The AI investment market of 2024 is characterized by extremes. At the foundation model layer, rounds are measured in billions — Anthropic, OpenAI, Mistral, Cohere — with capital from both traditional VCs and strategic investors who are effectively betting that the model provider becomes a durable layer of the stack. At the application layer, there is frothy seed and Series A activity across every vertical. But in the infrastructure middle — the tools, runtimes, and platforms that sit between models and applications — the landscape is more interesting and less picked over.
We are occasionally asked whether there is still a meaningful role for seed-stage infrastructure investors given the amount of capital in the market. The answer is yes, but the question is worth taking seriously.
What Seed Capital Actually Provides
The financial return argument for seed investing is straightforward and well understood. But I want to make a different case — about what seed-stage capital provides to founding teams, not to investors. A seed round from investors who have a specific thesis about the infrastructure stack gives a founding team more than money. It gives them a set of investors who are pattern-matching actively on the problems they are solving, who can connect them with early enterprise customers who have the same problems, and who have a strong opinion about what the right architecture is and why.
This matters at the infrastructure layer more than it does in consumer or application products. Infrastructure companies need to get the architecture right early; retrofitting fundamental design decisions at Series B is much harder than getting them right at seed. Having investors who think deeply about the infrastructure stack from day one changes the quality of the early architectural decisions.
The Thesis Check
The other thing that seed investing provides is honest thesis feedback. When we write a seed check into an AI infrastructure company, we are committing to that company's architectural bet. If we are wrong about the bet — if agent memory turns out not to be a standalone category, if event streaming for agents does not take off, if LLM observability commoditizes into model provider dashboards — we do not make money. The thesis discipline at the seed stage is different from the discipline at later stages, and it is a useful feature of the ecosystem, not a bug.