Kai Wu of Sparkline Capital joins Excess Returns to discuss his paper Surviving the AI CapEx Boom. In this episode, Kai breaks down the unprecedented level of investment in AI infrastructure, why today’s AI buildout mirrors past technology booms, and what it all means for investors. He explores the parallels between AI and historic bubbles, the implications of massive corporate CapEx spending, and where value might ultimately be captured as the cycle plays out.

Topics covered:

• Why big tech’s CapEx spending has exploded and how much they’re investing

• The trillions in revenue needed to justify AI infrastructure spending

• Historical parallels with the railroad and dot-com buildouts

• Why companies that invest heavily often underperform

• How the Mag 7 are shifting from asset-light to asset-heavy businesses

• The risks of “circular deals” and financial entanglement in AI

• Why the AI race resembles a prisoner’s dilemma

• Which layers of the AI stack may capture long-term value

• How early adopters and infrastructure players differ in capital intensity and returns

• Where investors might find opportunity beyond the obvious AI names

Timestamps:
00:00 Introduction and overview of AI CapEx boom
03:00 Why Kai researched AI investment cycles
05:00 Scale of big tech’s CapEx spending
07:00 Revenue needed to justify AI infrastructure
08:30 Market concentration and valuation risks
11:30 Historical parallels: railroads, internet, and AI
14:30 The capital cycle and overinvestment dynamics
17:30 “This time is different?” and lessons from bubbles
18:00 Factor investing and high-asset-growth underperformance
21:00 Sector and firm-level CapEx trends
22:30 Winner-take-all dynamics and competitive pressure
26:00 How the Mag 7’s business model is changing
30:00 Comparing tech CapEx to utilities
34:00 The circular deal problem and financial risk
37:30 The AI arms race as a prisoner’s dilemma
40:30 Will AI be winner-take-all?
43:30 Lessons from the railroad and dot-com eras
47:00 Where the value is captured in infrastructure vs adoption
48:00 Identifying early AI adopters and hidden beneficiaries
50:30 Sector and geographic AI exposure
54:00 Capital intensity and valuation differences between infrastructure and adopters