The a16z Show

Joel de la Garza × Barrett Lyon

Barrett Lyon argues the Internet’s open model is eroding as corporate servers, carrier-grade NAT, and policy-driven identity and age controls move communication and surveillance into the network layer.Docs.net owns the stack from software to wiring, using AI-operated hardware and private domains for peer-to-peer exchange; scaling under applicable laws while resisting telemetry-driven surveillance remains the key risk.

David George × Sarah Wang × Alex Immerman × Santiago Rodriguez

AI’s rally is being driven by earnings rather than multiple expansion, with the S&P 500 below 20 times earnings while growing about 15%.Capex is projected to rise from $416 billion in 2025 to about $780 billion in 2026 and exceed $1 trillion annually from 2027; enterprise adoption remains concentrated in pilots, with inference-cost declines and workflow deployment as catalysts.

Anish Acharya × David Pawlan

AI assistants may win by recovering “free money”—HSA reimbursements, fare credits, or 50% lower water bills—not by abstract productivity gains.Proactivity can be a moat when agents act invisibly, but one unauthorized insurance switch could destroy trust.At roughly $20 per user per day for ambitious browser-driven services, economics may require narrow agents priced at $200-$300—or eventually $1,000—a month.

Ben Horowitz × Martin Casado × Diogo Almeida

Jev targets AI’s missing software layer: a natural-language primitive that joins a state machine and selects actions with confidence, making software itself more capable rather than merely speeding code production.Its economic unlock is reliability—“the same intelligence every time”—which could let incumbent SaaS vendors automate high-return workflows across existing distribution, while state consistency, security, cost, speed, and strong guarantees remain unresolved.

Erik Torenberg × Aaron Levie × Martin Casado × Steven Sinofsky

Agent swarms expose a security gap: controls built for trustworthy, capacity-constrained humans fail when software exhausts APIs and confuses good tasks with bad ones.That implies upgrades across identity, authorization, observability, operating systems, networks, and granular data permissions.Meanwhile, Jevons-style option selection could shift innovation beyond frontier labs by making models faster, cheaper, and more accurate interfaces for traditional software.

Erik Torenberg × Amjad Masad × Gagan Biyani

Amjad Masad says AI can now perform machine learning itself—“RSI,” “vibe research,” or “autoresearch”—and Replit is building a product enabling any business owner with data to fine-tune a model.He sees this as the next inflection after Replit Agent’s 2024 coding shift, while AI agents may solve the technology bottleneck behind failed no-manager structures and move Replit toward a “self-driving company.”

Ali Ghodsi × Martin Casado × Sarah Wang

Frontier AI’s immediate enterprise risk may be cyberattack acceleration: CVE-to-exploit time fell from 2–3 years in 2018–19 to “basically hours” now, while most organizations lack automated detection and threat hunting.Ghodsi says recursive self-improvement shows none of four required conditions; enterprises need context and cost control more than smarter models, though a frontier freeze would be disastrous to the labs.

Jennifer Li × Gorkem Yurtseven × Batuhan Taskaya

Fal’s H3 Max turns MiniMax’s open-source H3 into a 35×-faster, order-of-magnitude-cheaper endpoint at the same Elo score, after external validation.Post-training and RL lift quality before systems optimization, while kernel work raises utilization from 30–40% to 70–80%; the model runs on one 8-GPU node.H3 Max became Fal’s most popular video model by nearly 2×, while the next 1–2 months target 99.9% controllability and Hollywood adoption could accelerate.

Alex Rampell × Joe Schmidt × Keith Peiris

Keith Peiris stopped Tome at 2 million monthly users because models lacked context about the presenter, audience, and their relationship, then built Lightfield as a “business world model” around a relationship log.Both pricing extremes failed, leading to platform fees plus seats for core CRM and consumption elsewhere.Speed remains the risk: ElevenLabs moved to Salesforce after dashboards took four months.

Ben Horowitz × Erik Torenberg × Greg Brockman

Greg Brockman says Astra is “pretty reasonable to call ...AGI,” with coherent 24-hour runs across domains and a discontinuous jump driven partly by computer use that reduces reliance on purpose-built connectors.OpenAI shifted 25% of production engineers to security, saturated Astra’s identifiable P0s, and committed $1B plus discounted CrowdStrike access, but compute distribution and safety, security, and alignment may become the next binding constraints.