Tempus AI: A (Tentative) Short Thesis $tem
Company name changes, reimbursement decisions, and clinical AI vs GitHub Copilot
Tempus AI, a self-professed ‘AI-enabled precision medicine company’, went public in June. There’s a lot to like about the business:
It recorded revenues of ~531mm last year, with 2024 revenue expected to be 700mm, representing 32% YoY growth.
Its S1 emphasizes that Tempus aims to leverage generative AI and large language models to ‘unlock the power of precision medicine.’1
The company’s put a significant amount of effort into collecting vast quantities of oncology data. At the time of Tempus’ S1 filing its oncology database was 50x the size of The Cancer Genome Atlas, the largest public cancer genomics dataset available.2 Given the importance of data for training large-language models, this is important!
Its data business counts 19 of the 20 largest public pharmaceutical companies as customers. This data business has two components: Insights, which licenses out Tempus’ clinical, molecular, and imaging data to aid pharma companies as they go through the drug development cycle, and Trials, which helps pharma companies more quickly enroll patients in clinical trials.3 Its Trials product appears to have real utility: in one instance, AstraZeneca used it to enroll 25% of a clinical trial’s participants.4 This should be quite appealing to investors. Almost 90% of drug development costs come from clinical trials, and delaying a trial because you can’t find relevant patient populations gets expensive quickly.5 Solving this is valuable! Selling into the largest pharma companies has the additional benefit of being a steadier business than selling into biotech startups: If a Pfizer clinical trial fails your product can still be used to run other trials. If a biotech startup clinical fails that startup will go under and you’ve lost a customer.
So in other words, you’ve got a company doing hundreds of millions of dollars in revenue, growing quickly, that’s figured out enterprise pharmaceutical sales, and that stands to benefit enormously from the AI-wave. Consequently, the company trades at ~10x EV/2025e sales, which the sell-side notes is reasonably justified and actually at a discount to data/AI peers.6
Unfortunately, things get more complicated once one reads beyond the first few pages of the company’s prospectus. For one, AI revenue amounted to 5.5mm, or 1% of total revenue, for fiscal year 2023. Additionally, Tempus offers little in the form of products that use generative AI. The oncology AI Applications it mentions are:
Tumor Origin Test – predicts where cancer may have originated in a patient
HRD, DPYD, and Tempus Purist Tests - At the risk of massively oversimplifying things, these tests help physicians determine the most effective course of treatment for certain cancer patients
To be clear, these four models can all be valuable, but it’s odd for management to emphasize their excitement about generative AI when the majority of its AI products aren’t leveraging it. Its Tumor Origin test was ordered on ~10% of solid tumor profiles for 2023, but no usage data is given for the other three. It’s possible that this usage data was kept confidential to keep competitors in the dark, but the most likely explanation is these usage figures are quite low. There isn’t anything wrong with this, except perhaps when a company pitches itself as primarily an AI business.
Tempus also applies AI to cardiology, using clinical data to flag care gaps, or gaps between a patient’s treatment and best practices, that may exist for those with various forms of heart disease. The company recently received FDA product clearance for its ECG-AF algorithm, which uses AI to flag patients who may be at increased risk of atrial fibrillation/flutter, but again, this isn’t something that leverages generative AI.7
This brings us to Tempus One and Olivia (which is currently in Beta), products that management explicitly pitch as leveraging GenAI. Tempus One serves as an AI assistant to doctors, enabling them to quickly pull up patient information and test results. That’s useful and saves doctors time! It’s also only mentioned once a reader is almost 200 pages into the company’s S1 filing, and gets no mention during management’s discourse on the potential of GenerativeAI nor during Tempus’ first earnings call as a public company. Moreover, Tempus One is not generative in a true sense. It absolutely helps doctors stay more organized, but it’s difficult to see it as a real analog to something like Intercom Fin or GitHub Copilot. Consequently, I’m skeptical either of these offerings are viewed by management as having significant implications for the terminal value of the business.
The reality is that, from a revenue perspective, Tempus is primarily a cancer diagnostics business. The company offers a variety of different cancer assays, as well as a neuropsychiatric test that predicts which medications a patient suffering from mental illness might respond best to. The company groups all these tests under ‘Genomics’ on the income statement, and it’s an impressive business! Revenue grew from ~198mm to ~363mm last year, and gross margins went from ~24% to 48%.8 Although it’s a competitive space, there’s nothing wrong with existing presently as a diagnostics provider.9 This is especially the case when its Genomics line so clearly feeds into both the fast-growing Data business and its, at this stage mostly theoretical, AI Applications business. Tempus is keen to emphasize Genomics’ spillover effects with its ‘Cohort Lifetime Value’ figures:
“We define “Cohort Lifetime Value” as the cumulative revenue attributable to a specific cohort of de-identified records, including revenue derived both from the initial sequencing (Genomics) and licensing and related services (Data and Services), less the initial sequencing costs incurred to generate the data ultimately licensed. In 2018, the first full year that we operated a laboratory, we sequenced samples from approximately 7,500 patients. From that 2018 cohort of sequenced patients, through December 31, 2023, we generated $66.2 million of combined revenue from sequencing, data licensing of de-identified data derived from those records, analytical services, and clinical trials matching, which is approximately 7.4 times the revenue we received from sequencing of that cohort in the initial year. The total cost to sequence the 2018 cohort was $17.4 million, of which $9.0 million was covered by reimbursement for the corresponding sequencing tests.”10
While sequencing revenue is technically one-time, the data collected from an individual test can be used again and again both as part of Tempus’ Data offerings and as an input into the company’s present and future AI offerings. This ability to turn historical sequencing tests into future revenue has the upside of meaning Tempus can worry less about the unit economics of the segment as compared to competitors who view themselves only as testing providers.11 If the real value of your business is going to reside in the AI Applications diagnostic tests are used to develop, then you don’t necessarily need to sweat it if medicare reimbursement rates come down.
If one has a sufficiently long-term view, the above two paragraphs are all well and good, as is this passage that waxes poetic on the potential of generative AI within healthcare:
“The ability to leverage generative AI on top of large, harmonized, multimodal datasets provides the opportunity to make diagnostic tests more personalized, and therefore more intelligent. Intelligent Diagnostics incorporate an individual patient’s longitudinal phenotypic, morphologic, and molecular data, including outcome data from the patient’s EHR, to give laboratory test results clinical context. In doing so, Intelligent Diagnostics can leverage generative AI to make laboratory tests more accurate, tailored, and personal. The test result itself is designed to be specific to each patient and their own unique patient journey. The result is also informed by our large dataset that enables association of clinical outcomes and therapeutic response for patients who are similar to the patient being treated.”12
The issue for investors is that this potential/opportunity very much exists as only potential. The secondary issue is that a very-heady revenue multiple has been slapped onto a business whose revenue barely consists of anything AI-related. Management is of course aware of this, and despite harping on about generative AI in the S1 has made it quite clear that this revenue will take a while to materialize:
“We as a country and as a healthcare system haven’t figured out how to pay for AI…..even though we’re excited that a lot of these codes are getting issued and physicians want these kind of algorithmic diagnostics. We still have a ways to go before we can figure out how to get it into guidelines, how to get it into routine practice and how to ultimately get reimbursed for these kind of tests. And I would suspect that’s a kind of multiyear journey.……Our cardiac algorithms to predict atrial fibrillation was also approved by the FDA, which was amazing. But, again, there’s not currently a reimbursement pathway for that test.”13
This multiyear journey calls into question the degree to which management can really afford to not focus on the unit economics of its diagnostics segment. Early on in the S1 one reads that the company doesn’t need to focus on gross-profit per test, but later on one learns that there’s a team focused on getting test reimbursement to levels that are materially higher than what they are today.14 Average reimbursement per oncology test rose from $916 to $1452 from 2022 to 2023, but this figure would need to increase substantially to get to the 5k per test that Guardant commands. A focus on boosting reimbursement suggests management isn’t actually all that eager to get in any sort of price war, and for good reason! Elon might be right that all of Tesla’s value resides in whether Full Self-Driving succeeds, but that doesn’t mean Tesla’s stock price responds well when the company cuts prices on its cars.
There is of course a sense in which this whole piece takes a very uncharitable view of Tempus. Yes, AI Applications make up only a small portion of revenue now, but that revenue will grow quickly, or at least grow quickly once the healthcare system figures out how to reimburse companies for AI tests. That may not happen over the next few years, but once it does Tempus sits in an ideal position. The company’s data asset is something other testing providers currently can’t compete with; Guardant or Grail could sink large amounts of time and money into developing similar data pipelines, but it’s not something they would be able to do overnight. The testing company with the most data should be the one that’s in the best position to develop Generative AI products on top of that business, and so far that company is Tempus.
This might all be true, but that doesn’t justify the revenue multiple that Tempus trades at. I have no view on whether the company will beat estimates over the next few quarters: Tempus operates two very real business lines, and the size of its contracts with pharmaceutical/biotech companies suggests it provides real value to its customers.15 The business could, and likely will, continue to grow nicely even if its AI Applications never really take off. But it’s not yet an AI company, and it doesn’t deserve to trade at an AI multiple. You can think of clinical AI applications as sitting on the opposite end of the spectrum from AI-coding assistants. GitHub Copilot very much has product market fit and does over a 100mm in ARR. Tempus doesn’t yet have a true generative AI product, and even once it does will have to wait on reimbursement decisions before that product drives any real revenue. Over the next few years, the company’s top-line won’t be driven by AI, but instead by big pharma budgets and how much its Genomics segment continues to grow. Not to put too fine a point on it, but it’s difficult to not get overly cynical about the business’ short-term AI prospects when Tempus was renamed from Tempus Labs to Tempus AI only in 2023.16 I’d say it’s a short, but given the AI hype it’s difficult to know over what timeframe the short will work; one could easily end up losing quite a lot of money even if right on the fundamentals.
Correction: An earlier version of this research said Tempus offered no GenerativeAI products. That’s incorrect! I don’t think Tempus One or Olivia have implications for my thesis, as explained above.
Disclaimer: The information in this post is not intended to be and does not constitute investment or financial advice. You should not make any decision based on the information presented without conducting independent due diligence
Ibid, pg 2
Insights currently makes up 75% of Tempus’ Data business and is also higher margin than its Trials product.
AstraZeneca 2022 Annual Report, pg 36.
From Morgan Stanley’s initiation piece on the company: “Diagnostics, Data, and Differentiation.” In fairness Tempus currently trades at $54 a share as compared to MS’s $46 price target, but I think both numbers are too high!
Importantly, FDA Product Clearance is not the same as a product being reimbursable by health insurance companies. More on that later!
The substantial increase in gross margins is not a trend that should be expected to continue, and was a result of its RNA-sequencing panel getting a CPT code and thus qualifying for reimbursement.
Competitors for Tempus’ tests include Guardant, Natera, Exact Sciences, Veracyte…..
S1 Filing, Pg 118
Ibid, Pg 118
Ibid, Pg 147
Tempus Q2 Earnings Call
S1 Filing, Pg 194
Over a multi-year period, AstraZeneca committed to spending 220mm on Tempus products, and GSK committed to spending 180mm. Tempus also has a deal with Recursion that would be worth 160mm if they remain partners through 2028. It’s worth noting that GSK and AstraZeneca both got discounted prices, and that Recursion can pay for at least some of Tempus’ services in shares of the company (S1 Filing, Pg 122-123).
In fairness to Tempus, the company did still pitch itself as an AI business before the name change, as you can see from the 2020 funding announcement.
oooooopphhhhhhhhhhhhh