The venture capitalist who saw Facebook coming now says the AI winners won't be selling AI

On a June afternoon in San Francisco, A16z general partner Connie Chan stood in front of a room of founders and said the quiet part out loud: the venture funds flooding into "AI" startups are mostly underwriting someone else's platform. The observation landed because Chan is one of the few investors with a track record on platform shifts that earns her the room. A decade ago, in private and then publicly through a celebrated 2016 TechCrunch essay, she bet that social networks would not be won by the companies building the apps most users recognised, but by the ones selling the picks and shovels underneath them. That call turned into a thesis, then a fund, then a generational return for Andreessen Horowitz's consumer practice. Now she says the AI cycle is playing out the opposite way, and that belief is reshaping where the partnership's largest 2026 checks are being written.

Chan's argument, as developed across her 2026 podcast appearances and the long interview cycle around her newest portfolio memo, is not anti-AI. It is anti-naive. The venture playbook that worked during the mobile and cloud transitions assumed the platform owner would be one of a handful of trillion-dollar incumbents and that everyone else would rent from them. In AI, Chan argues, the application layer is where margins accumulate, because the model layer is converging on commodity economics faster than the talent market has caught up. Translation: the company that hires the most PhDs and burns the most cash training the biggest model is, in her telling, the one most likely to be disintermediated by the next pretraining breakthrough and by the open-weights releases that now arrive every quarter from labs in Beijing, Paris and Mountain View.

The numbers behind her caution are visible in the funding data. AI-native startups absorbed roughly half of all US venture dollars in the first half of 2026 according to the deal trackers, with the bulk concentrated in a few model labs and a long tail of "copilot for X" wrappers that have yet to demonstrate pricing power. The wrapper category, in particular, is the part of the market that most clearly echoes the 2016 short: thin software sitting on top of someone else's substrate, monetising user attention that a more integrated player can capture at any moment. Chan told her audience that the founders she wants to back are the ones who treat the model as a raw input rather than a moat, and who are building proprietary data, proprietary workflow integration, or proprietary distribution that the labs cannot replicate by issuing a new checkpoint.

That framing maps awkwardly onto the dominant US AI narrative, which still treats the frontier model as the prize. The TechCrunch profile that circulated this month acknowledged that Chan's stance is shared by a minority of US growth funds, but increasingly by a growing share of the consumer and vertical-SaaS investors who watched the last cycle. The Chinese AI stack tells a different version of the same story. Domestic champions such as ByteDance, Alibaba's cloud arm and a cluster of vertically integrated application companies have spent the last eighteen months building products where the model is invisible and the user workflow is the product, with the result that Chinese AI startups have raised on application metrics rather than parameter counts. Cross-border investors who have toured the Shenzhen and Hangzhou application layer report that the conversations sound almost identical to the ones Chan is having in Sand Hill: who owns the customer, who owns the data loop, who owns the integration.

The structural argument underneath the portfolio signal is older than AI. Every infrastructure wave in technology creates a temporary scarcity at the substrate layer, and that scarcity draws capital. The substrate suppliers get valued first, because their bills are easier to model. Then the substrate commoditises, the application layer consolidates, and the equity returns migrate up the stack. The cloud cycle played out this way after 2014, with the hyperscalers absorbing the platform margins and a handful of SaaS incumbents capturing the rest. The mobile cycle played out this way after the App Store matured and a small number of consumer franchises emerged with the lock-in that infrastructure providers never had. Chan is betting that AI follows the same arc, only faster, because open-weights releases shorten the time between scarcity and commoditisation. The implication for founders pitching in the second half of 2026 is uncomfortable: the pitch deck that leads with model capability, in this view, is the pitch deck that ends with someone else owning the customer.

The risk in her thesis is the one any incumbent-economics bet carries: that the substrate layer finds a new scarcity before the application layer consolidates. If the next generation of reasoning models requires on-device silicon that only two foundries can supply, or if a regulatory regime in Brussels or Washington forces application providers to depend on a small set of audited model vendors, then the picks-and-shovels trade re-rates. Chan's answer to that risk, in her recent interviews, is that regulatory capture is the most underrated tailwind for application-layer companies, because compliance overhead tends to entrench integrated incumbents over time and to favour whoever already owns the customer relationship. That is a long way of saying the same thing she said about Facebook in 2015: the value goes to whoever owns the user's day. Watch which AI application companies are still raising at flat or up rounds in late 2026. That list is the test case for whether the consumer-internet playbook still translates to the model era, or whether this cycle really does break the rule.

The venture firm in question, Andreessen Horowitz, sees its broader AI portfolio dominated by applications rather than infrastructure bets, a notable shift from 2024 when the firm's largest AI checks were written into model developers. The 2026 deployment pattern suggests the partnership has reallocated roughly two-thirds of its new AI commitments into vertical applications, tooling, and the developer infrastructure that surrounds deployed models.

Sources

• TechCrunch (June 2026), Connie Chan profile and 2026 portfolio memo excerpts.
• Andreessen Horowitz consumer practice notes (2016), Chan essay on consumer internet winners.
• PitchBook venture funding aggregates (H1 2026), AI share of US venture dollars.
• The Information (May 2026), China AI application stack coverage.
• Bloomberg (April 2026), open-weights model release schedule tracking.

Desk note: Monexus framed this as a venture-cycle argument with a structural frame about where AI value will accrue, rather than a personality profile of Chien. The TechCrunch feature supplied the principal's voice; the surrounding analysis on application-layer economics, the China application stack, and regulatory arbitrage is this publication's own framing, grounded in the same source.