The company that built its name on video-game graphics now wants to teach robots to operate. In March, NVIDIA released more than 700 hours of surgical video as an open dataset, plus an AI model that turns a typed instruction into robot motion.
For medical students and residents, this is not just a tech story. It is a training story. The AI arriving on the wards in the next few years will draft notes, help plan radiation and guide surgical robots, and someone will have to check its work. That someone is you.
From gaming chips to the operating room
At its GTC conference in San Jose, NVIDIA launched what it called the first domain-specific physical AI platform for healthcare robotics, according to NVIDIA's announcement as reported by GamesBeat. Three pieces matter:
- Open-H: more than 700 hours of surgical video gathered with about three dozen collaborators. NVIDIA calls it the largest healthcare robotics dataset.
- Cosmos-H: models that generate realistic surgical video from text prompts, images or robot motion data, so a robot can rehearse on synthetic cases before it ever meets a patient.
- GR00T-H: a vision-language-action model, trained on Open-H, that converts a text command describing a clinical task into robot movement.
This is already more than a demo. CMR Surgical, which contributed most of the surgical data in Open-H, is using Cosmos-H simulation to train and validate robotic intelligence for its Versius system before clinical deployment. Johnson & Johnson MedTech is using Isaac Sim and Cosmos-based workflows to train and validate its Monarch Platform for Urology, according to NVIDIA's robotics press release.
The $1 billion bet on drug discovery
The surgery news followed a January splash. At the J.P. Morgan Healthcare Conference, NVIDIA announced a joint research lab with Eli Lilly, with $1 billion committed over five years for compute, talent and infrastructure, plus a collaboration with Thermo Fisher Scientific on autonomous laboratories, TechRepublic reported. NVIDIA also told the room that U.S. healthcare is adopting AI at nearly three times the pace of the broader economy. That is the company's own claim, not an independent measurement.
Berlin: AI that designs proteins and plans radiation
Then, on October 21, NVIDIA went further at GTC Berlin. In its announcement, the company unveiled:
- RFD4-Proteina, an open BioNeMo model built with the UW Medicine Institute for Protein Design that lets scientists and AI agents design and refine new proteins against chosen goals.
- A radiotherapy partnership with Siemens Healthineers, which is exploring NVIDIA's Nemotron models to help oncologists evaluate and refine treatment plans and plans to use NVIDIA Holoscan to adapt treatment in real time.
- AI scribes and clinical assistants in Europe, as companies including Corti, Dedalus and Sword build on Nemotron.
Kimberly Powell, NVIDIA's vice president of healthcare, framed it plainly: “Agentic and physical AI are opening new possibilities across medicine.”
Who is actually using this?
NVIDIA's own State of AI in Healthcare and Life Sciences survey, published in February, found that 70% of respondents were actively using AI, up from 63% in the previous survey. Generative AI use jumped to 69% from 54%, and 47% said they were using or assessing AI agents. The company says the report draws on more than 600 industry respondents.
Read the fine print, though. The survey comes from the company that sells the hardware, the answers are self-reported, and the respondents are industry professionals, not a random sample of hospitals. It shows direction, not proof that patients are better off.
Washington wants AI agents too
The agent idea has reached Medicare. At the HIMSS conference in Las Vegas in March, CMS Administrator Dr. Mehmet Oz said the agency had discussed giving an agentic AI tool to every Medicare beneficiary, Nextgov reported. Not everyone is sold. NPR reported that critics, including a University of Minnesota rural health researcher, warned that AI avatars for rural care could erode trust and turn rural patients into test subjects for unproven technology.
The catch: AI can be confidently wrong
Medicine has seen this movie. A widely used proprietary sepsis model, tested independently at Michigan Medicine, reached an area under the curve of just 0.63 and missed 67% of patients who developed sepsis (Wong et al., JAMA Internal Medicine, 2021). It was already in wide use. Faster chips do not fix a model that was never tested on patients like yours.
What it means if you are in training
- AI will draft; you will verify. Catching the anchoring error, the missed red flag or the bad drug interaction in an AI draft is clinical reasoning, and it is the skill that makes the tool safe.
- Ask who built it. NVIDIA supplies models and infrastructure. The device maker or software vendor is responsible for the clinical product, and you should know whether it is FDA-cleared for the use in front of you.
- Learn the robot's limits. Systems trained partly on synthetic video will reach operating rooms. Know what a system was trained on and how to take over when it misbehaves.
- New careers are opening. Clinical informatics and imaging AI need clinicians who speak both languages. Our guide to AI in medicine covers the informatics pathway.
Cheat sheet: NVIDIA's healthcare names
| Name | What it does | Where you may meet it |
|---|---|---|
| NVIDIA Clara | The umbrella brand for NVIDIA's healthcare and life sciences software | Behind imaging, genomics and device software |
| MONAI | Open-source framework for medical imaging AI | Radiology and pathology research |
| BioNeMo | Models for biology and drug discovery, including protein design | Translational labs and pharma |
| Holoscan | Real-time AI for devices that stream sensor data | Endoscopy, surgical video, radiotherapy |
| Parabricks | Accelerated genomics analysis | Clinical sequencing |
| Nemotron | Open language models used to build AI agents | AI scribes and documentation assistants |
| Cosmos-H, GR00T-H | Synthetic surgical data and robot models | Robotic surgery and simulation centers |
Definitions are based on NVIDIA's description of its Clara platform and the 2026 announcements above.
What to watch next
- Whether robots trained partly in simulation win regulatory clearance for new tasks.
- Whether CMS turns the beneficiary AI-agent idea into a formal pilot, and what safeguards come with it.
- NVIDIA's next GTC in San Jose, which the company usually holds in March. Find AI and informatics meetings on the MedRise conference list.
NVIDIA is building the machines. Medicine still needs people who can tell when the machine is wrong. You can practice that skill now with MedRise Viva.
MedRise is not affiliated with NVIDIA, CMS or any company named here, and this article is not sponsored. It is education only and is not medical or legal advice.
