🧩 NVIDIA's full stack computing platform for healthcare
I have spent my career thinking about where compute, software, and domain expertise intersect, and today that intersection is transforming healthcare. At the base of everything I describe is a simple fact. We build chips and systems, and we design coherent software and libraries that run across the full stack, from tiny edge devices to the largest AI factories in the cloud. That coherence matters. It is how breakthroughs in simulation, molecular design, multi modal models, inference, and robotics all move from research to impact.
When I talk about the NVIDIA full stack, I mean more than just hardware or one library. I mean a continuum where CUDA underpins acceleration, domain specific libraries accelerate critical operations, and model and deployment frameworks let teams run those models at enterprise scale. That continuum runs from a small single board computer to a DGX system to massive Blackwell-based clusters. The programming model is consistent, and that makes the engineer's job simpler and the scientist's job more productive.
Why is this important for healthcare? Because modern biomedical problems are deeply heterogeneous. Genomic pipelines rely on FFTs and high throughput sequence alignment. Molecular simulations rely on GPU-accelerated molecular dynamics and specialized kernels that have evolved from decades of research. Medical imaging needs multi gigabyte 3D volumes processed at high throughput for segmentation and reconstruction. Clinical workflows require speech, reasoning, and secure edge deployments. To address all of this, we cannot rely on a one-size-fits-all approach. We need a full stack that supports domain specific models and domain specific deployment platforms.
Why domain specific acceleration matters
Artificial intelligence is not a monolith. It is inherently domain specific. A transformer tuned to clinical notes behaves differently than one trained for protein folding. The kernels used in CT reconstruction are different from those that accelerate molecular dynamics. So I focus on building domain specific modules and libraries because they let subject matter experts move faster. When you give a lab the right primitives, they spend less time wrestling with the basics and more time making discoveries.
That is why we put resources behind libraries that you will find throughout the healthcare ecosystem. When you use a platform that has been co designed with the workloads of this industry in mind, the performance improvements multiply. That is the foundation we are providing for drug discovery, digital health, simulation, and physical AI.
🧬 Open models and the next era of biological design
We are living through a profound shift in how biology is studied and engineered. AlphaFold sparked a revolution by showing that AI can crack a fifty year old problem, and that moment made it clear that models will reshape the scientific method in biology. I believe the evolution we are now seeing can be described in generations.
The early generation was about pattern recognition and using convolutional networks to extract structure from data. The transformer introduced the ability to handle sequences and context at scale, enabling breakthroughs such as sequence to structure mapping. The second generation enabled us to do accurate structure prediction, and the third generation we are entering now is moving the field from prediction to design. These models do not just tell you what exists. They help you design what could exist.
To make that shift practical for industry and research, I have to talk about open models and open infrastructure. Open models not only accelerate discovery through shared knowledge, they also allow individual organizations to fine tune and integrate models with proprietary data in private deployments, which is essential for commercial R and D. To make open models truly useful, they must also be deployable at enterprise scale, and that is why I talked about packaging models as NIMs, a to go box for models optimized for extreme performance and enterprise deployment.
OpenFold3, La Proteina, and the NIM concept
One concrete example is OpenFold3. We worked with the OpenFold community and the OpenFold Consortium to turn a leading protein structure prediction model into a production ready NIM. The NIM concept is intentionally simple. It is a model packaged, optimized for performance, and ready to be deployed as a microservice. This lets researchers and enterprises run these powerful models reliably at scale without needing to become GPU optimization experts overnight.
Another set of open models I am excited about are those focused on biomolecular design. La Proteina is an all atom model capable of de novo protein design and controlled design starting from a given sequence. It can generate exquisitely detailed proteins up to 800 amino acids in length, which pushes the boundaries of protein design models that were typically limited to smaller sequences. Kodan FM is purpose built to optimize RNA sequences for design. Resin is targeted at synthesis prediction and route planning, crucial for chemistry and materials science. GenMol generates drug like small molecules to seed lead optimization efforts. We are releasing these models on Hugging Face with full training recipes because open collaboration drives progress.
From sequence to structure to engineering
When I describe these models, I emphasize the trajectory. At first, models gave us insight into patterns. Then they gave us structure prediction at scale. Today they are beginning to act more like design engines, enabling us to pose constraints and objectives and receive candidates that can be tested in simulation and experiment. Multi modal models that combine sequence, structure, and experimental signal let us incorporate more biophysical context into the design loop. That compounding effect between prediction, simulation, and optimization is what will make molecular design a true engineering discipline.
We are already seeing early winners. Groups like the Boltz lab at MIT announced Boltzgen for binder design that meets challenging benchmarks. Genesis Molecular AI revealed high performance models for protein ligand design. These developments are not isolated. They are part of a pattern where model architectures, compute scale, and domain knowledge converge to enable design at speed and fidelity that the field has not previously enjoyed.
🏭 Building AI factories with pharma: Lilly’s example
Drug discovery and development is compute hungry. Large pharma has rich internal datasets and decades of experimental history. When you combine that institutional memory with modern foundation models, you unlock novel ways to search chemical space and understand clinical outcomes. That is the thinking behind what I call an AI factory, a dedicated computational instrument that turns data into biomedical intelligence.
Eli Lilly is building what I believe will be the industry benchmark for a biopharma AI factory. They ordered infrastructure at scale, including Blackwell based systems, to power their next generation of foundation models that understand and simulate biology. This infrastructure is not simply about scale. It is a platform for modeling, simulation, and optimization that will accelerate the development of molecules, models for clinical development, and tools that can be shared through federated platforms like Lilly ToonLab.
What is important about the AI factory concept is twofold. First, AI factories enable organizations to train and run very large, domain specific models that can reason across multiple data modalities. That reasoning ability is central to capturing tacit knowledge that exists in chemistry and biology laboratories, including "what failed and why" which is often as important as knowing what succeeded. Second, when these factories are combined with federated sharing mechanisms, they can accelerate startups and collaborators by giving them access to curated models and tools without compromising proprietary data.
Why pharma will need on prem supercomputers
I often say that it would not surprise me to see many pharmaceutical companies announce their own billion dollar supercomputers. The work of exploring chemical and biological design spaces is fundamentally exponential, and the more compute you have the broader and deeper your search can be. That compute will be used not just for training models but for massive inference workloads, multi scale simulations, and closed loop design workflows that iterate between hypothesis, simulation, and experiment.
🧫 Accelerating virtual cell models with the Chan Zuckerberg Initiative
Understanding biology is not just about individual molecules. It is about systems of interacting cells and processes, and virtual cell models are an important frontier. The Chan Zuckerberg Initiative has been investing in virtual cells to enable richer multi scale modeling. These virtual cells need powerful models to capture dynamics across scales, and they require open datasets, benchmarks, and tooling so the community can iterate quickly.
We are working with CZI to bring models like Kodan FM into their virtual cells platform. The objective is to build open science infrastructure for multi scale biological modeling. That includes open datasets, synthetic data generation, benchmarks for model evaluation, and well defined pipelines so researchers can take a model from training to evaluation to interpretation without getting stuck on plumbing or compute.
Benchmarks are crucial here. The biological domain needs rigorous, meaningful ways to evaluate models. These evaluations must reflect functional or mechanistic performance as well as predictive accuracy. I am proud that we are partnering to define and operationalize those benchmarks so that everyone can measure progress reliably.
🩺 Digital health and agentic AI for care delivery
Healthcare systems globally are facing a basic supply and demand imbalance. Demand for care is growing due to aging populations, rising chronic disease burden, and patient expectations, while the supply of clinicians remains constrained. The result is long waits, administrative friction, and clinician burnout. That is where digital health agents can make an outsized difference.
I describe a path from administrative automation to higher level clinical augmentation. Start with agents that help with scheduling and navigation. These are the front door to care. When we reduce friction at the front door, more patients get timely access to needed services. The next layer is conversational agents that support documentation and authorization, turning free form clinical conversations into structured tokens and actionable data for downstream workflows. Finally, we have agents that can support care management and long term monitoring.
Speech is the most natural interface in care. Tokenizing voice, that is converting spoken interactions into discrete units that models can reason over, unlocks a cascade of downstream benefits. That includes more accurate documentation, better billing and authorization, improved monitoring, and the ability to build longitudinal patient representations. For many clinicians, the ability to auto capture the narrative of a patient visit and feed it into decision support is transformative.
Examples from startups and partners
One vivid example is Abridge, which provides clinicians with an app that captures doctor patient conversations and turns them into accurate, actionable records. Because the app is domain aware and clinically informed, adoption is rapid. Clinicians who try it often cannot imagine going back to the old way. That is a pattern I expect to repeat across well designed digital health agents.
Other startups are tackling wellness and specialized care. Companies focused on mental health, fertility and IVF, and chronic disease management are building agentic systems that extend care beyond clinic walls. They use multimodal inputs including speech, EHR data, and device streams to coordinate care, triage patients, and prompt interventions.
🔊 Inference at scale and the Blackwell performance leap
There is a practical constraint that sits between ideas and impact. Inference is not free. The long form reasoning that agents require, including planning, tool usage, and the repeated queries to multiple models, puts significant pressure on inference infrastructure. This is where Blackwell matters.
Every new GPU generation is a systems achievement, but Blackwell is a particularly large step forward. I have seen single generation improvements on critical inference workloads that are on the order of ten times or more. That kind of improvement is not just about running models cheaper. It is about enabling higher fidelity, longer horizon reasoning at acceptable cost, and that unlocks new agentic experiences in clinical environments where latency and cost matter.
For example, when a clinician uses an agent to summarize a complex visit and then ask follow up questions that require the agent to reason about past notes, imaging, and medication history, the system will run millions of inference operations across multiple models. Blackwell gives us headroom to make those interactions fast, safe, and affordable for healthcare systems to deploy at scale.
🔬 Precision health: partnering with Verily
Making complex biomedical data AI ready is a major bottleneck. Verily understands that problem at scale. Their workbench is a researcher centric platform that brings together curated cohorts, like the All of Us dataset, and makes it accessible for analysis and modeling. By integrating accelerated bioinformatics pipelines, multimodal model tooling, and optimized stacks for training and inference, we are reducing friction for thousands of researchers.
What excites me about the Verily partnership is the ability to combine curated population scale datasets with high performance compute and modular models. Researchers can preprocess data, run state of the art genomics pipelines like Parabricks, and build multimodal language models using frameworks like NeMo and Clara. The result is a platform that can support population scale analysis as well as translational research that moves from bench to bedside.
📈 Rewriting the digital health stack
What I see happening across digital health is not incremental change. It is a rewrite. For years, healthcare software has been tool centric, made for humans to use. The new generation of software will be agentic, software that uses tools on behalf of humans and orchestrates downstream workflows. That change touches every SaaS company, every EHR, every clinical workflow vendor. Epic and other incumbents are already integrating agentic capabilities, and startups are innovating at the edges with new paradigms for patient engagement and clinician augmentation.
When the entire stack becomes agentic, the way we build electronic health records, scheduling systems, and documentation software changes. Rather than forcing clinicians to adapt to software, the software adapts to clinical workflows and automates routine tasks. This shift will help close the gap caused by clinician shortages and will create new opportunities for clinicians to focus on higher value activities like complex diagnosis and patient relationships.
🤖 Physical AI and robotics: surgery and biomanufacturing born in simulation
Beyond digital agents and molecular models, a third major transformation is at the intersection of AI and the physical world. I call this physical AI. It includes surgical robotics, hospital automation robots, and AI enabled biomanufacturing. The pattern is the same across these domains. A digital twin and simulation first approach lets teams design and validate systems in software before building physical devices. That approach improves safety, reduces cost, and lets us iterate quickly.
To make real time robotics viable in healthcare you need a platform designed for the millisecond budgets of perception, planning, and actuation. IGX Thor is our enterprise grade physical AI platform for robotics. We designed it to be medically oriented, capable of low latency sensor processing, and to host the compute needed for vision language action models on the edge. With IGX Thor, you can run multiple sensor streams, do real time inference, and integrate with Holoscan for streaming data pipelines.
Diligent Robotics, Moxi 2, and hospital automation
Diligent Robotics adopted IGX Thor for their new Moxi 2 platform. Moxi 2 is built around a three computer architecture, combines simulation based training, and is deployed with the real time processing needed for hospital navigation and interaction. Before a robot shows up on a ward it can be trained in a digital twin, validated in simulation, and fine tuned for the specifics of the facility. That pipeline reduces the operational burden on hospital staff and increases the reliability of deployment.
Isaac for healthcare robotics and the SoArm demo
I want to emphasize how essential simulation is for practical robotics. If you only train robots in the physical world they will never generalize. They need to be born in code first. Isaac is a platform we created to make that pipeline feasible. You take a CAD model, simulate sensors and physics, generate diverse data, and then use that synthetic data to train policies. Once trained, policies are deployed to IGX Thor devices at the edge.
At the conference I showcased a low cost so called hello world of healthcare robotics, the SoArm. It is a $200 3D printed arm integrated with open models and tools. Developers can use the DGX Spark for training, Isaac for simulation and policy learning, and IGX Thor for edge deployment. The goal is to make experimentation accessible, so teams can build prototypes and then scale to clinical grade platforms.
Johnson & Johnson, Omniverse, and simulation first surgical platforms
Johnson & Johnson MedTech is building a simulation first approach for surgical systems with their Monarch platform. Inside Omniverse, CAD models are converted into physically accurate virtual operating rooms. Digital twins of the theater, instruments, and patient anatomies let the team refine surgical workflows, train clinicians, and generate synthetic data for vision models. That synthetic data feeds Cosmo style foundation models that provide rich perceptual inputs for robotic systems.
I often use a phrase from those projects. The first cut is not made with a scalpel, it is made in code. That encapsulates the idea. By designing and iterating in software, you can test scenarios, measure performance, and lower the risk before a device ever touches a patient.
🩻 Multimodal medical AI and Clara Open Models
Medical imaging is a domain where multimodality and domain knowledge are essential. Segmentation, 3D reconstruction, and chain of thought reasoning for imaging specialists are active areas of research and productization. Clara Open Models are our approach to providing pre trained, high quality, domain specific models for medical tasks. These include whole body 3D segmentation models that can automate labeling or assist interactive annotation workflows.
I want to call out MONAI, the community maintained medical AI framework. MONAI has become an industry standard for medical AI development. It has millions of downloads and supports clinicians and researchers in building competitive models for imaging tasks. Projects like Kitware have integrated these capabilities into tools used by radiologists every day, further closing the gap between research models and practical clinical tools.
We are also releasing models and training recipes for chain of thought reasoning in radiology. The goal is to capture the tacit reasoning radiologists use when reading studies, encode that into models, and then surface intermediate reasoning steps so clinicians can see how a model arrived at a finding. That kind of transparent reasoning is key to trust and adoption in clinical settings.
🌐 The ecosystem advantage: open source, benchmarks, and community
Throughout these announcements one theme is clear. Open source and community are the lifeblood of meaningful progress. Whether it is OpenFold, MONAI, Parabricks, or the many open datasets we support, open collaboration accelerates discovery. I believe most of the model work, and many of the training recipes and data curation efforts, should be shared openly when possible. That not only democratizes research but also sets an interoperable baseline for enterprise adoption.
However, openness must be balanced with safety, privacy and rigorous evaluation. That is why federated platforms, synthetic data generation, and thoughtfully designed leaderboards are so important. We need to create evaluation frameworks that capture clinical relevance, not just generic accuracy metrics. That requires domain experts — clinicians, chemists, and biologists — to help define benchmarks and to validate models in real world scenarios.
Benchmarks, pipelines, and ML Ops for biology
One of the biggest impediments to deploying biology models in production is the lack of robust ML Ops pipelines tailored to biological data. Making raw experimental data AI ready requires curation, normalization, provenance tracking, and integration across modalities. We are partnering with organizations to create those pipelines so that models can be trained, benchmarked, and deployed with reproducibility and compliance baked in. The result will be faster iteration cycles and better scientific reproducibility.
❤️ What this means for clinicians, researchers, and patients
The practical outcome of these technologies is straightforward but profound. Faster sequencing leads to quicker diagnoses, especially in acute settings like the NICU where every hour matters. Better molecular models shorten the discovery timeline and focus experiments on the most promising candidates. Digital health agents reduce administrative load and free clinicians to spend more time with patients. Robotics and simulation reduce procedural variability and improve surgical training. Together, these changes increase access, improve outcomes, and lower costs.
For example, faster genomic sequencing workflows are no longer a theoretical benefit. With optimized tools and libraries, sequencing analysis pipelines have set world records for speed. These improvements are not vanity metrics. They translate into diagnoses that can be delivered in time to change clinical decisions. That is where compute meets patient impact.
Digital agents are having a similar effect in clinic. By automating scheduling, intake, and documentation, we can increase throughput without compromising quality. That means fewer no shows, better preparation for procedures, and more consistent follow up care. It also means clinicians can spend their cognitive resources on complex decision making rather than clerical work.
Privacy, safety, and guardrails
All of this progress must be accompanied by rigorous safety frameworks. Agents must be designed to avoid delivering unauthorized diagnoses. Models must respect privacy and protect patient data. Federated learning and privacy preserving techniques will be a major part of how we scale adoption while maintaining trust. I want to be explicit about this. Technology without trust will not scale in healthcare. We must build systems that are auditable, transparent, and governed by clinical oversight.
🛠 How to get involved and practical next steps
If you are a clinician or researcher wondering how to get started, here are several practical steps you can take right now. First, explore domain specific open models such as those on Hugging Face. Try OpenFold3 as a NIM for structure prediction or La Proteina for protein design. Second, adopt community frameworks like MONAI for medical imaging. Third, experiment with Isaac and Omniverse if you are working on robotics or simulation. Finally, engage in federated platforms and consortiums to contribute data, help define benchmarks, and collaborate on real world validation.
If you are in pharma or biotech, start planning for the AI factory era. That means thinking strategically about compute investments, data curation, and the governance models that let you leverage proprietary and open models safely. If you are in a hospital or health system, pilot agentic systems for scheduling and documentation, and measure their impact on throughput and clinician satisfaction.
- Adopt open models and experiment with NIMs for rapid iteration.
- Use MONAI and community tools for imaging work, contributing back when possible.
- Start with simulation first for any robotics or device development.
- Invest in inference efficient hardware for agentic workloads and long tail reasoning.
- Engage with federated platforms to accelerate translational work while preserving privacy.
📖 Closing thoughts: a new chapter in medicine
We are at a unique moment. The confluence of domain specific AI models, high performance compute, agentic software, and simulation driven physical AI is rewriting the healthcare stack. This is not incremental. It is a re architecture of how we discover, manufacture, deliver, and plan care. That change will be felt across pharmaceutical R and D, clinical workflows, medical devices, and in the lives of patients.
I am optimistic because this transformation is happening in the open. Partnerships across industry, academia, nonprofit initiatives, and startups are accelerating progress. From OpenFold3 to Lilly's AI factory to Verily's research workbench to collaborative platforms for virtual cells, the building blocks are being assembled.
My invitation to you is simple. Think big. Build collaboratively. Use open models and public toolkits to accelerate your work, but pair that openness with rigorous evaluation and safety. Use simulation to de risk hardware and robotics development. And if you are designing clinical agents, put clinicians at the center of design so that tools augment professional judgment rather than obscure it.
We are building a future where biology becomes an engineering discipline, where digital health agents extend care beyond walls, and where physical AI augments clinical teams in safe, validated ways. The technology stack is ready. The community is engaged. The time to act is now.
"The first cut is not made with a scalpel. It is made in code."
Thank you for joining me in this work. Together we can build a more intelligent, accessible, and resilient healthcare ecosystem.
