Episode Summary
Executive Summary: The episode debates whether the AI boom is a bubble or a durable industrial transformation. Derek Thompson says agents and rising token demand weakened his bubble thesis, while Paul Khodrowski argues AI is still an infrastructure bubble driven by overbuilding, debt, and diminishing returns in core model improvements. They agree AI is transformative, but disagree on whether current spending can be sustained and where value will shift as commoditized tokens pressure software and chip demand.
Main Topics: AI as an infrastructure bubble (Priority: 5/5): Khodrowski frames AI spending as a classic capital-expenditure bubble like railroads or canals: spending races ahead of revenue, then a crash occurs before long-run gains emerge. He argues today’s AI buildout sits at the intersection of loose credit, real estate, tech, and policy. Agents, tokens, and revenue growth (Priority: 5/5): Thompson’s revised view centers on autonomous agents. He argues tools like Claude Code and Codex dramatically increase token usage, which boosts revenue and makes the gap between AI spending and revenue smaller, weakening the bubble case. Railroad analogy and historical bubbles (Priority: 4/5): Both speakers use railroads to show how a technology can be simultaneously revolutionary and bubble-like. Khodrowski emphasizes overbuild, financial crashes, and eventual long-term utility, while Thompson highlights that transformative tech can still produce a destructive speculative cycle. Hyperscalers, debt, and capital intensity (Priority: 5/5): Khodrowski argues the largest AI firms are not immune because AI capex is consuming free cash flow, limiting buybacks, and increasingly relying on debt and private credit. He sees this as a sign the sector is becoming utility-like and vulnerable to repricing. SaaS disruption and the token economy (Priority: 5/5): Khodrowski introduces the ‘SaaS Pocalypse’—AI tokens as a new industrial commodity that compresses software margins and lowers moats. He says many software businesses are being directly pressured by cheap, abundant tokens. Winners shifting toward energy and heavy assets (Priority: 4/5): The discussion ends with a view that value is migrating away from pure tech toward energy, transmission, transformers, HVAC, and other heavy-asset, low-obsolescence industries tied to data-center buildout and electrification.
Key Arguments: AI spending resembles prior infrastructure bubbles because capex is outrunning revenue and is increasingly financed by debt, buybacks, and private credit. Railroads were both a bubble and a foundational technology; AI can be the same—important long-term but still prone to overbuild and crashes. The market initially rewarded AI capex with higher valuations, then shifted to neutral, and now increasingly punishes more spending, especially among hyperscalers. Agentic tools create real usage and revenue growth, which could reduce bubble risk by expanding token demand and improving monetization. Khodrowski argues token consumption in coding is not representative of most white-collar work, so extrapolating current usage growth to the whole economy is misleading. He says AI model improvement is slowing on composite benchmarks, while orchestration layers, not base models, are driving recent perceived gains. The growing dependence on capex is turning hyperscalers into utility-like firms with lower margins and higher maintenance spending, reducing upside for equity holders. SaaS companies are exposed because AI tokens are a deflationary commodity that weakens software moats and makes products easier to replace or compress. Energy and infrastructure providers benefit from AI, but much of that upside is endogenous to data-center demand and may reverse if AI spending slows. If the bubble pops, it will likely show up first in credit markets, GPU overordering/unwinding, and then in power, HVAC, and related supply chains.
Data Points: Annual AI spending: about $700 billion per year - Thompson cites the scale of AI infrastructure spending across chips, data centers, and power Historical equivalent of AI spending: one Manhattan Project every 3 to 4 weeks - Used to illustrate the extraordinary size of private-sector AI investment Historical equivalent of AI spending: one Apollo program every 5 months - Another comparison underscoring the scale of AI capex SP 500 performance: up about 0.5% in the first two months of 2026 - Khodrowski contrasts the weak broad market with the AI-related stock decline Mag 7 performance: down 5% in the first two months of 2026 - Used to show reversal in the largest AI-linked stocks Microsoft performance: down 13% year to date - Example of a major hyperscaler under pressure NVIDIA performance: down 3% year to date - Example of AI market cap weakness despite continued relevance AI capex multiplier: $1 of AI capex once added roughly $2 of market cap - Khodrowski describes the early euphoric market response to AI spending Free cash flow exposure: more than half of free cash flow for some hyperscalers - Argument that AI capex is consuming a large share of available cash Stock buybacks: Oracle did about $800 million last quarter, down from $6–8 billion a decade ago - Illustrates reduced ability to offset dilution and support share prices Technology share of U.S. industry: around 60% of all U.S. industry - Khodrowski compares today’s tech dominance to railroads in the 1900 index Benchmark improvement (2022–23): about 12%+ year-over-year - Khodrowski’s composite benchmark reading for model progress in the earlier regime Benchmark improvement (2023–24): around 5%–6% year-over-year - Shows slowing pace of model improvement Benchmark improvement (recent): around 2%–3% year-over-year - Evidence used to argue model progress is decelerating Token cost decline: 70% to 90% year over year over the last five years - Khodrowski says tokens are a hyper-deflationary commodity NVIDIA gross margins: 73% gross margins - Cited as a sign of incumbent advantage that may be temporary Software employment: more than 2.5 to 3 million U.S. jobs - Khodrowski says software is among the most threatened occupations Energy sector contribution: 60% to 70% of recent EPS growth - Khodrowski says much of energy’s recent strength is tied to data-center buildout Productivity growth: around 2% to 2.8% - Thompson notes recent productivity data that some are attributing to AI GPU overordering: orders may be 2x to 3x actual need - Khodrowski expects a future unwind in Nvidia-related demand Inference hardware efficiency claim: about 75x more tokens per second - Example given for newer firms like Talus improving inference economics
Pivotal Quotes: "AI is a bubble. There's no question." — Paul Khodrowski: Direct answer when asked whether he still believes AI is in a bubble "Software eats the world. Software becomes the world. Software eats itself." — Paul Khodrowski: Summarizing his view that AI’s deflationary token economy pressures software from within "The bubble and the golden age are not one thing or the other." — Derek Thompson: Opening framing that technologies can be transformative and speculative at the same time
Implications: Listeners should expect continued volatility: AI may transform the economy while still destroying capital in the near term. Watch credit, hyperscaler margins, GPU demand, and SaaS pricing power; gains may shift toward energy, grid, and heavy-asset infrastructure.