Vijay Pande’s New VC Firm Built Its Own AI Agents Instead of Hiring Associates

Vijay Pande spent a decade building Andreessen Horowitz’s Bio + Health practice into a business managing close to $4 billion. In June 2025, he walked away from it to start something much smaller. His new firm, VZVC, co-founded with investor Zack Werner, runs on a handful of concentrated bets a year instead of dozens, carries no associates, and leans on AI agents to do work that would normally require a small team. In a new interview with TechCrunch, Pande laid out why he chose to go smaller just as the venture industry is racing to deploy AI at scale, and why he thinks the biggest obstacle to AI curing disease has nothing to do with model quality.

From $4 Billion to Five Bets a Year

Pande became a16z’s ninth general partner in 2015, when he founded the firm’s Bio + Health Fund after more than a decade as a Stanford professor of chemistry, structural biology, and computer science, according to his biography. He was already known in scientific circles as the creator of Folding@home, the distributed-computing project that turned millions of home PCs into a volunteer supercomputer for disease research. Under Pande, the practice raised a new fund in early 2022 and rebranded as “a16z Bio + Health,” after making more than two dozen new investments the prior year that nearly doubled the size of its overall portfolio, according to a16z’s own announcement of that raise, and it grew into a practice TechCrunch describes as managing close to $4 billion by the time Pande departed.

VZVC looks nothing like that. According to STAT News, the firm is currently raising its first fund, and a source told the outlet it could reach up to $400 million, a small fraction of what Pande managed at a16z. On the investment side, it is just Pande and Werner. “We were actually intending on hiring associates,” Pande told TechCrunch, “but it turned out, with the agents that we’ve built up, not to be something that we need to do.” The firm never filled the junior role it had planned for, because the AI tooling it built for its own workflow already covered the gap.

The bet sizing follows the same logic. “We’re [not] driving 30 bets per year,” Pande said. “We’re talking about probably five: not a lot of investments, very concentrated.” He drew a sharp line between that approach and how a typical fund operates: “Adding a company at a typical fund is like adding a Facebook friend,” something done quickly and without much deliberation. For Werner and him, he said, it is closer to “wanting to have another child.” That structure also changes who VZVC is up against for a deal. Rather than fighting for a hot, competitive round the way larger funds do, Pande said, “people make room for us.”

Why Biology Cannot Just Scrape the Internet

The more technical half of the interview centers on a problem Pande has spent years thinking about: unlike the text that trained large language models, biological data was never sitting on the open internet waiting to be scraped. Pande has described biology as moving from a “science of discovery” to something that can be engineered, and he told TechCrunch that shift comes from AI helping identify drug targets, design candidate molecules, and support clinical trials, which he called “the most expensive part of the process.”

He pushed back on the idea that this has already made trials cheap. “That’s, I think, very much an aspiration,” he said, noting a real trial can still cost hundreds of millions of dollars, and that only about 20 percent of drugs that enter human trials make it through all three phases. The reason so many fail, in his view, is not that researchers did something wrong: it is that the animal models most drugs are tested on beforehand are poor stand-ins for human biology. “The AI model is not going to be perfect,” he said, “but it’s going to be way better than any animal model would be.”

That is where the data problem bites hardest. Because biological data cannot be scraped the way text can, Pande said, “you don’t have any of this data that people can just all train the same thing, and your data can’t be distilled from one model to another.” Every company ends up building its own proprietary, walled-off dataset instead of drawing on a shared pool the way language-model developers do. Pande pointed to open-source biological foundation models as one path around that, similar to how open large language models have narrowed the gap with proprietary ones, though he stopped short of saying the data-sharing problem is close to solved.

A New Baseline for Precision Medicine

Pande also described a shift in what “normal” means in medicine. Blood test results and other diagnostics have traditionally been measured against population averages, he said, but the more useful comparison is against a given patient’s own baseline: whether a result is unusual for that specific person. He credited part of the shift to proteomics and other measurements that go beyond genomics, plus a wave of automated, robotic lab measurement that pairs naturally with AI analysis, over a decade of steady progress in both the biology of disease targets and the chemistry of drugs designed to hit them.

What Pande Thinks Is Overhyped

Asked what is overhyped in AI and biotech right now, Pande did not point to a specific product or claim. He pointed to the framing itself. “The thing that always gets tricky is when there’s this call that AI is going to cure all everything,” he said. “The reason for hesitance there is not because of any doubt about AI. It’s about doubt of the data.” Large language models work, in his telling, because there is an enormous amount of text to learn from; biology does not have that luxury yet, and no amount of model improvement changes that on its own.

It is a notably restrained message from someone whose new firm is betting its entire strategy on AI, both as an investment thesis and as the operating model replacing the associates VZVC never hired. Pande’s own account suggests the two halves of that bet are connected: the firms he is most interested in backing are the ones that treat AI as a tool for getting real biological data, not a substitute for it.