Vijay Pande Left $4B at a16z to Run a Tiny, AI-Heavy VC Firm

In June 2025, Vijay Pande left Andreessen Horowitz after growing its bio and health practice from scratch to nearly $4 billion under management. His new firm, VZVC, co-founded with investor Zach Werner, operates with no associates, leans on AI for operations, and makes only a handful of concentrated bets per year instead of the 30-plus that larger funds typically chase. In a TechCrunch interview published August 29, 2026, Pande explained the pivot and flagged what he sees as a structural problem unique to AI in medicine: biological data cannot be pulled from the internet or shared between models.

What happened

Detail	Fact
Previous role	Led a16z’s bio/health practice, ~$4 billion AUM
Departure date	June 2025
New firm	VZVC, co-founded with Zach Werner
Investment pace	A handful of concentrated bets per year (not 30+)
Staff model	No associates; AI handles day-to-day operations
Pande’s prior background	Stanford chemistry professor; creator of Folding@home

Pande was a Stanford chemistry professor best known for building Folding@home, the distributed-computing project that recruited millions of home PCs to simulate protein folding for disease research. Marc Andreessen and Ben Horowitz brought him in roughly twelve years ago after spending their firm’s first five years deliberately avoiding healthcare. He turned that bet into one of the larger dedicated bio funds in venture.

The departure surprised people who watched him build it. His explanation is simple: at scale, you spread attention thin. A concentrated portfolio with AI handling back-office work lets a small team go deeper on fewer companies.

Why biology’s data problem is different from every other AI category

Most AI progress in text, code, and image generation is built on data that can be scraped, shared, and distilled from one model into another. Biological data does not work that way. Patient records, lab assays, and clinical measurements sit inside hospital systems and private biotech databases. No one can train a foundation model on biology the way OpenAI trained GPT on the open web.

According to Pande, this means nearly every biotech company building AI tools ends up constructing its own walled dataset. Those datasets cannot be distilled between models the way text models can be compressed and shared. The result is a field where proprietary data is the competitive moat, not the model itself. That is structurally interesting for investors and structurally frustrating for the broader goal of sharing medical advances.

Pande also drew a parallel to a long-standing problem in medicine: specialist silos. A patient whose condition spans oncology and endocrinology may see two doctors who never coordinate. Proprietary AI datasets reproduce that silo problem at a data layer, not just a human one. Those tracking how AI agents handle access control will recognise the pattern: data boundaries that made sense for humans can actively block AI from delivering its promised value.

The drug development numbers that explain the opportunity

Pande put a specific number on the core problem AI is trying to solve: only 20% of drugs that enter phase 1 clinical trials ever make it through phase 3 and reach patients. Eight out of ten fail, and each trial can cost hundreds of millions of dollars. The amortized cost of all those failures is why drugs end up expensive.

The main reason for failure is not bad science. Most drugs are designed using animal models, particularly mice, and animal models are poor predictors of human outcomes. Pande’s argument is that AI models, even imperfect ones, will clear that bar. Once they do, the cost and time profile of drug development changes significantly.

He also pointed to precision medicine, which he calls “precision medicine” in industry jargon, as the next layer. Right now a doctor compares your blood test results to population averages. A better system would compare them to what is normal for you specifically, and pick the first drug that actually fits your biology rather than iterating through several that do not. Genomics was supposed to deliver this but, as Pande notes, your genome is “like the blueprint for your house on day one” while your body is the house after years of renovation. Proteomics and other newer measurement layers are closer to capturing current state.

Our take

The “small fund, concentrated bets, AI-for-ops” pitch is becoming familiar in venture. But the biological data point Pande raises is genuinely underappreciated outside biotech circles, and it has a direct parallel for any business building AI on proprietary data.

If your competitive advantage depends on a dataset no one else can replicate, that is a real moat. But it also means your AI only gets smarter as fast as your data collection operation grows, not as fast as the public models do. Businesses thinking about AI integration should take stock of which of their operational data is genuinely proprietary and whether they are capturing it in a structured way right now, before a larger competitor does.

The VZVC model itself is worth watching for a different reason. No associates, AI for operations, and a handful of bets is essentially a lean-agency structure applied to venture capital. It works if the judgment is sharp and the AI tooling actually replaces process overhead. Whether it scales to a second fund without adding headcount will be the real test.

What to do about it

1. Audit what proprietary data your business generates that competitors cannot easily replicate: customer behaviour, operational logs, service outcomes.
2. Start storing it in a structured format now, even if you have no immediate AI use case. Retrofitting structure later is expensive.
3. When evaluating AI tools for your industry, ask the vendor whether their model was trained on public data or domain-specific private data. The answer tells you how much of the heavy lifting you will still have to do.
4. Follow the VZVC portfolio companies as a signal for where serious money is placing concentrated bets in AI-driven bio and health, and consider whether analogous dynamics exist in your sector.

Proprietary data is the new distribution. Build the pipeline to collect it before you need the model to use it. If you want to talk through what that looks like for your business, reach out to the Lumien team.

Source: TechCrunch · AI