral much weaker in a downturn.
Why this is unlike older infrastructure busts
During the telecom crash, bankrupt networks still left durable fiber in the ground. In AI infrastructure, a failed operator may be forced to dump aging GPUs, which can drag down collateral values for everyone else using similar machines. The risk is a self-reinforcing cycle in which defaults lower asset prices, making refinancing even harder across the sector.
Demand remains high, but consolidation is likely
Major platforms still report that customer demand exceeds available computing supply. Microsoft said demand was outrunning capacity and pointed to delays in subcontracted infrastructure delivery. The likely outcome is not an end to the AI buildout, but a consolidation in which a smaller set of well-capitalized winners controls more of the market and pricing power.
CONCLUSION
The current upheaval in AI infrastructure is less a collapse in demand than a repricing of who can afford to build at scale. As financing tightens and local opposition slows projects, cash-rich hyperscalers are positioned to gain while debt-dependent subcontractors face a harsher survival test.
Ask a question