Tempus AI Investor Day 2026 | Precision Medicine Diagnostic Data Library Sells Fuel Growth Plans

Thank you. Welcome everybody. We're going to go kind of quick, but we'll have Q&A sections throughout. So, hopefully we'll be able to get to all your questions. We also had a few releases out this morning, so just giving people a heads up. One's already I think hit, the other one's coming shortly. 10 years ago, we started Tempus to solve a single problem, which is could you use AI to essentially unlock precision medicine. Okay, given our clicking capabilities are back, I'll start over. So 10 years ago, we started attempting to solve a single problem. Could we use the artificial intelligence to unlock precision medicine? In order to do that, you basically need two things. You need vast amounts of proprietary data to build models to bring the benefits of AI to healthcare and you need a distribution system to take those insights and deliver them to the hands of physicians and patients. Tempus is unique in that it has both. We have vast amounts of data. We have vast compute and modeling capabilities and a vast distribution system to essentially target the main use cases within healthcare which is can you match patients to the right drug, the right trial, the right therapy. But in order to do this, you have to build a sustainable operating system. And that I think is one of the first places that makes Tempus truly unique is that we have built this connected ecosystem that allows us to essentially generate vast amounts of data from the clinical workflow, turn that data into insights and then essentially feed those insights back into the US healthcare system which then makes people want to connect with us more and therefore feed us more data, generate more insights, deliver more applications and this whole ecosystem is now sustainable which is probably one of the most exciting parts about it. It's large. It operates at scale and it's sustainable, meaning we don't have to invest billions of dollars consistently in making it work. It works every day.

This integrated and sustainable system is essentially spanning vast amounts of healthcare data morphologically, typical and molecular data. So this is molecular data that you would generate from sequencing patients or producing molecular data across a wide range of data. It's DNA data, RNA data, it's methylomic data, imaging data from CT scans and MRIs. It's digital pathology slides. It's clinical data in structured and unstructured format. And all this data comes together in real time to essentially create this network effect system. Something will always not work and we'll just kind of bounce and hit it too. The ecosystem we built now exists at scale. So we're connected to about 65% or two-thirds of all academic medical centers in the US. The vast majority of oncologists in the US are now connected to Tempus in some way, shape or form, ordering our tests. We run a large volume of tests and so we generate lots of that molecular data at the beginning of that flywheel. We're connected to more than 5,000 institutions throughout the United States, which means we're connected to a significant percentage of the US healthcare system. There's maybe 8,000 hospitals or so. We generate enormous amounts of data and have built a world-class team to make sense of all this data. So in addition to being sustainable and running this operating system, we are sustainable at scale.

The output of all this connectivity where we're essentially trying to generate data as part of the clinical practice, turn that data into something useful, generate an insight and put in the hands of everybody who needs it, whether those are physicians, patients or researchers. This has now produced an enormous amount of data, over 500 petabytes of data. For people who've been watching the growth of our database over time, it's really quite extraordinary. I mean, not long ago, we were at 50 petabytes of data. And here we are some five plus years later at 500 petabytes of data. It spans over 45 million patients. There's over 9 million images in that data set which is just one of the largest digital imagery data sets that we know of in the world in terms of digitized pathology slides and radiology records also connected to clinical outcome response including a large volume of samples we've sequenced over four and a half million and then the very bottom of this data set or this funnel is over 400,000 of these really really rich multimodal records. These are records where we have typically DNA and RNA and clinical data and outcome data, response data, adverse event data. We have imaging data. It's all the totality of what you would need to basically interrogate real world data and figure out all the insights we're going to talk about in a minute.

On the bio arm side, in order to make this business sustainable, we've divided it into essentially two parts. We built a diagnostic business and then a data and applications business. So I want to first start with the diagnostic business. What makes our diagnostic business unique is the comprehensive nature. If you're going to generate vast amounts of molecular data and then be in the business of connecting that molecular data to clinical data so you can contextualize it. The question you'd ask is why am I going through all this effort to contextualize molecular reports? And the answer is that if you just run sequencing and generate like here's a patient, they have a mutation and I'm going to hand that to somebody at best you're basically in the business of targeted therapies or targeted medicine. You're not really in the business of precision medicine because you know nothing about that patient. And so the journey we set out 10 years ago was could we essentially make diagnostics smart? That we help contextualize them and basically wrap technology or AI around them to help physicians make really high quality decisions and help researchers do much more efficient research. On the diagnostic side, that begins at not just generating an insight in a comprehensive manner to get an answer or result, but then bring it all the way through. I know something, I've learned something. How do I connect it to clinical data so I can essentially figure out like if I find a mutation, I don't want to recommend a therapy that a patient just took in a prior line and failed. Recommending that again would be pointless. I don't want to recommend a clinical trial that that patient is not eligible for because the patient happens to be a smoker and one of the exclusion criteria of this trial I would recommend is you can't be a smoker. So by connecting rich molecular data or any kind of laboratory test result or diagnostic data to clinical data, you can contextualize it to go from kind of answer to insight. So I found something interesting through EHR connections. I'm now going to contextualize that so I can make a more intelligent decision which is really powerful for clinical care. But that same vast amounts of multimodal data where you have an insight connected to outcome and response is also what's needed for research. So all the research that people are trying to do to figure out how to make sequencing more useful for cancer patients, our technologies are empowering that at scale. And then once you're generating lots of molecular data and you're contextualizing it and you're powering a bunch of research, you end up with this last mile which is how do I put that in as many hands as I possibly can and so the investments we've made in connectivity and AI essentially allow us or will allow us in the future to distribute these insights at scale and our goal is not just to distribute them to all cancer patients but to all patients in the United States. If Tempus is successful over time we'll be connected to every hospital in the United States, almost every hospital across all major disease areas and every time there's a diagnostic insight we will be in the middle of that trying to figure out how to take that diagnostic insight, contextualize it, wrap a whole bunch of insights around it that only people like us have because of the nature of the data we have and then deliver those to every clinician in real time so that patients are always on the right therapeutic path.

It begins with having, we started in cancer. You have to pick a place to start. So we started in cancer. We started trying to make molecular testing in cancer as intelligent as we could or comprehensive profiling as intelligent as we could. And we made a decision early on that if we were going to be in the business of contextualizing tests, we wanted to be as comprehensive as we could be. We didn't want to just give somebody part of an answer part of the time. We wanted to give them the answer all the time. And so in cancer that begins if you look at the compendium it starts with who's at risk of getting cancer. It then translates into who has just been diagnosed with cancer and how do I essentially put them on the right therapeutic path and that kind of bifurcates into solid tumor profiling or liquid biopsy because not all patients have enough tissue to be sequenced. So you need both a liquid solution or blood solution and a tissue solution. And then post treatment how do I monitor these patients? How do I look for when their disease might be coming back or see if the therapies I'm giving are sustainable? And so Tempus operates across the entire spectrum. We are strong in hereditary profiling, strong in therapy selection, both solid tumor and liquid biopsy and strong in all the ancillary tests that come along with that and then strong in MRD and monitoring. So we'll cover all that shortly.

Here's just a quick snapshot of the comprehensive nature of that portfolio. We have a series of FDA approved assays. We just added to our portfolio this morning. We had a tumor only FDA approved. Historically, and we now have tumor only approved which for us is quite significant in that it expands the amount of FDA approved tests we can offer to essentially 100% of our DNA portfolio. And given that we have ad pricing that's quite powerful. So that was a big approval for us. And then we have a series of other, as Jim will talk about in a little bit. We have a series of LDT tests and those cover RNA, liquid biopsy, other areas, whole, something of that nature. Several of those are also going down this kind of FDA regulatory approval path. We have a series of pharmacogenomic assays we offer. Things that have become super powerful these days, whether it's DPYD or UGT1A1. They also have pharmacogenomic profiling of patients with neurological issues such as major depressive disorder, bipolar disorder. They have a whole bunch of algos that sit on top of these diagnostics. We'll talk about in a second on a series of tests that typically are ordered alongside these whether it's immunohistochemistry stains or other tests of that nature. And then we have obviously a fairly large and growing portfolio in rare disease and cardiology and obviously hereditary profiling. So a significant body of assays and essentially when you understand Tempus strategy it begins with what is the diagnostic that's going to be most commonly ordered in this disease area and how do I either offer that diagnostic or partner with somebody that's offering that diagnostic. Both of which are perfectly fine solutions. Generate that diagnostic data, begin to generate or consume that diagnostic data at scale in real time across a large percentage of the US market and then begin to collect clinical data that's connected to that diagnostic so I can figure out like what's happening, what drugs are patients going on, are they responding, are they not and essentially create a self-learning system to make that diagnostic better and to contextualize it to make it personalized so that when a physician orders that diagnostic, it actually helps them figure out what to do next. I'm going to bring up Mike to talk a little bit about some of these tests in greater detail.

So just for a second to kind of provide some context. So when Tempus is about 10 and a half years old, give or take and when we started Foundation Medicine was the leader by far in solid tumor profiling and had a liquid offering and Guardant obviously leader in liquid and at the time people thought it'd be very hard for us to kind of catch up and make progress and if you look at the progress we've made over the last and the lab's maybe eight or nine years old over the last let's say eight plus years where we become number one or number two in both of those spaces and have unit growth that is kind of best-in-class. It really speaks to not only the comprehensive nature of these assays and how kind of good they are. We are the gold standard I think or in that top tier gold standards in both solid tumor profiling and liquid biopsy but also the technology we wrap around these tests. We'll show you in a few minutes our main ordering system which is called Hub. But in addition to that, we don't just run these tests in isolation. We connect all these tests to each other and we connect all these tests to clinical data and outcome data. And so what ends up happening is the ecosystem, if you're a doctor ordering our tests, it just gets smarter and smarter and smarter, which has led to such a high new physician acquisition rate and such a high physician retention rate. We talked a little bit in the last quarter about the algos or algorithmic diagnostics that sit on top of these tests and we have a variety in market today. We have algorithms that predict homologous recombination deficiency algorithms that predict site of tumor origin algorithms that predict a whole bunch of other clinically relevant topics. But also one of the newest algos we've deployed is called our immune profile score which Ezra is going to talk about in a second. These algos now have an attachment rate of greater than 40%. Which essentially means a doctor's choosing to bolt on one of these algorithms more than 40% of the time they order with Tempus which means they find enough value to say I want that additional insight. Ezra, you want to talk a little bit about IPS.

And the benefit of these algorithms is that they're essentially providing insight on top of something else that might be previously known. So there are historic markers for IO response like tumor mutational burden or TMB. Unfortunately, it's just wrong too often. These are kind of coarse scores that might be right, might be wrong. And if you have vast amounts of data, you can kind of refine the score. And in this case, we've unlocked roughly 20% of patients who you wouldn't think would respond to immunotherapy that will and 20% of people you think should respond to immunotherapy that won't. And that journey, which we'll talk about a bit more when we get into the foundation model in a second, is what we expect to happen across all biomarkers and all diagnostics, not just in cancer, other disease areas. You generate some kind of diagnostic insight laboratory test result. You connect it to rich outcome and response state over time. You track what's really going on. You refine the diagnostic. It becomes smarter. Collect data. Track. Refine. So on and so forth. Before we leave therapy selection or comprehensive genomic profiling go to MRD, I want to talk a little bit about the growth drivers of CGP. We believe that the unit growth rates we're experiencing which are pretty extraordinary are sustainable for a long period of time for a few reasons. One, CGP is still not saturated. There have been a significant number of reports come out recently that estimates the percentage of doctors ordering these tests is in the kind of 40 to 50% range. There's a significant number of folks that still don't order these tests even though they should in areas where NCCN guidelines call for these kind of tests being ordered. That's not a small percentage of cases and so we suspect the whole market will grow and since we're going faster than the market that will accrue to our benefit. The second is that we're the beneficiary of 40 plus years of research where you're essentially tying a biomarker to some kind of therapeutic benefit. It started in cancer. It started with Nixon. It started with all the genomic work we've done in that area. And so if you've looked at the trend, we've been sequencing patients earlier and earlier in their diagnosis and so we suspect over time the vast majority of patients even into stage two and stage one depending on the disease area going to be profiled. So that's another benefit. The third benefit is that there's been a significant migration to more comprehensive profiling at therapy selection. We've broadly published on the benefits of doing solid tumor profiling and liquid biopsy profiling as have many of our competitors. So there's now a large volume of work around the benefits of doing both. There's an equal body of work about the benefits of doing DNA and RNA which is kind of having an explosive moment in terms of therapeutic relevance. And so we suspect that trend will continue. So you have a bunch of docs that aren't ordering that will. You have stage expansion into earlier stage and you have this trend of being more comprehensive ordering more tests. So we suspect that top companies in therapy selection or CGP will continue to do well. Well, we suspect we'll do better than that group or most of that group in large part because of the technology investments we've made which integrate our platform, contextualize it, make it really smart and so doctors have been flocking to our platform over the last five plus years and we don't see that trend slowing down.

With that I want to hit MRD and monitoring. Let me provide a quick overview and then I'll bring up Kate. When we decided just when in each area we have to think strategically we got into therapy selection that's where we began we began in solid tumor profiling had earned the right to go to liquid that was a conscious decision we made we then began doing hereditary profiling and realized that the best way for us to win in that space long term and really to tap into the bigger part of the market that's currently untapped today was to acquire Ambry who was the gold standard of that space and we'll talk a bit about that in a second. With MRD it was a bit trickier. It was an emerging space. It was a new space. And so we had to make a series of decisions. One is what did we think was going to win long term? Was it going to be tumor informed profiling or tumor naive? So and that is still I think up for debate. And the next question is how was the science going to evolve in terms of these assays and their ability to detect cancer earlier and earlier and the only thing we knew is that unlike other spaces this space was exploding in large part because of Natera very rapidly meaning what took Foundation Medicine let's say 10 years to do from a commercial perspective Natera was doing in like a year or two so this space was growing very very quickly getting to scale, burning lots of money and yet the space hadn't yet really kind of landed in the place it's going to land long term. So our approach was to be kind of derisk that by having multiple irons in the fire. And so we made a decision to partner with whom we thought was the best in class on the tumor informed side Personalis who had a whole genome assay at the time and to make investments developing our own tumor naive portfolio knowing that those investments would be significant over time. So I want to cover, I want to say one thing and I'll bring up Kate. We have a distinct advantage in this area. One is we generate enormous amounts of data. So that data the same data that's allowed us to build best-in-class assays in solid tumor profiling and liquid will allow us to build best-in-class assays in MRD. We just can, we've got enormous data points and can learn and refine over time. The other is that we're connected to a vast number of oncologists in the United States and so we're a partner of choice for most people that have emerging technologies. With that, I'll bring it over to Kate.

So before we get to hereditary, I want to cover a few things. So on the tumor informed side by partnering with Personalis, we have a best-in-class assay in market. Obviously, they're expanding the number of indications where they're getting coverage. And so we will be kind of unlocking volume over time as unit economics continue to improve on their side. They started like a year ago had zero approved. They now have three. So and there's more coming. It also I think speaks on the tumor informed side to the benefits of our platform. We went public. We told people that at heart we're a tech company and we didn't expect to run every single diagnostic in the world across every disease area. And so we would partner with people and essentially open up our platform much the same way Apple has a platform where you have apps that they make money off of. If you look at the unit economics of MRD and tumor informed, it speaks to that. We essentially generate the kind of profits at the present moment from that test that we would generate if we landed at scale on our own. So we're a bit agnostic as to whether from a profit perspective we're agnostic as to whether or not we partner with somebody and generate the EVA we would love to generate or whether we run the test and eventually generate that. So we happen to have, we're fortunate in that regard. On the tumor naive side, we launched our first version of this assay and it was performing pretty well in terms of its stratification of patients, but the market's moving so quickly that the limits of detection ppm you have to hit is just much lower. And so we were out there with an assay at I don't know 500 or 1,000 ppm and the market was at 100. And so we began working on a next generation of our tumor naive assay about a year ago I think and we are already hitting getting close to some of the levels we put on this slide. And so we intend to migrate the entire platform to this new version. At the same time we'll replace our CRC assay. We're also focused on the fact that given how fast things are moving, if we go one indication at a time, by the time we get to the third or fourth indication, the market will move again. So we have to kind of skip a bunch of that and really go from redoing CRC to focusing on pan cancer. So that's our strategy. Tumor naive given that it represents 2 or 3% of our volume. It just isn't material today, but eventually hopefully will be. On that note, let me hit hereditary for a second with Tom and then we'll come back. Okay, it's all yours.

Thank you. So just really quickly, if you'll notice, one of the common themes here is multiple tests generating multiple amounts of data connected to other forms of data, like data about this patient over time for life. And it kind of yields over time a more intelligent platform that can grow. And so that's just our playbook area by area.

You want to come? So really, I just want to cover one thing. We're going to actually give you a quick demo of Hub for a second. We haven't demoed our two main systems. Hub is our main system that physicians use, and Lens is the main system that researchers use on the biopharma sites. We'll demo those today for a second. But all this comes together right in this giant connected ecosystem. Like every part of this company is working on this platform that essentially generates rich diagnostic data, typically molecular data, connected to other different data modalities. Again, phenotypic data, morphologic data, some kind of text or image or model, whatever, puts it all into this big giant environment where we run compute, generate an insight, put the insight back into the test, put the insight back in the hands of a physician. And so this is all happening at scale. And because we're connected to over 5,000 hospitals where we have data connections and BAAs and legal agreements and IT connections and all these really complicated things, we're in a unique position to pull data out, generate an insight by augmenting it, and then putting it back in. So with that, let's talk a little bit about our main platform. The platform essentially, you're going to see this both today in Hub and in Lens. Hub is essentially an ordering tool that we use for physicians to connect with. You can go to it directly on your iPhone or iPad. It's integrated with most major EHRs in some way, shape, or form. And then sitting inside Hub is this brain which we call Tempus One or One. And this is essentially all the benefit of the agents we have built live inside One. And we have built a ton of agents. I don't know, thousands of agents that essentially take disparate, miserable, siloed, multimodal data and make sense of it. And the challenge with large language models, regardless of what model you're using, Claude, Google's models through Gemini, ChatGPT, these models were not trained on healthcare data. They were trained on internet data. So they just don't work perfectly with digitized pathology slides or DICOM files from CT scans or rich molecular data. If you dump in trillions of A's and B's and T's and C's, you don't get much out of a large language model. So sitting inside these products is our brain, One. And with that, Laura is going to give you a quick demo.

So just to provide some context. Sorry. So you're in a demo environment because we hand you the main production environment or someone shows up to five guns or whatever. But essentially what you can see here is that we have all this technology wrapped around the report itself. So we have the ability to take the clinical data we're connecting and have the system automatically generate a summary of that patient or summary of their current therapeutic regimen or whatever's going on. Off to the side, each one of these things in the main production system, you could click on these things and actually go into the raw note and read the note itself because we pulled the note out, we've de-identified it, and the note is there. And then the system is just consistently, as you make decisions of maybe you want to go from putting a patient on a combination of immunotherapy and chemotherapy to maybe looking at a target on the RNA side, it's keeping track of all that. So at the end of the day, these cases are going to get very complex and being able to summarize what happened, being able to ask questions becomes pretty powerful.

Cool. So just at a really high level, this technology just doesn't really exist other places. So other people have ordering systems where you can track an order, get a result, but here it's all brought together allowing you to contextualize it. And the challenge with cancer cases, and I think most of the things that kill us, heart attack, stroke, cancer, later stages, they're complex. The comorbidities, the amount of things you have to consider, just complicated. And so as an oncologist or a treating physician, you think you're treating one disease, but pretty quickly you're treating another disease or another complication. And here all that information is accessible. In the case of Tempus, you can order a whole variety of tests to begin with. You can track patients, but you can also communicate with the outside world and figure out like, 'Hey, if I put this patient on this drug, what's the most notable adverse event and what does that mean? And how should I think about it?' Maybe I'm going to bring somebody back sooner or dose them differently or be careful. And this is the reality of treating patients, right? You're treating a patient, but all of a sudden they have major weight loss and so you can't put them on the next dose of chemo or give them the next IO because they're just not doing well. And so being able to get ahead of that is super powerful. On that note, we're going to jump into the foundation model for a second. These systems, we began building software to kind of integrate all this, that Agent One that sits inside the system that allows you to basically bring all these different disparate healthcare data sets into one place, is the same technology we use for our large-scale foundation model. Just for people that don't know, I'll provide a quick overview. About a year ago, we made the decision to build a large-scale foundation model in partnership with AstraZeneca and Pathos. AstraZeneca was providing the majority of the funding. They invested about $200 million to build this model and then we began building it. The model is quite significant. It sits inside a cluster of about 1,080 H200s. So it's a fairly large compute cluster. We loaded in an enormous amount of de-identified data across all these major modalities. And we've begun generating insights from this data. We had to do significant pre-training, significant compute as in run this cluster for like 90 days at 100% capacity or thereabouts and then do a post-training. We announced this morning some of the first insights from that model. And at a high level, what's amazing about this model is you're taking enormous amounts of multimodal data like BAM files at scale, clinical data, billions of notes at scale, things of that nature, and then asking it to predict what's happening. And it's performing in many instances as well as super small, highly tuned models, which Kate will cover.

Yeah. So thank you. So really quickly, and then I'll bring Ezra back up. So this is our strategy: comprehensive tests, add on a bunch of algorithmic insights that make those tests better, which we are doing today, which is driving our unit growth rate to be so high. And then the kind of next level of that would be run large foundation model at scale. Have the system instead of generating one insight every six months or year, generate an insight a week. That's just things that people didn't know. Is a patient going to respond to an EGFR inhibitor? Are they going to respond to an ALK inhibitor? Does the NRG1 fusion matter? Are they going to respond to immunotherapy? What are the adverse events most likely? How long are they going to be on this therapy? Whatever it is, generate those insights at scale. Analytically and clinically validate them. We have a machine to do that. Put them into the report. And over time, you just become like, it's hard not to get those insights because over here you're ordering a test, putting your patient on a drug, not knowing whether they're likely to respond or not. And over here you can. And we suspect that's the future. Whether we're the only company that can offer this or other people offer it, I don't know. But I'm 100% convinced that old world of targeted therapy will die and this new world of precision medicine will show up. On that note, we want to talk to you a little bit about how these algorithms are also affecting our clinical workflow. We have another product called Page or Tempus Preview, which is essentially leveraging our digital pathology library to generate a whole bunch of insights. Some of those insights are making calls early, some of them are making calls when the doctor can't get that information. So I'll pass it off to Israel.

Okay, on that note, I'm going to turn over to Jim to talk a little bit about the financials of diagnostics.

You want to start with the first couple?

So good tee-up for what's going to come later. We'll talk about this in a minute, but at a super high level, we generate an insight from these foundation models. But one main, very large foundation model, lots of micro models. You generate the insight and then you essentially have to figure out if that insight works across data that you have that you didn't use to train the model. We're fortunate that we generate so much data that we have a huge bolus of data we've used to train the models and a large bolus of data that sits off to the side. So we can make a prediction using that data and then see if it holds up in other data that we have. And then once that's true, you know you have something that's working. Then you go seek to basically go to third-party data sets and validate it. Diagnostic insights will essentially for the near-term live on our diagnostic tests because that's where they'll live. And life science insights or biopharma insights are already being used by the people that have access to these models. There are half a dozen biotech and pharma companies today that have access to one or more of these models that are using those insights to interrogate their R&D portfolio, design more intelligent Phase 2s, so on and so forth. So they're already being used and we'll talk about that in this section.

We can both cover it. I mean, I'll start. I don't follow the street model. Some of the models I've looked at essentially have like high growth rate in '26. They had high growth in '24, low growth rate in '26, '27. Then they had high growth rate in '25, low growth rate in '27, '28. Now they have high growth rate '26, low growth rate 2029. So they're just nonsensical. I mean, they're essentially saying, 'You're growing really fast, but one day you won't.' And is there any logic behind that? I have no idea what that logic is. We look out at our portfolio. If we're going to tell the world we think we're going to grow at 25%, we don't want to look stupid. So you have to believe that we really think that's going to happen. So there is no street model that I would look at and be like, 'Oh, they know something we don't know. We have more information.' And at the present moment, we believe we're going to grow at 25%. Roughly, some of it will be our volumes are pretty healthy right now. We're growing in the low 20s. So that's awesome. Some of it's going to be ASP lift. We just had kind of two big drivers of that. Roughly half of that gain showed up this morning when we had FDA approval for TTO. That's $200 of lift across a massive number of tests. I don't know the number, but it's probably $75 to $100 million of gain. So that's a real number. XF when that's approved is another big piece. So part of it's ASP lift, part of it's volume lift. And then you're launching new tests and the other things are happening. There's always pluses and minuses. Everything doesn't grow equally up and to the right. So some tests will overperform, some will underperform, but as Jim mentioned, the benefit of having this kind of a comprehensive portfolio is we're big enough to absorb that and still deliver that 25% year growth rates. In terms of MRD, we chose Personalis because we thought they had a great test. We invested in the company, obviously as of this morning, that's been a great investment. I don't even know how much money we made, but it's a lot of money. And at the end of the day, they've been a great partner and we're executing that strategy. In terms of whether or not at some point we'll look to consolidate those companies, that's obviously not for this meeting here and I don't have a good answer for that anyway. But I will say this: our strategy of not needing to run every diagnostic is the right strategy. I know it sounds a little crazy to diagnostic investors, but I promise you, Amazon doesn't make every single product. And nor does Apple make every single app. There's going to have to be technology platforms that take the US healthcare system, which is very complex, and translates it to physician care and patient care across the board. And the companies that do that can't do every single thing themselves. They'll have to find ways to partner with third parties. So we've always been focused on not just running tests, but figuring out ways to make money partnering. And as I mentioned earlier, our net unit economics today are as good with Personalis as they would be if we owned the company. So if we at one day own the company, all you get is revenue gain. You get no net income gain.

Yeah. It's also worth noting when we went public two years ago, people were like, 'Oh, you're running DNA and RNA and this is not going to get paid for and it felt kind of like you guys were on the edge.' If we were on the edge two years when we went public, we're now in the middle of the bus, like maybe getting the back of the bus. I mean, you've got reimbursement rates from our competitors that are two times ours. You've got these portfolios being rolled out where we've got competitors that are like, 'Click this button and you have 12 tests. It just will keep showing up forever.' So we are not cutting edge in terms of like, 'Hey, order a bunch of stuff and is it going to get paid for?' There are people that are way further ahead of us that are driving all kinds of unit growth rate by being aggressive. We view ourselves as not being aggressive. We view ourselves as being comprehensive but not aggressive. And that's where we want to be. We want physicians to be able to logistically make a decision and order things in an administratively intelligent way, but we never want them ordering tests they don't want and we never want to bundle in five tests like the next year when they don't really want that. So I think to the extent we were on one end, we're not. In fact, I think, Kate, you don't know who that 9% is prior. That's the other thing to keep in mind when you're thinking about reimbursing these tests. As a provider, I don't know who falls into that 9%. So I have to order both in order to get that answer. The same is true with RNA. I don't know who's going to fall into the 23% that's only going to be surfaced by RNA. So I have to order both. So logically it makes sense to reimburse. And the trend I think is at this point certainly on the RNA side, that train has left the station. We could argue whether or not there will be rules over time about how often you can order an MRD test, how often they're going to get paid for, how often treatment response monitoring is going to get paid for. I do think that space is going to be, as I mentioned earlier on, I don't think all the puzzle pieces are in the right spot. So I don't know where that's going to land, but the RNA train's left the station. It just matters way too often there. We're in the middle of a bunch of CDx. Our competitors in the middle of a bunch. You come back five years from now, there will be dozens and dozens of RNA-based therapies that you will need RNA expression for. We'll do one more. We have more Q&A, so we'll do it in rounds.

Let's jump to the data business and then we will come back to Q&A if you don't mind. Okay. So to be sustainable, two main businesses: a diagnostic business and a data business. Our data business is the one that I think for a lot of diagnostic investors is unfamiliar. So we're going to spend some time trying to walk through how it works and why it's growing and so successful. First of all, if the question is whether or not data and AI and technology are going to permeate drug discovery and development and healthcare, the answer is like 100% yes. It can't not happen. Every industry who ever has said it's not going to happen has been washed away by technology. And just giving you one example, you can go back to 1960, 1970 when people were like trading on the New York Stock Exchange like orange futures and would have bet you their life that this could never be replaced by technology and it's all been replaced. So that's just the unstoppable nature of technology and healthcare will be a beneficiary of that. But it is coming. In our case, we have spent the last 10 years really building the piping to generate a healthy and sustainable data business. That piping is all about how do you pull data out of the US healthcare system at scale? How do you combine it with something else that makes it super interesting like molecular data? How do you produce an insight and then how do you package up that insight? If you think about it, we got two end customers. We have to package up an insight for a doctor: your patient is not going to respond to this particular EGFR inhibitor, you should know that. Packaging it up for a biopharma company is far more complex. They're designing trials. It could be novel discovery. It's a lot going on. So you have to give them the tools to interrogate this data. And in that regard we stand alone. If people are wondering, when we started the IPO process two years before we went public and people, I think if I would have asked nine out of ten investors they would have never thought our data business would be this big and they all would have thought we'd have massive competition. Both of those have proven the opposite. Our data business is big and growing and we have almost no competition. And it's really the technology and tools we've built that wrap around this data, the connected platform, the analytic capabilities, the ability to deliver data at scale, interrogate at scale, that's unique. And this data comes from many sources. You can't just be a sequencer that generates DNA and RNA data and be like, 'I'm in the data business.' We have data coming from our CareGap products, our clinical trial matching products, our real-time clinical trial matching products, our AI tools and technology, our diagnostic business, our radiology products, our pathology products, our cardiology. Many, many ways to get data which allows the data set to be real, contemporaneous, and useful. It's also part of this network effect that fuels our diagnostic business is also helping our data business. The more data we collect, the more insights we generate, the smarter our platforms get, the more people want that data, they're licensing it, which allows us to invest in those tools, and it becomes this really positive virtuous cycle. This is used across the entire...

Compendium of decision-making. What I think a lot of people don't understand is why are people licensing your data. And Ryan's going to get into some use cases and the tools around it in a second, but it's really the entire R&D, the discovery and development lifecycle. Do I have the right target? Am I going after the right indication? Do I need a biomarker? How do I design my trial? These kind of questions aren't worth a million dollars or 5 million to a big pharmaceutical company or big biotech; they're worth hundreds of millions. You get it wrong, you've got a billion-dollar failure. You get it right, you've got a $10 billion franchise. Our data and our modeling tools are used really across every aspect from early-stage R&D through clinical development now into commercialization. Our products touch really anywhere you could use data or AI or modeling to help make more intelligent decisions. We're certainly not in oncology at scale and will be in other disease areas over time. Ryan, you want to jump in?

Great. Thank you. Okay. So, at a high level, I just want to cover the kind of scale and scope of our data business. We're connected to, we're working with something like 19 of the 20 largest pharmaceutical companies in the United States. That metric is held pretty constant over the past several years. The good news being we're still working with all these folks. We work with over 250 biotechs. We've signed in excess of $2 billion worth of data licensing deals. Our revenue last quarter on the data licensing side was $87 million. We have now large partnerships in place with not one or two big pharma companies but lots and this number continues to grow as we sign more and more of these kind of multi-year, $10 to $20 or $30 million a year engagements with folks. We also deliver an enormous amount of data. As Ryan talked about, the Lens platform, these capabilities exist at scale not just to generate data, structure, harmonize, clean it, analyze it but also deliver those insights including the raw underlying de-identified data to biopharma. You think about delivering petabytes of data. It's just not a small task. Ryan got into a bit of the ROI that we are used to measure, but increasingly folks are looking at this probability of technical and regulatory success and whether or not we're actually generating ROI. If they're making investments, if they're licensing $20 million of data, are they generating $60 million or $100 million of gain? We've been through rigorous analyses over and over again with people who are increasing their spend where they're roping in finance, they're roping in biz dev, they're roping in other teams to validate the return on this data licensing. Over and over again it comes back that this is accretive and so people increase their spend. We have a long history now of people increasing their spend which we'll get into in the next slide. This is typically how it works. Very rarely does somebody say, 'Hey, I'd like to sign a $100 million deal and license $25 million a year of data for the next four years.' More often than not, they start with, 'I'll do a $250,000 project or a half-million-dollar project or a million-dollar project.' They get the data, they have to try it and test it. Multiple teams are involved. They have internal computational biology resources and bioinformatics resources and biostatistical resources and R&D teams and they interrogate this data and try to figure out is it good, is it clean, can they use it, is it representative. So you have to get through all these hurdles before you get to the next project and the next project. We have a long history of going from one program, one project with one asset to multiple assets to multiple assets over multiple years to expanded partnerships and ultimately strategic partnerships. As I have said for a long time, our pricing model is similar to the large cloud providers AWS or GCP or Azure. You don't have to sign a big deal to be on Azure or GCP or AWS. You can spend $100 a year with AWS and they're happy to have you as a client or you can spend a billion dollars a year. The only thing that you gain by making a longer-term commitment both in terms of years and dollars is a reduced price. So, if you think about all the people that sign these multi-year, very large agreements with Tempus, the only thing they're getting is a discounted price, meaning they value the data so much they're willing to make a multi-year commitment because they want to save that money. I think it just speaks to the value of the product we built. Obviously there's a great example of that. Our first strategic partnership with a big pharma. We have several strategic partnerships with big biotech but with AstraZeneca that was signed I think in 2021. Obviously that relationship is going strong and they're funding our foundation model and that project runs for the next several years and so we've got a longstanding relationship. GSK was another large partner that came on board. There's a few years left in that agreement and then Merck recently signed up as another large pharma who came on board in a strategic way and that partnership is just starting to kind of grow and prosper in every which direction and it just speaks to the fact that the biggest cancer companies increasingly realize they need our data to do all the things that Ryan talked about a minute ago. If you look at just the overall relationships and concentration, we work with about 240 companies in 2025. That number should be up in 2026. We worked with 35 in 2020. So over a 5-year period, we went from basically 35 people in total licensing our data to 240. Back then in 2020, 85% of our business came from our top five clients. Now it's 59 and shrinking dramatically. That number is just kind of on a free fall down. So the good news is the business is diversifying itself over time. Increasingly people don't just want our data, they also want models. It's actually what we call it data and applications, but in reality it's not a great name because what's happened to us over the last year is it's rare that people just want our data. More and more they actually want models. You could almost call it modeling and applications. Data in and of itself is interesting, but in a world of large multimodal models, which all of our companies have some exposure to, they want to know how they can use this data to build their own proprietary models or take their proprietary models and hyperscale them with more data and actually figure out how to build something that's proprietary and advantageous to us. Almost every conversation we have now is a blend of license some data and use our capabilities to build models. Build those models on the Lens platform. We have both CPUs and GPUs connected to that platform at scale or we'll partner with you to build models in some way, shape or form. I'm going to bring Kate up to talk a little bit about the benefits of the foundation model on the biopharma side. Again, this is a huge cluster, massive amount of data that's producing insights. Some of those insights have therapeutic relevance. Some of them have research and discovery benefits and Kate will cover that.

And again, the release this morning covers some of this. There's also papers coming out that go into much greater detail about how these models predict what they predict and you can take a look at it. Ultimately we've crossed the major hurdles we had to cross. When we entered into the first foundation model agreement with AstraZeneca they had essentially two criteria we had to meet. One was a C-index score for a trial that was out in the public while the other was a trial that they had data for we didn't have data for for which they had trained a very specific model. The question was could this large-scale model replicate trial outcome data that's publicly available and privately available both where there's no model that's predictive and a highly tuned model that's predictive and that was the hurdle we had to get over and at the time that hurdle was not seen as being easy to get through given this is the first time we were building a large-scale multimodal model in oncology with all of our data. These models just get much better over time. Think about ChatGPT 1 versus whatever 5 we're on now and so the fact that it performed this well this quickly I think is an indication of what is to come. Once you have these kind of models performing at scale or insights performing at scale, you end up saying to yourself okay and this is the last part of our business is what do you do with them, how do you distribute them to the broader ecosystem. We've long been focused not just on the diagnostic side of the business and the data side but also of the application side, how do you take these applications or algorithms and distribute them broadly. Given that we have this connected ecosystem, the 5,000 plus hospitals, we're in a unique position to be able to distribute AI into the US healthcare system at scale in ways other people can't. There's all kinds of questions that people are answering every day. What critical biomarker should I target? What are the therapeutic options I have? What clinical trials is my patient eligible for? Did I overlook something? And there are also questions that they're not asking like ambient in the background is a mistake occurring that no one knows about where a care gap is being broken. We built technology once we had these connected rails and we had data flowing in and out of all these hospitals and we had the ability to in real time take that data in, generate insights and put the insight back into the hands of a provider. We chose two starting places to focus on. One is could we use this technology to match patients to clinical trials and could we use this technology to close care gaps. The third is could we use this technology to develop entirely novel algorithm diagnostics and distribute those. In the first two, clinical trial matching we call Time and our CareGap program we call Next and both of these things are operating at scale. They just don't generate lots of money which I've said many times but they operate at real scale. This is not like we've got a few people using these things. We have many providers using them, multiple care gaps deployed, millions of patients being screened. We are enrolling lots of patients in trials. We are closing lots of care gaps in real time. These things operate at scale. If they were paid for like I suspect they will one day be paid for, this would already be a large business. We're fortunate that our two main businesses, diagnostics and data, generate enough gross profit and enough investment, enough dollars we can invest that we're able to really lean into some of these forward products like our applications business that we think will one day be quite big. In cardiology, 60 plus algorithms deployed across multiple conditions. In oncology, a whole body of algorithms. In radiology, we've got a few in market today including our IPN module that operates at scale. In terms of clinical trials, we have dozens of providers enrolled in our program, I can't remember, it's 80 or 70, some very large number of providers that spans a thousand plus, 2,000 plus oncologists and we have at any given moment in time a nice basket of trials that we're able to enroll patients in in a rapid manner. This program is starting to really scale as we are moving from the stage of kind of we've been over the last 3-4 years proving that it works and now all the conversations are about enterprise engagements where big pharma is like I want to give you 15-20 trials and we now are in a position where we're like it's too many. So we're now fortunate that we're able to say to people like we can't take your 15 or 20 trials which is creating a really unique market condition. All of these technologies that allow you to scan clinical and other forms of data in real time and then generate an insight essentially are part of this future where there will be ambient AI-enabled co-pilots that exist in the wild that essentially make good doctors great doctors and great doctors superhuman. That is essentially what these products are designed to do. They're designed to kind of be there in the background, structure all this data that historically was unstructurable, generate these insights and make sure that they're in the hands of people. Not only does this require a connected ecosystem, which we have, which we've invested enormous amounts of time and money in building, but it requires really intricate technology. We have an entire technology stack surrounding this. Products like Edge and Air and Locker which are enormous in scope. We can allow a healthcare provider to give us access to this data without it losing security and protection of that data. We can bring in different forms of multimodal data that don't even reside in their EHR, DICOM files, CT scans, pathology slides, other forms of data like for example when you get a 12-lead ECG it sits in a different wave file from GE, it's not even in Epic. So we can basically with our Edge server pull all these different data forms in, structure that data and then Air is our technology platform that allows us to basically take the insights and put them back in the hands of the doctor through Epic or through other EHRs that they're using. So all of these technologies are necessary to kind of pull off this how do you listen in the background, find an insight and deploy it. In addition to our applications business that is growing rapidly, we also have an algorithmic business which we call ALOS. We believe that in the future these algos will be pervasive. We've talked about that historically. We are now not alone in that conversation. We have over the last several months I think been dragged to Washington several times. People have been here. I think the entire world now is thinking about how are we going to manage a world where we don't just generate wet lab diagnostics but dry lab diagnostics, how do we make sure these things can get validated and paid for, how do they get ordered, all of that. They can't be stopped. There's going to be more of them. Patients are going to want them and the US healthcare system will have to adapt and we want to be front and center in that change. We started in a couple of areas where we're most pervasively engaged. One is pathology. We have a variety of pathology algorithms. Some which are FDA approved like our Paige Prostate algo which actually has FDA approval. Others that are in flight with the FDA, for example our Pan-Cancer suite has a breakthrough designation, is going through that process now. We have a variety of cardio algorithms we've talked about historically. Two which are FDA approved, our ECG-based atrial fibrillation or AFib algo and our ECG-based low ejection fraction or LowEF algo. Many more in flight, many more coming. We envision a world where Tempus is able to basically generate hundreds or thousands of these algorithms. We are convinced they will one day be paid for by the normal reimbursement process. And this will eventually be probably the largest of all of our businesses by far. I'll go through one use case which we talked about a little bit historically which is the ECG-based use case because I think people kind of understand maybe oncology or pathology but let's just go to a new area. Obviously heart attack is the number one killer of people in the United States. One of the most common diagnostic insights you would get is from an ECG. This is the thing that's kind of pervasively in primary care especially for older patients. You go in and you see your doctor and you might get this as part of routine care. And yet this test that comes back is effectively wrong 3% of the time. So we run a few hundred million ECGs a year in this country and 3% of the time we're just telling people something is wrong. We're saying you're fine but you're not fine. You're likely going to have a heart attack or stroke within a year and we don't know it. As you can imagine the technologies that were built that are most in market today are now 30 or 40 or 50 years old. So they didn't use AI to make these decisions. We've taken millions of ECGs, connected them to outcome response data and other critical diagnostic data like echocardiograms and the like. We just built a really powerful portfolio of ECG-based algorithms that are either FDA approved or in the middle of being FDA approved. These things are also deployed at scale. We have lots of algorithms deployed at lots of hospitals, 140 plus hospitals touching millions of patients, not small but again doesn't generate meaningful revenue yet. Here's the basic use case of how these things will get big. We have a hospital that recently rolled out our ECG platform. It's one of the top academic medical centers in the United States. It's rolled out, they see a large number of ECGs a year. There is currently a code to reimburse part of this world at $128 per ECG. That code relates to part of the population. The code will be expanded to we think the entire population. So you could imagine as these things get to scale that one hospital alone could generate a few million dollars of revenue and multiply that by lots of hospitals and lots of ECGs and you get to several hundred million dollar business just on our AFib predictor. LowEF is even bigger. There's other algorithms coming and so I would imagine that we will for sure be running some kind of algorithmic diagnostic on all ECGs in the future and somebody will generate a billion or $2 billion dollars of revenue just from that product alone. And that same thing is going to happen with echoes and CAT scans and MRIs and mammographies and digital pathology slides. Each one of these data modalities that's being generated is an algorithmic diagnostic opportunity that's being missed today. And even at small dollars, $50 or $100 per algorithm, the impact you get is enormous. So you're going to spend even if you spend a billion dollars generating an algorithmic insight, you're likely going to save the US healthcare system 50 or 100 billion dollars of mistake because as these patients don't get caught, as they don't get found, they show up with complex disease and the most money we spend is in the last 90 days of life. Just to summarize, the data and applications business is growing rapidly. The bellwether of that business is our data licensing and modeling business which is having a moment and growing quickly and we're excited about the future of apps. On that note, you want to hit the financials?

I'll take the first, Ryan, you can take the second. Look, I think where this is going, we talked a little bit about it a minute ago. I'm convinced where this is going is over the next several years we will begin to layer real-world data insights on top of every biomarker that is really therapeutically relevant. So it will no longer be I sequence a non-small cell lung cancer patient to see if they have EGFR, ALK. It will be once I know someone has EGFR, ALK, what does it mean? Are they going to be in the one quarter of patients that have almost no response, the one quarter of patients that'll be on that drug for 5 years, or the 50% of patients that'll be somewhere in between. And I will want to know that because that will determine

Yeah. And the second question, you know, not all data is created equal. And so there are many different facets of that data market where there are existing players that have been selling data for decades, right? Those groups are actually addressing a different type of question, really around what is happening in the real world, but it's very descriptive, like looking to see what happened, playing back the news. The reason why our business has grown and why we see no competitors in our space is that we're addressing why is this patient not responding to this existing standard of care. And that question, that why question, is what's at stake when they're designing that clinical trial. So when people are making a phase three global investment decision, again 200 to 500 million in that decision, that's at stake. I really need to understand why are these patients not responding, and my hope is that my drug is going to address that. And so really we're providing data to those types of customers and those use cases and others, but that ROI, delivering that type of value and helping those biopharma companies increase the success rate of those investment decisions is why our data business has grown to a scale that hasn't really been seen in our space. And so that maybe addresses the kind of what we see in the broader market.

And I've said this earlier, like, you know, I think provocatively said like there will come a time when no phase 3s ever fail and some company will be responsible for that. If you think about it, it's not, I know it sounds crazy, but it's just like from a tech perspective, of course it's going to happen. I mean, if a large phase three fails, you either didn't understand the mechanism of action that drives your drug or what you saw in a phase two is not being seen in a phase three. There's no other reason, right? Like literally, it's like that. So both of those are solvable with real world data at scale. You can understand what drives people to respond to your drug and then you can look at and interrogate at a comprehensive level the population that's in your phase one and phase two and then look at the real world to see is that representative of what I'm going to see and then track it as you start to enroll patients in your phase three. And in almost every instance when you have big phase 3s that fail, we've gone back and looked at several massive phase 3s that have failed and we basically said ahead of time we could have predicted this failure. Like here, you can see it, I could have predicted it from just digitized H&E, let alone more complex data. So this, I think the R&D spends are going to get very efficient as AI becomes pervasive in R&D.

Yeah, I can start. So I think there's no big seismic change that we see coming. The big seismic change I think in terms of just a general business was we sold this very large foundation model deal to Pathos and that came with both a very large data license and some compute, and the compute is at a lower margin, right? So that's like, oh wait, that feels a little different than what is normally there. And there was a time when I thought people would sign a bunch of these very large deals. The way the market has evolved is they're actually signing, they'll sign lots of smaller deals, but not I think these giant big deals. Like you're seeing it now, people are saying I want to build a lung cancer model, I want to build a prostate cancer model, I want to build a digital pathology model, I want to work with you to find a new biomarker using scan data. And so I think it's just moving so fast that that's the way it seems to be moving from lots of teams across these big pharma. So I think what you're going to see is the margin profile of our data business will look similar to what it is today, if not better. I think you'll see a blend between models and data that looks at some point almost you won't be able to tell which is what. Are you paying me $2 million to build a model or licensing $2 million of data? It won't matter and most deals will have some of both. I think the most sophisticated biopharma companies that we see that are really embedding AI into those critical decisions are actually using the data now at a higher level, which means they're just going to consume more, right? So the AI systems and the use cases you deploy within a pharma company will need that, especially if you're seeing high ROI use cases for not just one phase three but you want to see it for every phase three and every phase two or every trial. That's what we've seen for the most sophisticated. I think for a majority of the market though, they're not all as sophisticated as the leading early adopters.

But similar to like the big compute guys, like you don't go to Azure or GCP and say, let me see your menu. How much of this is storage? How much of it compute? How much of it's large cluster compute, small cluster compute? How much of it's ingress? How much of this egress? You're like, I don't care. Just give me the number, okay? Because the margin profile is all pretty similar. Same thing with us. We're doing a deal right now that's we're just licensing embeddings, modeling embeddings. No files are moving, just the insights from those files that exist from a model we built on top of those files. So it looks and smells and feels like a data license but it's really just someone giving somebody something they can use to build a model. So these things are going to merge together.

Yeah. I mean I think you've already seen a piece of it, you've already seen which is we've had three quarters in a row of hundred million plus, you know, bookings, some significant way higher than that, and TCV growth. So TCV is just going up and up and up even at our scale where we're licensing a lot of data. I think that certainly is a part of it, but I think it's nothing compared to what's going to come. I think the first step is you have these CEOs saying this really matters and for that to translate into a signed contract takes time and we're in the middle of that world. But I suspect over the next, you know, one to whatever three, four, five quarters, you're going to see a lot of these folks that are realizing they have to jump in with two feet. Jump in with two feet. And I would say what I would say to you is a year ago we had a pool with like one or two people thinking about diving in. Now we have a pool with like 20 people thinking of diving in. So the question is, how many are going to dive?

Yeah. So I mean let's just take, we'll use, you know, let's take Northwestern. So you know we have spent an enormous amount of time building pipes between us and Northwestern, building an infrastructure where that data can flow freely, deploying these algorithms at scale. Oh by the way which took like, I mean years and years of effort, legal, it's not small. Then once you deploy these algorithms, when you bring AI into the healthcare system, you inevitably break some part of it. These systems weren't designed for AI. So we began running our AFib algorithm at Northwestern and all of a sudden we were producing an enormous number of patients that needed an echo or needed to wear a patch. Like they don't have like doctors lying around being like, hey, great. So you have then you're like, whoa, whoa, wait a minute. How do we now, what's the change management side of this? And so you have to go through that. And so that's why it's going to take time for these businesses to actually really scale. You have to lay the pipes or do all that kind of foundational work for years and then as the revenue starts to come, it really can be like a river. The failure of almost every AI company in our space is that the revenue is always way further than you think and the cost is always way higher than you think. And so if we just had a hundred AI companies and looked at their decks, you'd be like, oh my god, every one of them was too optimistic and failed. And most go out of business. They just can't keep investing in that horizon that's always further out. We're just super lucky that we have a business that allows us to make those investments and still generate incremental EBITDA improvement and that's just like compounding. I mean, our, our, our, our, the approval this morning adds like a ton of additional revenue that we either drop to the bottom line or choose to invest. But if you look at the kind of EBITDA generation of the business today and what's coming in 2027, we just have a lot of money that we're able to invest and still be EBITDA positive and cash flow positive where most of our competitors that also make a lot of investments are just burning money.

Yeah, so really early discovery and clinical development are kind of merged together under the R&D budgets of these pharma companies. And so that's the bulk of our business. Again, we're addressing the why aren't certain patients responding to the existing therapies. Every discovery team needs to know that. Every development team needs to know that. And that's kind of the lion's share of the budget that's being allocated to our contracts. Now that being said, we have customers that are also working with us in commercial, but that's a growth area for us, right? And so our data can be used there as well, but again, no one has been able to address the why aren't these patients responding to therapies. And that's why we haven't really had to compete with others and why we've been so focused on that because even though we're working with 19 of the top 20, we're not working at the same level for all 19 of those companies. And so our growth expansion still has opportunities just within R&D. And then you kind of have expanded customer segments that can grow even beyond just the R&D teams.

Yeah. I'll say one thing and we'll jump to the financials, we'll do Q&A again in a minute. So if you have a question, but like this, this is, I think people have not historically understood our data business and they don't understand the moat around it. So here's a great example. We have been one of the largest abstractors of cancer patient data for the last decade. I mean literally when ASCO decided to partner with somebody, we were one of their two partners. Like we've been doing this for a long time at scale. We also have invested, I don't know, a few billion dollars in technology. We have like 600 or 700 software engineers that have been focusing on this problem. So just assume we're a massive abstractor and assume we've built unbelievable technology and we only know oncology. Like that's if you think of our last 10 years. Okay. Now the world of large language models shows up and you're like, can't I just do this in an automated way? Why do I have, why do I have distractors? That journey, and we are now crossing that journey in two of the largest indications. That journey has taken us all this time to be in a world where we can be like we have the ability now to take 100% of our, you know, lung cancer patients, breast cancer patients and do automated abstraction at scale. So we're, which we're doing right now. So all of a sudden I as a data client can now access 100% of all the notes that exist across these millions of patients. But that's taken a decade of unbelievable time and energy just to get there. And even then you still need humans to ensure that it's right. And so it becomes another powerful tool as we leave early stage R&D and move deeper into development and commercialization.

Yeah.

You want to take this one or you want me to take it?

On that note,

I would think it's, I would suspect, right, I would suspect it's actually on this kind of a curve. In other words, I didn't know that. So if you said it went from 200,000 to 500,000, then I would say, okay, well, in the next 5 years, I bet it goes from 500,000 to 3 million. Like, in other words, it's especially with biotechs that have no money or smaller or pharma companies that are more budget conscious, they're far more conservative than, you know, than a giant global pharmaceutical company that can make a $20 million bet and it isn't the end of the world if it's not a great bet. To a company that's got $50 million, they've raised $100 million, you know, those bets are. So I would suspect it goes way up and you end up, if we fast forward 5 years from now, you probably have whatever 500 clients spending 3 to 5 million instead of 200 clients spending 500,000.

Yeah. Another way to think about it, next question, is like you could almost imagine a world where we're, we're probably there now. I go to every single biotech in oncology in the United States, market cap let's say sub a billion or sub 500 million and be like, hey, we'll give you access to our data just give us like 20% of your company. And I would think like almost every would be like, great, like it's that valuable, right?

Yeah. So let's talk about the ark for a second. So our compute capacity is probably equal to all of pharma combined. Like literally we have for our H200s we have a roughly equal size cluster of GB200s which are like 4x the H200s. It's like just I mean based on the numbers you just gave we probably have comparable compute to like all biotech and all pharma in oncology and then some. That doesn't even include all of our other compute which is equal to or bigger than the two clusters we set up for the foundation model. So we have a ton of compute and we have a ton of data and so I think that is the ark. I think the challenge for us in terms of both valuation and growth is we have always been focused on long-term sustainable growth. We've had opportunities in the past to kind of make decisions that would accelerate our growth in the short term but may have then hurt us in the long term and we just had not to choose those. And so I think we feel very comfortable that the aggregate business will grow 25% or so. We've called out that the data business will grow faster, that could be in the 30s, that could be even faster. There could be other parts of that business that grow slower like for example our CRO business is I think shrinking or relatively flat. So we have other parts of the business that are not growing because we're not investing in them. And so it all kind of, and we don't get into all the micros of these different businesses. But at the end of the day our two largest businesses, our oncology sequencing business and our data licensing and modeling business are growing faster than everything else because everything else is growing slower and you can see our growth rate. So, and we don't think that's slowing down. So, we're not forecasting like something to actually decelerate materially. We're just saying, hey, for 2026, expect 30 plus percent growth in the data business and the three-year growth rate is going to be called 25. And we don't benefit by saying the three-year growth rate is going to be 30 or 35. Our stock won't go up because of the first part you mentioned which is we're caught in this middle ground where technology investors who tend to invest in and think about AI don't understand diagnostics and diagnostic investors who are deeply scooped in next generation sequencing don't understand the data and AI. And so you end up in a world where somebody's always worried about the thing they don't understand intimately and so they just don't know what to. And so I don't know when that solves itself. It may solve itself. It may not solve itself in which case we could do things to solve it. But at the end of the day we've been focused this year on just kind of getting these businesses in the best spot we can, making sure that ark is as good as it can be. And if the market doesn't ultimately recognize the value of our data and apps business, which if we took the data and apps business public tomorrow, I would suspect it would trade higher than the entire market cap of Tempus and could trade at 2x it. So it's not, I mean it's one could argue it's got negative value. So we just kind of look at it and say eventually that'll either solve itself or we'll solve it.

Yeah, I'll cover the first. So the number of people currently focused on the MRD portfolio, I don't see Laura, I think she was here earlier but I'm going to say call it 15 to 30 somewhere in that range and we have about 200 plus folks in the field I think in just core oncology CGP. So think of it in terms of like if you were trying to understand like percentage, it's 10 to 15% or something somewhere in that range, maybe sub 10. So it's highly gated. And, you know, Personalis is public. So you could look at their financial statements. But if we sent them, let's just say we sent them 20 times the volume of orders we were sending them today, how much cash would they burn? I don't know how much cash they would burn. I don't know how much cash they have, but I don't think that math is sustainable. So it's all been tightly orchestrated to make sure that, you know, we are growing as they are growing and that the whole thing works so we don't end up either breaking their labs or breaking the financials of the business. And so it's in a great spot. I mean what I would, what we have said historically is the demand is way more robust than we would have thought a year ago. There doesn't seem to be a cap on it. So I'm comfortable that as we continue to invest here and expanding, we'll expand in terms of...

I'll take the first. There's no, so it's a two-part answer. There is no gate needed and we for sure will run those studies. There's no gate needed because physicians are free to make decisions based on the data you present them and we don't get paid for these insights. So if this was Oncotype DX for example and I wanted to get paid, I'd have to run a very large study to then convince somebody to pay me and put me in a guideline. But in my case, if I can predict EGFR response or ALK response, I get paid for it. And so doctors can make a decision based on what I published as to whether or not they believe that that is predictive or prognostic and whether they want to pay attention to it. So I think there's no gate. These things will come at scale, but I suspect all of them will turn into, and they'll have to be analyzed and clinically validated before they get on the report. And I suspect all of them will turn into really cool studies over time to figure out how good are they and eventually they'll weave their way into guidelines. But the river will come and we will not gate it. And I think it will be material in terms of how people have to react to it.

I think the way you have to think about this space is first of all a significant percentage of the market is CRC and in colorectal cancer you have lots of tissue. So it's not a great offering to say, hey, switch from a tumor informed assay that's really sensitive and come to a tumor naive assay that's less sensitive when you've got a lot of tissue. So we've all long suspected tumor informed would win the day in CRC. In other areas like for example lung cancer where tissue is far more scant, we think okay this is a great assay for tumor naive. The problem is those assays haven't performed that well. So tumor informed is winning the day there as well. So I think we're in this kind of weird zone where including our own, whereas the tumor naive assays aren't performing well enough to win the market at scale and the tumor naive assays need to go through this consistent R&D process getting from 500 ppm to 300 ppm to 200 ppm to 100 ppm to 50 ppm. There's some zone. I don't think you need to get to like 10 or 2, but there is a zone where you need enough sensing specificity and enough low enough limit of detection that you're actually like, okay, this thing can play against the market leading assays like Signatera. We are, you know, getting very close to there now. And so we're migrating from version one of our naive assay to version two. And now we're going to kind of roll that out at scale in terms of these different studies. But one more comment to make there, just like something that differentiates us in the space. We've talked a lot about this kind of multimodal data and what we start to see not just in Tempus but in the field at large is when you add additional modalities of data, you're able to increase signal to noise. And so as we think about the really broad data sets that we have that we're continuing to generate, part of the belief is that we will be able to then layer in these other modalities much faster and be able to then improve upon the assays and the technologies that we have. So lung cancer or whatever the indication, we'll be able to leverage imaging and all of the other things to add to that signature.

Yeah, I will say one last thing on MRD. I'm a big believer in MRD. I think it's an awesome space. We want to play in it. It is unclear to me how this whole thing shakes out. Remember, I think you have to really understand the adoption, right? You turn this test on and then you turn it on in bundles, you get doctors to basically order it in bundles on a recurring basis at scale. And so all of a sudden you go from like not a lot of volume to what looks like a lot of volume. And if you get even some minimal level of reimbursement, 500 bucks, 10,000, it's like not small. But I don't really know how that all shakes out. I feel very good that comprehensive genomic profiling has shooken out. But I think on the MRD side, we have to let this thing play out for another couple of years to figure out what's the cadence. What's the cadence doctors are going to want to order? What's the cadence that's going to get paid for? How does it really work? It will be a great space, but I don't think you can just look at the trend lines today and be like, oh, they're going to continue for the next 10 years. I think there's going to be some movement there to figure out how to rationalize what's going on. You've got some doctors ordering these things every month and some doctors never ordering them. And that is a unique paradigm relative to CGP there. You just had people that believed or didn't believe in the genome. Like, you know, I don't think it matters but once they're like, oh, it matters, the ordering patterns were pretty normalized.

Yeah. What are the logistics on that?

Thanks everyone, everybody.