The venture capital industry has spent the past four years funding a gold rush, but one of its most prescient voices thinks the prospectors are digging in the wrong place.
Chi-Hua Chien, the Kleiner Perkins partner who backed Facebook before most of the industry understood social networks, is now making a contrarian case about artificial intelligence: the companies that will capture the most value from the technology are not the ones building foundation models or selling API access. They are the ones quietly using AI to transform their existing businesses—retailers optimizing supply chains, insurers automating underwriting, manufacturers predicting equipment failures before they happen.
The argument is not that OpenAI or its competitors will fail. It is that the economics of selling AI as a product are brutal, while the economics of deploying AI as a capability inside a business with existing customers and distribution are extraordinarily attractive.
The picks-and-shovels trap
Silicon Valley loves the picks-and-shovels metaphor—the idea that during a gold rush, the smart money sells equipment to miners rather than mining itself. Nvidia has been the poster child for this thesis, and its stock price reflects it. But Chien's point is subtler: even the picks-and-shovels layer is becoming commoditized faster than expected.
Foundation models are converging in capability. Open-source alternatives are closing the gap. Inference costs are plummeting. The moat that seemed so wide in 2024 looks narrower every quarter. Meanwhile, the companies with proprietary data, established customer relationships, and domain expertise have something that cannot be replicated by training a larger model.
Where the value actually accrues
Consider the difference between a startup selling an AI-powered legal research tool and a law firm that deploys similar technology internally. The startup must acquire customers, compete on price, and watch margins compress as rivals proliferate. The law firm captures the productivity gains directly, charges the same hourly rates, and faces no new competitive pressure from the technology itself.
This dynamic plays out across industries. The AI layer becomes infrastructure—valuable, necessary, but not where the economic surplus concentrates. The surplus flows to whoever controls the customer relationship and the data that makes AI actually useful in a specific context.
Our take
Chien's thesis is uncomfortable for an industry that has poured tens of billions into AI startups, but discomfort is often a sign of insight. The most transformative technologies rarely make their inventors the richest people in the room—they make their best deployers wealthy instead. Electricity did not mint fortunes for power plant operators; it minted them for manufacturers who electrified their factories. The same pattern is emerging with AI, and the investors who recognize it early will look as prescient as Chien did in 2007.
