General Partner, a16z (infrastructure practice) · ex co-founder/CTO, Nicira · Expert Watch · Slow Watch, monthly scan · our rolling analysis
Why we watch him. Martin Casado is a16z's clearest voice on where value actually accrues in AI, and he argues the market, not model benchmarks, decides it, which cuts straight to Dominic's enterprise-AI and coding-vs-knowledge-work questions. His thesis that AI is now "a capital problem, not an engineering one" (inference, compute and energy economics, the raise-train-ship-raise-bigger flywheel) is a useful counter-baseline to model-centric hype. He is also a leading advocate for evidence-based, marginal-risk AI regulation over existential-angst policy, relevant to Dominic's AI-policy lens.
Current headline view: "This feels like 1996": the AI boom is early, with years to run, and not a bubble peak. Value accrues to those who win real markets, not to whoever holds the best model at a given moment. And AI is now a capital-and-compute problem more than an engineering one, with a floor and ceiling set largely by energy constraints.
Latest 1-page summary (as at 11 Jul 2026) 🟠 new thinking
The core theses we track him against:
Value accrual: markets decide, not benchmarks. Value flows to whoever wins the market, not whoever has the best model; apps, infra and brand moats beat a benchmark lead ("Where Value Will Accrue in AI", a16z LP Summit).
The capital flywheel. Today's AI rounds are effectively compute contracts, and frontier-model firms can out-raise the whole app ecosystem built on top of them (Latent Space, "Bitter Lessons").
Evidence-based policy. Regulate on measured "marginal risk", not speculative existential fear; the GPT-2 "too dangerous to release" fears never materialised ("Base AI Policy on Evidence, Not Existential Angst", ~Dec 2024).
AI-optimist. "The future of AI is amazing"; technology is neutral and net-beneficial, and doom framing misleads.
Different, but not magic. The binding constraints are capital, power, data centres and compute, not core engineering; in his words "it is not that hard to build AI models".
Cycle timing: this is 1996, not the peak. We are early in the cycle, with a real floor and ceiling set largely by energy constraints, not at a bubble top.
AI coding as a multi-trillion market. Software development could become a multi-trillion-dollar market as the barrier to coding collapses.
Anti-AGI-framing. AGI debates encourage "lazy thinking" and obscure how technology actually creates economic value.
Method note. This v1 baseline is distilled from his a16z writing and talks, the AI + a16z podcast, and long-form interviews (the Gelsinger fireside, Latent Space, The Generalist), with dates marked ~ where not confirmed on-page. It is not yet a full X history read. The monthly scan refines it from here and flags any deviation.
Analysis backlog (newest first)
Deep source docs are held locally in the workspace at expert-watch/, not on this page.
11 Jul 2026 · v1 profile built. Identity verified (Martin Casado, GP at a16z, ex-Nicira/VMware). Eight core theses distilled from his a16z writing, the AI + a16z podcast, and long-form interviews (the Gelsinger fireside, Latent Space, The Generalist), not yet a full X read.
Notable reactions (operators & researchers)
As at 11 Jul 2026.
None logged yet; this v1 baseline captures Casado's own positions. The monthly scan will populate reactions from operators and researchers as they surface.
Recent media
As at 11 Jul 2026. Dates marked ~ where not confirmed on-page.