Tempus AI, Inc. [TEM]

Tempus is a 25%-growing precision-medicine platform with a genuine, filed operating-leverage mechanism — Diagnostics gross margin 45.3% to 58.9% to 61.3% across five quarters — trading 58.4% below its October-2025 high while every diagnostics peer rallied 85% to 228%. None of that is an edge. The binding failure is Gate 2 and it is structural rather than marginal: consensus FY2026 revenue of $1,592.9m across 18 analysts spanning a 1.2% range IS company guidance of $1.59-1.60bn transcribed, the house forecast is +1.1% against it, and EPS estimates have been revised UP monotonically for 90 days. There is no variant available to anyone until guidance itself moves. The duration variant on MRD fails 2 of 4 legs — most damningly on the transcript leg, where MRD mentions per 10k words run 7.1, 0, 0, 0, 3.9, 0, 3.5, 4.5 across the company's entire public history, so the PEAK IS THE FIRST QUARTER and there is no emergence at all. Gate 3 fails because the 30 July print tests execution against a guide consensus has already copied, while the house-Street gap is entirely an exit multiple (4.59x vs 7.02x on FY2027E revenue). Gate 6 fails hard and idiosyncratically. Valuation carries almost no information: the methods span $8 to $145, a 17x range, and the DCF is 71.9% terminal value. The strongest genuinely differentiated finding is bearish and generatively sourced — SBC compounding at +129% against +36% revenue growth, and equity issued at ~5.2x forward revenue to buy Personalis at ~16.7x on a 13.7% gross margin, roughly 19% dilution for 5.6% of revenue — but it does not translate into a below-consensus forecast, so it is not a short either. Expected return is +2.1% net against a 4.7% cash hurdle. Cash is the better position.

Key findings

• DATA DEFECT RESOLVED — TEM's verified share count is 179,404,620 outstanding (Class A 174,360,831 + Class B 5,043,789, filed Q1 2026 10-Q balance sheet at 2026-03-31) and 178,964,000 weighted-average DILUTED. An automated scan resolved it to ~0.2m shares, implying an $8m market cap against ~$7.7bn of real equity value. The cause is a CHAIN OF THREE traps, not one bad tag: (1) dei:EntityCommonStockSharesOutstanding 404s from the SEC companyconcept API because TEM tags it dimensionally per share class and the aggregation surfaces only undimensioned facts — TEM's entire dei block contains exactly one concept, EntityPublicFloat; (2) TEM's us-gaap weighted-average share tags are SCALED IN THOUSANDS despite a 'shares' unitRef, so 178,964 means 178,964,000 and reads as $7.7m of market cap; (3) the only point-in-time us-gaap share balance is TreasuryStockCommonShares = 183,229, a correctly-scaled but wrong number that also reads as $7.9m. Both plausible fallbacks land on ~$8m. Proof the right numbers are right: filed net loss -$125,919k / 178,880,000 basic = -$0.7039 against filed basic EPS of -$0.70.
• NEW, FROM THE MERGER AGREEMENT, AND NOT IN THE ORIGINAL RESEARCH FILE — Exhibit 2.1 defines a LOWER FLOOR PRICE of $46.00 and gives PERSONALIS a termination right under section 8.1(h) if TEM's Parent Stock Price (a 15-consecutive-trading-day VWAP ending before the last trading day prior to Closing) is finally determined below it. TEM closed at $42.91, 6.7% BELOW that level. This completes rather than contradicts the research file's reading that below the $48.42 Floor Price 'Personalis holders bear the downside': they bear it only to $46.00, below which they can walk. It also supplies the missing explanation for the ~17% gross merger spread the research file flagged as unresolved — the spread is not diffuse deal risk, it is the market pricing a specific, filed, price-contingent termination right that is currently in the money for the seller. Note the asymmetry: the $76,806,179 termination fee runs FROM Personalis TO Tempus and is not triggered by 8.1(h).
• GENERATIVE FINDING — 'stock comp' is the single most-mentioned term in ALL EIGHT quarters of TEM's public history (69-93 per 10k words), and it was not looked for. The reason is that it is quantitatively enormous and management structures the entire non-GAAP presentation around excluding it: Q1 2026 SBC was $52.7m against $23.0m a year earlier, +129.4% YoY against revenue growth of +36.1%. FY2026 guided Adjusted EBITDA of ~$65m sits against ~$210m of annualised SBC. Both 'Adjusted EBITDA inflects positive' and 'GAAP operating loss is unchanged at ~-$250m on 25% more revenue' are true simultaneously, and which one is treated as the earnings determines the entire valuation.
• GENERATIVE FINDING, AND A CORRECTED FALSE POSITIVE — 'Genomics' and 'Data and services' fell to EXACTLY ZERO in 2025Q4 after six stable quarters, which looked like the most dramatic signal in the dataset. It is a RELABELING ARTIFACT: TEM renamed the lines to 'Diagnostics' and 'Data and Applications', verified against the Q1 2026 10-Q which uses the new labels for both current and restated prior-year columns. The corrected combined series is flat. This is reported because it is exactly the class of false positive the mention-frequency method is prone to, and it was caught only by reading the surrounding text rather than trusting the count.
• NEGATIVE RESULT, RECORDED DELIBERATELY — MRD is NOT an emerging term. Per 10k words across the company's entire public history: 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 near-zero-to-material emergence. This looked prospectively like the obvious high-conviction candidate and the data does not support it. What did change is a DISCLOSURE — a quantified MRD volume (~4,700 tests, +56% QoQ) first appears in Q4 2025 — which is a weaker and different signal. Recording the negative result is the point.
• THE SEGMENT DIVERGENCE THAT ONE-SEGMENT REPORTING CONCEALS — Q1 2026 Diagnostics gross margin 61.3% (+500bp YoY) against Data & Applications 71.1% (-350bp YoY). The consolidated +310bp improvement is a large Diagnostics gain netted against a real deterioration in the higher-margin, more software-like segment that carries the AI narrative and the entire premium-multiple argument. It is growing faster at a FALLING gross margin. Blending them, as a single-segment presentation invites, conceals this entirely.
• THE ACQUISITION ARITHMETIC — TEM's own enterprise value is ~5.2x forward revenue; it is paying $1.5bn for ~$90m of annualised Personalis revenue, or ~16.7x, and that revenue carries a 13.7% gross margin against TEM's 63.8%. At the capped 0.3356 ratio the all-stock case issues ~35m shares, approximately +19% on 179.4m outstanding, for ~5.6% of FY2026E revenue. The honest counter is stated rather than strawmanned: Morgan Stanley notes ~80% of Personalis clinical volume already runs through Tempus channels, the low gross margin is a scale artifact consolidation addresses, TEM removes a third-party economics split on a product it already resells, and Personalis brings ~$194m of net cash. None of that is refuted by the multiple arithmetic — and none of it has been quantified by the company, which has disclosed no synergy target.
• THE INSIDER BUYS ARE NOT WHAT THEY APPEAR, AND THE CHECK MATTERED — all 12 open-market 'P' purchases in the window are dated 2026-02-13 at $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 five months before announcing the acquisition, not an insider buying TEM. THERE IS NO INSIDER BUYING OF TEM STOCK. Under Cohen/Malloy/Pomorski the informative signal — clustered non-routine insider purchases — is ABSENT, while $189.9m of sales (Lefkofsky $171.3m, 90% of the total) is scored as weak negative evidence only, because sales are mostly noise.
• VALUATION CARRIES ALMOST NO INFORMATION ON THIS NAME, and that is the finding rather than a hedge. The methods span $8 to $145 — a 17x range. DCF at a defensible 13.63% WACC returns $12.80 with 71.9% of value in the terminal; the market price requires an ~11% WACC, a 20-24% terminal EBIT margin against -15.5% today, and a ~16% nine-year revenue CAGR. The superficial comp read (TEM 6.16x vs NTRA 14.4x and GH 18.4x = a 60% discount) DOES NOT SURVIVE the EXAS datapoint: Abbott, a strategic acquirer with full diligence, paid 6.65x revenue for a $3.25bn-revenue 69.7%-gross-margin franchise in March 2026. Against the only completed control transaction in the set, TEM is roughly fairly valued.
• THE STREET'S BASE CASE IS THE HOUSE'S BULL CASE — decomposed on a shared revenue base, the entire $42.91-to-$66.57 gap is the exit multiple, 4.59x versus 7.02x on FY2027E revenue, with house and Street within 1.1% on the fundamentals. The house bull multiple (7.00x) and the Street average (7.02x) are the same number. Street targets have drifted DOWN (~$72 to ~$66-67) while ratings have not changed — the classic pattern of a sell side marking to market rather than changing its mind.
• THE MRD TRIAL THAT IS CITED AS THE STRONGEST INDEPENDENT EVIDENCE WILL NOT READ OUT IN ANY UNDERWRITABLE HORIZON — NCT07211178 (MRD longitudinal ctDNA, n=900, RECRUITING, started 2025-10-27) has a registry PRIMARY COMPLETION of 2032-12 and full completion 2033-12. It corroborates that MRD investment is underway; it is not a catalyst. This reinforces rather than undercuts the research file's 2029-2032 time-to-revenue estimate. The nearest-dated MRD-relevant readout TEM owns is NCT05234177 (CRC surveillance) at ~2027-02.
• CALIBRATION — TEM is the FIRST name in coverage that genuinely tests item B1/B11. That item records that the 'does Gate 6 wrongly block beaten-down names' question cannot be tested opportunistically and requires names actually trading BELOW their 200-day; the GH/TXG/TWST experiment failed because all three turned out to be top-decile momentum names. TEM trades 29.1% BELOW its 200-day at a 52-week low, -58.4% from its high. Gate 6 is a genuine blocker here rather than a tailwind. HOWEVER, TEM is a WEAK test of the charitable case, because Gate 2 binds independently and deleting Gate 6 entirely would not change the verdict. Recorded as the first data point of the right type, with that caveat stated.
• CALIBRATION — item B9's exculpatory reading gains an OUT-OF-SAMPLE CONFIRMATION. B9 records that 16 of 16 house targets sat below spot at a median -46.1%, and offers the exculpatory split that 'handed a name that has not run, the house prices it at fair value, not at a discount.' TEM's probability-weighted target of $43.93 sits +2.4% ABOVE spot — the first above-spot house target in coverage, on the most beaten-down name yet underwritten. This is counter-evidence against the framework being systematically too strict, which is the side of the calibration file that most needs feeding.
• CALIBRATION, RAISED AS A STRUCTURAL QUESTION AND NOT ACTED ON — on a name where 18 analysts transcribe company guidance into a 1.2%-wide consensus band, GATE 2A IS UNSATISFIABLE BY CONSTRUCTION for the guided year. Nobody can hold a material variant against a number that is definitionally the guide. That is arguably correct behaviour (there genuinely is no variant), but it means Gate 2 will systematically bind on tightly-guided names regardless of merit, and it interacts with existing item B12 on the unwritten materiality threshold. Flagged for the user to rule on; no gate was loosened.

Sections

Disclosed limitations

• SCOPE — Task 4 (chart pack) was NOT run for this name. There is no charts/ directory and the published site has no Chart Pack page, unlike most other names in this coverage. Tasks 1, 2, 3 and 5 were authored 2026-07-27 under an explicit instruction limiting scope; the catalyst calendar, model notes, manifest and publication were completed 2026-07-29.
• THE MENTION-FREQUENCY CORPUS IS A WRITTEN RELEASE, NOT A TRANSCRIPT. Alpha Vantage EARNINGS_CALL_TRANSCRIPT was not used — this run was allocated a hard budget of ONE Alpha Vantage call against the shared 25/day cap, spent on EARNINGS_ESTIMATES. The corpus is 8 consecutive quarters of SEC EDGAR 8-K Exhibit 99.1 earnings releases (2024Q2-2026Q1), one source used consistently, all counts normalised per 10,000 words. It contains NO Q&A, so the prepared-versus-prompted split the method requires CANNOT BE COMPUTED AT ALL, and no inference about analyst interest can be drawn. Every count should be read as the equivalent of prepared remarks.
• THE WINDOW IS ONLY EIGHT QUARTERS LONG because TEM listed in June 2024. Claims of the form 'first ever' are claims about the company's entire public history — unusually strong, but that history is only two years, and several other names in this coverage have 19-22 quarter windows.
• MRD REVENUE, ASP AND UNIT ECONOMICS ARE NOT DISCLOSED and were only partially closed. Volume appeared for the first time in Q4 2025 (~4,700 tests, +56% QoQ). The ~$2,157 revenue-per-test used throughout is DERIVED from Personalis' Q2 2026 preliminary figures ($22.4m / 10,384 clinical tests) in the merger press release — it is not a TEM disclosure and rests on a single quarter.
• THE TAM BUILD RESTS ON TWO UNSOURCED ASSUMPTIONS — the share of new US cancer diagnoses with resectable/curative-intent disease (~50%) and the number of surveillance tests per patient (~8-12). Both are flagged HIGH uncertainty in the research file. The resulting $18-27bn band contains the company's '$20 billion opportunity' claim only at ~100% penetration, which is the elastic definition the reference warns is gameable. Current Personalis penetration is ~0.4%.
• INSIGHTS CUSTOMER CONCENTRATION IS NOT CLOSED. TEM operates as ONE reportable segment and discloses no customer concentration. Named biopharma relationships found across the releases are Merck, Gilead, AstraZeneca and Pathos; count and concentration remain undeterminable from public sources.
• SEGMENT-LEVEL OPERATING INCOME IS NOT OBTAINABLE. One reportable segment; the CODM reviews consolidated results only. Gross profit IS separable by line and is analysed as such, but no opex allocation exists and none was invented — the forecast is consolidated, a genuine loss of resolution that is a direct consequence of the company's disclosure choices.
• ORGANIC VERSUS ACQUIRED GROWTH IS DISCLOSED EXACTLY ONCE — FY2025 Q4, '83.0% year-over-year with 33.5% organic growth (excluding Ambry)'. No organic figure is given in any other quarter, which makes the organic trajectory non-reconstructable.
• THE MODEL HAS NO BALANCE SHEET OR CASH FLOW STATEMENT. The DCF is driven from income-statement drivers plus D&A/capex/NWC ratios rather than a linked three-statement build, so THERE IS NO BALANCE CHECK TO RUN — the usual integrity test does not exist here.
• THE MODEL MIXES A STANDALONE DCF WITH POST-DEAL SCENARIOS. The IS and DCF tabs are standalone TEM; Personalis enters ONLY in the Scenarios tab as three hardcoded post-deal inputs. This is a real inconsistency in the workbook and is why the DCF's $12.80 and the scenario grid's $43.93 are not directly comparable.
• THE MODEL HOLDS THE SHARE COUNT FLAT at 179,405k in the IS and DCF tabs while expensing SBC at ~13% of revenue. This UNDERSTATES future dilution, and it does so in the direction that flatters the DCF per-share value.
• A SECOND MODEL IMPRECISION, FOUND ON 2026-07-29 AND NOT PREVIOUSLY FLAGGED — the company-add-back row is held at a flat 1.2% of revenue for every forecast year, making FY2026E company-definition Adjusted EBITDA read $83.7m against management's ~$65m guide, a $19m (29%) overshoot. Scope is PRESENTATIONAL ONLY: the DCF computes EBIT from its own driver rows and never references that cell, and the Scenarios tab runs off hardcoded revenue, so no published number changes. Recorded rather than repaired silently.
• A LABELLING DEFECT IN THE MODEL — on the IS tab, rows 25-29 hold the forecast driver values in columns E:I while column A carries text belonging to the block above. The VALUES ARE CORRECT and every formula references the right cells; only the row labels are misaligned. Flagged rather than repaired, because repairing it would change the workbook after the valuation was published off it.
• THE DCF IS 71.9% TERMINAL VALUE and FY2026E unlevered FCF is -$200m. The point estimate is a restatement of the 3.5% perpetuity and 16% terminal-margin assumptions rather than independent evidence; the reverse DCF is the more informative construction. Separately, the equity weight in the WACC is computed from spot x shares, so a falling share price mechanically LOWERS the WACC — standard, but named.
• THE EXIT-MULTIPLE ANCHOR IS UNIDENTIFIED AND THIS IS THE LARGEST UNRESOLVED WEAKNESS. CALIBRATION_WATCH item B15 records that the anchor set for high-growth life-science names is unidentified — the clean anchors all grow 1-7.5% while the disputed names grow 15-39%. If the correct anchor for a 63.8%-gross-margin, 25%-growing multimodal data platform is the NTRA/GH trading range (14-18x) rather than the EXAS transaction comp (6.65x), then a 4.59x anchor is systematically too low and this analysis will miss for as long as that regime holds. That criticism applies directly here and is NOT resolved by this memo.
• PIOTROSKI F-SCORE AND ACCRUALS ARE NOISY FOR A LOSSMAKER. Accruals of -1.18% are near-meaningless with both terms of the ratio negative, and estimate revisions on a name where consensus transcribes guidance measure how management guides rather than how the business performs. The two quantitative signals supporting a long are precisely the two the framework flags as least reliable here.
• INSTITUTIONAL OWNERSHIP IS A PARTIAL PULL. Per references/edgar-pipeline.md there is no single EDGAR endpoint for all 13F holders of a ticker. Named holders (Baillie Gifford ~8.74m shares, ARK ~7.56m) and the 52.53% aggregate come from public aggregators, not a paid feed. Short interest, peer multiples and the analyst count likewise come from stockanalysis.com as of 2026-07-28, and the analyst count and average target VARY BY SOURCE (16 analysts/$66.57 stockanalysis; 18/$66.06 S&P Global via MarketScreener; a third reporting $67.20) — that variance is disclosed rather than hidden behind one false-precise number.
• 10B5-1 STATUS OF THE INSIDER SELLING WAS NOT CONCLUSIVELY ESTABLISHED from the filings reviewed. $171m from one executive over twelve months against a $7.7bn market cap sits at the boundary of the Cohen/Malloy/Pomorski carve-out for sales 'extreme in size or unusually timed', and it is scored as WEAK NEGATIVE EVIDENCE, not as a mechanism.
• BORROW COST AND AVAILABILITY FOR THE SHORT SIDE ARE NOT VERIFIABLE from available sources. Stated rather than assumed away — it is one of the reasons Gate 5 fails for a short rather than the only one.
• THE CATALYST CALENDAR CONTAINS NO FDA OR CMS DECISION DATES, no company-confirmed conference dates, no Personalis shareholder-vote date and no Form S-4 filing date, because none exists in any source read. TEM's MRD product is currently RESOLD Personalis NeXT Personal and TEM has no pending PMA, 510(k) or NCD identified. Rows that cannot be dated say so; earnings dates beyond 2026-07-30 are marked [NOT ANNOUNCED] and derived from the company's own historical 8-K cadence rather than guessed.
• THIS MEMO IS DATED ONE DAY BEFORE THE Q2 2026 PRINT (2026-07-30, 4:30pm ET) and is NOT VALID PAST IT without a refresh including a PEAD check. Prices, the E[R] calculation and every flip point are as of the 2026-07-27 close.
• THIS WAS A USER-DIRECTED TICKER, NOT A SCREEN ADVANCE. TEM would almost certainly not have survived idea-screener, which filters on momentum and trend after liquidity — it is bottom-quintile on 12-1 momentum, 29% below its 200-day and at a 52-week low. The funnel that produced this name is 'someone named it', which carries no prior of positive expected return, and that is base-rate information rather than a complaint.