Nvidia bets on agentic AI to turbocharge biotech discovery

Artificial intelligence played a prominent role at this week’s Bio International Convention in San Diego, the largest biotech event with vendors spanning the full ecosystem of companies in this industry. Today in a special address, Kimberly Powell (pictured), vice president and general manager of healthcare and life sciences at Nvidia Corp., made the case that agentic AI is about to do for biotech what it just did for software — and the company’s BioNeMo is the stack that turns generic large language models into working “AI scientists” that are both faster and cheaper to run. Nvidia wants to make ‘AI scientists’ mainstream in biotech Powell opened her presentation by outlining where the industry is now. “We are witnessing the fastest platform shift the life sciences industry has ever seen,” she said. She compared AI to the microscope, X-ray crystallography, and gene sequencing, calling them a new class of scientific instruments. This time, the instrument doesn’t just see or measure; it reasons, plans and acts. At the event, Nvidia announced its BioNeMo Agent Toolkit, a software stack that turns large language models into domain-specific AI agents capable of executing end-to-end biology and chemistry workflows — from literature review to protein design to lab automation — while optimizing for performance and cost. From generative to agentic AI for science Powell’s core thesis is that the life sciences, a $300 billion annual pharmaceutical budget (global R&D is reaching $3.8 trillion), have quietly been preparing for this inflection for a decade. On one side, there has been an explosion of AI research in biology, chemistry, imaging and genomics. On the other hand, Nvidia has been building the infrastructure to operationalize that research: GPUs, networking, CUDA-X libraries and domain platforms such as MONAI, Parabricks, cuEquivariance and BioNeMo. What has changed in the last 12 to 18 months is the emergence of agentic AI, systems in which a large language model “brain” is wrapped in a harness that manages tools, memory, security policies and multistep workflows. Nvidia’s NeMo Curator and NemoClaw framework and open-source harness are generic versions of that pattern; the BioNeMo Agent Toolkit is the life-sciences-optimized edition. “Agents are becoming the modern application layer in life sciences,” Powell said. “Every single one of the thousands of companies in life sciences is about to become an agent builder.” That’s a very different framing than “just another model.” It says the next application tier in biotech won’t be GUIs and pipelines, but rather networks of specialized agents coordinating work across digital and physical labs. BioNeMo as the scientific toolbox — tuned for speed and cost Nvidia’s announcement positions BioNeMo as the science that sits behind those agents. In practice, the BioNeMo Agent Toolkit does three important things for biotech teams: Packages proven life-science models, such as protein folding, molecular docking, generative chemistry, genomics and imaging, into agent-callable tools with clear schemas: what each tool does, what inputs it requires, what outputs to expect and how to troubleshoot. Exposes those capabilities via NIM microservices that can run on-premises, in the public cloud or across hybrid environments, so pharma and biotech can place compute where data and regulatory constraints demand. Optimizes for token efficiency and computational cost, not just raw accuracy, by giving agents access to highly accelerated libraries and models, so they spend fewer tokens and less wall clock time hunting for the right tool or rerunning failed steps. Powell specifically addressed the historical cost-performance trade-off. She described BioNeMo’s skills and tools as “the knowhow” that lets agents complete complex workflows with “strong task completion, workflow accuracy, and reduced token expense — that means less compute, more reliable results.” In other words, a BioNeMo-enabled agent doesn’t just produce better science; it does so with fewer LLM calls and more efficient graphics processing unit usage, making cost and performance optimization possible at the same time. Powell emphasized that BioNeMo is agent-agnostic. The same toolkit can serve agents built on OpenAI, Anthropic, in-house LLMs or Nvidia’s own Nemotron models. That matters for buyers who don’t want their next decade of drug discovery workflows locked to a single model vendor. What an AI ‘co-scientist’ looks like in practice To ground this in something beyond architectural diagrams, Powell walked through a protein-binder design workflow targeting MCL1, a protein that helps tumor cells survive. Traditionally, that path — understanding the target to generating binders, predicting structures, scoring candidates and deciding what to synthesize — takes months of specialized human effort. A generic agent can attempt that workflow but will burn time and tokens “searching for the right tools, figuring out how to call them and oftentimes completely failing to complete the task.” With BioNeMo, Powell said, a scientist gives a single goal such as “Design a binder for MCL1,” and the agent: Retrieves or predicts the target structure and its binding region. Generates candidate binders using BioNeMo generative models. Folds the target and binder together, then evaluates docking poses using accelerated structural engines. Ranks and returns the top candidates for human review — “all done without human intervention.” This is the “AI scientist” pattern many startups are pursuing. The key nuance is verification. Panelist Andrew White, co-founder and chief technology officer at Edison Scientific, noted that as agents improve, “the era of humans writing questions and agents taking the test is over. We really do need this kind of lab-in-the-loop.” His takeaway: The true bottleneck is shifting from reasoning about existing literature to running new experiments, which is exactly where closed-loop digital and robotic labs come in. Why this matters for biotech and pharma For biotech leaders, the strategic implications are less about any single toolkit and more about the operating model shift Powell and the panelists described: Compression of timelines. Powell argued that agents will “take scientific discovery and shrink the timeframe” — work that took years moves to months, and months to days. Josh Meier, CEO of Chai Discovery, gave a concrete example. Antibody design success rates have risen from one in 1,000 to 10% to 15% in just a few years, driven by improved models and faster iteration. Rising expectations on wet-lab speed. As in-silico design compresses from months to hours of GPU time, lab workflows become the new bottleneck. Meier pointed out that many assays were never optimized for speed because there was no incentive; now, tightening that loop is a competitive necessity. New collaboration patterns: Powell sees pharma shifting from primarily “deep scientific relationships” to partnerships that integrate frontier AI labs, tool providers, and platform companies within closed-loop systems — where every experiment feeds back into proprietary foundation models and agents. Benchling CEO Sajith Wickramasekara echoed this, arguing that electronic lab notebooks are evolving from retrospective records into “systems of action” co-authored by AI. Lowering barriers and de-siloing science. Powell believes tools like BioNeMo will let biologists tap into advanced modeling “in a natural language way, instead of having to get into any type of coding at all,” breaking down silos between disciplines and making modern AI tools accessible to more of the bench. That last point is worth watching. If AI agents can reliably orchestrate highend modeling and workflow automation behind a conversational front end, the practical distinction between “computational biologist” and “wetlab biologist” starts to blur. Reading the signal for the road ahead From an industry watcher’s perspective, BIO 2026 is less about Nvidia “entering” life sciences, since it has been here for a decade, and more about standardizing the agentic stack for biotech before others do. The BioNeMo Agent Toolkit turns Nvidia’s existing beachheads, such as MONAI, Parabricks, cuEquivariance and BioNeMo models, into a coherent runtime that any agent harness can plug into, with clear value props for speed, accuracy, and cost. The open-source angle is also notable. Powell made it explicit that the toolkit is available on GitHub and is designed to work with both open- and closed-frontier models, giving pharma and biotech the option to build their own domain-specific “brains” on top of Nvidia’s toolbox. In a world where IP, data residency and regulator trust are existential concerns, that flexibility will matter. Powell closed with an ambition that neatly captures Nvidia’s posture: “Agentic AI has revolutionized coding — that’s a done deal. Now this ecosystem is assembling to revolutionize science as we know it.” For biotech leaders, the question is no longer whether AI can help science, she argued, but “does AI have the right instruments to run science?” With the BioNeMo Agent Toolkit, Nvidia is betting that the answer for a growing slice of the industry will be yes. Zeus Kerravala is a principal analyst at ZK Research, a division of Kerravala Consulting. He wrote this article for SiliconANGLE. Photo: Zeus Kerravala A message from John Furrier, co-founder of SiliconANGLE: Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities. 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network. About SiliconANGLE Media