What you should know:
– Nvidia today unveiled a massive expansion of its healthcare AI initiatives, implementing new open models as part of the NVIDIA Clear Platform and announce key partnerships with industry giants like lily and Innovaccer.
– These advances will accelerate scientific discoveries, transform complex surgical procedures using digital twins, and incorporate real-time intelligence into clinical and operational workflows.
Pharma 4.0: Lilly builds the world’s largest artificial intelligence factory
lilyalready recognized as the number one most AI-ready pharmaceutical company, is deploying the world’s first NVIDIA DGX SuperPOD AI factory, owned and operated exclusively by a pharmaceutical company.
- Immense computing power: Built with 1,016 NVIDIA Blackwell Ultra GPUs, the AI Factory delivers more than 9,000 petaflops of AI performance, poised to compress drug discovery timelines and enable advances in personalized medicine and genomics at industrial scale.
- Drug discovery and development: The AI factory will be used to train large-scale biomedical foundation models. Some models will be available from Lilly TuneLab, which is now the first drug discovery platform to offer open-base Lilly and NVIDIA Clara models.
- Digital twins for biomanufacturing: Lilly is leveraging NVIDIA Omniverse to create digital twins of its manufacturing lines. This allows teams to model, stress test, and optimize entire supply chains before making physical changes, accelerating quality control and getting medications to patients faster.
- Economic impact: This initiative is a cornerstone of Lilly’s $50 billion commitment to expand its U.S. manufacturing and R&D footprint, poised to create 13,000 high-wage jobs in manufacturing and construction.
Surgical and imaging innovation with physical AI
NVIDIA is extending its AI to the operating room and imaging suite to solve operational stress and complex diagnostic challenges.
- Robotic Simulation for Surgery: Johnson & Johnson Medical Technology is using NVIDIA Isaac for Healthcare and Omniverse to advance the MONARCH platform for urology. Teams use virtual environments to simulate device setup and patient interaction, optimizing design and ergonomics. Digital twin simulation can compress design reviews that previously took months into hours.
- AI agent for radiology: GE HealthCare is develop a diagnostic AI agent image assistant integrated into devices to address the critical shortage of radiologists. It will process scans, enable interaction through natural language and create interactive reports, leveraging LLM, VLM and agents.
- Incidental findings: GE HealthCare is piloting an AI agent solution to detect and report incidental findings on CT scans (which occur in more than 47% of abdominal CT scans). The AI will identify high-risk lesions and recommend follow-up images, while the radiologist retains full control.
Open models and data ecosystems
NVIDIA is accelerating open science and practical implementation of AI through new models and partnerships focused on data integrity and accessibility.
- Open models for discovery: NVIDIA is deploying open models like CodonFM (co-developed with the Arc Institute, used by the Stanford RNA Medicine program) that learn RNA rules to improve therapeutic design. Another model, La-Proteina, creates three-dimensional protein structures atom by atom twice as long and faster than previous models.
- Explainable AI (XAI) in Radiology: Clear Reasona vision language model, advances explainable AI. NIH is integrating Clara Reason models into radiology workflows to assist in report writing and support physician training.
- Multimodal AI integration: Innovaccer has adopted NVIDIA’s complete AI platform, including NeMo Guardrails and Triton Inference Server, to accelerate voice, text and multimodal intelligence in healthcare workflows. This forms the basis of Innovaccer’s Agents of Care™, which automates repetitive tasks using multimodal data for greater speed and accuracy.
- Data efficiency: GE HealthCare is conducting research on energy-efficient neural networks for tomographic imaging, and early results show a reduction in reconstruction iterations from 40 to just six, reducing computational power needs.
