Why Vijay Pande left a $4B bio practice to make five concentrated bets
Vijay Pande built and ran the bio practice at Andreessen Horowitz and is reported to have overseen close to $4 billion in that role. In June of last year he left to co‑found VZVC with Zach Werner, and deliberately shrank the playbook.
His thesis is concise: in biology, careful execution and proprietary data often matter more than a high‑volume investment cadence. So instead of “30 bets a year, ” VZVC aims for a handful, Pande says “probably five, ” and a lean operating model powered by AI agents.
Why five bets, not thirty?
Drug development is capital‑intensive and failure‑prone. Pande summarizes the arithmetic bluntly: “The probability of a drug going successfully from the first trial to the end of the third trial is just 20%.” He adds, “If 8 out of 10 fail, and these things cost hundreds of millions of dollars, the amortized cost gets really high.”
Those statements reflect widely reported industry realities: late‑stage trials account for the bulk of development costs and program attrition has long been a drag on returns, a point documented in analyses from groups such as the Tufts Center for the Study of Drug Development and trade organizations like PhRMA and BIO. Faced with that financial math, VZVC trades breadth for depth. They back fewer companies, build deeper partnerships, and offer active help with clinical and go‑to‑market execution.
Operating lean, and why AI agents matter
VZVC runs with a very small human team: Pande notes the firm has “no associates.” Instead they lean on software agents to handle routine workflows so partners can focus on high‑value work with founders.
Common, realistic agent use cases in a bio VC: automated literature synthesis that extracts relevant biomarkers and trial endpoints; continuous monitoring of clinical registries for recruitment or competing trials; automated data‑quality checks on vendor assay readouts; and alerting on IP or regulatory changes that affect a portfolio company. VZVC says these agents handle day‑to‑day operations, freeing partners to mentor founders and solve commercialization bottlenecks.
That model makes sense when the portfolio is small. If you commit to five long‑horizon investments and expect founder relationships to last “5, 10 years plus, ” as Pande does, it’s efficient to automate low‑signal tasks and focus human attention where it moves the needle.
AI helps, but data drives results
Pande’s scientific background, as a Stanford chemistry professor who built Folding@home, shapes his view of AI in biology: the field has seen “very significant advances” over the past decade, but biology is not text.
“The thing that always gets tricky is when there’s this call that AI is going to cure all everything. The reason for hesitance there is not because of any doubt about AI, it’s about doubt of the data.”, Vijay Pande
Large language models grew because massive text corpora are publicly available. Biology depends on experimental datasets, sequencing runs, high‑throughput screens, imaging, and clinical readouts that are costly to produce and often proprietary. Public repositories such as GenBank or GEO exist, but many of the highest‑value clinical and assay datasets become “walled‑off” assets. That difference shapes who can train foundation models in biology and who captures downstream value.
Pande expects biological “atlases” or foundation models, large pre‑trained models trained on multi‑omic and phenotypic data, to emerge and have major impact. He also anticipates open‑source versions of those models will matter, similar to how open‑source LLMs influenced language AI. The caveat: open foundation models are only as useful and fair as the data behind them and the governance around access.
Clinical trials remain the bottleneck
Even when models improve early‑stage predictions, clinical validation stays expensive. Pande warns that late‑stage trials can still cost “hundreds of millions of dollars, ” and the high failure rate means expected costs per approved drug stay large unless failure rates decline materially.
AI can reduce specific risks, such as better patient stratification, earlier toxicity signals, and more predictive preclinical models, but it is not a substitute for prospective human trials. Expect incremental improvements folded into established regulatory pathways, not a sudden elimination of phase III testing.
What VZVC is focused on
- AI for healthcare delivery: Tools that change how care is delivered at scale, decision support, workflow automation, and better triage for therapies.
- AI for clinical trials: Design, recruitment, endpoint selection, and monitoring, areas where better prediction and operational tooling can reduce time and failure risk.
Pande emphasizes founder quality above flashy tech. He looks for long horizons and integrity, and says, “For Zach and I, it’s more like . . . wanting to have another child. This is a big deal for us, ” signaling deep, emotional and operational commitment to the companies they back.
Practical actions for founders, investors, and AI teams
- Founders: Prepare a two‑page clinical execution plan and a data‑governance memo for fundraising conversations, investors will scrutinize how you will access, curate, and protect high‑value datasets.
- Investors: If you choose a concentrated strategy, list three operational capabilities you will provide (e.g., clinical ops, regulatory strategy, data engineering) and be ready to commit time over years, not months.
- AI teams: Prioritize partnerships to access at least one longitudinal or prospectively collected clinical dataset and plan for prospective validation on held‑out cohorts, retrospective performance alone won’t persuade regulators or clinicians.
Signals to watch
- Availability of well‑benchmarked, multi‑omic atlases with reproducible held‑out test sets and public leaderboard-style evaluations.
- Evidence that AI‑augmented trials measurably reduce failure rates or shorten timelines in a repeatable, peer‑reviewed way.
- Emergence of sustainable governance and business models that reconcile proprietary datasets with open foundation models, clear licensing, privacy protections, and incentives for data sharing.
Key questions, and short answers
-
Why did Pande leave a large bio practice to start a tiny firm?
He wanted deeper, long‑term engagement with a small number of companies and believes a lean team augmented by AI agents can support very concentrated investments, Pande says “probably five” a year.
-
Can AI replace clinical trials?
No; AI can improve prediction, patient stratification, and trial design, but clinical validation remains required and expensive, the industry still faces high late‑stage costs and attrition.
-
How is biology different from LLMs?
Biology depends on experimentally generated datasets (sequencing, screens, clinical readouts) that are often proprietary, unlike the massive public text corpora that enabled LLMs.
-
Will open‑source foundation models in biology matter?
Pande expects them to have broad impact similar to open‑source LLMs, but their usefulness depends on data access, governance, and who controls high‑value experimental measurements.
-
What founder traits does VZVC prioritize?
Long time horizons (5-10+ years), high integrity, and the ability to execute clinical and go‑to‑market strategies, not just build better models.
Pande’s move is a corrective to two extremes: high‑volume investing with shallow engagement, and the notion that AI alone will instantly make drug development cheap and fast. His approach responds to biology’s economics and the current distribution of data, and when the stakes and costs are this high, quality of execution, data access, and long‑term partnership often beat quantity.