The venture capital veteran who called Facebook in 2004 now says the AI economy is being mispriced

On 17 June 2026, TechCrunch published a long profile of Chi-Hua Chien, the Kleiner Perkins partner who in 2004 became one of the earliest venture investors in what was then a small Palo Alto social network called TheFacebook. The piece, by Connie Loizos, makes a single argument that runs against the prevailing Silicon Valley mood: the firms best positioned to capture lasting value from artificial intelligence are unlikely to be the ones selling AI itself.

That is a contrarian claim at a moment when capital is flooding into model developers, inference platforms, and "AI-native" applications promising to replace entire categories of white-collar work. Chien's career is the reason his dissent carries weight. He backed the social web before it was obvious that social would become infrastructure. He is now telling anyone who will listen that the lesson of that cycle is being forgotten.

The long view from Kleiner

Chien joined Kleiner Perkins Caufield & Byers in 2008 after a stint at the firm that took him through the 2004 TheFacebook investment, which the firm made while it was still run by the office of former US Vice President Al Gore. As Loizos recounts, the bet was made on the conviction that identity online would eventually become a fundamental layer of the consumer internet — a thesis that looked speculative in 2004 and obvious by 2010. Kleiner was, in Loizos's telling, early because the firm was willing to underwrite a behaviour change rather than a product feature.

That frame is what Chien now applies to AI. In the TechCrunch interview, he describes the current environment as one in which a great deal of money is being routed toward infrastructure — chips, model labs, data labelling, inference providers — and comparatively little toward the question of where AI gets embedded into existing industries in ways customers will pay to keep. The article quotes him on the limits of horizontal AI products and on what he sees as the structural difference between a tool that is bought once and a system that becomes part of a firm's operating cost.

The cultural-anthropologist framing in the headline belongs to Loizos, not to Chien. She argues that his instinct has always been to study a technology the way an ethnographer would study a village: who lives there, what they trade, who holds the power, what the rituals are. That posture, she writes, is what made Kleiner's TheFacebook bet look like a study of identity before it looked like a financial decision.

Why the AI bets look different this time

The pitch of the current AI cycle, as reported across 2025 and 2026, is that the model is the product. The large labs — OpenAI, Anthropic, Google DeepMind, and a long tail of well-funded challengers — sell access to a reasoning engine. A second tier sells tooling around that engine: evaluation, observability, fine-tuning, agent orchestration. A third tier, where most venture capital has actually landed, sells applications that wrap the model in a vertical workflow: legal review, sales prospecting, customer support, code generation.

Chien's argument, as the TechCrunch profile renders it, is that the first two tiers are commodity businesses in waiting, and the third tier is structurally vulnerable to the first. If model prices fall — and the broader industry's pricing curve has been falling for two years — the wrapper application loses its margin the moment a customer can replicate the workflow with a raw model call and a well-written prompt. The durable moat, in his reading of the social-web precedent, belongs to companies that own a workflow that the AI merely makes cheaper, not companies that are the workflow.

This is, in plain language, a restatement of an old venture heuristic: own the customer relationship, not the substrate. It is the lesson that an earlier generation of investors extracted from the platform shift to mobile, and the generation before that extracted from the shift to the web. The reason it is worth restating in 2026 is that the cost of compute has fallen far enough that the substrate is no longer the bottleneck. The bottleneck is distribution, regulatory permission, and trust — all of which are properties of the application layer, not the model layer.

The structural frame

What the AI cycle is doing, in capital-allocation terms, is replaying a familiar pattern. A new substrate is built at great expense by a small number of firms funded by patient capital. The substrate becomes cheaper. A wave of "X for AI" applications is funded, on the thesis that the substrate's fall in price will translate into their rise in margin. The wave consolidates. The survivors are the firms that owned a workflow the substrate could not displace. The substrate firms become utilities.

This pattern is not unique to technology. It is the same shape that the cloud-computing cycle took in the late 2000s, the same shape the personal-computer software market took in the 1980s, and the same shape the electricity market took in the years after the first commercial grids. The investors who made durable returns on each of those transitions were, almost without exception, the ones who underwrote the application layer once the substrate layer was no longer scarce.

Chien's claim, sharpened by Loizos's reporting, is that the AI substrate will follow the same arc. The question for the next eighteen months is whether the capital currently chasing the substrate — and the very large valuations attached to it — is being priced for that arc, or for a longer monopoly in which the model layer itself captures most of the economic surplus. The venture veteran who called the social web in 2004 is betting it is the former.

Stakes and what remains uncertain

The stakes are concrete. If Chien is right, the next decade of durable AI returns will accrue to firms in healthcare logistics, legal services, construction procurement, mid-market accounting, and industrial maintenance — sectors that are unglamorous, regulated, and require long sales cycles. If he is wrong — if model-layer firms manage to sustain their pricing power and lock in their customers through switching costs — then the substrate firms will retain a much larger share of the surplus than the venture class is currently pricing in.

What remains genuinely uncertain, even after Loizos's long sit-down, is the timing. The social-web transition took roughly six years from TheFacebook's founding to the moment when it was obvious that social identity was infrastructure. The AI transition may move faster, because the substrate is digital-native and the distribution channels are already built. Or it may move slower, because the workflows that the AI is meant to displace are tangled into compliance regimes, professional licensure, and incumbent procurement contracts. The TechCrunch profile does not pretend to resolve that timing question, and neither can a reader of it.

What is clear is that the venture industry's most senior cultural anthropologist has now placed a public marker on the question. The rest of the cycle will be priced against it.

This publication treats Chi-Hua Chien's framing as one informed read of a still-unresolved cycle, not as a forecast. The TechCrunch profile is the only source consulted for the direct claims and quotations above; readers seeking primary documents on the underlying AI economics should follow the model-lab earnings releases and the inference-pricing public benchmarks directly.