NVIDIA’s Healthcare AI Stack in 2026: From GPUs to AI Agents

Posted on by IE Contributor

For most of the last decade, “NVIDIA in healthcare” meant one thing: fast GPUs that helped radiologists and researchers crunch images and genomes. That picture is now badly out of date. Over 2026, NVIDIA has rebuilt its healthcare business around open models, AI agents that can run scientific software on their own, and real-time computers that sit inside medical devices. The latest round of announcements, made at GTC Berlin on October 21, 2026, shows where this is heading: from tools that help scientists to systems that work alongside them.

Here is what NVIDIA’s healthcare platform looks like today, what changed this year, and what it means for hospitals, drug developers and the companies that build for them.

Entrance to NVIDIA GTC 2025
NVIDIA’s GTC conference, where many of its healthcare platforms are announced. Photo: Jack145945, CC BY-SA 4.0, via Wikimedia Commons.

The Stack: Five Platforms, One Strategy

NVIDIA still uses the Clara name for its family of healthcare models, but the platform has grown into several distinct layers. According to NVIDIA’s healthcare and life sciences platform page, the core pieces are:

  • BioNeMo, the development platform for AI-driven biology and drug discovery, bundling open models, libraries, datasets and NIM microservices. NVIDIA claims 2x faster biofoundation model training and 6x faster inference.
  • Nemotron, NVIDIA’s family of open models, which the company now positions as “the foundation for digital health AI,” with ambient clinical agents and deep research agents as the headline use cases.
  • Isaac for Healthcare, a robotics and simulation platform for surgical autonomy, autonomous ultrasound and hospital automation.
  • Holoscan, a real-time runtime that ingests high-bandwidth sensor data, runs AI on it and drives deterministic actions inside medical devices.
  • Parabricks, GPU-accelerated genomics software that NVIDIA says analyzes whole genomes more than 100x faster at roughly half the compute cost.

The common thread, as HPCwire observed after GTC in March, is an interconnected stack in which progress in one layer, such as faster protein modeling, is meant to reinforce progress in the others, such as simulation or robotics.

Drug Discovery Goes Agentic

The biggest shift of 2026 is in drug discovery. In January, at the J.P. Morgan Healthcare Conference, NVIDIA announced a major BioNeMo expansion, including new Clara open models such as RNAPro for RNA structure prediction and ReaSyn v2, which checks whether AI-designed molecules can actually be synthesized. The same release revealed a co-innovation lab with Eli Lilly, part of an investment of up to $1 billion over five years in talent, infrastructure and compute, and a collaboration with Thermo Fisher to make lab instruments smarter and labs increasingly autonomous.

In June, NVIDIA went further with the BioNeMo Agent Toolkit. The idea is simple but significant: general-purpose AI agents struggle with scientific software because they have to guess which tools to use and how to interpret results. The toolkit packages more than a decade of NVIDIA life sciences libraries and models as “agent-callable” skills for tasks such as virtual screening, protein binder design, genomic analysis and medical imaging analysis. NVIDIA says more than 50 companies were already using it at launch, with Anthropic and OpenAI integrating it and Lilly, Natera, Schrödinger, Databricks and Snowflake among adopters. Jensen Huang summed up the pitch: frontier models are the brains, and BioNeMo is the scientific toolbox.

GTC Berlin added the most eye-catching piece yet. NVIDIA and the University of Washington’s Institute for Protein Design released RFD4-Proteina, an open BioNeMo model that combines programmable protein design with search. In the research draft cited by NVIDIA, it beat the evaluated baselines on design success rate and throughput across protein binding, small-molecule binding, enzyme scaffolding and DNA binding benchmarks. It is available through partners including Amazon Bio Discovery, Google Cloud, Hugging Face, Benchling, Databricks, OpenAI and Anthropic. Nobel laureate David Baker framed it as a move from discovering the proteins nature made to designing entirely new ones.

From the Lab Bench to the Treatment Room

NVIDIA’s ambitions extend well past the research lab. At GTC in March, the company introduced a set of surgical robotics resources, reported by HPCwire: Open-H, a dataset of more than 700 hours of surgical video; Cosmos-H models for generating synthetic surgical data; GR00T-H, a vision-language-action model that turns clinical instructions into robot actions; and Rheo, a framework for simulating hospital environments. CMR Surgical, Johnson & Johnson MedTech and Moon Surgical were among the companies contributing data or testing control policies in simulation.

In cancer care, the October announcements highlighted a collaboration with Siemens Healthineers on adaptive radiotherapy. According to NVIDIA’s GTC Berlin release, Nemotron models are being explored for “agentic co-planning” to help oncologists refine treatment plans, combined with CUDA-accelerated dose calculation and Holoscan to connect live sensor data with AI during treatment. NVIDIA notes that roughly 35 million people a year are projected to be diagnosed with cancer by 2050, which is why specialist capacity matters so much.

Hospital operations are the third front. European companies including Avelios, Corti, Dedalus, Dawn Health and Sword are building AI scribes and patient-support tools on Nemotron, while Apian and VitVio are using Cosmos world models and Omniverse libraries to build digital twins of hospital operations.

Infrastructure Still Matters

None of this works without serious compute, and NVIDIA’s customers are buying it. HPCwire reported in March that Roche is expanding to more than 3,500 Blackwell GPUs across hybrid cloud and on-premises environments, supporting foundation model training, drug discovery, diagnostics and manufacturing digital twins. Lilly’s DGX SuperPOD, which NVIDIA describes as the most powerful in biopharma, is the base for its new co-innovation lab, which plans to adopt next-generation Vera Rubin systems.

At the other end of the scale, Holoscan targets the edge. NVIDIA says the runtime scales from a Jetson Nano to a DGX and can move tens to hundreds of gigabits per second of sensor data directly into GPU memory, which is the kind of throughput needed to process raw ultrasound or imaging signals in real time.

Practical Takeaways

  • For biotech and pharma teams: the BioNeMo Agent Toolkit and open models like RFD4-Proteina are available now on GitHub and through major clouds, so experimenting with agent-driven screening or protein design no longer requires building the tooling from scratch.
  • For hospital IT and digital health builders: Nemotron is NVIDIA’s bet for clinical documentation and patient-facing agents. Open weights make on-premises deployment possible, which matters for patient data.
  • For medical device makers: Holoscan and Isaac for Healthcare provide a path from simulation to real-time deployment, but clinical validation and regulatory clearance remain the gating factors.
  • For investors and strategists: NVIDIA is positioning itself as the shared layer beneath frontier AI labs, cloud providers and life sciences software vendors, not as a competitor to any of them.
  • For everyone: read benchmark claims carefully. Results such as RFD4-Proteina’s are computational and come from a research draft; laboratory experiments still decide whether a design works.

Conclusion

NVIDIA’s healthcare story in 2026 is no longer about selling faster chips to hospitals. It is about building an open, connected platform where AI agents design molecules, robots learn surgical tasks in simulation and real-time systems adapt cancer treatment to each patient. As HPCwire cautioned, biological complexity, safety standards and regulation will slow the path from demo to bedside. But with Lilly, Roche, Siemens Healthineers and dozens of AI-native biotechs building on the stack, NVIDIA has made itself very hard to ignore in the future of medicine.

Sources