“We’re not doing 30 bets a year”: Vijay Pande on betting small after running $4 billion at a16z
Vijay Pande is trading a $4 billion war chest for a leaner, AI-native fund — and he argues the shift mirrors a fundamental change in how biology itself gets done. After stepping down from Andreessen Horowitz’s biotech practice to launch VZVC, Pande says the industry is finally moving from a "discovery" mindset, defined by serendipity and high-risk bets, to an "engineering" discipline where AI models design therapies with predictable properties. But he warns that the bottleneck has moved upstream: clinical trials remain brutally expensive, and the proprietary datasets hoarded by incumbents are the wrong fuel for the models that could fix the pipeline.
AfterQuery reportedly becomes Y Combinator’s fastest-ever unicorn, now valued… Xebia: Why AI agents fail without the right data… OpenAI aligns safety practices with EU AI Act’s GPAI…
Background: From Portfolio Theory to Platform Engineering
For years, biotech venture capital operated on a portfolio theory borrowed from pharma: place dozens of high-risk "bets" knowing most will fail, hoping a few blockbusters cover the losses. Pande’s former shop, a16z Bio + Health, deployed roughly $4 billion under this logic. VZVC, by contrast, is deliberately smaller and built around an AI-native thesis. Pande’s argument is structural: biology is becoming programmable. Advances in foundation models for protein folding, molecular generation, and cellular simulation mean researchers can increasingly *design* biological function rather than *screen* for it.
This shift changes the unit economics of early-stage R&D. When you are screening millions of compounds hoping one binds, you need massive capital for wet-lab throughput. When a model proposes a handful of high-probability candidates, the capital intensity drops — but the demand for high-quality, standardized training data rises. Pande contends that the venture model must shrink and specialize to match this new loop: faster iteration, lower upfront burn, tighter feedback between compute and bench.
Why It Matters: The Data Wall and the Trial Tax
The engineering metaphor holds only if the models see the right data. Pande identifies a critical mismatch: the most valuable biological data — clinical outcomes, longitudinal patient records, high-resolution omics — sits in silos. Hospitals, CROs, and large pharma treat these as competitive moats. But AI models trained on narrow, proprietary slices fail to generalize. Pande argues that open, shared datasets — think of the Protein Data Bank or GenBank for the clinical era — are the prerequisite for AI to actually de-risk drug development.
The second constraint is harder to code around. Even with perfect *in silico* design, clinical trials remain the dominant cost center. Phase 3 studies routinely run hundreds of millions of dollars and take years. AI can optimize protocol design, patient stratification, and site selection, but it cannot eliminate the regulatory requirement for human safety and efficacy data. For business owners and AI adopters, this means the ROI curve for "AI drug discovery" stays flat until the clinical bottleneck is addressed. Companies that pair generative models with novel trial designs — adaptive protocols, synthetic control arms, decentralized enrollment — will capture value that pure-play software shops cannot.
What's Next: Incentives, Infrastructure, and the New Fund Model
Pande’s bet is that a smaller, technically fluent fund can align incentives better than a massive generalist vehicle. VZVC plans to invest not just capital but compute infrastructure, data engineering support, and regulatory strategy — effectively acting as a co-developer rather than a passive check-writer. This mirrors a broader trend: specialized "studio" funds that embed operating teams alongside portfolio companies.
For the ecosystem, three signals matter. First, watch whether public-private partnerships (like the NIH’s Bridge2AI or EU’s EHDS) produce genuinely open, ML-ready clinical datasets. Second, track FDA guidance on AI/ML-enabled drug development — especially around qualification of in silico evidence for IND-enabling studies. Third, monitor if LPs start favoring concentrated, thesis-driven funds over broad biotech index exposure. Pande’s move is a signal that the "spray and pray" era of bio VC may be peaking.
Takeaway for Business Owners
If you are building or buying AI in life sciences, don’t optimize for the discovery headline. Optimize for the data loop: can you access diverse, longitudinal, high-fidelity clinical data? Can you run faster, cheaper trials — or partner with those who can? And does your capital structure match the new iteration speed? The winners won’t be the ones with the biggest models; they’ll be the ones who turn biology’s new engineering discipline into a repeatable, capital-efficient process.
Source: Original Article