Tempus AI, Inc. [TEM] — Company Research (Task 1)

CIK 1717115 · Exchange Nasdaq · SIC 7370 (Services — Computer Programming, Data Processing) Valuation date 2026-07-27 · Last trade $43.11 (Alpaca last daily close $42.91, 2026-07-27) Memo framework investment-memo v1.4.2, run unmodified Scope of this run Tasks 1, 2, 3, 5 only. Tasks 4 (charts), 6 (catalyst calendar), 7 (assembly/publication) and 9 (ledger) are out of scope by instruction.

0. Mention-frequency analysis — RUN FIRST, BEFORE ANY VIEW WAS FORMED

Per references/mention-frequency.md, this was executed before the model, before the sell side, and before a thesis existed. Its purpose here is hypothesis generation, not confirmation. Every emerging and decaying term is listed below as an open question with no interpretation attached; the investigation of each follows in §0.3.

0.1 Corpus — and why it is not Alpha Vantage

Source: SEC EDGAR 8-K Exhibit 99.1 quarterly earnings releases, 8 consecutive quarters, 2024Q2 → 2026Q1.

The Alpha Vantage EARNINGS_CALL_TRANSCRIPT endpoint was not used: this run was allocated a hard budget of one Alpha Vantage call in total (shared 25/day cap across concurrent agents), which was spent on EARNINGS_ESTIMATES for the consensus bridge. references/mention-frequency.md explicitly sanctions the fallback used here — "8-K Ex-99.1 earnings releases from EDGAR (complete and near-constant length) both worked well on the MU and NET re-runs."

One source is used consistently across the entire series, per failure-mode #3 in that reference. The window is bounded by the IPO: TEM listed in June 2024, so 2024Q2 is the first earnings release that exists. Claims of the form "first ever" are therefore claims about the company's entire public history — which is unusually strong here, but the history is only two years long, and that is stated rather than implied away.

Limitation, stated plainly: this corpus is a written release, not a transcript. It contains no Q&A, so the prepared-vs-prompted split that the reference file requires cannot be computed. Every count below should be read as the equivalent of prepared remarks — management's own unprompted framing. That is the stronger of the two signals, but the weaker signal is simply absent, and no inference about analyst interest can be drawn from this data.

Length is stable (σ/μ ≈ 11%), so the artifact that corrupted the first ISRG pass is not a live risk here — but every series below is nevertheless reported per 10,000 words, per failure-mode #1.

0.2 The counts (per 10,000 words)

0.3 Open questions generated — and what the independent corpora returned

Listed as questions first, per the method. Provenance of each hypothesis is recorded.

Q1. Why did Genomics and Data and services go to exactly zero in 2025Q4 after six stable quarters? Generated by the raw run; initially looked like the most dramatic signal in the dataset. Answer: it is a relabeling artifact, not a decay. In the 2025Q4 release TEM renamed its two revenue lines — "Genomics" → "Diagnostics", "Data and services" → "Data and Applications". Verified directly against the Q1 2026 10-Q income statement, which uses the new labels for both the current and restated prior-year columns. This finding is reported because it is exactly the class of false positive the method is prone to, and it was caught only by reading the surrounding text rather than trusting the count. Corrected series (combined old+new labels) is flat — see the "Diagnostics line" and "Data line" rows above.

Q2. Why is stock comp the most-mentioned term in all eight quarters, at 69–93 per 10k? Generated. Answer: because it is quantitatively enormous and management structures the entire non-GAAP presentation around excluding it. Q1 2026 SBC was $52.7m vs $23.0m a year earlier — +129.4% YoY against revenue growth of +36.1%. FY2025 SBC was $124.7m; Q1 2026 alone annualises to ~$211m. SBC is excluded in full from Adjusted EBITDA, the only profit metric the company guides to. This is the single highest-conviction finding of the generative pass and it was not something looked for in advance.

Q3. Why does acquisition rise 9x and stay there? Generated. Answer: TEM has become a serial acquirer. Ambry Genetics (Feb 2025, ~$381m cash), Deep 6 AI (Mar 2025), Paige AI and OneOme (2025), and now Personalis (announced 20 July 2026, $1.5bn enterprise value). Corroborated independently by the balance sheet: total assets grew +145.6% in FY2025 (to $2,275m) and goodwill grew from $73m to $470m.

Q4. Why does volume emerge from zero, and gross margin decay by half? Generated. Answer: a deliberate shift in the disclosure frame — from margin-percentage narrative toward unit-volume narrative. Q4 2025 was the first release to disclose an MRD volume (~4,700 tests, +56% QoQ) and to lead with "Oncology volume growth of 29%". Two readings are possible and the data does not discriminate between them: (a) unit growth is the genuinely better metric for a scaling diagnostics business, or (b) volume is a more flattering frame than margin. Noted as unresolved.

Q5. Is MRD an emerging term? Generated — and the honest answer is NO, which is why it matters. The per-10k series is 7.1, 0, 0, 0, 3.9, 0, 3.5, 4.5. The highest reading is the first quarter in the window, and there is no clean near-zero→material emergence. This looked, prospectively, like the obvious "GI-at-ISRG" candidate — and the data does not support it. What did change is not the word frequency but the disclosure: a quantified MRD volume metric first appears in 2025Q4. That is a weaker and different signal than a mention-frequency emergence, and §3 of the valuation document treats it as such. Recording this negative result is the point — the ISRG post-mortem exists because three of four claims in an excited first pass did not survive checking.

Q6. Why did reimbursement/coverage and the xT/xF/xM product codes decay to zero? Generated. Product-code disappearance tracks the shift to portfolio-level ("Oncology", "Hereditary") reporting after Ambry. The reimbursement decay is not explained by anything in the corpus and remains an open question — it is a topic that matters enormously to a lab business, and management stopped writing about it after 2024Q4.

0.4 Independent corpus — ClinicalTrials.gov (the healthcare pre-revenue corpus)

Queried clinicaltrials.gov/api/v2/studies?query.spons=Tempus (free, no key). 23 studies returned. Fields were opened and read before citing, per failure-mode #2.

• NCT07211178 — "Evaluating Minimal Residual Disease (MRD) Through Longitudinal Circulating Tumor DNA", start date 2025-10-27, RECRUITING, 900 patients, conditions: Breast Cancer / TNBC / HR+ HER2−. This is a genuine, dated, independent corroboration that MRD investment is underway, and its start date (Oct 2025) precedes the first MRD volume disclosure (Q4 2025) and the Personalis acquisition (Jul 2026) — the correct leading order.
• NCT05234177 — Tempus CRC Surveillance Study, ctDNA biomarker profiling, RECRUITING since 2022, 300 patients.
• NCT07154823 — AML/follicular lymphoma molecular biomarker registry, start 2026-01-27, 550 patients.
• NCT06911632 — AI-assisted referral study in pulmonary hypertension/ILD, start 2025-07-23.

Honest characterisation: the portfolio is overwhelmingly observational biomarker registries and biobanks, not a proprietary therapeutic or device pipeline with binary readouts. These generate data assets (which is TEM's actual business) but they do not create the kind of dated, value-inflecting catalyst that a drug or device trial does. The registry corpus tells you TEM is investing in MRD and in AI-assisted care-gap detection. It does not tell you what either is worth.

0.5 What is NOT disclosed — and which corpora were attempted

Per references/unpublished-scoping.md, "not disclosed" is a research task, not a limitation to report. For each item below, the alternative corpora attempted are named with what they returned.

1. Business

Tempus is a precision-medicine company with two disclosed revenue lines and, per its own filings, one reportable operating segment. The instruction for this run was to keep the segment economics separate rather than blending them; that is possible at the gross-profit level (revenue and cost of revenue are disclosed separately) and impossible below it (no opex allocation is filed). The analysis below respects that boundary rather than inventing an allocation.

1.1 Diagnostics (75.0% of Q1 2026 revenue)

Lab services sold to physicians, genetic counselors and research institutions. Comprises: - Oncology testing (the original Tempus genomics business — xT/xF/xM tissue and liquid profiling) - Hereditary testing (acquired with Ambry Genetics, closed February 2025) - MRD (currently commercialised via the November 2023 Personalis partnership, reselling NeXT Personal)

1.2 Data and Applications (25.0% of Q1 2026 revenue)

Licensing of de-identified multimodal datasets to biopharma ("Insights"), clinical-trial matching and related services, plus the newer AI application layer (Lens analytics platform, Tempus Next care-gap detection, Paige Predict digital pathology).

1.3 The economics are genuinely different by line — and diverging

This is the most important structural fact in the business and it cuts both ways. The consolidated +310bp margin improvement is composed of a large Diagnostics gain and a real Data & Applications deterioration. The higher-margin, more "software-like" segment — the one that carries the AI narrative and the entire premium-multiple argument — is growing faster but at a falling gross margin. Blending them, as a single-segment presentation invites, conceals this entirely.

Longer arc for Diagnostics gross margin: 45.3% (9M 2024) → 58.9% (9M 2025) → 61.3% (Q1 2026). That is a ~1,600bp improvement in five quarters and it is the strongest evidenced operating-leverage fact in the file.

1.4 Key operating metrics (as disclosed)

2. The Personalis acquisition (announced 20 July 2026) — the dominant recent fact

Filed on Form 8-K, 2026-07-20 (Items 1.01, 9.01), with the merger agreement as Exhibit 2.1.

2.1 The arithmetic that matters

The exchange-ratio cap of 0.3356 implies a TEM reference price of $16.25 / 0.3356 = $48.42. TEM now trades at $42.91 — below the cap, so Personalis holders bear the downside from here and TEM shareholders are protected below $48.42. PSNL itself trades at $11.98, versus $14.40 of stock value at the capped ratio — a ~17% gross spread that prices meaningful deal and/or further-TEM-downside risk.

Dilution vs. revenue added. At the capped ratio against ~104.7m PSNL shares outstanding, the all-stock case issues roughly 35m TEM shares — approximately +19% on 179.4m outstanding — in exchange for roughly $90m of annualised revenue (Q2 preliminary × 4), i.e. ~5.6% of FY2026E revenue. On the disclosed figures this is ~19% dilution for ~5.6% revenue.

The multiple paid vs. the multiple received. TEM's own enterprise value is ~$8.33bn on FY2026E revenue of ~$1.60bn = ~5.2x forward revenue. It is paying $1.5bn for ~$90m = ~16.7x forward revenue. Personalis' trailing gross margin is 13.7% (stockanalysis.com), against TEM's 63.8%. TEM is issuing equity at ~5.2x sales to buy revenue at ~16.7x sales, and that revenue carries a gross margin roughly one-fifth of its own.

The honest counter-argument, stated because it is real and not a straw man: Morgan Stanley's published note observes that ~80% of Personalis' clinical volume already runs through Tempus channels, so the cost synergies and channel duplication savings are unusually concrete for a diagnostics deal; Personalis' low gross margin is a scale artifact that consolidation directly addresses; and TEM removes a third-party economics split on a product it already sells. Personalis also brings ~$194m of net cash. None of this is refuted by the multiple arithmetic — but none of it has been quantified by the company either, and no synergy target has been disclosed.

2.2 The market's verdict

TEM fell −7.7% on the announcement day and −18.6% over the following four sessions (17 Jul $52.47 → 24 Jul $42.69), making a new 52-week low. Sell-side coverage was largely unmoved (Needham reiterated Buy, $75). The equity market and the sell side disagree sharply about this transaction, and that disagreement is unresolved as of the valuation date.

3. Financial position

All figures tied to the filed 10-K (FY2025, filed 2026-02-24) and 10-Q (Q1 2026, filed 2026-05-05).

FY2024's $534m SBC and $691m operating loss are IPO-vesting distortions (June 2024 listing) and are not a run-rate.

Three facts stand out.

1. GAAP operating loss is not improving. −$252.9m in FY2025; Q1 2026 at −$84.7m versus −$68.7m in Q1 2025 — 23.3% wider year on year on 36.1% more revenue. The FY2026 guide (revenue ~$1.595bn, Adjusted EBITDA ~$65m) implies a FY2026 GAAP operating loss of roughly −$250m, i.e. essentially unchanged from FY2025 despite +25% revenue.
2. SBC is compounding faster than the business. +129.4% YoY in Q1 2026 against +36.1% revenue. At the Q1 run-rate SBC is ~13–15% of revenue and is excluded in full from the only guided profit metric.
3. The balance sheet has levered up. Cash and marketable securities $643.8m against total debt of ~$1,233m (convertible senior notes $729m net, convertible promissory note $199m, Ares term loan $205m, revolver $100m) — net debt ~$589m, versus a net cash position two years ago. Operating cash flow remains ~−$73m per quarter.

3.1 Adjusted EBITDA — read the definition

Adjusted EBITDA is defined (FY2025 release) as net loss excluding interest, D&A, taxes, equity-method losses, fair-value changes on warrants, warrant assets, marketable equity securities, contingent consideration and holdback liabilities, stock-based compensation, employer payroll tax on SBC, acquisition-related expenses, IPO franchise taxes, other tax, loss on debt extinguishment, and — subtracting — amortisation of deferred income from the SB Tempus IP licence.

FY2025 bridge: net loss −$245.0m → EBITDA −$136.7m → Adjusted EBITDA −$7.4m. The $129m of distance between EBITDA and Adjusted EBITDA is predominantly the $124.7m SBC add-back.

The company explicitly declines to reconcile guided Adjusted EBITDA to net loss. The FY2026 guide of ~$65m of Adjusted EBITDA sits against ~$210m of annualised SBC; on the same disclosures GAAP operating income remains roughly −$250m. Both statements are true simultaneously, and which one is treated as "the earnings" determines the entire valuation.

4. Ownership, insiders and control

4.1 Control

Dual-class: Class A (1 vote) and Class B (30 votes per share). At 31 March 2026: 174,360,831 Class A and 5,043,789 Class B. Class B is convertible 1:1 at the holder's option. Founder/CEO Eric Lefkofsky holds the Class B economics. Insider ownership is reported at 35.38% (stockanalysis.com, 28 Jul 2026). Public Class A holders have limited ability to influence outcomes.

4.2 Insider transactions (SEC Form 4, all filings since 2025-07-01, parsed from EDGAR)

Top sellers: Eric Lefkofsky (CEO & Chairman) $171,255,149 — 90% of all insider selling. Then Jim Rogers (CFO) $4.2m, Ryan Fukushima (COO/CEO Data) $6.7m across roles, Andrew Polovin (General Counsel) $3.7m, Theodore Leonsis (director) $1.5m.

Applying the required empirical asymmetry (Cohen, Malloy & Pomorski 2012): insider sales are mostly noise — diversification, taxes, 10b5-1 plans — and should not be presented as bearish evidence. The reference explicitly carves out sales that are "extreme in size or unusually timed." $171m from one executive over twelve months, against a $7.7bn market capitalisation, is at the boundary of that carve-out; the filings reviewed do not conclusively establish 10b5-1 status for the full programme, and that is stated as a limitation rather than resolved by assumption. It is scored here as weak negative evidence, not as a mechanism.

The insider buys are not what they appear, and the check mattered. All 12 "P" transactions are dated 2026-02-13 at prices of $7.67–$11.00 — prices that correspond to Personalis, not TEM (which traded near $60 then). The filer is "Tempus AI, Inc." itself: this is TEM accumulating Personalis shares in the open market five months before announcing the acquisition, not an insider buying TEM. There is no insider buying of TEM stock in the window. Under Cohen/Malloy/Pomorski the informative signal — clustered, non-routine insider purchases — is absent.

4.3 Institutional ownership

Institutional ownership 52.53% (stockanalysis.com). Named holders include Baillie Gifford (~8.74m shares, growth-mandate active) and ARK Investment Management (~7.56m shares, thematic active). Total institutional shares rose ~5.06% over three months. Notably, the largest disclosed holders here are active thematic managers rather than index capital — which is the opposite of the usual pattern and means the marginal holder is a discretionary seller, not a mechanical one. Full 13F reconstruction was not attempted; per references/edgar-pipeline.md §4 there is no single EDGAR endpoint for all holders of a ticker, and this is flagged as partial rather than presented as complete.

4.4 Short interest

30.36m shares short = 26.45% of float, 16.86% of shares outstanding; short ratio 5.03 days (stockanalysis.com, 28 Jul 2026). Ortex reported ~30.5% of float in June 2026 — described in press coverage as a record high, and reached before the Personalis announcement. Days-to-cover rose from 4.8 to 6.5 on that print. This is among the most heavily shorted mid-caps in US healthcare and is load-bearing for Task 5 Gate 5.

5. Industry, competition and TAM

5.1 Competitive set

Precision oncology diagnostics: Foundation Medicine (Roche), Guardant Health, Caris Life Sciences, Natera, Exact Sciences (acquired by Abbott, March 2026), NeoGenomics. Hereditary: Myriad Genetics, Invitae (estate), Ambry (now TEM). MRD specifically: Natera (Signatera — the category leader by a wide margin), Guardant Reveal, Personalis NeXT Personal (being acquired by TEM), Foundation Medicine.

Notably, TEM's earnings releases name none of these competitors in any of eight quarters — competition appears exactly once per release, in boilerplate. Competitive positioning is not discussed in management's own written framing.

5.2 TAM — the required bottom-up build, and the required implied-penetration statement

The Personalis press release asserts MRD is a "$20 billion opportunity." Per references/tam-sizing.md this is precisely the unusable form — a top-down, vendor-published headline. Built from units instead:

The company's $20bn sits inside this band — but only at 100% penetration, which is the elastic definition the reference warns is gameable.

SAM today is far smaller: Personalis has Medicare coverage in three indications. At ~41,500 annualised tests (Q2 2026 × 4) against an ~10m-test pool, current penetration is ~0.4%.

Required implied-penetration statement: TEM's $1.5bn price for Personalis implies capturing roughly 3–4% of the US MRD testing pool — about 300,000–420,000 tests a year at ~$2,157 — to generate the ~$700–900m of revenue that would justify the price at a 2x exit multiple. Personalis runs at 0.4% today. The purchase therefore requires an 8–10x increase in penetration, against a category in which Natera's Signatera has a multi-year head start, an entrenched reimbursement position and materially greater scale.

Time to revenue: the deal does not close until late 2026/early 2027. Meaningful MRD revenue at the required scale is a 2029–2032 event on any reasonable adoption curve. Discounted at the 13.6% cost of capital derived in the valuation document, a 2031 revenue pool is worth roughly 45c on the dollar.

Base rate for adoption: new molecular diagnostic categories with favourable reimbursement have historically taken 6–10 years from first Medicare coverage to 20% penetration (Signatera's own CRC trajectory is the closest analogue). Assuming faster than that is permitted but must be named as the above-base-rate assumption it is. This memo does not assume faster.

TAM expansion vs. share gain: MRD is a genuinely new addressable pool for TEM (it currently resells rather than owns), so this is TAM expansion at the company level — but it is share competition at the industry level, against an incumbent with a large lead. These have very different margin implications and the company's framing conflates them.

6. Risks

1. Dilution as a standing feature. ~19% from Personalis, on top of SBC compounding at 129% YoY and prior stock-funded deals. Revenue per share is growing far more slowly than revenue.
2. Non-GAAP framing risk. The only guided profit metric excludes the largest real expense. If the market re-anchors on GAAP, there is no earnings support at any multiple.
3. Leverage and burn. Net debt ~$589m, operating cash outflow ~$73m/quarter, Ares term loan maturing September 2027, converts due 2030. A cash election on Personalis draws the revolver further.
4. Integration. Five acquisitions in eighteen months, one reportable segment, no disclosed synergy targets, and goodwill of $470m against $416m of book equity.
5. Competitive position in MRD is second at best, against Natera's Signatera.
6. Governance. 30:1 voting Class B; $171m of CEO selling; no insider buying.
7. Crowding. 26.45% of float short with 5.03 days to cover — squeeze risk in both directions and a violently reflexive tape.
8. Reimbursement. Coverage/reimbursement language has decayed to zero in management's written disclosures, which is the topic most likely to determine Diagnostics gross margin.

7. What the generative pass produced, and what came from priors

Recorded per the method, so the ledger can eventually score whether generatively-sourced hypotheses outperform.

Sources: SEC EDGAR (CIK 1717115) — FY2025 10-K, Q1 2026 10-Q, 8-K Ex-99.1 ×8, 8-K 2026-07-20 (Items 1.01/9.01) and Ex-2.1, Forms 4; XBRL company facts. ClinicalTrials.gov API v2. Alpaca Markets (daily bars, options). Alpha Vantage EARNINGS_ESTIMATES (one call, cached to data/av_estimates_TEM.json). stockanalysis.com (share count, short interest, ownership, peer multiples). Every figure is either sourced above or explicitly labelled an assumption.