"We're not doing 30 bets a year": Vijay Pande on betting small after running $4 billion at a16z | TechCrunch

Vijay Pande managed nearly $4 billion in capital at Andreessen Horowitz before walking away in June 2025 to start something dramatically smaller — and the reasoning behind that decision turns out to have as much to do with AI agents as it does with investment philosophy.

What Changed

Pande arrived at a16z more than a decade ago as a Stanford chemistry professor, best known outside academia for Folding@home, the distributed-computing project that recruited millions of home PCs into a volunteer supercomputer for disease research. Marc Andreessen and Ben Horowitz had spent the firm's first five years deliberately steering clear of healthcare and life sciences; when they finally decided to bet on the category, they handed Pande the keys. Over the following decade-plus, he built the practice into one of the largest healthcare-focused venture portfolios in Silicon Valley, eventually overseeing close to $4 billion in assets.

Then, in June 2025, he left. The new firm — VZVC, co-founded with Zach Werner — is built on a deliberate rejection of the high-velocity model that defines most large venture funds. Where a major biotech practice might underwrite dozens of investments per year, VZVC targets roughly five. That's the "30 bets a year" reference in the interview published by TechCrunch on August 29, 2026: Pande is explicitly choosing concentration over breadth, conviction over coverage.

The fund's size and staffing reflect that philosophy directly. VZVC operates with two people. That headcount is not a temporary state while recruiting; it appears to be the intended steady state. The firm considered hiring associates to handle the operational load that typically requires a larger team at established funds — and then chose not to. AI agents absorbed that work instead.

Multiple reports confirm the same basic profile. A Toldrop summary of the interview notes that VZVC uses AI "extensively," covering research, screening, and operational tasks that would typically fall to junior investment staff. The Yahoo Finance version of the story offers the same framing: a deliberately lean structure, concentrated biotech bets, and an AI stack standing in for headcount.

How It Works

The practical mechanics of running a fund on AI agents rather than humans aren't fully detailed in the interview, but the shape is clear enough. At most large VC firms, associates handle deal sourcing, market mapping, competitive landscaping, due diligence document preparation, and portfolio company monitoring — high-volume, research-intensive work that demands consistent attention but rarely requires partner-level judgment on every task. These are precisely the categories where current AI agents, given good tools and well-defined scope, perform at or near junior-analyst level.

Pande's background is relevant here. Folding@home was fundamentally about replacing centralized, hierarchical compute with distributed, asynchronous work across many nodes. The mental model transfers: instead of a team where tasks flow down and results flow up through layers of staff, VZVC runs a flat structure where agents handle throughput work and Pande and Werner apply judgment at the decision points that actually require it.

There is a harder constraint underneath all of this — one that Pande identifies as the real bottleneck in AI-driven biotech: biological data. Unlike code, legal text, or news, the datasets that matter most in drug discovery cannot be scraped from the public internet. Proprietary clinical records, genomic datasets, and experimental assay results sit behind institutional walls. That means model weights alone don't constitute a durable advantage in this category. The organizations with proprietary biological data hold an edge that can't be closed simply by fine-tuning a foundation model on what's publicly available. Pande frames data access — not AI capability — as the fundamental constraint in the space he's investing in.

What It Means for Developers

Our read is that this story matters well beyond the venture capital world because it is one of the first credible, institutional-scale confirmations that a small team plus an AI agent stack can functionally replace a much larger one — at a firm handling real capital and making real decisions under fiduciary pressure. It's not a startup demo or a thought experiment. VZVC is operational, making investments, and explicitly choosing not to hire the people it originally planned to hire.

For developers building in the agent infrastructure space — workflow orchestration, long-running task management, memory and retrieval systems, agentic evaluation — this is meaningful validation that the use case is real. Someone with deep domain expertise and institutional credibility is betting his post-a16z career on AI agents handling the operational layer of a professional services firm. That signals genuine demand, not just early-adopter enthusiasm.

The biotech-specific data moat observation also carries weight for developers working on enterprise AI or vertical applications. Pande's argument is that in data-scarce domains, the scarce resource is proprietary structured data — and whoever controls it sets the floor of what AI can achieve in that vertical. This has direct implications for anyone building AI tools for clinical research, genomics, or drug development: access agreements and data partnerships may matter more than model selection, fine-tuning strategy, or inference speed.

What to watch is the broader pattern. VZVC's thesis — five concentrated bets per year, backed by deep expertise and an AI-augmented two-person operation — will take years to validate against the spray-and-pray funds it's implicitly competing against. But if the approach holds, expect more established investors to follow. The associate-track model that has defined VC operations for decades is facing its first serious structural pressure from AI, and the pressure is coming from someone who built one of the largest practices in the business.

Sources