Goldman Sachs 47th Annual Global Healthcare Conference

Great. Good afternoon, everyone. I'm Evie Kozlowski, the Life Science Tools and Diagnostics Analyst here at Goldman Sachs. I'm joined here with Eric Lewkowski, the CEO of Tempest AI. Thank you, and thank you so much for being here.

Thanks for having me.

I guess to kick us off, maybe start by walking us through Tempest's business model. How are you structured differently for most diagnostics companies? And then, I guess, what was the genesis of your unique kind of connected data strategy?

Well, that's the unique part. So I think, you know, we didn't, like other people, Tempest is almost 11 years old, and like other people, we didn't set out to just be a lab. We set out to kind of solve a problem, which was how do you contextualize the sequencing results in a way that would allow you to figure out what's the right path for this patient, What drugs should they take? What adverse events might they expect? What clinical trial can they actually enroll in? And so to do that, we needed clinical data. And so from our earliest onset, we said we have to connect molecular data and clinical data at the same time. So the challenge back then was we would go to hospitals and say, we want to sequence your patients. We just opened up this lab, so we were already kind of not established in a world where foundation medicine kind of dominated. But, oh, by the way, you have to give us your clinical data because these things don't make a ton of sense, and we want to use technology to have them make sense. And so it was a very friction-filled sales process. But, you know, fortunately, people realized that it was kind of a mess, and they needed someone to help them make sense of all of it. And so people started giving us their data, and we were sequencing, and these things just began to scale. And then, obviously, over the last 10 years, we've become the biggest players out there.

Conversation becomes a little easier. I think diagnostics represents the majority of your revenues today, but I guess how do you see that changing over time? Should we expect diagnostics to be the core engine longer term, or should we expect to see the data and applications business become more essential to the P&L?

Yeah, I mean, I think we have long said that they're both great businesses. Now, you know, they're both spaces in flux. Our data business is in flux because AI is having such a catalytic effect on the business and how you can structure data, what you can do with the data. You know, years ago we were just selling data for the purpose of kind of analytics. Today, we sell most of our data sold for the purpose of model building. But the diagnostic business is equally in flux. What's happening is pricing which people thought, let's say, when we opened up our lab maybe eight or nine years ago, people generally thought prices in next-generation sequencing would be commoditized down. They've actually gone the other way. and with companies like ours and Garton and others that are getting FDA approved and ADLT status, the pricing has actually gone up and most recently gone up quite a bit. The last two prices of big assays comparable to some of the things we have in market aren't in the $5,000 range. They're really in the $8,000 range. And so even we who we just got approval for our tumor-only assay, so we'll have kind of 100% of our solid tumor portfolio under FDA ADLT status, we're also in the midst of getting FDA approval on our liquid biopsy product. We would have thought a month ago that that product would have been priced at, let's say, $5,000 or so, but it's now likely to be priced at $7,500 or $8,000. And so you have this kind of precipitous rise in prices in sequencing, and let alone the fact that still the majority of sequencing is paid for by Medicare and Medicaid, not private payers. And that also is starting to turn a crack. So I think you're going to have this really significant rise in ASPs for sequencing companies over the next several years, contrary to what people thought. I suspect, like, in the aggregate, the margins of NGS companies will be, like, really extraordinary over the next, let's say, three to five years. That may at some point normalize, but it's going to be a pretty extraordinary ASP story for the next three to five years.

Great. And then I guess just digging more into diagnostics, so one of the differentiators of your strategy is connecting the dots for physicians and patients and then unique insights that provides them. maybe walk through these competitive advantages in the market and how much health systems and doctors weigh that aspect when deciding to use a Tempest test versus another provider?

I mean, one of the hardest things we've had to do, when I talk about the difficulty of convincing people to sign up with us, it's also not just giving us their data, but it's connecting to our ecosystems so we can get the data. And that entails this really long process of getting through legal, signing the appropriate BAAs where it needs to be in place, then getting through the IT implementation, which can take forever. And so it's a really long race. When we were public, I would equate it to mowing 3,000 lawns. Like you just have to, there's no way to cheat the system. And so we've done that with something like 5,500 hospitals in the United States. So we have built connections to a significant percentage of providers, and those connections allow us to sequence patients, collect data, generate some kind of diagnostic insight, and put the insight back in the hands of treating physicians at real scale. And not just one form of data, but typically many forms of data. So we're able to move digitized files like pathology slides or radiology scans in addition to structured and unstructured notes. So I think it's a huge advantage if you want to be in the business of kind of wrapping AI around a laboratory test result and connecting to hospitals. And so the people that want to do what we do will have to in some way, shape, or form build all those connections.

Okay. Okay. And I guess within your oncology portfolio, I mean, you have offerings across the entire cancer care continuum. Maybe starting with therapy selection, though, that space is growing really well broadly. And then you recently got FDA approval for the XTCDX. So what are you seeing there from a competitive standpoint? And how do you expect the FDA approval to change volumes or ASP?

I don't think the FDA approval will change ASP. So it could add $7,500 million of revenue and obviously with no cost, and so that part is great. But I don't think it changes adoption. Our adoption is really driven off of the connectivity we talked about a second ago. I mean, if you look at our unit growth, which has kind of been in the low 20s, somewhere in that range, versus others who might be in the low teens or not growing at all, it's predominantly because we offer this contextualized, connected, technology-enabled platform that physicians really like. And I would say this years ago, there's only 14,000 or 15,000 oncologists in the U.S., and they have extraordinary influence and power. They don't have to use anybody. And so, you know, you can just follow their ordering patterns to know what is or isn't working for them. So when they really like a company and order a lot more, that company must be doing something well. And I think that's what we've been able to achieve in therapy selection, not to mention the fact that the overall space continues to grow both as we move from sequencing only later stage patients to sequencing earlier stage patients and as the guidelines call for more comprehensive profiling more often. So you have, like, therapy selection is a good space that's growing. We're growing faster by virtue of those connections, and that's driving our units. The FDA stuff just drives ASP, but the culmination of both for us means that we expect kind of 30% revenue growth of our main diagnostic business for some time.

Okay, great. And then how should we think about the conversion to the XTCDX throughout the remainder of the year? At this point, the conversion of XTCDX

is for us kind of largely irrelevant because the only benefit of the conversion was that we were moving from, let's say, $29.23, which is the typical price we get paid as an LDT to the ADLT price of, let's say, $4,500. But now that both assays are approved under the same ADLT status, it doesn't want to make a difference. So it's less about migration. It's more about flicking the switch, and we expect the switch to be flicked kind of January 1, and so for all of 2027, we would expect the higher price.

You also recently launched the XH whole genome test for heme. I mean, maybe just talk us through the whole genome approach. How does this improve the workflow of the test? And then do you plan to do this, add WGS for solid tumors? Yeah, so there are some real benefits to whole genome sequencing.

There's obviously massive workflow benefits because you don't have panels that have all kinds of complexity. The historic challenge of whole genome sequencing was that you needed to sequence at a high enough depth of coverage to get those really critical genes like EGFR and ALK and things that you just needed. You couldn't afford to be at low pass and miss. As sequencing costs continue to come down, and they've been dropping at, let's say, a 40% reduction of total sequencing costs ex-labor for the past, like every three to five years for the past, you know, I don't know, decade or more. You're in a place where if you're planning kind of your next generation of assays, there's no reason not to plan them to be whole genome in orientation. So we will migrate our entire solid tumor portfolio over some period of time to whole genome at a relatively high depth of coverage because it's now cheap enough that it's roughly what we're paying for our panels. The challenge won't be that we could turn that switch on tomorrow. That's a relatively easy switch to flick. The challenge is that the process of getting these assays to be approved in New York, in Moldex, with the FDA, that's a really long process. So even at the point where we said, okay, we've got a whole genome assay that we would like to bring to market, it could take three years or five years to get it to market because you have to replace each one of these products. Otherwise, you'd have disruption in pricing, which we don't want to have. So I think whole genome's coming. The technology and the cost is there. It's just going to take some time.

Okay, okay. And then touching on MRD, maybe can you talk through your strategy to go after both the tumor-informed approach and the tumor-naive, and why do you think it's important to have both within your portfolio?

We had this approach from the onset. When we first got into this space like three years ago, I think people thought that approach was crazy, but now I think everyone's got the same approach, so clearly it's resonating. Our approach is always the same, which is we try to think like an oncologist. So how would an oncologist view this problem? And that's going to be the winning solution. So oncologists don't think about informed or naive. And the vast majority of oncologists aren't fixated on purely limits of detection or things of that nature. They really want a comprehensive solution that's going to kind of meet their needs, that's easy and simple and helps them make decisions. And so the challenge is like, especially in practices where the majority of patients are treated, sometimes you've got lots of tissue, sometimes you have very little tissue. Just bouncing between, for example, colorectal cancer and lung cancer, you'd find very different kind of tissue repositories by virtue of how tissue is collected. So to say to somebody, oh, when you have lots of tissue, use me. When you have a little bit of tissue, use somebody else. It's not a great experience. So we've always thought there are places for tumor-naive and places for tumor-informed, and you should have both in market. We chose to begin developing our own naive product, developed the first generation, got an assay in market, realized that the limits of detection and the PPM of that assay was not competitive given how fast the market was moving, chose to work on a next generation of that product, which we're working on now, not just in colorectal cancer, but across a whole variety of indications. So it will really be a one-chassis pan cancer. We'll still have to get approved each disease at a time, but one-chassis pan cancer. And that is going on now, and we expect those assays to be in market maybe in 27 and maybe paid for in 28. At the same time, the vast majority of our volume is actually tumor-informed and comes through a partnership. We have a personalis where we distribute that assay. I think we have through like the next three years or something where we're the exclusive distributor of that assay. And that has had really great adoption because it's a well-developed product that performs super well in the market.

Can you maybe dig into a little bit the reimbursement on the personalis test and where you stand today, where you would like it to be before you start kind of driving more volume through your sales engine?

The issue for us is, so we get paid, we have like a marketing and distribution deal with them. So we get paid, I think something like $470 today per test. And I think our ASP is about $350 because there's some accounting that moves around. But at the end of the day, as I said during our investor day a few weeks ago, Our net margin on the test is probably what it would be even if we were one company. And we're collecting totally normalized reimbursement. So we have a very healthy net margin on the test. Even if our ASP was $1,000 per test, we probably wouldn't be any better off than we are today based on the net profit we generate from that test. So for us, that's not why it's gated. It's gated because as personnel gets reimbursement one indication at a time, its ability to run the test at greater scale goes up. So if we sent them 200,000 orders tomorrow, and let's say their ASP is 300 bucks, and their cost is three times that, whatever it is, I have no idea, they would burn an enormous amount of cash.

So it's not so much like what we want to do.

It's really more that we're in this partnership with them, and we don't want to flood them with too many orders. So I think, again, we also discussed this in our investor day. I think we have about 15%. The Salesforce selling MRD is about 15% the Salesforce selling a tumor therapy selection. So we could wrap up our efforts probably six or sevenfold if we wanted and generate way more volume.

Okay, okay. And then timing of when you think from there front they might be ready to kind of have you generate more of this revenue for them?

They're making great progress. I mean, they now have three indications that are approved. They've made some progress in breast, lung, and eye response. And so obviously CRC is a big piece of the market, so that is a piece of the puzzle that they'll need to solve. But, you know, I would think over the next year or so their reimbursement is going to be a lot healthier and our ability to really ramp up volumes will keep going up. And so I think every quarter you'll see us ramping up volumes and it'll just be this steady incline. I mean, our growth rates now are extreme. It's like, I don't even know what it is, like hundreds of percent annual growth rate. So it's not like it's not growing quick.

Yeah, definitely. And then on the hereditary side of your portfolio, you know, you've talked in the past with the largely untapped hereditary cancer testing market. You recently acquired AMBRI. I guess, what initiatives do you have in place to start penetrating that market going forward? And what were some of the challenges that previously prevented some of that market penetration in the

past? So we began on, we've long, again, we've long felt you have to be comprehensive. And comprehensive is, are you at risk of getting cancer? You have cancer, now you need to be treated. That could either be, I have tissue or I have blood, and then I'm post-treatment. How do I monitor you, and how do I detect disease when it comes back? So you need to be comprehensive. So we began offering an inherited risk assay in cancer a while ago. We had several different partners eventually chose AMBRI because they're the gold standard, and then acquired them early last year. I think it's taken us a while to really understand that market. So our thoughts on how they would grow have moved around a bit. We originally thought low to mid-teens, then we thought like high teens, low 20s, then we thought like low to mid-teens. In part, it's because in the middle of that, Invitae went bankrupt, and all of a sudden that created like massive market fluctuations. But it's a really interesting market that for sure is going to grow dramatically as the market is unlocked. The craziest part of the market is there are something like, I don't even know, 1,500 to 2,000 genetic counselors who order something like a million and a half to two million tests a year. And that is really the gate that keeps the volume from growing. And yet, based upon private payers and coverage policies in place and Medicare-Medicaid, about 70 million Americans are eligible for testing and would be fully reimbursed. So it's one of the very few markets I've ever seen where you have, I don't know, 68 million tests you could run that aren't ordered and so like that's going to change i don't know how fast but that's going to change and especially in a world where the trend among individuals that is also unstoppable is i want to understand risk and i want to take control of my health care so like if we fast if i come back a decade from now do i think we'll run 10 million tests a year for sure. Do I think Ambry will win that market? I don't know, but we're in a good spot.

Yeah, a very exciting opportunity. And then maybe also talk through your exposure in rare disease. I mean, how big is that market? How are you positioned in it? And then any recent trends you're seeing within volumes or ASP there? It's a new market for us, so we're quite

small in it today. Essentially, it was a market that Ambry was in, and then they got out of it because reimbursement was terrible. And then GeneDx began getting, and others began getting good reimbursement, and so they got back into it. And then just as they got back into it a year ago, the market moved from exome to genome. So it's been kind of also a bit of a bumpy road. But they have a good product, and I think those volumes will start to grow up, will start to grow at a good clip toward the back half of this year. And, again, long term, it's a market we like a lot because we don't want to just run a test. We want to contextualize it. We want to use AI to contextualize it. If you're going to run a test, somebody else can run the test, and it's fine. But if the test requires contextualization, we have a dramatic advantage by virtue of the fact that we have, you know, 700 software engineers, you know, 500 petabytes of data, you know, as much compute capacity as the entire rest of all of health care, like times multiples. So we want to be in that. We want to be in that business. And there's almost nowhere I can think of a better place for a Tempest than in rare disorders where the clinical journey and the phenotypic story is honestly equally, if not more important, than the genetic story or the genomic story. So, you know, we should do really well there. It's just a matter of building out those products over time.

Okay, great. And then I guess shifting to the data and applications business, maybe start by walking us through kind of your solutions, trials, next, and then algos, and what customers you serve and then how you actually get paid.

Well, we don't get paid for a lot of that stuff much, which is why those are fairly nascent businesses. The cool part of our story is that our main businesses, our data licensing and AI modeling business and our diagnostic business, especially in therapy selection and hereditary. Those are like good businesses that grow and kind of pay for everything else. But the revenue story of AI applications in health care is quite small, and it's quite small because there are literally structural problems to getting paid, And namely, Congress, the American Medical Association, and the entire infrastructure doesn't yet know how to pay for AI. So, like, this isn't a Tempest problem. This is a general health care problem. And so the good news, I think, is that it feels to me like, and certainly this administration is more innovative than others, but it feels to me like people are really working hard to get their heads around, how do I pay, why wouldn't I pay, for example, for a digital pathology algorithm that can call ER and PR and HER2 status for a patient for $20, why would I pay $100 and have it be done worse? That doesn't make any sense. Why wouldn't I pay for an algorithm that's going to predict a heart attack? Why would I let someone have a heart attack? So it feels to me like there's actually movement there, a bunch of really good movement. But I don't know if those are going to be big revenue businesses in 27 or 31. Like there's no doubt in my mind one day they're going to be very, very big businesses. But I don't know when. And so we have a long history of, like, you know, I think trying to be very candid with our forecasts. And so our whole point is, like, if we can't see it, we don't forecast it. So these are all kind of cool businesses. They actually, all of them operate at real scale. So, like, we have our cardio algorithms deployed at, like, 50 or 100 hospitals. We have our care gap algorithms deployed at another 50 or 100, which means they're touching, like, millions of patients, thousands of doctors, running, like, you know, an enormous volume of these things. We just don't collect lots of money.

Okay. And then I guess when you were building out your business model, why was it important to own the data? What moat does this provide to you relative to maybe someone just joining the market?

I think that if you think about the value of kind of AI businesses, businesses that in theory are going to leverage the kind of prevailing frontier models, meaning ChatGPT, Claude, Gemini, Grok, Llama, pick a model. They basically have to have two things. You have to have two things, one or the other or both. One is proprietary data to train and then proprietary distribution so that if you generate it, you can put it in the hands of somebody who will pay you. And what's weird is we had the same mental framework 10 years ago when we started Tempest. We just were thinking about the benefits of machine learning and big data and optical character recognition and natural language processing. we weren't thinking about the benefits of large language models because they didn't really exist at scale back then. But the same logic was there, which is we want to be in the business of, like, training something proprietary and distributing it in sight. And so you have to own the data or at least have proprietary access to data. Otherwise, you do run the risk of, like, somebody else being able to do it and they don't need you.

Yeah, yeah. And touching on some of your large pharma contracts, I mean, I think, yeah, you recently mentioned that you're seeing increased penetration within some of these accounts. What's the feedback been when they decide to expand the contract? And then are there any particular use cases that they're finding with your data that maybe they didn't expect initially?

I mean, I think we're just becoming at a really rapid pace. is what's been interesting for me over the last six months is how fast we're becoming an indispensable partner to big pharma. And I think the best example, I posted this a few hours ago. In the last 10 days, we had two CEOs, the CEO of Merck and the CEO of AstraZeneca, both of whom, without our knowledge, we had no idea, independently commented on the strategic importance of Tempest to their AI initiatives, one in an earnings report and one on CNBC. So, and, you know, you just don't, I mean, if you follow Big Pharma, that doesn't happen. So I think it just speaks to the fact that, like, and this is just the beginning. I mean, I really believe that the value we can provide is extraordinary. I mean, another example is we had a big pharma company came to us and they wanted to load 25 or 30 trials onto our network. And we have this just-in-time network that now is like it's, we have too many people that want in, so there's like a long waiting list, okay? So we looked at the trials that they had and we said to them, we'll take these 13 trials, but these 13 or whatever we're not taking because they won't enroll well. And it was this aha moment for the pharmaceutical company saying, how do you know they won't enroll well? And we can just show them literally in real time the fundamental changes to care that exist today that probably didn't exist five years ago when they designed the trial that make this phase three problematic. So in a world post-antibody drug conjugates, in a world post-her-too-low, in a world post-whatever, things have changed. And so all of a sudden, your funnel now of patients is just different. And more and more, I think, big biotech and big pharma are like, I need to have that. I can't be blindsided when I have a billion-dollar franchise or a $5 billion franchise that unravels on me at the last minute because I wasn't collecting real-time molecular and clinical data in oncology.

Yeah, I mean, that's great. The other kind of question that comes up with pharma in their pipelines is, you know, AI is helping to drive better return profiles on their R&D dollars, I guess, and, you know, speed up time to market for therapies. Like, do you think eventually pharma would look to build out data sets of their own, and how do you weigh that?

I mean, I don't know if this is right. This is from our investor day, so I'll take it. I mean, I haven't verified it. I'm sure it's right because it was from an analyst who I think covers the space. And they had said that they were kind of amazed that Eli Lilly had announced it was building its own model and that it was like kind of touting its 700 terabytes of data that it had collected over the past century. I mean, I think we generate 700 terabytes of data every 12 hours or something. I don't know, every 16 hours or something. So, like, it's that crazy. So, like, so I don't know. I mean, I just think at the end of the day, we're adding something like 20 or 30 petabytes of data a month at this point. And so, and we purge data. We're adding so much data, we're purging data. Because we're like, we just, it's like the storage cost is so extreme. So I think ultimately we're just in a unique position in that. even if big pharma wanted to build these models or if the big foundation model people wanted to build these models, they're going to have to come to somebody like Tempest that has this kind of unrestricted data, at least in a de-identified format, and try to license it.

I did want to touch on the foundation model with AstraZeneca. I mean, how significant are the initial findings from this model, and then how do you see pharma utilizing these in future workflows?

Yeah, I mean, the good news is, again, I mean, Pascal talked about some of this on CNBC a couple days ago, so you can hear it from him. But it's a big deal. I mean, we took an enormous amount of data and disparate forms of data, molecular data, clinical data, imaging data, digitized H&E's, and we essentially trained a model based on all these different data modalities and then said, can you predict the outcome of this public trial and the outcome of this private trial? We had set up with AZ two benchmarks. One was we had to cross a CNDX score on a public trial and one was a private trial that we didn't know. They had trained internally a very sophisticated small model to predict that trial. Highly tuned, highly trained years. and our large model was able to hit both benchmarks, both the public trial and the private trial. So, you know, that was the massive hurdle that they needed us to get through to be like, okay, this is going to work at scale, which we obviously thought it would. And so now it's just a matter of like, now you go. You make model two, model three, model four. You refine your post-training and you try to get really predictive. And I think we're building these digital twins. And so I believe we will be able to predict with high fidelity the likelihood of patients that are going to respond and not respond in a trial. And we're predicting these things blindly. So we'll be like, okay, here are 42 patients. We think those eight had a super response. We think those eight had no response. And these 24 or whatever were in the middle. And the righter we get, the more important we become.

Wow. Okay. And with just about a minute left, I mean, what do you feel is the most underappreciated part of the Tempest story from investors? And I guess, what are you most excited about as you look at your business over the next several years?

I mean, I'm excited about all of it, both the diagnostic business, I think it's a ton of tailwind and the data business. I think the challenge for us, which I have, you know, at times not appreciated, is just how difficult it is for these two investor worlds to commingle. I mean, you have the diagnostic investors just hate the technology and data business that they don't understand, and the technology investors just hate the diagnostic business. And they just have spent no time trying to understand it, so it feels very destabilizing for them. And I imagine the same thing happened with other companies that had kind of similar dual business models going, but that has been, I think, that's the most underappreciated part of the story. If we took our data business public under normal terms, it would probably be worth significantly more than the entire, than Tempest in totality. So that will either correct itself or will at some point do as many to correct it.

Okay, great. Thank you so much. Thanks for being here.