Tempus AI, Inc. (TEM) Earnings Call Transcript & Summary

Awesome. Well, I appreciate everyone being here. I'm Matt Strom, Morgan Stanley Investment Banking. I lead the health care data and AI practice. Great to see you all and great to have Eric with Tempus here with us again this year. And maybe just before I start quickly, for important disclosures, please see the Morgan Stanley research disclosure website at morganstanley.com/research disclosures. If you have any questions, please reach out to your Morgan Stanley sales representative.

So with that out of the way, I think we want to jump right in. And today, we're going to really focus on a unique sort of health care data and AI story with Tempus and want to dig right in with Eric. So maybe, Eric, just off the start, I think a lot of people think of Tempus in some ways as a genomics company. It's obviously still a large portion of your revenue. But if you zoom out, you've built a really large multimodal longitudinal clinically annotated data set and now you're sort of layering this AI on top of it. Maybe for the audience, just start with, is Tempus a diagnostic company? Is it an AI company that happened to start in oncology? Maybe just start with that framing.

I think it's -- at our root, we have always been a technology company. 10 years ago, you really couldn't be an AI company other than in theory, today, if you're a technology company, you're probably an AI company at least in some way, shape or form. So today, we are absolutely an AI company. One of the challenges we've always had is that we don't come at this in terms of like, oh, we're going to build models, and that's our core business. We really come at it by saying we're going to garner access to proprietary data, use that data, whether it's enhanced by our own models or other people's models to generate insights and deploy those insights back into the clinic. And so that's our -- has always been our core business. In order to get the data, we had to basically open up a lab and start sequencing patients because that was the data that we needed in oncology to kind of generate these insights. And so that business over the last 10 years has -- we're now maybe 8 or 9 years ago, has become the biggest part of our revenues, about 3/4 of our revenues, give or take. And so we live in both of these worlds where we are both a NGS company sequencing patients and generating proprietary data and a data AI company that takes that data and generates insights, licenses the data, licenses the models, all that. So we really have these 2 different businesses going, which makes it complicated because diagnostic investors are always kind of afraid of our data business because they don't get it, they don't invest in AI or tech. AI investors are scared of the diagnostic business because they're like, I don't do diagnostics. And so we've had to straddle both of those worlds as other companies like us, like Tesla and Amazon, other people have straddle both of those worlds.

Maybe just double-click into where your data actually comes from. You mentioned the genomic side, but there's a lot of people that produce genomic data. But where does it come from? And sort of what had to be true operationally as well as culturally to ingest that data, get the trust of your customers and be able to actually use that data?

Yes. The first thing we had to do, which was unique is when we began -- because we're a tech company that began sequencing patients, from our earliest inception, we would go to people and say, "Hey, we'll sequence your patients, but you have to give us the clinical data for these patients because we're not just interested in sequencing your patients. We're interested in understanding whether or not the insights we produce from these -- from the sequencing is working. Like if we found a mutation and we recommended a drug, did your patient go on that drug and how did they respond? And so the big hurdle that is both cultural and logistical and administrative was saying to people, like we'll sequence your patients, but you got to give us all this data. And we -- not only is it give us this data, we have to be able to de-identify the data and use it for any lawful purpose we want because not only are we interested to generate insights that make our reports better, but we want to generate insights that make drug companies better. And this might sound like reasonable today, but 10 years ago, this was like heresy. Like you'd mentioned the word drug company and to most providers, they would be like, I'm never going to talk to you again. But we would go into these meetings and say to people, unless we're missing something, you people don't make drugs. And so why would we not want to make drug companies smarter. So I think culturally, from our earliest inception, we weren't just sequencing patients, we were collecting clinical data for those patients longitudinally over time. And so very quickly, we ended up amassing a very large data set of rich molecular data, DNA, RNA, other insights, connected to rich outcome and response data. And it's the combination of that data set that was -- that is and was so valuable. However, once we began amassing huge amounts of data, and you're talking hundreds of petabytes of data, we realized, okay, it's now time to start to license this data to biopharma. But when we just handed them a bunch of data, they couldn't find real value in it. So we had to build a whole array of tools around that data that make it useful, not just harmonizing and structuring the data, but really allowing people to interrogate the data, build cohorts of interest, refine those cohorts, unpack the data, unpack insights. And so if you look at our financials relative to other people in our space, especially on the diagnostics side, the most glaring standout is we have a very -- and have always had a very large technology team, I think like 700 software engineers and product folks, people like that, and enormous investments in cloud. 5 or 10x other people in our space. So we've always invested a lot in making the data useful. And that's -- we announced -- we may get to it. But if we look at our recent deal we announced with Merck, which is another very large strategic collaboration for us, you just don't have people like AstraZeneca and GSK and BMS and Merck and others signing these $100 million-plus deals unless the data is both incredibly useful and they can generate real insights from it.

And so how does that differ from -- so take the Merck example, you guys just announced a deepened collaboration with them today. It strikes me that those are, again, gearing more towards real deep collaboration relationships rather than, as you said, just access to data or something. But how does that sort of -- the relationships that you've got with your pharma customers differ on the data side from what they could go -- theoretically go find in other parts of the market, whether it's rollout data, rollout evidence, et cetera?

Yes. I mean I said this at JPMorgan a few months ago, we -- sorry, I'm just choking for a second. But we saw a few years ago, we saw people entering the data market, especially our competitors talking about how they were going to launch data businesses. And so we had some of that noise. And today, that noise has really dampened. I mean we just -- when we're working with big pharma, they're either licensing our data at scale where they're really just not licensing this kind of data. At the present moment, we just have a unique product. And so we're never in a situation where -- or at least if I think about the last year or 2, we're never in a situation where someone says, "Hey, we want to license your data, but we're also looking at somebody else, another big sequencing provider in the space, whether that's Caris or Guardant or whoever, and we're kind of -- this is their price and this is your price. Like that never happens. They either want the kind of data. They either believe that the kind of data we have can be transformative to their oncology programs, what assets to pursue in early R&D, how to design a more intelligent Phase II, how to manage site selection to ensure you're enrolling the right patients, all that. They either believe that data is transformative or they don't. If they do, they -- we're the partner of choice. And then what ends up happening is these deals all kind of start small. Merck is a great example. They all start relatively small. Somebody licenses whatever, $1 million of data, and they want to solve -- answer 1 or 2 questions and then they want to answer more questions. And then at some point, they realize they want to answer lots of questions. And if you look at our data business, any biopharma can license one file for a few thousand bucks. Like so there's no -- we don't mandate that you have to license lots of our data over multiple years. So when a client signs up for -- and in the case of Merck, it's a 5-year agreement but 4 years are committed. So when someone signs up for like 4 years of locking into lots of data, all they're getting is access and a discount, right? They're essentially getting access to our tools at a discount on the data. And so our pricing works very similar to AWS or GCP or Azure, where you can buy a little bit or a lot and all that varies is really price. And so I think it speaks to the fact that as people -- they might start small, but pretty soon, they realize like I'm going to need a ton of this data and it's integral to my programs, and I want the best price I can get. And so I'm happy to sign up for a multiyear commitment.

It strikes me that the data business for you all, maybe partially because of the more health care-focused investor base has always been a debate. The debate when I first started working with you guys was, oh, you can't produce revenue out of this, pharma is not going to pay for data. I think at least part of the debate in the market now is what's facing a lot of tech companies, which is the data is going to come from somewhere else or you can Vibecode your way into some sort of solution that's going to work, which seems ridiculous in the pharma context, but so be it. I'm just sort of curious, as you guys look at the data today, is it the scale? Is it the density? Like is it the size of the asset that makes the difference? Is it the tools you built around it? Like what are some of the moats that you feel like you're building up with your customer base besides the uniqueness of the product itself?

Yes. I mean, I think -- look, the -- I think if you think about the existential threat these days more and more is that the large foundation models are going to get so smart that they can do a lot of things other people can do. This is the whole like AI eating software. One of the challenges those models have by their own admission is that at some point, they run out of kind of free public data to train the models on. And there's varying estimates of when they run out of that data. But I think there's pretty good consensus in like '27, '28, they're hitting the ends of that. So more and more of those companies are coming to people like us saying, what data do you have? And I think the next frontier of fun is going to be the big frontier modelers trying to garner access to more and more proprietary data like the kind of data Tempus has to train their models. In our case, the data we have is really hard to replicate. First, you have to go to, in our case, I think, 5,500 of the roughly 8,000 hospitals in the United States and convince them they should give you their data, which is not quick. Then you have to get through legal, which is even slower. And then you have to get through IT, which is even slower because these people have Epic or Cerner or these large systems, they have an enormous road map of work that have to get done. And in order for us to get the data, we typically have to integrate at scale. And it has to be longitudinal. You can't just get one time point. You got to get multiple time points and not just one kind of data, you need structured data, you need unstructured data, physician progress notes, you typically need other forms of data. So we built up this really large data set. It's approaching 500 petabytes. It's connected to lots of hospitals. And so I think -- and it also, by the way, is connected to our own proprietary sequencing. So even if somebody could get their hands on the clinical data, they can't get their hands on the VCFs and BAM files, all that rich molecular data that a company like ours has unless you partner with some company like ours and try to marry it all up. And one of the flaws of other people that I think have tried to compete with us is you've had people who have lots of molecular data trying to cobble together clinical data or people with clinical data trying to cobble together molecular data, and it just doesn't work. So -- or it hasn't worked up until now. So I think we're in a unique spot. And I would suspect that it's only a matter of time. I'm running a blog post on this, so I want to give that away. But we're in regular contact with the world's largest modelers. And I would say -- and technology. And I would say their interest in this kind of data on a scale of 1 to 10 was a 1. I would say it's now like a 5. And interestingly, every one of these companies that we're engaged with, again, this is coming out in a week or 2, is asking the exact same question. They want longitudinal patient histories at scale. And if you think about it, the reason they want longitudinal patient history is not to digress is, these models are very good at predicting the next likely word. They're so good at predicting it that you can ask almost any question and they give you incredible insights, right? They become that good. And it's just because they've been trained to predict the next likely word, see spot and the next likely word is run. I think these folks believe as we do that with enough data, like the kind of data we have, you can predict -- instead of predicting the next likely word, you can predict the next likely drug or the next likely therapy that would work for a patient. And my guess is that we're relatively close to being able to train these very large models that could be truly predictive. That can start to say, like if you're on 5 milligrams of statin, should you be on 10 or if you're on this antidepressant and this hypertension medicine, is it bad for you, not for the whole world, but for you.

Individualized.

Individualized. And so I think at that point, that use case, I think, for these folks is very compelling because if you're paying $20 a month to like write an assay or to write an e-mail or like whatever and something else comes up that's nearly as good, you might stop paying $20 a month. But if you're paying $20 a month to figure out like how you're -- what drugs you should be taking and how to protect your health, it's a pretty durable use case.

It strikes me in the example you just gave, though, that the model is being tuned and trained with your data, you could see that use case. But -- and so in that case, your data is very valuable. But conversely, if you're going to go back to trying to impact the patient at the point of care, your pipes in and out of the hospitals are very valuable, too. It's sort of a go-to-market partner essentially.

And I also think -- yes, I also think we very much view it as our data is going to be central, not just in oncology, but we've got large data sets in cardio and neuro. Our data is going to be invaluable to build models and generate insights. On the consumer side, I suspect those models will be delivered by the big consumer companies. Like we have no aspiration to be that company. So they'll be delivered by Apple and Google and OpenAI and Anthropic or whatever. On the provider side -- on the pharma side and provider side, I suspect those insights will be delivered by companies like ours, both because in order to connect to the U.S. health care system, you have to be a covered entity. It's complicated. There's all kinds of logistical issues. So I think at the end of the day, we have a moat. And then in terms of pharma, they're not just interested in like asking, at least at present, asking like superficial questions. They're interested in asking incredibly detailed questions that are influenced, and this is the key part, by their own data. And in our case, we are a trusted provider, both to 8,500 oncologists in the United States and most of big pharma. And we have their data and data is moving back and forth. And I just don't see a world anywhere in the near term where the biggest pharmaceutical companies are uploading their critical clinical trial data to OpenAI or Google or whoever. I just think it's too invaluable. So I suspect we've got a pretty good moat on both sides.

Maybe we could move from the data level to the sort of intelligence or AI level for a second. You guys have had some announcements around foundational models in the space. What does that actually mean in health care? What are you referring to when you're talking about building those for -- in partnership with your customers?

Yes. So I think I'll give you the most tangible example because I think it's relatively close to being at a point where this is public as well. So like if you think about it, the foundation model we're building, and we're building 2. We're building one with AstraZeneca and Pathos. We're building a second that's pan disease on our own, 2 different compute clusters that we've established. One is about 1,000 H200s, one is roughly that size, but GB200s. And what's happening is we're building these models so we can generate multimodal insights that you just can't see unless you have enormous amounts of data. So let's just take one of those insights. So if I'm a non-small cell lung cancer patient, the standard of care is that I would be profiled for 2 particular biomarkers, EGFR and ALK. And if I'm EGFR positive, I would go on an EGFR inhibitor. That would be my guideline therapy. And like most drugs in cancer and like most drugs in many other disease areas like diabetes and cardiac patients, these drugs tend to work in episodic in different ranges. So some percentage of the population, the drug doesn't really work at all. You'd go on the drug and within 3 months, you need to go off the drug because it's not working. Some percentage of the population, you're going to be on that drug for 5 years and to have an incredible response. And then there's a big part that's in the middle. So it's very hard to take all the different clinical characteristics of patients and build models that are predictive because as you can imagine, patients that get non-small cell lung cancer are quite varied, a ton of heterogeneity. So -- but when you have a large model like we have, you can begin to train the models to look for those outliers and build predictions. And so I think we're not far away from -- on our tests, unlike other tests, not just saying this patient is EGFR positive, but also providing context. This patient is EGFR positive, high, EGFR positive, mid, EGFR positive low.

And that means do X? Is that the...

High would mean something like this patient is going to -- we predict this patient will be on -- will do very well, taking an EGFR inhibitor, whereas EGFR low would be we predict this patient will not do well. Like if you give the patient an EGFR inhibitor and tell them to come back a year later, don't be surprised they had metastatic disease. So -- and I think you will see that like we're about to open that Pandora's box. And I think it just is the beginning of an entirely new era of precision medicine, where you can collect vast amounts of data, train very big models and be unbelievably predictive so that you start to have this end of one contextualization of every drug instead of the way we are today, which is, oh, your cholesterol is high, go on 5 milligrams of a statin. Like really, should be 5, 10, 20, this, should I come in and have a calcium score? Should I whatever stress test or an [indiscernible]. And you don't know because we can't -- we're not good at stratifying risk. But these models can stratify risk. And so I would suspect that, that -- and so that I would think will be highly catalytic to our diagnostic business because we're just -- our tests are smarter and more personal than others and also highly catalytic to our data business because every pharma company over time is going to need to know where does the drug work and where does it not work because physicians are going to know that and ultimately, patients are going to know that.

Does that change the -- is there a regulatory infrastructure that needs to change for you to deploy those specific insights, the EGFR example and a reimbursement regime that needs to change? Or does that fit into the current sort of world?

I think it fits into the current world of oncology because in the current world of oncology, we give oncologists a great deal of latitude to make decisions as to how to treat these patients because that's just the world of oncology. Other disease areas are far more rigid. And also because most of these tests are LDTs, they're not FDA-approved tests. We have an FDA-approved test as a few others do, but the vast majority of tests in the market are just non-FDA approved, there's a different regulatory structure to modify those. If you want to append a medical device that's FDA approved, you have to go back to the FDA. So like our ECG algorithms, we have to get FDA approval because GE got FDA approval for its electrocardiogram. But for laboratory diagnostics, you can say all kinds of insightful things on top of that because these tests go through an alternative pathway. And they have to be reviewed by a physician in order to take action. So I think there's a pretty wide amount of latitude. I would suspect, though, over time, our competitors on the diagnostic side will want or need similar tools that help them quantify their tests. We have a test out in the market now called Immune Profile Score, which basically modifies another test called Tumor Mutational Burden, which is wrong about 20% of the time on both ends, meaning it misses patients that should get an immunotherapy and it captures patients that shouldn't. And other -- we have other competitors that have similar algorithms, and I suspect more and more coming.

Just on the model side, maybe one last question. So I think you've talked today and you've certainly talked a lot publicly in the past around the sort of integration of a lot of different types of data, genomics, pathology, clinical notes, et cetera. Are there sort of other large data sets or forms of data you need to either produce yourself or get your hands on to improve these models and improve what you guys can sort of deliver in the future?

Yes. I mean I think at the present moment, no. But in the near term, I think, yes. And we -- in our most recent letter that Jim and I wrote, we called out the fact that we were fortunate that the business was generating more gross margin because we're running at whatever, I think our growth last year was like 33% or something, but we're growing at around 30% and the cost we need to run the business are much less. And so we're generating lots of leverage in the core business. And so we made a decision to not just handle that EBITDA -- incremental EBITDA gain to the bottom line, but to hold back some of it to invest in sustaining that growth. One of those buckets of investment is new data sets, both outside of oncology in areas like immunology, but also in new data modalities in oncology, in particular, single-cell sequencing, spatial transcriptomics, epigenetic data at larger scale. So I think there's other data sets that will become important proteomics beyond base level proteomics. But right now in oncology, we have an enormous amount of data and still even with people like Merck coming on board, which is amazing, joining the ranks of some of our other large strategic partnerships, there's still -- I don't know what the total number is, but we still have well more than 50% of the biggest oncology companies, top 20 that aren't strategic clients of Tempus. Maybe we have 5 and there's 15 to go. And so I suspect over time, all those folks will also sign up.

If you think about...

At that level, they're all clients just now at that level.

Right. Expansion of that opportunity. Right. If you think about health care AI, where do you see the long-term value accruing? There's all this debate right now. Is it the data layer, the model layer, is it the application, the workflow layer. Where do you sort of see it accruing in health care as you play out the next sort of phase here?

I think we're still at the part of the curve where the data is the scarcest asset to train the models that change both patient and physician behavior. So we're at the part of the curve where those who have access to the data at scale likely have the proprietary asset. Over time, we'll move to what you do with the data becomes more important. I think there's a fork in the road, as I mentioned. There will be consumer companies that dominate one side of it, and then there will be enterprise companies that dominate the other side. I tend to think they'll be different. And so our focus is on dealing with providers of biopharma. There's -- and that's just -- so we've always thought of our business in kind of 3 buckets. There's a data generation part of our business, we're very lucky that the data generation side of our business is both high growth and generates really high margins, like 65% margin. So that's a healthy business in and of itself. And then that provides all this data that has an even higher margin, 75% and is growing even quicker. And we think both of those businesses are kind of multibillion-dollar businesses over the next, whatever, several years. And then I think the real interesting part of the story, which I think you're getting at is, look, when there's -- when data is pervasive and there's all these models out there, whether Tempus is the leader in that or one of the leaders, you're going to be generating all kinds of insights and how do we pay for those insights. And I don't have an answer for that. I think it's through like AI-enabled applications or some kind of algorithmic diagnostic that's paid for, but I can't tell you that for sure because I don't know. But it feels to me like that business eventually is the really big business. Like if these are big businesses, that's the mega big business because -- and I just use our ECG algorithm as one example. Like we have this ECG algorithm that predicts undiagnosed AFib and undiagnosed low EF, about 70% of heart attack and stroke are one of those 2 in terms of normal ECGs. So in theory, we can predict about 2% of the total error of ECGs in the United States just off our 2 FDA-approved algorithms. We run 200 million, 300 million ECGs a year in this country. That algorithm currently has partial reimbursement for a subset of that at about $128 an algorithm. But like at some point, you could imagine there being universal reimbursement at $50 to $100 for that algorithm. If somebody runs $100 million, it's a big number. And I think that is going to be repeated over and over again where we just have these algorithms that will predict mistakes that are made at scale, which type 2 diabetes medications you go on? Are you -- should you be on a statin and ACE inhibitor or some other cardiac -- pick an algorithm that -- where you've got people on the wrong drug, the wrong time, wrong dose. And so I think algos is a big business down the road.

Maybe just 2 last questions to close. I think you've talked a lot about in the short term, AI and health care is probably overhyped in the long term, it's probably underhyped. You may have played a little bit of that vision out today, but what is it that you think investors and maybe especially technology investors who don't spend as much time in health care are maybe sort of misunderstanding or could understand better about that paradigm?

Yes. And I think -- by the way, I think it's interesting because when I said that, there was -- and I think it was probably a year ago or I don't know, 7, 8 months ago, there was all this kind of euphoria around just AI more broadly. And that seems to have dissipated, at least in our case, it seems to have dissipated quite a bit. I think there's still a bunch of that private euphoria as it relates to maybe Anthropic and OpenAI. We'll see how they trade in the public market. There's certainly still a bunch of euphoria around NVIDIA. But some of the -- like everything with the word AI in it a year ago trading high, I think that has certainly gone away. And one could argue, I think probably -- I don't know if that has anything with Bitcoin, but you had some of these asset classes that felt like they were kind of risky and retail-driven that were trading high about a year ago that have all come way down. So now someone said to me, is AI overhyped in health care in the short term, I would say, no, if anything, I think we've actually crossed that chasm where the opportunity of AI in the near term is probably underappreciated. It's way underappreciated in the long term, but it's probably also now underappreciated in the near term or the short term. And I think it's because we're about to start to see -- and I suspect in '26, I believe in '26, you will start to see very tangible evidence that AI is going to impact health care at incredible scale, both from companies like ours on the provider and pharma side and companies like OpenAI or Anthropic or Google or whoever or Apple on the consumer side.

So if you played out the next -- Tempus is, I think, around 10 years old right now, a little over. If you play out the next 5 years from the company, it strikes me that there's sort of a big transformation ahead of you. So if you guys execute well on that next 5-year journey, what does the company look like at that time? Where do you think the real drivers of the business sort of sit at that point?

Yes. We've projected 25% growth for the next 3 years. But if things go well, I would suspect or I would hope we beat that pretty materially, especially on the data side. It's harder to predict the diagnostics side only because now I'm getting into like long-term trends of NGS. But I think the data long-term prediction and the AI long prediction is much higher than 25%. So I think if things go well, the business is just significantly larger. If you compound something at around 30% for 5 years, it's a much bigger number. We're starting on a base of about $1.6 billion. And so just kind of you can do the math. And so we're focused on that. We're focused on building a business that is growing rapidly, that generates lots of leverage that allows us to reinvest in ways other people can't to compound our leverage so that 20 years from now, not 5 or 2, but 20, that's how you -- I think that's how you build a very big company, right? When you look at companies like Amazon or whatever, they've just been compounding for a long, long, long time. And that's what we want to build.

Awesome. Well, thanks a lot for coming, Eric. I appreciate you being here, and we'll look forward to what you guys do in '26.

Thank you. Thanks for having me.