The AI Bubble: Why the System is More Fragile Than You Think

This conversation with Paul Kedrosky explains that the current AI boom is not just a technological cycle. Instead, it is a systemic event tied to credit, real estate, and policy. The hidden risk is an over-determined system where failure is likely, regardless of how good the technology is. For investors and operators, the advantage lies in recognizing that current market signals, such as GPU scarcity and GDP contributions, are distorted by hoarding and debt-fueled financing. Those who understand these feedback loops can distinguish between genuine utility and the deflationary pressures that will soon reshape the sector. Ignoring these downstream effects creates a blind spot that will become costly when the cycle of external financing hits the reality of diminishing returns.

The Illusion of Scarcity and the Reality of Hoarding

The idea that GPUs are in short supply and therefore perpetually valuable is being undermined by the system's own logic. Kedrosky points out that utilization rates for high-end GPUs are low, often hovering between 35% and 40%. This paradox exists because of widespread double and triple ordering driven by fear. Companies are hoarding hardware to avoid being caught short during a sudden market shift, creating a massive, latent supply overhang.

"There's colossal amounts of hoarding going on, double and triple ordering going on because everyone's terrified that if something happens, some inflections, some parabolic moment, even more parabolic than what we've seen that they'll be caught short."

-- Paul Kedrosky

When this hoarded capacity eventually enters the market, it will act as a deflationary shock. This could hurt the margins of frontier companies that are already struggling to maintain growth as token prices fall.

The Debt-Financed Flywheel

A shift has occurred in how data centers are funded. Initially, these projects were powered by internal cash flows, which signaled operational health. However, as of mid-2026, most financing is external, sourced from private credit, asset-backed securities, and sovereign flows. This creates a dangerous disconnect: lenders are funding data centers based on the creditworthiness of the hyperscalers backing the leases, rather than the economic utility of what happens inside the centers.

This financialization creates a feedback loop where capital continues to pour into construction even as the underlying application layer faces diminishing returns. Because the debt is now tied to the investment-grade status of the parent companies, the system has effectively turned technology firms into utilities, but with higher valuation multiples and ongoing, non-discretionary capital expenditure requirements.

The Over-Determined Nature of Failure

Kedrosky argues that at current valuation levels, failure is over-determined. Much like the dot-com era or the decline of Nike from its peak, there are dozens of low-probability events that could trigger a collapse. When a system is priced for perfection, the specific cause of the eventual downturn, whether it be regulatory shifts, supply gluts, or model convergence, matters less than the fact that the system has run out of room for error.

"At high valuations, failure is over-determined in a statistical sense, meaning that there are so many ways to fail all of which are low likelihood... that when you combine them all and turn it around and say, given 20 different ways this could fail... there's greater than a 60% chance of failure."

-- Paul Kedrosky

The emergence of vibe chipping, which uses AI to accelerate chip design, will likely speed up this process by flooding the market with new designs. This will erode the competitive moats that incumbents currently rely on to justify their premiums.

Key Action Items

• Audit your exposure to AI-as-Utility: Re-evaluate tech holdings that are heavily tied to data center CapEx. Over the next 12 to 18 months, expect a rerating as the market begins to view these companies less like high-growth software firms and more like capital-intensive utilities.
• Monitor the 2028 supply tsunami: Watch for the influx of capacity from Taiwanese and Chinese chip manufacturers. This will likely be the point where the current scarcity narrative flips to an oversupply crisis.
• Ignore the harness noise: When evaluating model performance, look past the harnesses, such as wrappers like Claude Code or Decodex, that mask the reality of model convergence. If you are building, focus on proprietary data and workflow integration, as the models themselves are rapidly becoming commodities.
• Re-evaluate GDP causality: Do not rely on current GDP growth figures as proof of policy success. Recognize that much of the recent growth is a result of data center investment, not structural economic improvement. This insight is necessary for long-term macro positioning.
• Prepare for vibe chipping disruption: If you are in the hardware space, anticipate a collapse in the barriers to entry for ASIC design. The tribal knowledge that protected incumbents is being digitized and automated, which will lead to a surge in niche, low-power inference chips.