Here's How The AI Bubble Bursts — With Paul Kedrosky

Summary

Paul Kedrosky makes the case that today’s AI infrastructure surge is unprecedented when normalized against GDP, fixed investment, bond markets, and the speed of deployment. He argues data centers should be viewed as capital-intensive factories with ongoing replacement cycles, not durable real estate-like assets, and says the economics are being squeezed by rapidly deflating token prices and short GPU lifespans. Kedrosky is skeptical that AGI rhetoric justifies unlimited spending, calling it a ‘God of the gaps’ argument, and he thinks model convergence will push competition toward price and marketing rather than durable product differentiation. He also warns that the unwind could spread through debt markets, index funds, insurers, and banks, especially if higher rates, export controls, or public-market scrutiny slow the funding machine. In the end, he sees AI as potentially world-changing technology that still may produce a disastrous investment cycle.

Notes

• The AI capex wave now exceeds prior Western infrastructure booms on multiple normalized measures, including GDP share, non-residential fixed investment, GDP-growth contribution, and bond-market presence.
• The AI buildout is happening much faster than electrification, railroads, interstates, or fiber, which reduces the time investors have to pause and reassess returns.
• Alex frames the discussion as a question of how an AI bubble could burst even if the technology keeps improving.
• Investors should underwrite AI data centers from the capital-provider side, comparing returns to alternative project-finance opportunities rather than using simple spend-to-revenue arithmetic.
• The relevant benchmark is cap rate versus comparable commercial real estate projects, not a naive ‘put in a trillion, get out a hundred trillion’ framing.
• Data centers have continuing capital requirements and a duration mismatch: the asset looks long-duration, but the underlying hardware must be upgraded or replaced frequently.
• Some GPUs in modern data centers fail on an 18-month cycle, and upgrade churn also comes from new GPU generations.
• The revenue widget for AI infrastructure is tokens, and Kedrosky says token prices have been falling 70% to 80% year over year on a constant-performance basis for at least four years.
• That combination—upfront capex, ongoing capex, and hyper-deflationary token pricing—makes the commercial-real-estate analogy fragile for investors.
• GPU useful life depends on how the chip was used early on; training-heavy usage is much harsher than inference-only workloads.
• Inference-only chips should last longer than training chips, and the market may shift toward inference-specific ASICs instead of GPUs.
• Model quality is converging, which lowers switching costs and makes it easier for customers to change providers.
• As model differentiation narrows, competition shifts toward marketing spend and price, which is bad for returns.
• Public-market investors are likely to push frontier model companies to move up the stack once IPO scrutiny arrives.
• Kedrosky argues that SaaS buyers purchase software to solve a problem, not because vendors are inherently defensible or beloved.
• The AGI-as-call-option argument is, in Kedrosky’s view, an undiscountable bet that should not justify paying anything.
• He rejects that logic as a ‘God of the gaps’ argument: when the case is weak, people insert AGI to justify unlimited spending.
• Large investors often sound bullish to LPs while privately comparing AI projects to other infrastructure-like assets.
• AI is being pitched to local officials as ‘AGI factories’ in exchange for tax, water, power, and real-estate concessions.
• Major capital allocators keep participating while the cycle is hot, because they fear missing out and face LP pressure.
• For very large pools of capital, check size can matter more than project quality; sovereign wealth funds may seek $100 billion-sized allocations.
• Kedrosky says the AI bubble sits at the intersection of technology, real estate, loose credit, government policy, and geopolitics.
• The national-security framing around winning against China helps sustain the spending cycle.
• More than half of data-center capital now comes from external financing, so funding costs matter a lot.
• A move from mid-6% hurdle rates to high single digits would sharply weaken AI data-center economics if token prices keep falling.
• Export controls could balkanize the market, shrinking the addressable opportunity set.
• A government co-investment stake would change valuation math and the multiple an investor is willing to pay.
• Much of the recent AI progress has come from harnesses and post-training layers rather than bigger pre-training runs.
• That puts pressure on the case for billion-dollar training runs if RLHF, quantization, and similar techniques drive most gains.
• The unwind could propagate through S&P 500 index funds because hyperscaler and AI names are such a large share of the index.
• Kedrosky thinks the spillover would resemble the 2007-2009 financial crisis, but with AI assets instead of U.S. real estate.
• He says the AI race is uniquely crowded: too many well-funded participants are starting from the same line.
• Kedrosky says he avoids new passive index exposure because he is already indirectly overexposed to AI through prior commitments.
• China is also overbuilding data centers because provincial leaders compete to impress the central government.
• Kedrosky compares China’s pattern to prior overbuilding in batteries, solar, hydropower, and real estate.
• He thinks current LLMs are not truly dynamic systems because they cannot update weights in real time.
• His long-run view is that AI will become a utility layer, like electricity, with utility-like returns.