Walk into any AI venture pitch and you'll hear the same metaphor: pick-and-shovel investing. The story goes that during the California gold rush, the people who made the most money weren't the prospectors — they were the merchants selling tools to the prospectors. By analogy, the safest AI investments are infrastructure, tooling, and platforms that every AI company will need regardless of which application layer wins.
It's a tidy story. It's also leading a lot of funds into bad investments.
Why the metaphor breaks
The picks-and-shovels metaphor breaks down for one specific reason: the picks and shovels in AI are being given away for free, or near-free, by the same companies that make the gold. OpenAI ships its own developer tools. Anthropic ships its own evaluation framework. Google bundles its agent runtime into Vertex. Every layer of the so-called "infrastructure stack" is being absorbed into the foundation model providers as fast as third parties can build it.
We've watched a dozen well-funded picks-and-shovels companies in the past eighteen months either get acquired for parts or quietly pivot to applications. The vector database wars are mostly over. The orchestration framework wars are mostly over. The evaluation tooling wars are being fought by companies that will be irrelevant in two years because the underlying models will evaluate themselves.
Categories under structural pressure
- •Vector databases — being absorbed into Postgres, into the model APIs themselves, and into managed cloud services.
- •Generic agent frameworks — being commoditized by OSS releases from the labs every quarter.
- •Evaluation and observability — model providers are shipping their own, with privileged access to internals.
- •Prompt management — features inside every AI dev platform; not a company.
- •Generic RAG-as-a-service — the model providers will do this for free to keep you on their API.
The historical analogy that actually fits
The historical analogy that actually fits AI is not the gold rush. It's the early internet. In 1996, the conventional wisdom was that the safest internet investments were infrastructure: the routers, the protocols, the hosting providers. Those companies absolutely existed and some made money — but the trillion-dollar businesses turned out to be applications. Google. Amazon. Facebook. Netflix. Companies that solved a specific user problem on top of commoditizing infrastructure.
We expect the same pattern in AI. The trillion-dollar businesses will not be vector databases or evaluation tools. They will be applications — agents that solve a specific business or consumer workflow so completely that they replace the existing software category. That's where we invest, and that's why we structure our fund to hold positions long enough to capture that compounding.
When infra investing still works
There's nothing wrong with infrastructure investing in principle. There's plenty wrong with it in 2026, when the companies you're funding are competing directly with the same labs whose models they depend on. We tell founders this directly: if your moat depends on the foundation model layer not building what you're building, you don't have a moat. You have a head start, and the labs will catch up faster than your runway lasts.
The infra companies we'd still write a check to
- •Solutions that span model providers and have no incentive to be absorbed by any single one.
- •Tools rooted in regulated or compliance-heavy domains where lab-built generic solutions can't reach.
- •Hardware-adjacent and edge-deployment plays where the labs aren't operating.
- •Workflow-specific orchestration that ships with vertical depth, not a generic SDK.
What durable AI wealth looks like
The picks-and-shovels investors will look right for another year or two. Some will exit profitably. But the durable wealth in agentic AI will be created at the application layer, by founders who picked a specific workflow, owned it end to end, and built a real business around it. That's the gold. We're going for it.
