Tempus AI: The Ark Is Already Built — FDA Approval, a $200M Foundation Model, and a Data Moat Pharma Can't Replicate
Inaugural Investor Day, Chicago — May 29, 2026
Tempus AI held its first-ever Investor Day on Friday, and the company used the occasion to deliver two genuinely new pieces of information that change the near-term financial calculus: a same-day FDA approval for its tumor-only xT CDx assay that management says adds roughly $200 of incremental average selling price across what is likely a $750 million revenue base, and the first public readout from its large-scale multimodal foundation model built in partnership with AstraZeneca on approximately $200 million of external funding. Taken together, the event made clear that Tempus is no longer arguing it will one day become a data and AI company — it already is one, and the diagnostics business is financing the build.
The FDA Approval That Changes the ASP Math
The single most financially concrete development of the day arrived the night before the event. Tempus received FDA approval for its tumor-only xT CDx assay, completing what CEO Eric Lefkofsky described as full FDA coverage across 100% of the company's DNA portfolio. The significance is straightforward: tumor-normal sequencing requires a matched blood draw that is not always obtainable in a clinical setting. The tumor-only approval removes that constraint entirely. Chief Financial Officer James Rogers said the approval enables the company to capture an incremental $200 of ASP beginning in early 2027, and management estimated the addressable test volume at roughly $750 million, making this a real and quantifiable revenue event rather than a speculative one. The companion liquid biopsy assay, xF, has already been submitted to the FDA and is expected to add a further ASP lift when approved, though Rogers was explicit that it will not affect 2026 numbers.
Rogers also noted that clinical oncology volume growth of 28% in Q1 has continued at a similar pace through April and May, and the company reaffirmed full-year guidance of $1.59 billion to $1.60 billion in revenue with $65 million of adjusted EBITDA. The 25% three-year diagnostic revenue CAGR — which would put the diagnostics business at roughly $1.9 billion — is now supported by a more concrete ASP pathway than existed before today.
The Foundation Model: Performing at or Above Expectation
The second major disclosure was the first substantive public update on the AstraZeneca-funded foundation model. Tempus is running a compute cluster of approximately 1,080 H200-class GPUs, and Lefkofsky disclosed on stage that the company also operates a roughly equivalent cluster of GB200s — which he characterized as approximately four times the performance of H200s. His framing was pointed: "Our compute capacity is probably equal to all of pharma combined." For context, he referenced recent public disclosures showing Eli Lilly with approximately 1,000 GPUs and Recursion with roughly 500, suggesting Tempus has structurally outinvested the entire pharma sector in oncology-specific compute.
The model was trained on what Tempus describes as the largest multimodal oncology dataset in existence: over 500 petabytes of data spanning 45 million patients, 9 million digitized pathology and radiology images, 4.5 million sequenced samples, and more than 400,000 richly annotated multimodal records containing DNA, RNA, clinical, imaging, outcome, and adverse event data. The pre-training run consumed the cluster at near-100% capacity for approximately 90 days. AstraZeneca set two specific performance hurdles — replicating outcomes from a publicly available trial using no trial-specific training data, and matching a highly tuned narrow model on proprietary trial data. Lefkofsky said both thresholds have been cleared, calling the performance "at or above our expectation."
Kate Sasser, who leads MRD and the foundation model program, explained the practical implication for pharma partners: "Future state is that we're going to have models that will be able to surface those insights very rapidly in a more automated fashion. What we're really talking about is the ability to uncover insights that would take perhaps weeks or months to generate to now be able to do that in a very quick fashion." The immediate commercial use case is biopharma R&D — six biotech and pharma companies already have access to one or more of the models and are using them for target identification, trial design, and portfolio prioritization.
The Data Business: Concentration Falling, Revenue Per Customer Rising
The data and applications segment grew 41% in Q1, with the core data licensing and modeling business — which Tempus calls Insights — growing faster still, partially offset by slower growth in smaller adjacencies including the CRO business, which Lefkofsky acknowledged is "shrinking or relatively flat" because the company is not investing in it. Total contract value stood above $1.1 billion at year-end 2025, with approximately $350 million allocated to 2026. Net revenue retention for 2025 was 126%, a metric the company disclosed publicly for the first time today.
Customer concentration is declining rapidly. In 2020, the top five clients represented 85% of data revenue across a base of 35 total relationships. By 2025, that concentration had fallen to 59% across 240 companies, including 19 of the 20 largest pharmaceutical companies globally. The company expanded strategic collaborations with Merck, Gilead, and BMS in recent months. Lefkofsky was asked point-blank whether the roughly 25% three-year data CAGR represents a deceleration given the current momentum, and his answer was essentially that it does not — the stated range is conservative by design, the Insights business is already tracking above that threshold, and the drag comes from smaller segments the company is not prioritizing.
Ryan Fukushima, who leads the data business, walked through three specific pharma case studies that illustrated why customers keep coming back. One global biopharma company received a 5,000-patient pre- and post-treatment dataset — one that "didn't exist before" — that surfaced four novel immunotherapy targets, with the customer calculating an ROI of 30 to 50 times their investment on a single project. A second company used Tempus data to stress-test inclusion and exclusion criteria for a Phase III colorectal cancer study, deriving net present value of "somewhere north of $500 million" from avoided trial failure. A third used the platform to resolve a first-line versus second-line positioning decision for an ADC program based on real-world PD-L1 distribution data. Fukushima's framing of the competitive moat was crisp: "Others are addressing what is happening in the real world — it is very descriptive. We're addressing why is this patient not responding to existing standard of care. That why question is what is at stake when they're designing that clinical trial."
The Immune Profile Score and the Algorithm Attachment Rate
One of the more clinically interesting disclosures was the commercial traction of algorithmic add-ons, which Tempus calls algos. The attachment rate — the percentage of orders where a physician elects to add at least one algorithm — has now exceeded 40%. The flagship example is the Immune Profile Score, or IPS, developed over roughly two and a half years by Chief Medical Officer of Oncology Ezra Cohen and his team. IPS is a quantitative immunotherapy response predictor derived from Tempus's linked DNA, RNA, and longitudinal clinical outcome data. Cohen's description was direct: "IPS high predicts a benefit to immunotherapy, IPS low predicts a patient that will not benefit." Lefkofsky added the clinical stakes — the score reclassifies approximately 20% of patients who would not be expected to respond but will, and another 20% who appear to be good candidates but won't. The existing standard biomarker, tumor mutational burden, "is just wrong too often," he said.
The foundation model is designed to industrialize this process. Rather than generating one new algorithm every six months to a year, the model is expected to surface novel biomarker-outcome relationships at much higher velocity, which are then analytically and clinically validated before appearing on reports. Lefkofsky's strategic framing: "The old world of targeted therapy will die and this new world of precision medicine will show up."
MRD: Gated by Personalis's Capacity, Not by Demand
The MRD business generated more nuance than clarity. On the tumor-informed side, Tempus's commercial arrangement with Personalis is working — Personalis has now received three Medicare coverage decisions in roughly six months, covering non-small cell lung, breast, and IO monitoring — but volume is deliberately constrained. Lefkofsky acknowledged that if Tempus sent Personalis 20 times its current order volume, the financial math for Personalis would be unsustainable. The ramp is being orchestrated carefully. Only 10% to 15% of Tempus's roughly 200-person field force is currently promoting MRD products. Demand, he said, is "way more robust than we would have thought a year ago."
On the tumor-naive side, the picture is more complicated. Tempus's first version of its proprietary naive assay launched at detection limits of roughly 500 to 1,000 parts per million — well above the market standard of approximately 100 ppm set by Natera's Signatera. The company has been developing a second-generation assay for about a year and is "getting close" to competitive sensitivity levels, with a plan to migrate the entire platform and skip from a single-indication CRC launch to a pan-cancer approach. Lefkofsky was unusually candid about the competitive landscape: tumor-naive assays across the industry, including Tempus's own, have not yet performed well enough in lung cancer — where tissue scarcity should theoretically favor them — to displace tumor-informed alternatives. "We have to let this thing play out for another couple of years," he said, cautioning against extrapolating current MRD growth trends into a decade-long trajectory.
Hereditary and Rare Disease: Ambry's Untapped Population
The Ambry Genetics integration added hereditary cancer testing at meaningful scale. Tom Schoenherr, who leads that business, noted that over 70 million Americans meet NCCN criteria for hereditary cancer testing, against a current market of roughly 1.5 to 2 million tests per year annually — an order-of-magnitude gap. Ambry's automated high-risk screening platform, called Care, is active at approximately 250 sites and is integrating directly into Epic this summer. One pilot site identified 5,000 breast cancer patients in a 60-day look-back who should have received germline testing but did not. The hereditary business is expected to return to mid-teens growth in the back half of 2026 after lapping outsized share gains in the comparable periods of 2025.
Rare disease remains a secondary priority. A whole genome sequencing test is launching this summer and is expected to increase diagnostic yield to just over 40% from roughly one-third of patients today. Rogers was measured about the opportunity: "It's still not a significant driver today." Lefkofsky's more expansive view is that the clinical profile of rare disease — a molecular finding embedded in a long and complex diagnostic odyssey — is precisely the type of problem Tempus's data contextualization infrastructure is built to solve.
The Valuation Disconnect, Acknowledged Directly
In a notably direct exchange during the Q&A, analyst Brad Bowers from Mizuho framed the core investor problem: Tempus trades at roughly four times sales if the genomics business is pressure-tested, implying near-zero value for the data and AI platform. Lefkofsky did not deflect. "If we took the data and apps business public tomorrow, I would suspect we'd trade higher than the entire market cap of Tempus and could trade at 2x it. So it's not — one could argue it's got negative value." His explanation for the discount was structural: technology investors who understand AI do not understand next-generation sequencing reimbursement dynamics, and diagnostic investors who understand sequencing do not know what to do with a foundation model. "You end up in a world where somebody is always worried about the thing they don't understand intimately." He indicated the company may eventually act to resolve the disconnect if the market does not do so organically, without specifying what form that would take.
The balance sheet improved materially in recent weeks. A convertible note issuance was used to retire the remaining term loan, saving approximately $30 million annually in interest and putting positive free cash flow in reach by year-end 2026. The long-term financial model — reinvesting two-thirds of incremental gross profit and dropping one-third to EBITDA for the next three years, then inverting the ratio — implies an acceleration of free cash flow generation beginning in year four that management believes will be self-funding for the applications business that currently generates negligible revenue despite operating at genuine scale across 140-plus hospitals and millions of screened patients.