How AI Will Finally Make Healthcare Deflationary
AI in healthcare may be entering a new chapter, one where the biggest question is no longer whether the technology works, but who is willing to deploy it, measure it, and take responsibility for the risk.
This week, Steve sits down again with Eric Larsen to revisit his predictions from last year’s Webby-winning episode on generative AI in healthcare. Eric argues that the first wave of AI has been inflationary, reinforcing the old payer-provider payment model, but that the next wave could be deflationary as automation moves into revenue cycle, administrative work, clinical reasoning, and drug development. They discuss why incumbents still have a narrow window to co-develop the future, why clinical AI may move faster outside the US, and why liability may become the deciding factor in who wins.
We cover:
Why healthcare is still the sector most exposed to AI-driven change
How AI has reinforced fee-for-service dynamics so far, and why that may soon reverse
What makes some healthcare work more automatable than others
Why liability may determine how fast clinical AI gets adopted
Which health systems, payers, and life sciences companies are moving fastest
What will change across providers, payers, and pharma over the next year
—
About our guest:
Eric Jon Larsen is President of TowerBrook Advisors and a member of the healthcare leadership team at TowerBrook Capital Partners, a $30 billion AUM investment firm based in New York and London. TowerBrook invests across private equity, structured minority, and growth opportunities, with a strong focus on healthcare, partnering with health systems, payers, and other strategics. Notably, TowerBrook is the first mainstream private equity firm to achieve B Corp certification, reflecting its commitment to responsible business practices.
Eric is a nationally recognized healthcare strategist with a global advisory portfolio spanning CEOs and boards of leading healthcare organizations. He spent 25 years at The Advisory Board Company—five of those as President—advancing best practices in healthcare delivery worldwide. Following the firm's 2017 acquisition by Optum (UnitedHealth Group), Eric co-led strategic partnerships and market development efforts at UnitedHealth. He is also a Venture Partner at Thrive Capital and SignalFire, and serves on several digital health boards, including Somatus and Contessa Health.
—
Tune in
Full Transcript (AI edited for clarity)
Eric, welcome back to The Heart of Healthcare.
Eric Larsen: Steve, it’s awesome to be here. I re-listened to our pod from April of last year this morning, and it was super fun. The Webby was nice, but we’ve got a big bar to clear today.
Steve: I know. Hopefully we can run it back, and maybe we’ll keep doing this for the next decade.
Let’s dive in with a scorecard. You called this the most important moment in U.S. history and said that the U.S. healthcare industry has the greatest surface area exposure for GenAI to disrupt.
Now that we’re a year into it, did your predictions hold up? Let’s start with the macro one. Do you feel vindicated, or does a year of watching the actual deployment make you want to walk back anything you said?
Eric Larsen: I think we were pretty right. I don’t say that in a self-aggrandizing way. I say it in a humble, observational way.
If anything, the intuition is getting corroborated. This is the most powerful technology in history. It is the multiplication of intelligence. Twelve months later, almost thirteen months to the day, I think we know a lot more about the neurology of this intelligence. We can describe it with more precision than we could a year ago.
In terms of healthcare’s susceptibility to disruption, I think we’re very vindicated, for better or for worse.
Last year, I articulated why I thought healthcare was so exposed. When you strip away the mesmerizing generative elements of this technology, it is about brute force productivity augmentation. U.S. healthcare has the greatest labor addiction and labor intensity, with 23.8 million people. Since then, all we’ve done is add humans to healthcare.
Steve: That’s exactly the point. We talked about how we would have to reshape the labor force, and yet so far, I think this has been inflationary, not deflationary. We’ve added the AI, the cost of the AI, and while there are probably outcomes that have been improved, total cost seems to be going up because I don’t see a lot of humans being removed.
Eric Larsen: No. Quite the opposite. We’ve added 750,000 people to the labor ranks. U.S. healthcare is essentially holding up the entire hiring economy.
Steve: I think it drives about 50% of net new hires in the labor statistics, if I’m remembering correctly.
Eric Larsen: You are right. If you take away those 750,000 healthcare jobs, the labor force lost 200,000 employees. Can you imagine the convulsions that the equity markets would go into? You’d have emergency Fed proclamations, probably a 150-basis-point cut to the Fed funds rate. It would be catastrophic.
Healthcare has been buttressing the entire labor economy. I think it’s just a matter of time before we start to see labor substitution.
When we were together a year ago, I said the early applications, ambient listening, summarization, documentation, and coding, were pretty unimaginative.
Steve: I think you called them a little boring.
Eric Larsen: I did. I was uncharitable, and some friends in the industry, like Shiv Rao from Abridge, who we love, told me that was very uncharitable. But as we know, Shiv has big ambitions and has done great things since then.
I would say that is not true anymore. There were two milestones, one in December and one on February 5, where we saw the emergence of real non-linearity in the models. We saw the emergence of Anthropic’s Claude Code in December, and then a lot of the engineering and programming class disappeared from their families during the holidays to vibe code. You had luminaries like Andrej Karpathy say this was a massive magnification of capability.
Then on February 5, we saw another step-function jump with an almost cosmically coordinated release between OpenAI and Anthropic. OpenAI released ChatGPT Codex, and Anthropic released Claude 4.6. That was another massive jump in capability.
When I think about the first chapter in the deployment of the application layer, it was inflationary because incumbents weaponized the existing fee-for-service model and coding optimization. It reinforced the old payment hydraulics. You saw it benefit providers more than payers.
HCA hit over $115 billion in market capitalization. The seven publicly traded managed care companies went from about $1 trillion in market capitalization down to roughly half of that. Almost half a trillion dollars was incinerated. Two out of three Blues plans lost money last year. If trend for the country was 9%, trend for the Blues was 13%. A lot of CEOs, in unguarded moments, would say explicitly that this was about coding optimization.
I’m not going to weigh in on that debate, but I will say the first chapter was inflationary because it was used to reinforce the old adversarial model. The next chapter is going to be deflationary.
The way I think about it is: if you’re using the technology to amplify your reimbursement mechanism, the next chapter is automating the humans who do that. We spend $5.3 trillion on healthcare every year. $2.9 trillion of that is labor.
We talked last year about automating, augmenting, or eliminating. In the new paper, I’ve tried to decompose the 23.8 million jobs in U.S. healthcare, looking at 7.2 million jobs for hospitals and health systems, 611,000 jobs for payers, and another 209,000 jobs for PBMs and TPAs. The question is: how much is augmentable, and how much is substitutable?
It passes the intuition smell test that if you begin to automate some of those functions, especially in areas of functional verifiability, you are going to see major impact.
Steve: Say more about functional verifiability. I actually haven’t heard this yet, and I want to learn more.
Eric Larsen: This is a bit of my rubric. Andrej Karpathy said something like this: in Software 1.0, if you could specify the function, you could automate it. In Software 2.0, which is AI, if you can verify the function, you can automate it.
That is one of the reasons we’ve seen such a parabolic takeoff in coding. It’s why Cursor was reportedly acquired by Elon for $60 billion. It’s why Anthropic has gone from $9 billion in revenue in December to a projected $80 to $100 billion by the end of this year.
The functional verifiability point matters because anything with math, coding, a right or wrong answer, or provability is automatable. That is one of the reasons we’ve seen step-function jumps in revenue cycle management.
TowerBrook and CD&R own R1, and our CEO, Joe Flanagan, is one of the most prescient technologist CEOs in deploying this technology into areas of functional verifiability where you get massive augmentation.
I think the next chapter will be deflationary because, reductio ad absurdum, my bots are going to fight your bots. The providers were quicker off the draw than the payers, but the payers are smart, and they are industrializing their own agentic capabilities quickly. Eventually, it is going to cancel each other out.
Steve: I want to come back to that, but let’s stick on the application layer because I think that was where you were the least charitable.
The application layer is where we’ve seen the most explosion and growth of these companies in healthcare. If AI lets you build software really quickly, what do you have to do to stay ahead of the game long term?
If you were the CEO of an application-layer company, what do you think would be protectable versus not? What are the moats in this new world, especially in healthcare?
Are we going to have another scare for vertical AI application-layer companies, or is there a way to build real defensible moats against an Epic incumbent that might use AI, or against an Anthropic that can presumably do almost anything?
Eric Larsen: I’ll be a little provocative. I think software is largely uninvestable at this point.
I look at the inexorability of the God models. Their generalized capabilities are spilling over into every verticalized domain. If your moat before was the difficulty of programming your software, that is obliterated. There is no moat.
I’m a technophile, but I have zero coding capability, and yet I am vibe coding all the time with Claude Code. The barriers to entry and the activation energy are very low. You can get escape velocity in natural language.
If the moat before was the complexity of your tech stack, that does not exist anymore.
In healthcare, there is still some defensibility because of the regulatory enclosure and hyper-litigiousness of the sector. The regulatory moat is our friend in the short term.
Then the question is: how do you get to network effects, where each incremental user adds to the relevance of your platform? How do you develop such subatomic expertise in your vertical that you can encode it into your platform? How do you create a UX/UI that is sticky and that frontline users will clamor for?
That is a very high bar. I’m a big fan of Hippocratic and what Munjal has built. I’m a big fan of Shiv and what he has built. I’m a big fan of Adam and what he has built at Athelas. There are going to be great companies that emerge from this.
But there was $202 billion invested into startups last year.
Steve: And 80% of that was in AI, right?
Eric Larsen: I think it was $250 billion in total, and a disproportionate amount was in AI. But a lot of that went to the God models. That consumed a lot of it.
I think we are going to see a lot of capital incineration in this vintage. A lot of malinvestment. It is a Sand Hill Road conundrum right now. What is investable?
Everybody is throwing money into the God models for obvious reasons. Harvey is an interesting example, not in healthcare, but in legal.
Steve: Legora, for instance, which is the Harvey equivalent in law.
Eric Larsen: The question is how much is defensible? They had a $12 billion valuation. They built their own foundation model that was tuned exclusively to this esoteric legal domain. They had great uptake. But then Satya put a version of that into Copilot.
So now they are wisely pivoting more toward adapting on top of the foundation models.
I listen to Jensen, who I have incredible respect for, and he talks about the five-layer cake: power, the GPU stack, the God model stack, the application layer, and so on. Where does the monetary value accrue? It’s pretty unambiguous.
When we were together a year ago, Nvidia had a $3 trillion valuation. Today it has a $5 trillion valuation. You are seeing massive industrialization in CapEx. Last year, the hyperscalers had invested about $370 billion in CapEx the year before. It turned out to be about $420 billion. This year, we are projecting $800 billion in CapEx. Next year, $1.2 trillion. At $1.2 trillion, hyperscaler CapEx would be 3.3% of U.S. GDP.
Steve: That’s crazy.
Eric Larsen: The value is accruing at the God model layer and at the hyperscaler layer. The Mag Seven is now worth $22 trillion. I do not see it accruing as much on the application layer.
Healthcare will not have the same mass extinction rate we saw during the pandemic, when we had 13,000 diagnostic, therapeutic, care augmentation, clinical, and nonclinical workflow companies with 95% attrition. But it is hard to be a VC right now.
Steve: Thanks for showing me a little sympathy. I agree. Where you place your bets is hard to see. It is so dynamic right now, which also makes it really fun. Strategically, it’s hard, but it’s also a crazy, fun moment.
Let’s tackle another subject. Last year, we talked about incumbents versus insurgents and who would win. You sort of sat on the fence last year. You said it had to be a marriage, but you were skeptical about incumbents’ track record as co-development partners. What’s your take now?
Eric Larsen: I don’t think I was hedging as much as I was making a declaration that it has to be incumbents and insurgents together. I still believe that in the short term.
I said last year that incumbency is not an invariant law of nature. It is a head start. I still believe that.
U.S. healthcare is even more oligopolistic than it was a year ago. The top 10 health systems now represent $455 billion out of the $1.6 trillion sector. The top 100 health systems represent $1.2 trillion out of that $1.6 trillion.
Incumbents in our industry have to co-develop. They have abdicated every tech phase shift of the past generation. With each turn of the crank, they lose more agency, jurisdiction, and self-determination.
In the year since we were together, memory improved. Reasoning emerged toward the end of 2024, and with the advent of O3 in December 2024, it took a step function across 2025. We saw real augmentation in memory, then tool control. We moved from the scaling hypothesis to test-time compute and inference-time compute, where models got smarter the longer they thought. Then we got to agentification.
With each turn of the crank, from reasoning to memory to tool control to agentification, incumbents lose agency.
Over the last year, I’ve seen real stratification. A few payer, provider, life sciences, and medtech CEOs have decided to be autocratic about this. This is not a democratized thing. You can’t make this a gentle persuasion process. It has to come unambiguously from the CEO.
Steve: You have to red-pill it.
Eric Larsen: You have to red-pill it. One of the lessons from history is that ever since the first Industrial Revolution, we have seen a flourishing of humanity. But people have internalized the wrong lessons from history.
They think the inventor of the technology wins. If you invented the steam engine in Great Britain, that is why they got the military, political, economic, and sociological benefits. In the second Industrial Revolution, Germany sprinted forward in chemical capabilities, and the U.S. sprinted forward because it diffused electrification. In the third Industrial Revolution, the U.S. beat Japan because we invented more on the computer side.
But it is not invention. It is diffusion.
The country, company, or CEO that spreads and embeds the technology most horizontally and installs it the fastest wins.
The same is true for the 150. What is funny is that last year you called it the F-150. I had never said that in my life. Now everybody refers to it as the F-150.
Steve: There we go.
Eric Larsen: So now I have to use the F-150.
The insight here is that the CEOs among the payers, providers, medtech, and life sciences companies that have a blueprint for diffusion are the ones not abdicating their co-development responsibility. They are shaping the technology.
Healthcare is too sacred and too idiosyncratic to let some 20-something genius techno-solutionist design its installation. We need the incumbents.
Steve: Who do you think is doing it best in healthcare right now? If you had to point to one CEO or one company, who is the leader today?
Eric Larsen: I think Warner Thomas at Sutter is sprinting ahead.
Steve: Is that showing up in Sutter’s business, or is it too early to tell?
Eric Larsen: If you asked Warner, he would tell you that he has been unambiguous with the board and senior leadership team that Sutter is going to be an AI learning organization.
They are diffusing a base model and empowering frontline staff and tens of thousands of Sutter associates to become facile with the technology.
They are using a tight-loose-tight regime. The first tight part is that they are clear about what they are going to do, while making sure they are HIPAA compliant, maintaining data sovereignty, and preventing data exfiltration. The loose part is getting creativity at the individual level. This is democratizing technology. It empowers the individual, but the productivity benefits have not yet accrued to the enterprise. They have accrued to the individual.
So how do you capture that creativity from people who are going to be inventive and come up with productivity improvements? How do those diffuse into the enterprise?
The final tight part is rigorous measurement, codifying the wins and putting them into monetary terms. What is the reduction in SG&A? It’s not necessarily about reducing labor. It’s about mitigating the need to hire more.
I also think Brian Pieninck at GuideWell and Florida Blue is a real pioneer. They bought GPUs, and that started under the prior team. But Brian is one of the best leaders among the Blues, and even among the for-profit payer CEOs, in thinking about optimizations on the clinical augmentation side, administrative simplification side, and consumer empowerment side.
They’re partnering with Brett Taylor and Sierra on some conversational AI work.
I also admire what Lilly is doing and what Dave Ricks is doing. They are building what I think is the largest supercomputer ever in life sciences with Jensen, focusing on AI-engineered biomolecules and biologics, and simplifying the machinery around drug discovery and development.
Those are three companies and three CEOs I appreciate.
Steve: One company you didn’t mention, which may be the largest software company in healthcare, is Epic. They are great at creating FUD in the market.
Where do you think Epic is in its journey? How would you grade them? And related to our application-layer question, is there a world where other application and agent layers come around and strangle Epic? My colleague Sophia calls it the octopus strategy.
You said software is no longer that special, but Epic has many other advantages. Do we ever see Epic go away? Maybe not under Judy’s regime, but under someone else’s. Where does Epic stand as one of the largest incumbents?
Eric Larsen: I have cognitive dissonance on this one. On one hand, I am sympathetic to the entrepreneurial and startup ecosystem that is dealing with a juggernaut in Epic. On the other hand, I think very highly of Judy, Sumit Rana, their president, and Seth Hain, who is brilliant.
I admire what Epic has done in mobilizing a platform. Cosmos has something like 300 million longitudinal patient records, so it is the biggest structured and unstructured data repository in the industry.
I admired when they built their own foundation model. At first, they called it Comment, and I think Judy later came up with a much better name. In terms of an incumbent mobilizing, Epic has done a fantastic job.
But I do think you are going to see an “arm the rebels” strategy. We have seen a proliferation of brilliant founders focusing not just on administrative simplification, but also clinical AI across differential diagnosis, imaging, treatment protocols, care protocols, and specialty-specific immersion.
It is hard for an incumbent to survive every tech paradigm shift. We have not really seen that across history. So I would bet that insurgents will displace incumbents outside of domains with three things: regulatory enclosure, narrative warfare, and drawing up the drawbridge on interoperability.
Across history, incumbents preserve their advantage through those three strategies. When railroads threatened canals, when internal combustion threatened stagecoaches, when the PC threatened the minicomputer, you saw the same playbook.
You see it from monopolists today. Healthcare has a little bit of time because of those three properties: regulatory enclosure, narrative warfare, and interoperability constraints.
I am a little agnostic. The future is not written for Epic. I think we are going to see creative destruction in every sector of the economy. I agree with Vinod that half the Fortune 500 may turn over in the next few years.
Healthcare is artificially preserved for the moment, but I think the regulatory walls will be breached.
Part of the counterfactual here is what is happening in the Gulf states and China. One disagreeable fact of history is that autocracies and monarchies are better at implementation than messy Western pluralistic democracies.
Especially in clinical AI, where you can only go at the speed of liability and blame allocation, you are going to see clinical AI diffuse much faster and broader in China and the Gulf states than in the United States. We may end up having to reverse-import some of the longevity-expanding and deflationary advances from those regions. That worries me.
Steve: Let’s stick on that. A lot of money has been spent on operational, administrative, and patient communication layers of healthcare. I believe some of that should ultimately be deflationary, though it isn’t right now.
But I think the more interesting layer is clinical AI. When I talk to health system CEOs about clinical AI taken to its fullest extent, some of them roll their eyes. Some get it.
I tell them, if you don’t get ahead of this, other nations are going to do it. Consumers are going to demand this. We already see consumers using AI for diagnosis, recommendations, and mental health companionship.
Five years from now, maybe sooner, we may be at a place where people use these technologies to get the right diagnosis. Then the question becomes: how does our regulatory system, reimbursement system, and care delivery system adopt it? How do those gates come down, and when do we get to a place where clinical AI is driving more patient care in this country?
Eric Larsen: I am a huge evangelist for clinical AI, and not in a sequential way. I don’t think we need to start with administrative simplification and then cautiously inch toward clinical AI. I actually think that is a moral and ethical failure.
That reticence comes from a fallacy: that the current system is great. The current system is not great. In many ways, it is intolerable. We have hundreds of thousands of deaths attributable to medical error each year, and many more injuries.
I have reverence for U.S. healthcare and its practitioners. I would not want to be treated in any other country on earth. But this preservationist instinct, that we cannot deploy this too fast because it is too risky, is flawed.
This is also a broader societal question. We are evincing characteristics of a late-stage declining empire, where we are trying to be protective. Less than 30% of Americans are optimistic about AI. Around 80% of Chinese citizens are optimistic about AI. That is an interesting juxtaposition.
China is an engineering state. The United States is a lawyerly society. We have 1.3 million lawyers, whose profession is to be proceduralists and a source of the eternal no. I think we have overcorrected because the promise of clinical AI is not just deflation. I subscribe to Dario’s view that it will add decades to human longevity.
Today, Thrive announced that we led a $2.1 billion round into Isomorphic. Isomorphic won the Nobel Prize for Chemistry for what I would assert is the greatest scientific achievement in the last 50 years: AlphaFold.
If you think about diffusing clinical AI, and the fact that the God models are already superhuman in differential diagnosis and care, treatment, and protocol pathways, it is almost unethical not to diffuse them.
Steve: How does that play out practically? I agree with what you just said, but you talk to so many CEOs across the industry. How do you actually see this playing out?
Eric Larsen: It will go at the speed of blame allocation. Whoever underwrites the medical malpractice and product liability will win 100 to zero.
The analogy I use is BYD in China. BYD is the largest EV and autonomous vehicle company in the world. It has Level 4 autonomous parking and driving in Shanghai and Beijing. The company assumed product liability for the use of its autonomous driving capability.
In this country, we hold technology to a hypocritical standard of infallibility and perfection, not human equivalence or superiority.
Think about autonomous driving. We have 45,000 fatalities in this country every year. The top causes are drunk driving, distracted driving, and texting. Human, human, human. Yet one error from an autonomous vehicle can bankrupt the company. Cruise had a very unfortunate incident in San Francisco, and it bankrupted the company.
My belief is that whoever underwrites their product from a medical malpractice and product liability point of view will win.
There are precedents for this. A company called IDx, now Digital Diagnostics, created a product for diabetic retinopathy. It was the first FDA-approved autonomous diagnostician. They underwrote the product liability and medical malpractice.
This is a clarion call to startups: if you have absolute conviction in your product, can you assume the liability?
It is also a clarion call to hospitals and payers: who is willing to underwrite the liability?
Steve: That sounds like a great opportunity for a med-mal AI insurer. The risk may be much better than humans, and given the risk-off environment around new areas, there may be money to be made underwriting this.
Eric Larsen: Amen. I think you’re exactly right.
This is a re-emergent moment for the hyperscalers in healthcare. Look at Google, which is about to become the most valuable company in the world. Under Demis, whose Nobel Prize was not a coronation but a starting gun, Google is mobilizing across many vectors.
The vector they have not explored yet is assuming liability. If they are a $5 trillion company with $150 billion in free cash flow, can they underwrite liability for clinical AI? This is a challenge to whoever has the highest conviction.
Steve: That’s fascinating. I’ve always thought Google had the right to win medical search, even pre-AI, because consumers go to Google for healthcare information. But they never fully exploited that, maybe because they were worried about the negative Wall Street Journal headline or liability.
Historically, their cash cow was the ad business. But with changes in generative engine optimization and agents interacting with agents, those dynamics are changing. Maybe the tables have turned, and they are more willing to risk the core business to assume liability on the healthcare side.
Eric Larsen: I don’t want to be glib about liability, but with reinsurance and other mechanisms, I don’t know that it is an existential company strategy question. It is about insinuating yourself into the 18.3% of U.S. GDP that is healthcare in a central way.
The God models have decisively entered healthcare. ChatGPT has 950 million weekly active users and 40 million healthcare users per day.
We talked about this last year: humans lie to other humans to avoid stigma, judgment, or discrimination. But they will tell the truth confessionally to ChatGPT, Claude, or Gemini.
With OpenAI, Google’s co-clinician work, and Anthropic buying Coefficient Bio, you are going to see the God models step into the space.
I think Dario has been the most astute at recognizing that healthcare in this country is not a consumer business. I’m sorry to puncture the rhetoric of the last decade, but healthcare is not mediated through the consumer. Eventually it may be, but as long as reimbursement reinforces the entrenchment of the 150, you have to go through the 150.
Dario has been very astute about aligning with decision-makers on the establishment side of healthcare.
The field is wide open, and the liability question will be determinative.
The other thing is that we need a win. We are in a race between souring public perception, regulatory capture by incumbents, and the exponentials in the technology. AlphaFold was the greatest scientific achievement of the last 50 years, but it is too esoteric to make its way to kitchen table conversations.
We need to let the labs rip.
When we do our follow-up podcast, I hope we can talk about the center of gravity moving away from the gerontocracy, the 60- and 70-year-olds who lead peer-reviewed journals and Ivy-covered academic institutions on the East Coast.
The creators are no longer looking over the Atlantic Ocean. They are looking over the Pacific. It is Silicon Valley, 20-something insurrectionists who are tech-facile. They have the GPUs and the algorithms.
The center of gravity for scientific advancement is moving from east to west, from the gerontocracy to the young Turks.
The reasons we have these institutions mediating healthcare come from scarcity: scarcity of cognition, credentialing, and expertise. Multimodal LLMs democratize expertise. There is now superabundance of data and intelligence.
The nucleus of biomedical advancement is shifting geographically. Dario is aggressively getting into this. Sam is aggressively getting into this. Demis is probably the furthest along.
You and I both love Dario’s essay, “Machines of Loving Grace.” Do you know who he dedicated it to?
Steve: Demis.
Eric Larsen: Demis. He dedicated it to Demis.
I think the 150 have limited time to get in the game, diffuse this, and co-create. A good number of them have. In the last year, I have probably brought another 75 of the top 150 with me to Silicon Valley for immersive days with Dario and others, to get really smart and become facile with the exponentials.
I don’t think there are too many of the 150 that are not red-pilled at this point. But there is a small, stratified number that have gone from situational awareness to urgent mobilization.
Steve: A couple last things. I’m curious about your take on regulation. Our industry is so governed by regulation.
On one hand, this administration seems pretty hands-off, and maybe that helps the insurgents and creators. On the other hand, you cited some scary stats. Hyperscalers at 3.3% of GDP is a lot of power concentrated in very few hands.
How do you think about this administration’s view? Where would you put us as a country in terms of being prepared for this moment?
Eric Larsen: There is a lot in this administration that I don’t love. I don’t love the 17th-century mercantilist tariff nonsense. I don’t love everything about foreign policy, though there are parts I do like.
But the two areas I do appreciate are AI and healthcare. David Sacks and Sriram Krishnan have done a fabulous job, although I do want to talk about labor dislocation before we end, because I think Sacks is perpetuating a view that this won’t have labor dislocation, and that is shortsighted.
The other domain I respect is the healthcare team: Dr. Oz, Chris Klomp, Amy Gleason, Abe Sutton, Steph Carlton. That is an all-star team.
They are not as technocratic as their predecessors at CMS and HHS. I appreciate that. They are not weaponizing the instrumentality of government to over-engineer. They are wisely laissez-faire, but they have an enlightened view of stimulating deployment, even in clinical AI, not just administrative simplification.
They are a bit hamstrung by the Chevron ruling, which limits what departments can do unilaterally. They are also limited by the crazy quilt of 50 states and endless municipalities with their own regulatory patchwork.
There are 895,000 laws, regulations, and guidances across the country at the state, local, and municipal level.
This leadership, especially Chris Klomp, is articulate and thoughtful. I think they are going to create a conducive environment for stimulating this.
But the reality is that most innovation should come outside of, and irrespective of, what happens in Washington. That is why whoever assumes liability will have a real tailwind.
We also have $1 trillion of administrative spend that is automatable, amenable, or liable. I look at the $600 billion due to the adversarial payer-provider system and think there will be a lot of de-escalation.
The first chapter was inflationary because people weaponized the existing system. The next chapter will be deflationary.
I also think there is an Orwellian conspiracy of silence on job dislocation in this country.
Steve: I agree.
Eric Larsen: It is understandable. People are alarmed about job dislocation, and I am compassionate about that.
But healthcare is a $5.3 trillion sector. $2.9 trillion is labor. We have 23.8 million people employed. It is the only industrial vertical to see negative productivity growth.
Healthcare is great in this country if you are rich, educated, white, and urban. It is not great if you are not in one of those privileged categories.
We have $250 billion in medical debt. Medical debt is the leading cause of bankruptcy in this country. Deflation has to happen in healthcare.
I predict we will see 500 to 700 basis points taken out of U.S. GDP allocated to healthcare over the next several years, if we can penetrate the regulatory walls and begin to see labor substitution.
Right now, there are 1.8 million unfilled jobs in healthcare. And if you lower the cost of something, Jevons Paradox kicks in: the cheaper a commodity becomes, the more people use it. The cheaper healthcare becomes, the more people will use it.
If we add a couple decades to human longevity, we will need more healthcare. If we finally synchronize medical management with pharmacologic data, behavioral insights, and SDOH, we will need more healthcare.
But we are still going to see job dislocation. We are already seeing it in areas of automation, starting with functional verifiability. Then it will move to verticals that are codifiable or rules-based. Anything you can verify, you can automate. Anything you can decontextualize or take out of the workflow, you can automate.
That goes to BPO and offshoring. If you can take a process out and put it in India or the Philippines, you can give it to algorithms or agents. We will repatriate those jobs, but not give them to humans.
It is shortsighted not to talk about job disintermediation, because if you cannot acknowledge it, you cannot begin to solve for it.
The solution will be some combination of universal basic income or what Elon calls universal high income. Elon is projecting we will 10x global GDP in a decade. If current global GDP is $127 trillion, he is talking about $1.2 quadrillion in a decade, to the point where GDP and the economy become obsolete.
Then it becomes not universal high income, but universal provision. We provide goods and services because AI is massively deflationary.
Labor is about $60 trillion of global GDP. We have to find the bottlenecks. You can only go at the speed of your bottlenecks. If you can partially automate a process, the bottlenecks are where you need to double down on human labor.
Overall SG&A and salary, wages, and benefits may go down, but superstar employees in the bottlenecks may need to be paid more.
Automation is not monolithic. If you automate low-skilled tasks, compensation goes up and employment goes down because you need higher qualifications. If you automate high-skilled tasks, employment goes up but compensation goes down because lesser-trained employees can do the job. If you automate both, you enter the period we are talking about now.
We need an enlightened conversation about what is automatable, amenable, or liable, and how we retrain humans to do the irreducibly human things.
The fault I find with Jensen, David Sacks, and others who refuse to acknowledge job dislocation is that it prevents us from having an open conversation about reskilling and upskilling into those bottlenecks. This is going to be a big task.
Steve: It does a disservice to the technology and to the conversation. Every technology has goods and bads. We have to be honest about the bads and deal with them.
As a country, we have not been great at reskilling our workforce. That is a conversation we will have another time.
Let me close with one last question. If we are here a year from now, which I hope we will be, what is one prediction you have? What will be different a year from now in healthcare and AI?
Eric Larsen: I think we are going to enter the first deflationary period in healthcare, but it will be a jagged frontier.
Right now, intelligence itself is a jagged frontier. It spikes in certain areas with clear objective functions. But where there is squishier subjectivity, irreducibly human work, or data scarcity, it is not as good.
For health systems, I think we are going to see the biggest market share shift in a generation. We will see a resurgence of non-contiguous M&A across geographies. The biggest determinant of who is the aggregator versus the aggregated will be AI diffusion.
With a liberalizing FTC and DOJ, you will see a lot of non-contiguous $50 billion health system mergers. The top 10 systems currently control $455 billion, and that number will go up.
I think it will be deflationary. Hospitals spend $1 trillion on labor. $270 billion is administrative, and $730 billion is clinical. You will see real stratification in performance. The 1% operating margin for AI-adept organizations will go way up.
For payers, I think you will see the realization that we are entering a period of structural secular decline. It will not just be about V28, anemic rate notices, star ratings, or Wall Street Journal investigative reporting on insurer-driven coding.
You will see the creation of an AI-enabled sector with the margins of a utility. Again, there will be real stratification based on which payers move fastest to automate SG&A and use emerging medical superintelligence to predict when patients will decompensate, focus on longitudinal care, and close gaps in care.
For life sciences companies, you will see AI-engineered biologics and biomolecules enter Phase 2 and Phase 3 clinical trials. We may even see our first FDA approval and commercialization.
Instead of spending $2.6 billion on a molecule over a 10-year odyssey with a 90% attrition rate, we may see it done in one-twentieth of the time and one-twentieth of the cost.
Steve: I can’t wait to talk about that.
Those are the three big subsectors. You heard it here first. A year from now, we’re going to come back and see whether Eric was right: whether there has been deflation, particularly among the payers, providers, and life sciences companies that have been red-pilled, and whether we see the early green shoots of AI actually becoming deflationary for the most expensive industry in the world, which is healthcare.
Eric, as always, this was a masterclass. I love our conversations. It was great to look back. Thank you for joining us.
Eric Larsen: All right, my friend. Thanks, Steve.