Tempus AI, Inc. [TEM] — Financial Model Notes (Task 2)

Workbook: TEM_Model.xlsx · Built: 2026-07-27 · Notes written: 2026-07-29 Four tabs, live formulas throughout · All $ in thousands unless noted

0. THE VERIFIED SHARE COUNT — read this before using any per-share figure

An automated scan on 2026-07-29 resolved TEM's share count to ~0.2m shares, implying an ~$8m market capitalisation for a company with a real equity value near $7.7bn — a ~1,000x error. This section records the correct figures and the exact mechanism of the failure so downstream tooling has something trustworthy to read.

0.1 The authoritative numbers

THE NUMBER TO USE

Diluted share count: 178,964,000 (Q1 2026 weighted-average diluted, the GAAP EPS denominator). Shares outstanding: 179,404,620 at 2026-03-31 (balance sheet), or 179,564,767 at 2026-04-30 (cover page).

The model uses 179,405 thousand = 179,404,620 — the balance-sheet outstanding figure. At the $42.91 valuation-date spot this gives a market capitalisation of $7.70bn, consistent with the $7.73bn used in the comparables table and the $7.73bn cited at Gate 5.

Any share count below 1,000,000 for TEM is wrong by roughly three orders of magnitude.

Cross-check that proves the figures are right rather than merely quoted. Filed Q1 2026 net loss is −$125,919k. Divided by basic weighted-average shares of 178,880,000 that is −$0.7039, and the filed basic EPS is −$0.70. The arithmetic closes. It does not close against any ~0.2m share count, which would imply an EPS near −$700.

Latest available. No 10-Q or 10-K has been filed since the Q1 2026 10-Q (2026-05-05); the most recent filings are the 2026-07-20 merger 8-Ks and a 2026-07-21 SCHEDULE 13D/A. Q2 2026 results are due 2026-07-30 and will supersede these figures.

0.2 What actually went wrong — three separate traps, all real

The brief suggested dei:EntityCommonStockSharesOutstanding was "unreliable." It is more specific and more dangerous than that. The tag is not wrong; it is invisible, and the fallbacks are booby-trapped.

Trap 1 — the dei tag is absent from the SEC's structured API entirely.

GET https://data.sec.gov/api/xbrl/companyconcept/CIK0001717115/dei/EntityCommonStockSharesOutstanding.json
  -> HTTP 404  NoSuchKey

The cached data/tem_facts.json (SEC company-facts) confirms it: TEM's entire dei block contains exactly one concept — EntityPublicFloat. The share count is missing because TEM tags it dimensionally, once per share class, each fact carrying a class-axis contextRef. Both facts are present in the raw inline XBRL of the 10-Q (verified by reading data/tem_10q_q1_26.htm: 174,520,978 and 5,043,789, each in its own <ix:nonFraction name="dei:EntityCommonStockSharesOutstanding">), but the SEC's companyfacts/companyconcept aggregation surfaces only undimensioned facts. A dual-class registrant can therefore have a perfectly correct cover-page tag that no companyfacts consumer will ever see. Any scan that silently falls through a None here — and the coordinator reports the scan was reading tags at the wrong nesting level and returning None — lands on a fallback.

Trap 2 — TEM's us-gaap share tags are scaled in THOUSANDS. The obvious fallback is a weighted-average share concept. WeightedAverageNumberOfDilutedSharesOutstanding for Q1 2026 has unitRef of shares and a value of 178,964. Read literally that is 178,964 shares. The true figure is 178,964 thousand = 178,964,000, as the EPS cross-check above proves. XBRL share facts should be reported unscaled; TEM's are not. 178,964 × $42.91 = $7.7m — an "$8m market cap."

Trap 3 — TreasuryStockCommonShares is genuinely ~0.2m. The other natural fallback is a point-in-time share balance. TEM's only such us-gaap concept is treasury stock at 183,229 shares — a real, correctly-scaled, point-in-time share count that is simply the wrong one. 183,229 × $42.91 = $7.9m — also an "$8m market cap." There is no CommonStockSharesOutstanding or CommonStockSharesIssued concept in TEM's company-facts at all (verified: zero us-gaap concepts beginning CommonStock* carry a shares unit).

Both fallbacks produce ~$8m, which is why the error looks like one bug and is actually a chain of three. The correct resolution order for TEM, and for any dual-class registrant:

1. Read the filed balance sheet or the 10-Q/10-K cover page directly and sum the classes. This is the only source that is both complete and correctly scaled.
2. If using XBRL, take WeightedAverageNumberOfDilutedSharesOutstanding and sanity-check the scale against net loss ÷ shares ≈ filed EPS. That single check catches Trap 2 unconditionally.
3. Never fall back to TreasuryStockCommonShares.
4. Sanity floor: reject any resolved market capitalisation that disagrees with a price-times-shares estimate by more than ~10x. An $8m market cap for a Nasdaq name with $1.27bn of revenue and $11.0bn of reported public float (TEM's own dei:EntityPublicFloat, 2025-06-30) should be impossible to accept silently. EntityPublicFloat is present in TEM's company-facts and would have caught this on its own.

1. Structure

Formatting convention (standing user preference): the top line of each statement block carries the $; rows beneath are plain. Blue = hardcoded input, black = formula.

2. Verification — how it was actually checked, and what it caught

Excel automation was unavailable in this environment. The workbook was verified by loading it into the formulas Python package — an independent engine that re-parses and re-evaluates every formula from scratch — and reading back computed cells, rather than trusting the formula strings openpyxl wrote. This is arguably stronger than an Excel readback because it does not share Excel's evaluation engine. The check is reproducible: data/verify.py.

Re-run on 2026-07-29 for these notes. The workbook compiles clean and every published figure reproduces exactly.

2.1 Tie-out to the filed statements (v1.4.2 requirement — internal consistency is not enough)

2.2 The defect this verification caught

The first build carried a single line labelled "Adjusted EBITDA" computed as operating income + SBC + D&A. For FY2025 that returns −$25,801k. The filed figure is −$7,385k. The model was internally consistent and wrong by $18.4m, because TEM's definition also adds back employer payroll tax on SBC, acquisition expenses, loss on debt extinguishment, franchise and other taxes, fair-value changes and equity-method losses, while subtracting amortisation of the SB Tempus technology licence.

Fix: the original line was relabelled Adj. EBITDA (simplified: OpInc + SBC + D&A) (row 21), a company-add-back row was inserted (row 22, FY2025A = 18,416 hardcoded to the filed reconciliation), the true company-definition line became row 23, and row 25 carries =D23−D24 as a hard tie-out check that must read 0. It reads 0.

This is the general lesson the framework already records: a zero balance check verifies internal consistency, not input accuracy. Every actuals line here is tied back to the filed statement individually.

2.3 A second imprecision, found while writing these notes and NOT previously flagged

The row-22 company-add-back is held at a flat 1.2% of revenue for every forecast year — a proxy calibrated so FY2025A ties. Carried forward, it makes FY2026E company-definition Adjusted EBITDA read $83.7m against management's guide of ~$65m — a $19m (29%) overshoot.

Scope of the error: presentational only. It does not touch any valuation output. The DCF tab computes EBIT from its own revenue, gross-margin and opex driver rows and never references IS!E23; the Scenarios tab runs off a hardcoded FY2027E revenue and an exit multiple. Nothing downstream of IS row 23 exists. No published number changes. It is recorded because an unflagged 29% miss against guidance on a headline metric is exactly the class of defect that becomes load-bearing the moment someone reuses the tab. Treat IS row 23 for FY2026E–FY2030E as indicative, not as a forecast of the company's reported metric.

The GAAP line is unaffected and is the one the memo actually leans on: FY2026E operating income reads −$249.6m, reproducing the −$250m the research and valuation documents derive independently from guidance (~$65m Adjusted EBITDA less ~$210m annualised SBC less ~$105m D&A).

2.4 A labelling defect in the driver block

On the IS tab, rows 25–29 hold the forecast driver values in columns E:I, but column A on those same rows carries text belonging to the block above (A25 reads "Check (must be 0)"; A27/A28 hold a two-line explanatory note about the Adjusted EBITDA definition). The values are correct and every formula references the right cells — E25:I25 are revenue growth, E26:I26 gross margin, E27:I27 opex % of revenue, E28:I28 SBC % of revenue, E29:I29 D&A % of revenue. Only the row labels are misaligned. Flagged rather than silently repaired, because repairing it would change the workbook after the valuation was published off it.

3. Key assumptions and drivers

3.1 Market and capital-structure inputs (Assumptions)

The WACC is high and it carries the answer. It is the honest rate for a business with 65.7% trailing realised volatility, a 2.50 regressed beta and negative operating cash flow — but a sensitivity across 11.5–15.5% is published in TEM_Valuation.md §2.2 precisely because this single input dominates. Note the circularity that is standard but should be named: the equity weight is computed from spot × shares, so a falling share price mechanically raises the debt weight and lowers the WACC.

3.2 Operating drivers (IS, columns E:I)

Anchoring. FY2026 revenue growth of 26.6% is essentially management's own guide ($1.59–1.60bn); the model reads $1,610m, +1.1% against the 18-analyst consensus of $1,592.9m. FY2027 of $1,948m is −0.7% against consensus of $1,960.9m. The model is deliberately not a variant — that is the finding, and it is why Gate 2 fails.

Required base-rate check (Chan, Karceski & Lakonishok 2003). Growth persistence beyond chance is close to nonexistent, and sustained 20%+ growth for 5+ years at multi-billion revenue scale is rare. This base case assumes ≥19% growth for three consecutive years past $1.6bn of revenue — a top-quartile, arguably top-decile outcome for the reference class. It is adopted anyway because FY2026 is management's guide and FY2027 is below Street. The prior is being overridden, and this sentence is where that is named.

GAAP breakeven is not reached until FY2029. That is not editorialising — it is what the company's own guidance implies, and the model reproduces −$249.6m for FY2026 against FY2025's −$252.9m on 25% more revenue.

SBC is modelled as a real expense in the GAAP lines and is the reason breakeven is so distant. It is held at 13.0% of revenue in FY2026 (≈$209m, consistent with annualising Q1 2026's $52.7m) fading to 9.5% by FY2030. Q1 2026 SBC grew +129.4% YoY against +36.1% revenue growth; the fade to 9.5% is an assumption of normalisation that the last four quarters of data do not yet support.

3.3 DCF drivers (DCF)

Ten explicit years, FY2026E–FY2035E, mid-year discounting (periods 0.5 … 9.5).

Output (read back from the evaluated workbook):

Read the DCF as a statement of the assumptions, not as a measurement. 71.9% of value sits in the terminal, and FY2026E unlevered FCF is −$200m — the explicit period contributes mostly negative cash flows offset by a large Gordon tail. The reverse DCF in TEM_Valuation.md §2.3 is the more informative construction: $42.91 requires an ~11% WACC, a 20–24% terminal EBIT margin against −15.5% today, and a ~16% nine-year revenue CAGR. Demanding, but not absurd — and emphatically not "priced for perfection."

3.4 Scenario drivers (Scenarios)

All three states assume the Personalis deal closes, and all three run off the same revenue base so the exit multiple is isolated as the only variable.

Gate 4: scenario E[R] +2.39% gross, +2.1% net of ~0.3% costs, against a 4.7% cash hurdle — a −2.3pp shortfall that the workbook computes directly (Scenarios!J11 = −0.0231).

The exit multiple is the entire model at this point, and it is the weakest link. The bull 7.00x is set below the NTRA/GH trading range (14.4x/18.4x) and close to the only completed control transaction in the set — Abbott paid 6.65x revenue for Exact Sciences in March 2026. The bear 3.00x is where decelerating, cash-burning diagnostics assets have historically cleared. CALIBRATION_WATCH.md item B15 records that the exit-multiple anchor set for high-growth life-science names is unidentified, and that criticism applies directly here and is not resolved by this model. The 4.75x base multiple is one defensible choice among several, not a measurement.

Note the sensitivity this creates. Because all three scenarios share one revenue base and one share count, the model is a pure multiple-sensitivity grid. Every dollar of the $27.38-to-$66.33 spread comes from the 3.00x–7.00x range. The probability-weighted target of $43.93 sits 2.4% ABOVE spot — the first above-spot house target in this coverage, recorded in TEM_Valuation.md §3.3 as out-of-sample confirmation of CALIBRATION_WATCH.md item B9's exculpatory reading.

4. What the model deliberately does NOT do

Stated so the omissions are auditable rather than mistaken for oversights.

1. No balance sheet or cash-flow statement. This run was scoped to Tasks 1, 2, 3 and 5; the DCF is driven from income-statement drivers plus D&A/capex/NWC ratios rather than a linked three-statement build. There is therefore no balance check to run — the usual integrity test does not exist here, and that is a genuine loss of resolution, not a passed check.
2. No segment forecast. TEM reports one reportable segment. Revenue and cost of revenue are disclosed separately for Diagnostics and Data & Applications, so gross profit is separable — and the FY2024A/FY2025A actuals carry that split. Below gross profit no opex allocation is filed, so forecasting by segment would require inventing one. The forecast is consolidated. The divergence this conceals is material and is analysed in TEM_Research.md §1.3: Q1 2026 Diagnostics gross margin +500bp YoY against Data & Applications −350bp. The higher-margin segment carrying the premium-multiple argument is growing faster at a falling gross margin.
3. No Personalis pro-forma consolidation. The deal is unclosed. Its effect enters only in the Scenarios tab, as three hardcoded post-deal inputs (revenue, net debt, share count). The IS and DCF tabs are standalone TEM. Mixing a standalone DCF with post-deal scenarios is a real inconsistency in the workbook and is flagged here rather than smoothed over — it is why the DCF's $12.80 and the scenario grid's $43.93 are not directly comparable.
4. No dilution modelling beyond the Personalis share issue. SBC is expensed in the GAAP lines but the share count is held flat at 179,405k in the IS and DCF tabs. With SBC running at ~13% of revenue this understates future dilution, and it does so in the direction that flatters the DCF per-share value.
5. No explicit MRD line. MRD is not separately disclosed by TEM — volume was first given in Q4 2025, revenue and ASP never. It is inside consolidated Diagnostics. The bottom-up TAM in TEM_Research.md §5.2 is a sizing exercise, not a model input.

5. Data provenance

Model: TEM_Model.xlsx. Every actuals figure ties to a filed statement; every forecast driver is listed above; every assumption is labelled as one. The three defects found in this workbook — the $18.4m Adjusted EBITDA definition error (fixed), the $19m FY2026E add-back overshoot (flagged, presentational only), and the misaligned driver-row labels (flagged, values correct) — are recorded in §2 rather than repaired silently.