Nvidia and Abridge Build Healthcare AI Model for Clinical Workflows

Nvidia is collaborating with health tech startup Abridge to develop a foundation model specifically trained for clinical conversations, the companies announced Thursday. The model aims to enhance accuracy and reliability across documentation, workflow automation, and clinical reasoning support.

The new model will be built on Nvidia's Nemotron open model family, which makes both model weights and training data available to developers. Training will take place on Nvidia Blackwell AI infrastructure using de-identified clinical data across pre-training, mid-training, and post-training stages.

Training clinical knowledge from the ground up

Abridge executives emphasized that embedding clinical knowledge throughout all three training stages will improve the model's accuracy and precision across different medical specialties and care settings. By adapting the model to healthcare domains early in the training lifecycle, the company aims to build a system that can reason clinically from its foundation rather than having medical knowledge added as an afterthought.

The Nemotron architecture gives Abridge flexibility to optimize for quality, cost, and efficiency at different layers, allowing the deployment of appropriately scaled models for specific workflows. Nvidia is also an investor in Abridge through its NVentures venture capital arm.

Beyond AI scribes to clinical intelligence

Abridge has rapidly expanded beyond its initial AI medical scribe product. This week, the company announced platform integrations with payer and life sciences workflows, positioning itself as an "AI-native clinician intelligence platform" that connects care delivery, payment, and evidence-based treatment.

The company now serves 300 health systems and supports more than 100 million conversations annually. As it scales, Abridge CEO and co-founder Shiv Rao said controlling latency and accuracy across multiple use cases requires deeper involvement in the technology stack.

"Since we have a very maximalist sort of attitude around sprinkling intelligence everywhere we possibly can, it means we do have to reach down lower into the stack and control our destiny," Rao said during a keynote event in New York City on Thursday. "Latency is a big issue for us because we do want the clinician to be able to stop, swivel the chair, and have all of those artifacts there."

Why it matters

Generic large language models lack the clinical reasoning and domain expertise required for complex healthcare workflows. By training a foundation model specifically on clinical conversations from the ground up, Abridge and Nvidia are addressing a fundamental limitation in applying AI to medical settings. This approach could set a template for how vertical AI applications in regulated industries move beyond adapting general-purpose models to building domain-specific intelligence at the foundation layer.

Nvidia's healthcare expansion

Kimberly Powell, vice president of healthcare at Nvidia, said the company views healthcare as poised to become "one of the largest technology industries." She noted that AI has evolved through three major breakthroughs in the past 18 months: from generating content to reasoning through problems to performing work.

Nvidia has formed similar AI infrastructure partnerships with Verily, Innovaccer, Eli Lilly, Roche, Thermo Fisher Scientific, and Qiagen as it expands deeper into life sciences and medical technology.

"Every technological breakthrough requires you to look at the full stack," Powell said Thursday. "What we're recognizing together is it's time to go deeper in the stack—a clinical conversation foundation model, so that the complexity of healthcare and all of the workflows and the connectivity of this amazing ecosystem can be realized."

These details were first reported by Fierce Healthcare.