Descriptions:
Anj Midha — one of Anthropic’s founding investors, former head of AI investments at Andreessen Horowitz, and now founder of AMP — joins Harry Stebbings on 20VC to discuss the early days of Anthropic (rejected by 21 of 22 VCs), his current investment thesis, and his framework for understanding where AI progress is actually constrained.
Midha identifies four bottlenecks: context feedback (the continuous data loop needed for frontier research), compute, capital, and culture — arguing that culture may be the most important and least discussed. His view is that if an organization solves culture, algorithmic innovation follows naturally, because mission-driven researchers stay flexible on architecture rather than becoming ideologically committed to transformers or diffusion models. He pushes back firmly on claims of saturating scaling returns, pointing to his Menlo Park incubation Periodic Labs, a 30,000-square-foot facility using LLMs to predict novel superconductors, robots to synthesize them, and X-ray diffraction machines to validate predictions — a loop he says is seeing super-exponential gains per compute dollar.
Midha also makes the case for coordinated frontier model inference as a kind of “Iron Dome” for AI competitiveness, drawing an analogy to the historical buildout of electrical grids. The conversation covers VC decision-making under uncertainty, how to invest like a scientist by running parallel hypotheses, and why he views human alignment — not AI alignment — as the harder near-term problem.
📺 Source: 20VC with Harry Stebbings · Published April 14, 2026
🏷️ Format: Interview
