Vijay Pande Left $4B a16z for a 5-Bet AI-Native VC

A dozen years ago, Vijay Pande was a Stanford chemistry professor best known for Folding@home — the distributed-computing project that turned millions of home PCs into a supercomputer for disease research. Then Marc Andreessen and Ben Horowitz, who had deliberately avoided healthcare and life sciences for their firm's first five years, handed him the keys to launch a biotech practice. Over the next decade-plus, Pande grew that bet into an operation managing close to $4 billion.

So it raised eyebrows when, in June last year, he walked away from it all to start something much smaller. His new firm, VZVC — co-founded with longtime investor Zach Werner — makes about five concentrated bets a year instead of dozens. It has no associates. And it runs heavily on AI for its day-to-day operations.

In an interview with TechCrunch this week, he explained the reasoning. The details are interesting for two different audiences: investors watching the pendulum swing back toward concentration, and developers watching AI agents quietly replace junior staff.

Five Bets, Not Thirty

When asked how concentrated "concentrated" really is, Pande was blunt:

We're not driving 30 bets per year... we're talking about probably five, not a lot of investments — very concentrated. Adding a company at a typical fund is like adding a Facebook friend — that's something you do pretty quickly. For Zach and I, it's more like wanting to have another child. This is a big deal for us.

The point isn't just portfolio size. It's a shift in relationship depth. At a typical fund an investment is a transaction; at VZVC it's a long-term commitment. Pande wants relationships that run "5, 10 years plus" across a founder's next company too, and he looks for founders with high integrity who think in terms of "how do we win together" rather than beating competitors.

That operating style also changes who they compete against. "The funny thing about this model is that typically we're not trying to compete for a hot round — people make room for us," he said. Teams want VZVC for the hands-on involvement, not for a logo on the cap table.

The Associates Got Replaced by Agents

The most interesting operational detail: VZVC was going to hire associates. It didn't.

We were actually intending on hiring associates, but it turned out, with the agents that we've built up, not to be something that we need to do.

It's the same pattern software teams have seen all year. The junior-analyst role — sourcing deals, summarizing research, drafting memos, tracking pipelines — is exactly the kind of workload that LLM agents can absorb. A two-person fund running a compact set of bets doesn't need headcount; it needs leverage, and agents provide it without the overhead of salaries, offices, and management.

The founder-level judgment still needs a human at the top — Pande and Werner — but the funnel work in between can be automated. For engineers this is a familiar division: the humans own the decisions, the agents own the busywork. It is fundamentally the same argument for keeping a human review step in AI pipelines: agents compress the work, humans keep the judgement.

That said, agents can also run away from you, as the OpenAI/Hugging Face agent swarm incident showed. The honest read: agent-native firms get all the upside and all the new failure modes.

Biology Is Now an Engineering Discipline

Pande's core thesis goes beyond firm structure. Biology, he argues, is shifting from a "science of discovery" to something you can actually engineer.

Drug development used to have a strong fortuitous element. Now, he says, AI and machine learning let computers grasp something very complicated: figuring out which targets a drug should hit, making the drug, and even helping with clinical trials — which remain the most expensive part of the process.

The numbers make the point. The cost and time to reach a trial keeps shrinking with AI, but actually running one can still cost hundreds of millions of dollars. And here's the brutal part: the probability of a drug making it from the first trial to the end of the third trial is just 20%.

If 8 out of 10 fail, and these things cost hundreds of millions of dollars, the amortized cost gets really high. The reason they fail typically is not that the biologist did something wrong; it's that all the experiments these drugs were designed on were on animal models like mice, and in the end, animal models are just not very predictive of humans.

Animal models are a proxy that doesn't generalize. AI models won't be perfect either, Pande argues — but they only need to beat mice.

The "Right Drug for Me" Problem

The next phase is precision medicine, which Pande describes as "is the drug the right drug for me?" Most medicine today works on population averages; your blood-test values are compared against everybody else's. The real comparison, he says, should be: is this result weird for you?

His favourite metaphor on why genomics alone isn't enough: your genome is like the blueprint of your house on day one, but your house is fairly different now compared with the moment it was built. Proteomics and other measurements capture where your body actually is now, which is more relevant to disease.

And over the last decade there's been a steady clip of progress on both sides of the stack — AI for biology ("how do we treat this disease?") and AI for chemistry ("how do we make a drug for that specific protein?").

The Moat Nobody Wants to Talk About

Here's the part that should interest every AI engineer: biology is one of the few places where AI can't just scrape data off the internet. With text, a model trains on the entire open web. Biology doesn't work that way — there's no single public corpus of disease data, and your data can't be distilled from one model into another the way you can re-derive an LLM's knowledge. Every company ends up building its own walled-off dataset.

That data scarcity is a real problem, and Pande is direct about it. The healthcare "fix" depends on data sharing, but founders and investors have good reason to protect their findings — creating the same territorial silos that already fragment medicine (an oncologist and an endocrinologist treating the same patient often don't sync well).

The reason for hope, he says, is open source. He points to a larger trend of "atlases of biological information" — which, from a technology standpoint, are foundation models. As they become more common, he expects the same dynamics that let open-source LLMs compete with the corporate ones:

I think we'll see the same thing that's happened with open-source LLMs, which do very well against the corporate ones: open-source foundation models in biology having a very broad impact.

That's a genuinely developer-adjacent thesis. The models winning in biology will be the ones with open benchmarks, accessible code, and shared data — not the walled gardens. If you've ever built on an open agentic model, you already know the playbook.

What He Got Right and Wrong

Pande's honest retrospective is worth reading in full, but two points stand out.

First, the long game pays off. Ten-plus years ago he was arguing that AI plus machine learning plus medicine mattered, and got plenty of resistance: "Oh, that's never going to happen. That's never going to be useful." That resistance is largely gone now, and the arc has been fulfilling.

Second, technology alone isn't the business. He admits it took him time to appreciate that "as seductive as the coolest technologies are, it really always comes back to go-to-market." His advice to science- and product-side founders is pointed: take all that brilliance and apply it to go-to-market, because it's at least as hard as—or harder than—the technology.

The Reality Check on AI Curing Everything

Finally, Pande says the thing AI hype-skeptics have been waiting to hear. The overhyped claim right now in AI and biotech is the "AI is going to cure everything" narrative. Not because he doubts AI — he doubts the data.

LLMs work because there's so much data to learn from. When the data is just simply not there, then AI can't magically solve that problem.

That's the honest summary of the whole conversation. AI is genuinely unlocking drug discovery, clinical trials, and precision medicine. But unlike text, where the answer to "more data" is "just scrape the web," biology demands that companies build, share, and open datasets first. The five-bet firm and the open foundation models are both reactions to the same reality: focus on what you can control, and build where the data exists.

Sources:

• TechCrunch — "We're not doing 30 bets a year": Vijay Pande on betting small
• VZVC — vz.vc