Dan Sundheim - The Art of Public and Private Market Investing
Inspired by Dan Sudenheim on Invest Like The Best
Worth Stealing
The edge in markets has migrated to long-termism
Public markets haven't got more competitive across the board. They've got more competitive in the short term and less competitive over longer horizons. As quants and multi-managers hoover up every short-term signal, fundamental long-horizon analysis has become less crowded. There's lots of people playing different games in the public markets, creating opportunity.
Private markets are less efficient, and the competition is social not analytical
Private markets have fewer eyes on any given situation, but the limiting factor isn't intelligence as most people are doing similar analysis. The best companies choose their investors.
The airline analogy for LLMs
Transformative technologies don't automatically produce good businesses. Air travel changed the world and airlines destroyed capital. The question for LLMs isn't whether the technology matters, it's whether the economics work at the model layer.
LLMs as Netflix plus Spotify
Like Netflix, the model is fixed cost and front loaded, so spend heavily to train, then sell at near zero marginal cost. Revenue funds the next model, which funds the next. Like Spotify, the underlying content is a commodity. What creates pricing power and switching costs is personalisation. The moat, if it exists, comes from knowing the user, not from the model itself.
Focus beats breadth
Logic tells you to spread your fixed costs across multiple markets, reducing concentration risk. In reality you end up being spread too thin. One reason Anthropic has done very well is they chose enterprise and stuck to it.
Hyperscalers are financing mechanisms, not partners
The hyperscalers are functioning as a combination of venture backer and infrastructure provider simultaneously. The LLMs get the compute they need without building data centres; the hyperscalers get anchor tenants for their AI infrastructure and a stake in the upside. It works for both parties right now. The LLMs currently need cloud compute because they are cash flow negative. Once they are profitable, the incentive to insource is enormous.
Systems of record will outlast the hype
LLMs aren't building their own ERP systems. They're buying them. The software businesses most at risk are the productivity layer. The systems that run the core of a business are harder to displace than people assume.
The Taiwan semiconductor risk is underpriced
Over ninety percent of advanced semiconductor production is concentrated in one island. There is no scenario where China, Taiwan, and the US are all satisfied with how that resolves. The supply chain is easy to destroy and very hard to replicate.
The best businesses are low cost compounders
The most durable moat isn't brand or network effects. It's structural cost advantage with a positive feedback loop. Low cost drives volume and volume drives lower cost. SpaceX in launch, Costco in groceries.
Conviction scales with clarity of writing
The discipline of writing down an investment thesis isn't process for its own sake. The act of writing forces you to stress-test assumptions you'd otherwise skip past in your head.
My Thoughts
There is a version of the Dan Sundheim interview that is just another AI takes piece. It is worth reading past that.
The observation about public versus private markets sounds simple, and that's why it's worth paying attention to. In public markets, participants are playing different games: passive, quant, long-short, retail, each with different horizons and objectives. In private markets, especially venture, everyone is playing the same game. Which means the differentiation cannot come from the analysis. It has to come from you: your relationships, your reputation, your ability to get into the room before anyone else does.
The Netflix-Spotify framing for LLMs is useful but imperfect. Netflix holds at the level of capital structure: front loaded fixed cost, then distribute at near zero marginal cost. But Netflix makes a film once and streams it identically to every viewer. LLMs run inference on every query, and every query is different. Marginal cost is low but not negligible. The Spotify parallel may be the more durable one. The underlying model is increasingly a commodity. What creates pricing power and switching costs is personalisation, knowing the user, not the technology itself.
Where does this end in 10 years? The LLMs are using AWS and Azure as a financing mechanism today because they are cash flow negative and the hyperscalers have strong balance sheets. When that changes, the logic for insourcing is difficult to ignore. The question is what AWS and Azure look like on the other side of that.
The Taiwan observation is the one that stays with you. A single point of failure for the technology that powers everything, with no resolution that satisfies China, Taiwan, and the United States at once.