LLM Competition and Model Jockeying

In this episode of Excess Returns, Kai Wu of Sparkline Capital returns to discuss his latest research on AI adoption, ROI, and what it all means for investors.

Building on his prior work on the AI CapEx boom, Kai tackles the trillion dollar question at the center of today’s market: Is AI generating real, measurable economic returns across the broader economy, or are we still in an infrastructure-driven bubble?

Using a systematic analysis of earnings calls, patent data, and adoption trends, Kai lays out a framework for identifying which companies are truly benefiting from artificial intelligence and how investors can position portfolios accordingly.

Find the Full Paper Here:
https://etf.sparklinecapital.com/

Main topics covered:

• Satya Nadella’s AI bubble framework and why broad economic diffusion matters

• The AI adoption S-curve and where we are in the technology diffusion cycle

• A new AI ROI taxonomy based on earnings call analysis and quantified economic gains

• Real-world AI productivity, revenue, and cost-saving examples across industries

• Infrastructure vs early adopters vs laggards and how companies were categorized

• AI-driven outperformance and excess returns across different adopter groups

• Valuation dispersion between AI infrastructure stocks and AI early adopters

• The risk of overcapacity and lessons from railroads and the dot-com telecom boom

• Competition among large language models and the durability of AI moats

• S&P 500 exposure to AI infrastructure and hidden concentration risk

• The case for AI early adopters as a middle ground between growth and value

• Intangible value investing and the concept of AI yield

Timestamps:

00:00:00 The trillion dollar question and what “real ROI” means
00:03:19 Nadella’s bubble framework: diffusion vs a narrow CapEx trade
00:06:08 The classic tech diffusion S-curve and where AI is on it
00:32:25 Why infrastructure is being rewarded even if the ROI story is different
00:33:04 The key chart: adoption vs valuation shows “basically no relationship”
00:38:00 Why early adopters and laggards should separate
00:38:26 The “25% ROI” example and how it could show up later in fundamentals
00:39:03 Railroads and fiber: builders go bankrupt, users capture the value
00:39:45 Telecom index fell 95% and never recovered (dot-com bust parallel)
00:40:00 The application layer captures profits; infrastructure becomes a utility
00:41:00 The punchline: transformative tech, but builders can still be bad investments
00:42:57 Overcapacity question: where are we on the line?
00:43:17 The buildout: another $5 trillion of data centers “or whatever the number is”
00:44:00 If there’s no ROI, companies cancel orders
00:45:01 Moat and LLM competition discussion begins
00:49:00 The big one: adding infrastructure names gets the S&P to 46% AI infrastructure
00:50:00 “Alternative indices” swing you to laggard risk
00:51:00 The “false choice” and the “middle ground” framing (early adopters)