Nvidia, Abridge to build Nemotron AI for clinical conversations

Nvidia and medical technology company Abridge are developing an AI system based on Nvidia’s Nemotron models to analyze doctor‑patient conversations captured on Abridge’s platform. The model will be trained on de‑identified clinical recordings and run only within Abridge’s service. Abridge expects the system to join other AI tools on its platform later this year.

The partners plan to adapt the open Nemotron family with Abridge’s clinical expertise and data from its ambient‑listening technology, which records clinical visits and generates notes. Nvidia is already a backer of the Pittsburgh‑based startup. Abridge was founded in 2018 and raised $300 million in a recent funding round that set its valuation at $5.3 billion.

Nvidia’s healthcare lead, Kimberly Powell, described the work as an effort to add clinical knowledge to general AI models earlier in their development cycle. The training focuses on medical terms, the flow of clinical dialogue and elements of clinical decision-making that general-purpose models may miss.

Dr. Shiv Rao, Abridge’s co‑founder and chief executive, said clinical intelligence requires training and testing on real clinical data: “Generic models are powerful, but clinical intelligence still has to be trained, it has to be shaped, and it has to be evaluated against real‑world conditions. That’s a lot of our early experiments and work with the Nemotron team.” He added the company will use de‑identified content from its platform to customize the models.

Abridge’s director of applied science, Davis Liang, explained the company previously built its own speech‑recognition model because available third‑party services did not work well in hospital settings or with medical vocabulary. Liang also noted that smaller open models can be less costly than proprietary alternatives and easier to deploy on Abridge’s infrastructure.

The model will be deployed exclusively inside Abridge’s software rather than licensed to other vendors. Abridge says it will operate multiple underlying AI systems, with the Nemotron‑based model acting as one component of the updated platform.

Abridge’s software is already used at large health systems. Emory Healthcare rolled the platform out to more than 3,000 physicians. Emory’s chief executive, Dr. Joon Lee, said he expects more capable AI tools to support clinicians during visits and help capture information from conversations in near real time.

Abridge has stated it will use privacy techniques to separate personal identifiers from clinical content and will follow applicable regulatory and privacy requirements when training models on de‑identified data.

Nvidia has positioned its open model families for customization by partners in areas including drug discovery, medical devices and digital health. The Abridge collaboration adds a clinical‑conversation use case to Nvidia’s healthcare initiatives.

The companies provided the timeline as the platform update is planned for later this year and did not disclose further technical details or commercial terms.