The AI Bubble Is Widely Misunderstood | Steve Hou
Steve Hou, a quantitative researcher at Bloomberg Indices with a PhD in macroeconomics, argues that the AI bubble is real but widely misunderstood in its scale and duration. He contends that agentic AI and recursive model-calling have created nonlinear compute demand that most analysts underestimate, largely because most people don't code. He also examines AI's complex and uncertain effects on productivity, inflation, labor markets, and the US debt problem.
Summary
Steve Hou, a quant researcher at Bloomberg Indices and former AQR Capital researcher with a PhD in macroeconomics from the University of Michigan, joins the podcast to discuss his early conviction in AI's transformative potential and its macroeconomic implications.
On the AI bubble, Hou argues that while it is unquestionably a bubble, the more important questions are how large and how long it can last. He distinguishes the AI bubble from the dot-com bubble by noting that AI has been adopted and used almost immediately by consumers and enterprises alike, unlike the internet boom where massive unused infrastructure sat idle before being built out. He compares the AI development timeline to the movie 'The Core,' where a succession of chain-reaction explosions amplify each other — each wave of progress (ChatGPT, DeepSeek, agentic coding tools like Claude's Opus) re-energized the cycle rather than ending it.
A key inflection point Hou identifies is the emergence of agentic AI, particularly with tools like Claude Code and Opus. Unlike a single chatbot query representing a compute demand of roughly 1x, AI agents calling other AI models recursively can amplify compute demand by a hundredfold or more. He argues that most people underestimate this because most people don't code — making it difficult to appreciate just how much agentic AI has changed the productivity and demand landscape for compute.
On macroeconomic impact, Hou explains that AI's most direct GDP contribution so far has been through capital investment (the 'I' in C+I+G+NX), making it the fastest and largest capex cycle in modern history. He notes that this investment has buoyed the US economy during a slowdown from the 2022 rate hike cycle, with spillover effects visible in South Korean and Taiwanese export economies. He also describes an indirect wealth effect from rising AI-related equity valuations boosting consumption among wealthier households.
Regarding measured labor productivity, Hou is skeptical that recent strong productivity prints reflect genuine AI-driven gains. He suspects composition bias — as labor-intensive sectors shrink and capital-intensive AI investment grows, measured productivity appears to rise artificially. He also cites COVID-era over-hiring being unwound as a contributing factor. While anecdotal evidence from companies suggests meaningful productivity gains, he stresses the absence of clean causal identification in the data.
On monetary policy, Hou is skeptical of arguments that the Fed should preemptively cut rates based on anticipated AI-driven disinflation. He notes that productivity-driven disinflation from technological change has historically been a slow J-curve phenomenon, and that AI's most immediate macroeconomic effect is actually inflationary — driving up demand for hardware, electricity, and construction labor. He argues that any genuine AI disinflation would show up in observable data like wage growth and rent inflation, and policymakers should watch those metrics rather than speculate.
On the US debt problem, Hou frames it as an arithmetic issue: federal tax receipts have historically been capped near 18-20% of GDP regardless of tax policy, while spending and interest costs have grown secularly. The only viable lever is growing the GDP denominator. However, he notes that AI investment can paradoxically raise interest rates by competing for capital, and potential social spending on displaced workers could worsen primary deficits. He is cautiously optimistic about AI's long-run growth impact but emphasizes that Baumol's cost disease — where labor-intensive sectors like healthcare, childcare, and skilled trades resist productivity improvements — will remain a stubborn constraint.
Finally, Hou expresses strong enthusiasm for AI's impact on economics and econometrics specifically, predicting richer models, better agentic policy simulations, and broader accessibility to sophisticated analytical tools. He argues that AI lowers the barrier to using powerful econometric methods and could improve central bank communication, which he believes contributed to recent Fed policy confusion.
Key Insights
- Hou argues that the AI bubble is real and was always going to be a bubble, but insists the more important questions are how large and long it can last — and that most critics dismissed AI's significance the moment they labeled it a bubble.
- Hou contends that agentic AI, where models recursively call other models, increases compute demand by potentially a hundredfold compared to simple chatbot queries, fundamentally changing the demand outlook for AI infrastructure.
- Hou argues that unlike the dot-com bubble — where vast infrastructure sat unused before being adopted — the AI bubble is characterized by immediate, widespread adoption, making it structurally different and potentially more durable.
- Hou claims that most analysts underestimated the agentic AI acceleration because most people don't code, making them unable to fully appreciate how much AI has expanded what coders and non-coders alike can accomplish — analogous to watching Roger Federer play tennis without ever having held a racket.
- Hou is skeptical that recent strong US labor productivity data reflects genuine AI-driven gains, attributing much of it instead to compositional shifts in the economy and the unwinding of COVID-era over-hiring.
- Hou argues that AI's most immediate macroeconomic effect is actually inflationary, not disinflationary — by driving up demand for hardware, electricity, and skilled construction labor — and that any AI-driven disinflation should be observable in wage and rent data before the Fed acts on it preemptively.
- Hou frames the US debt problem as fundamentally arithmetic — tax receipts are historically capped near 18-20% of GDP regardless of policy, so the only real lever is GDP growth — but cautions that AI investment can simultaneously raise interest rates by competing for capital, making its net fiscal impact ambiguous in the near term.
- Hou predicts that AI will transform economics by enabling richer agentic policy simulations, broader access to sophisticated econometric tools, and improved central bank communication — comparing the ideal economist-AI relationship to a Michelin chef being asked to make fried rice, where the advanced capability is deployed on accessible tasks.
Topics
Full transcript available for MurmurCast members
Sign Up to Access