a16z dropped three AI bets this week, zero foundation model plays. They're betting on deployment infrastructure: data sovereignty, agent networking, and offensive security.

Conway Research (Underdog): On-device personal AI running email/calendar/meetings locally. Founder Sigil Wen hit the trifecta—Thiel Fellowship (sub-1% acceptance, alumni built Figma/Scale AI), Spearhead (Naval's program giving founders $2M to manage + carry), and Airchat (Naval's voice-first social app). This isn't just product validation, it's network access coded into one person. a16z led the round because Wen carries the right credentials to pitch data sovereignty.

Armadin: Autonomous offensive security using AI attacker clusters to generate verified kill chains. Founder Kevin Mandia built Mandiant, dropped the APT1 report in 2013 attributing seven years of breaches to PLA Unit 61398—turned anonymous hacking into named state adversary. He testified to Congress, got acquired by FireEye for $1B+, then absorbed into Google. This isn't a security startup, it's a national-level capability with a founder who literally defined what "Chinese cyber threat" means in US policy. a16z + Accel co-led Series B.

doxx: Agentic Defined Networking—parallel private network for humans + their agents, P2P browsing/device mesh. Founder Barrett Lyon built Prolexic (early DDoS mitigation, sold to Akamai for $370M), then Defense.Net (sold to F5). New Yorker profiled him hunting botnets, FBI used his work to jail Russian extortion crews. Two exits to infrastructure giants, clean law enforcement track record. He's selling the infrastructure for "unseen but not uncontrolled" networking.

The pattern: a16z isn't buying tech, they're buying founders with the legitimacy to own specific narratives. Post-foundation-model world, scarcity shifted from compute to "who gets to define the threat model and sell the solution." Mandia owns attribution, Wen owns youth-builder-immigrant legitimacy, Lyon owns parallel net credibility. Three sectors, one thesis: political positioning matters more than benchmarks.