Latest / Becker’s Healthcare Podcast / AI Built for Healthcare: A Keynote with Abridge & NVIDIA

Transcript

AI Built for Healthcare: A Keynote with Abridge & NVIDIA · Becker’s Healthcare Podcast

• 0:00
  - Thank you so much for tuning in to
• 0:01
  - the Becker's health care podcast series. We are
• 0:03
  - so excited to share a segment from Abridges
• 0:06
  - keynote recorded live at the New York Times
• 0:08
  - Center on 06/11/2026.
• 0:11
  - This conversation features doctor Shiv Rao, the CEO
• 0:14
  - and cofounder of Abridge, alongside Kimberly Powell, vice
• 0:18
  - president of health care at NVIDIA.
• 0:20
  - Together, they explore why generic off the shelf
• 0:23
  - AI fall short in medicine
• 0:25
  - and why they've partnered to build a clinical
• 0:27
  - conversation foundation model purpose built for health care,
• 0:31
  - touching on open models, the engineering behind speed
• 0:33
  - and accuracy,
• 0:35
  - and what it takes to put trustworthy AI
• 0:37
  - in the hands of every clinician.
• 0:39
  - We hope you enjoy the conversation.
• 0:41
  - So I I'm curious maybe to start things
• 0:44
  - off. Kimberly, like, why does, you know, you
• 0:46
  - why do you why does NVIDIA
• 0:49
  - care so much about health care and, like,
• 0:50
  - the set of problems that we focus on?
• 0:53
  - Yeah. I mean, this goes back almost twenty
• 0:55
  - years. I've been telling people I've been at
• 0:57
  - the company about eighteen where, actually, you know,
• 0:59
  - Jensen, our CEO, had an encounter with radiologists
• 1:02
  - early on who were trying to use gaming
• 1:04
  - chips to do really hard math, to do
• 1:06
  - inventory construction. And it was it was just
• 1:09
  - at that moment that we knew we created
• 1:11
  - something,
• 1:12
  - accelerated computing and now artificial intelligence,
• 1:15
  - that was going to be really helpful for
• 1:17
  - some of the world's hardest problems. And we
• 1:19
  - have done nothing but invest,
• 1:22
  - since.
• 1:23
  - And, you know, we're at this incredible moment
• 1:26
  - in time, and we felt for a long
• 1:27
  - time health care will be one of the
• 1:29
  - largest
• 1:30
  - technology industries,
• 1:32
  - and you're,
• 1:34
  - the antithesis and and a complete example
• 1:37
  - of how that is becoming complete reality.
• 1:40
  - And it was also at the middle of,
• 1:42
  - call it, the 2015,
• 1:44
  - you know, it became very clear,
• 1:47
  - that
• 1:48
  - intelligence is domain specific,
• 1:51
  - and we're in this phase of AI where
• 1:54
  - it's moving so fast.
• 1:56
  - How do we make it our mission statement
• 1:58
  - in the health care practice
• 1:59
  - to make technology that moves so fast
• 2:02
  - accessible
• 2:03
  - to the health care industry? Mhmm. So just
• 2:06
  - take a minute and and talk about what's
• 2:08
  - happened in AI in the last,
• 2:10
  - eighteen months. Yeah. We went from
• 2:13
  - AI that
• 2:14
  - can generate things,
• 2:16
  - AI that can reason,
• 2:18
  - to AI that can do work.
• 2:21
  - So we've gone through
• 2:24
  - three huge technological breakthroughs in just the last,
• 2:28
  - call it eighteen months.
• 2:29
  - And what we're talking about in our announcement
• 2:31
  - is,
• 2:32
  - you think about every technological breakthrough
• 2:36
  - requires you
• 2:38
  - to look at the full stack.
• 2:39
  - We think of AI as not a model.
• 2:41
  - It's a full stack computing problem. It starts
• 2:44
  - with energy.
• 2:45
  - It's then chips and systems.
• 2:48
  - It's then AI data centers and clouds.
• 2:50
  - Then it's core foundation models.
• 2:53
  - And finally, it's the application layer, which you
• 2:56
  - squarely have put
• 2:58
  - forward that vertical AI application layer has incredible
• 3:03
  - value,
• 3:04
  - effectiveness, and a an ability to change in
• 3:06
  - industry forever. And so I think what we're
• 3:09
  - we're recognizing together, and you've obviously come to
• 3:11
  - this conclusion yourself, is
• 3:14
  - it's time to go deeper in the stack.
• 3:17
  - And so a clinical conversation foundation model
• 3:20
  - so that the complexity of health care and
• 3:23
  - all of the workflows
• 3:25
  - and the connectivity of this amazing ecosystem
• 3:28
  - that you've brought to bear
• 3:30
  - can be realized
• 3:31
  - because it has to become much, much more
• 3:34
  - domain specific.
• 3:35
  - Yeah. I love it. Maybe just to sort
• 3:37
  - of unpack that a little bit more, like,
• 3:39
  - specifically
• 3:40
  - your philosophy
• 3:42
  - on open models Yeah. On the stack.
• 3:45
  - I'm curious if you can unpack that, and
• 3:46
  - I can certainly share how how we think
• 3:48
  - about things. Yeah. Yeah. That's right. You know,
• 3:50
  - it goes it goes into the this idea
• 3:53
  - that, you know,
• 3:54
  - imagine you try to throw a generic
• 3:57
  - AI model
• 3:58
  - into health care. It doesn't understand the clinical
• 4:01
  - language. It doesn't have the clinical reasoning,
• 4:04
  - and it surely doesn't have the domain expertise
• 4:06
  - of all of the long running tasks and
• 4:09
  - interconnected work that has to happen for workflows
• 4:12
  - to be completely transformed.
• 4:15
  - And so if that is the case that
• 4:16
  - generic AI is just not going to work,
• 4:19
  - we recognize this long ago when we started
• 4:21
  - investing in an open model strategy
• 4:24
  - so that,
• 4:25
  - it could be a professor at a university.
• 4:27
  - It could be a clinician
• 4:29
  - who has this
• 4:30
  - amazing unlock that they could imagine,
• 4:33
  - has a well invested platform starting point of
• 4:37
  - open language models. We work on open models
• 4:39
  - for biology.
• 4:41
  - We work on open models for, World Foundation
• 4:44
  - models so that we can help, you know,
• 4:46
  - realize robotics and automation in the physical world.
• 4:49
  - And we give that to developers and application
• 4:53
  - makers all over the world across every industry
• 4:57
  - to realize this great potential
• 4:59
  - and to go down more domain specialized
• 5:02
  - application spaces where the real value is delivered.
• 5:05
  - And you put the power of AI
• 5:08
  - in the hands
• 5:09
  - of the domain experts
• 5:11
  - who can be
• 5:12
  - extremely thoughtful in, one,
• 5:15
  - making sure it's built so that it can
• 5:17
  - be trusted, making sure it's built that it
• 5:19
  - it has traceability,
• 5:20
  - auditability.
• 5:21
  - It has, you know, all of the
• 5:24
  - built in knowledge of some very exquisite domains.
• 5:28
  - How could any generic AI company do that?
• 5:31
  - But our role we can play in that
• 5:33
  - is by giving you a huge foundation to
• 5:35
  - stand on of open models and then start
• 5:38
  - to domain adapt them for these really complex
• 5:41
  - workflows that we're we're gonna you're going to
• 5:44
  - build. Yeah. When we think about workflows, when
• 5:46
  - we think about our product,
• 5:48
  - really, you know, from our standpoint, we just
• 5:51
  - have to deliver the best possible experience.
• 5:53
  - We have to be able to move all
• 5:55
  - the needles that we were talking about earlier
• 5:56
  - today.
• 5:57
  - And to an extent, for sure, it means
• 6:00
  - that we are gonna partner with the Frontier
• 6:02
  - Model companies, and we do. We have incredible
• 6:04
  - relationships there.
• 6:06
  - And, also, since we have a very,
• 6:08
  - you know,
• 6:09
  - maximalist sort of attitude around sprinkling intelligence everywhere
• 6:13
  - we possibly can, it means we do have
• 6:15
  - to reach down lower into the stack and
• 6:16
  - control our destiny. And latency is a big
• 6:19
  - issue for us because we do want the
• 6:20
  - clinician to be able to stop, swivel the
• 6:22
  - chair, and have all of those artifacts there.
• 6:25
  - And, you know, that's where this partnership for
• 6:27
  - us has really, you know, opened our eyes
• 6:30
  - to all that we can do.
• 6:31
  - Even, you know, over the last, you know,
• 6:35
  - several sort of training runs, what we've already
• 6:37
  - been able to see, for example, f one
• 6:39
  - scores related to that CDI product and being
• 6:42
  - able to predict the right ICDs have gone
• 6:44
  - from,
• 6:45
  - I think, 69%
• 6:46
  - to 85%,
• 6:48
  - and there's a significant delta compared to anything,
• 6:51
  - you know, that we could get off the
• 6:52
  - shelf. Similarly,
• 6:53
  - with a a separate type of model that
• 6:56
  - tries to infer the context of the specialist,
• 6:58
  - so it understands that I'm the cardiologist. And
• 7:00
  - thus, off the bat, the cardiology specialty specific
• 7:03
  - note should look a little bit different. We're
• 7:06
  - raising that floor so there's a lot less
• 7:09
  - uncertainty in terms of what we we need
• 7:11
  - to do above it to make sure that
• 7:13
  - the output sort of meets the mark. We're
• 7:15
  - seeing similar results.
• 7:17
  - This is the second part of why open
• 7:19
  - models are so incredibly important.
• 7:22
  - We
• 7:23
  - want to help you create the experience
• 7:25
  - through all these very complex workflows
• 7:28
  - that is,
• 7:29
  - it's behind the scenes. It's very fast. It's
• 7:31
  - it's second nature. It's just somebody working on
• 7:34
  - your behalf. And so optimization,
• 7:36
  - if you think about our we have Nemotron.
• 7:38
  - Our open model is going from two to
• 7:40
  - three. You're seeing, you know, five times the
• 7:43
  - performance because
• 7:44
  - we're taking a full stack approach on these
• 7:47
  - models. They're at the frontier from their capabilities,
• 7:50
  - and then we understand
• 7:52
  - the technology today, and it will be ready
• 7:54
  - for, you know, VeriRubin, which is coming out
• 7:56
  - next, which has a whole
• 7:58
  - other gigantic
• 8:00
  - leap in terms of latency
• 8:02
  - as well
• 8:03
  - as cost performance. So we think about,
• 8:06
  - we want this capability
• 8:08
  - to be abundant throughout health care, providing the
• 8:10
  - right information, doing work behind the scenes, taking
• 8:13
  - away the administrative burden. We want it to
• 8:15
  - be accessible to every clinician treating any patient.
• 8:19
  - And so that means
• 8:20
  - we have to be thoughtful about the token
• 8:23
  - economics.
• 8:24
  - And so every day we wake up in
• 8:26
  - the morning as an accelerated computing platform to
• 8:28
  - think about not only the model, but these
• 8:30
  - complex agents and multitask,
• 8:33
  - long running tasks, always running agents,
• 8:37
  - right, who are going to do something I'm
• 8:39
  - super excited about,
• 8:40
  - in the future, which is,
• 8:42
  - you know,
• 8:43
  - think about,
• 8:45
  - hearing something in that conversation as a preventative
• 8:47
  - measure. Mhmm. Because it's always running, and I
• 8:50
  - heard that thing and it associates this across
• 8:52
  - a whole bunch of other data points, and
• 8:53
  - we're gonna start to get into hopefully much,
• 8:55
  - much earlier detection even. But that requires that
• 8:59
  - these agents are running around the clock. So
• 9:02
  - we've gotta be very, very thoughtful about the
• 9:04
  - token economics, and that is, again, giving
• 9:07
  - amazing,
• 9:09
  - applications and platforms like yourself the ability to
• 9:11
  - control all of that. Yeah. We are in
• 9:14
  - year one of, what will absolutely be for
• 9:17
  - us a multiyear investment. And
• 9:19
  - already, we've found a way to collaborate, and
• 9:22
  - it's it's very humbling for us, obviously.
• 9:25
  - NVIDIA is NVIDIA.
• 9:26
  - And our science team,
• 9:28
  - being able to spend time, you know, with
• 9:30
  - your researchers, but also our product team. You
• 9:33
  - know, over time, being able to service up.
• 9:36
  - Like, what are the requirements? What do we
• 9:38
  - actually need to do in in order to
• 9:40
  - serve all of our health system partners? Can
• 9:42
  - you speak a little bit to how the
• 9:43
  - partnership works, how what the mechanics are from
• 9:46
  - your side? I mean, it it's been just
• 9:47
  - an amazing
• 9:49
  - two way,
• 9:50
  - partnership.
• 9:51
  - NVIDIA is not a health care company. We
• 9:54
  - will never be. We'd never intend to be.
• 9:57
  - But in order for us to
• 9:59
  - tackle
• 10:00
  - some of the hardest work in the world,
• 10:03
  - work that has huge impact, and we convince
• 10:06
  - ourselves we do have a unique ability to
• 10:08
  - contribute.
• 10:10
  - And we find
• 10:11
  - a company or a thought leader or a
• 10:14
  - CEO or your team,
• 10:17
  - who sees the future differently
• 10:19
  - and understand technology is going to be an
• 10:22
  - absolute core component. I mean, you talk about
• 10:24
  - the platform, it's AI native.
• 10:27
  - That is where
• 10:29
  - two companies coming together, working at. We've been
• 10:32
  - working on speech research together. We've been, you
• 10:34
  - know, working in allowing you to train all
• 10:36
  - these already domain specialized models that has allowed
• 10:40
  - for your your platform to be so widely
• 10:43
  - adopted and the joy and,
• 10:45
  - everything that you've brought back to the clinical
• 10:47
  - practice and the patient outcomes and become this
• 10:50
  - trusted provider
• 10:51
  - across the ecosystem.
• 10:53
  - So we're working on everything from research to
• 10:56
  - that hardcore engineering so that things are running
• 10:59
  - really fast and that they're,
• 11:01
  - the cost control is there. So now we're
• 11:03
  - on this amazing new endeavor.
• 11:05
  - You've called it a treadmill.
• 11:07
  - NVIDIA often calls it the data flywheel.
• 11:10
  - Because you've been able to scale at such
• 11:13
  - a pace and because you are so thoughtful
• 11:15
  - in collecting
• 11:16
  - all of the good good feedback as well
• 11:20
  - as, you know, negative feedback or corrective feedback
• 11:22
  - to the system,
• 11:24
  - this is all intelligence.
• 11:26
  - You need the good, bad, and the ugly
• 11:27
  - so that you can build it back into
• 11:29
  - the system, and it becomes
• 11:31
  - incredibly smarter. And once you have a foundation
• 11:34
  - model like that, it becomes a core building
• 11:36
  - block to then go make all of these
• 11:38
  - downstream processes
• 11:40
  - that much more accurate
• 11:43
  - and effective in the work that needs to
• 11:45
  - get done. And so it's been a pleasure.
• 11:47
  - You've pushed us on our road map. We
• 11:49
  - push you on your road map. And I
• 11:51
  - think this is you know,
• 11:53
  - it's it's such
• 11:54
  - an an honor to be included in the
• 11:56
  - process, and
• 11:58
  - we wake up every single day.
• 12:00
  - And
• 12:01
  - because your application your your platform is so
• 12:05
  - it's in my doctor's office.
• 12:06
  - As a person who lives lower in the
• 12:08
  - stack in technology,
• 12:10
  - it's it's so motivating
• 12:12
  - and and just such an honor to to
• 12:14
  - be a part of it, and, we're excited
• 12:17
  - about the next couple years together. This is
• 12:19
  - a this is a time the health care
• 12:21
  - industry has never experienced
• 12:23
  - in history,
• 12:24
  - and you're making it happen, Shiv, and your
• 12:27
  - team. Thank you so much, Kimberly. Really, everyone,
• 12:29
  - all of us are doing it together.
• 12:31
  - So grateful to you. So grateful for our
• 12:34
  - partnership. Thank you so much for being here.
• 12:36
  - Thank you, Shiv. Thank you all. Thank you.
• 12:38
  - Thank you.