Phase Space AI

Financial Model Notes

NVIDIA [NVDA]

NVIDIA [NVDA] — Financial Model Notes

Every figure used in this memo, its source, and every place a source turned out to be wrong.

---

1. Source hierarchy applied

2. Defects found, with magnitudes

(a) .cache/universe_scan/NVDA_analysis.json carries a 23.2x revenue error. revenue_ttm: 10,918,000,000 with revenue_as_of: "2020-01-26" — that is NVIDIA's fiscal 2020 revenue, 2,375 days stale, against a true TTM of $253,491m. Understatement: $242,573m, −95.7%; the true figure is 23.2x the stored one. At that revenue NVDA would screen at 417x sales rather than 18x. Mitigating: the record set revenue_stale_days: 2375, wrote "revenue STALE (2375d)" into notes, and set status: INDETERMINATE. It failed safe — this is the recency assertion from valuation.md working as designed (the GOOGL 485-day precedent). The momentum block in the same file does not depend on revenue and is used. The revenue figure is unusable and was not used.

(b) AV commonStockSharesOutstanding = 24,391m is the diluted weighted-average, not shares outstanding. Filed: 24,221m (10-Q shareholders'-equity statement, Apr 26 2026; cover page "24.2 billion" as of May 15 2026). Error +170m, +0.70% — small in percentage terms, +$32.3bn of market cap in absolute terms at $190.01. Exactly the documented defect class (WDC +9.0%, COHR −20.5%). Percentage-small, dollar-large: at this market cap a 0.7% share-count error is larger than most of the corpus's entire enterprise values.

(c) AV shortTermDebt = 1,466 against a filed 1,000. +$466m, +46.6% — apparently the current operating-lease liability folded into debt. Immaterial to EV (0.01%) but it is the same class of error as the documented longTermDebt = 0 cases, in the opposite direction, and it would matter on a smaller name.

(d) av_vs_edgar.py returned UNVERIFIED_NO_EDGAR_OVERLAP — zero periods compared — on a name where AV and EDGAR agree to the dollar on every line. The tool reported cmp 0, dis 0, VERIFIED 0, UNVERIFIED 1 and "NVDA: no overlapping period," and separately reported names needing fiscal-calendar tolerance to be checkable at all: 0/1. AV normalises NVIDIA's period ends to calendar month-ends (2026-04-30, 2026-01-31) against EDGAR's true ends (2026-04-26, 2026-01-25) — gaps of 4 and 6 days, inside the ±7-day tolerance the tool claims to apply. The tolerance did not fire. Per DATA_DEFECTS the correct reading of UNVERIFIED is "not tested, NOT clean," so the entire nine-line verification in Research §0 was done by hand against the primary documents. Result: exact agreement on all nine lines across both the 10-K and the 10-Q. The tool's silence was a false negative, and it is the second recorded failure mode of this script.

(e) reverse_dcf.py's CLI does not expose --fcf-margin, so the documented fix is unreachable from the documented command line. DATA_DEFECTS records the terminal-only bias as "Fixed; pass fcf_margin." project_ev() and solve() both accept the argument; argparse does not define it, and the CLI exits with unrecognized arguments: --fcf-margin. Any caller following the usage string in the module docstring silently gets the biased terminal-only answer. Measured on NVDA at a 46.97% FCF margin:

+6.40pp is the largest instance of this bias measured in the project (prior maximum +6.04pp at a 35% FCF margin). The bias scales with cash generation, so it penalises the highest-quality names hardest — which is the mechanism the module's own docstring identifies. Worked around by importing the module and calling project_ev directly.

(f) Every stored terminal margin in the scans is the universe-wide 14.4% constant and was not used. Confirmed by inspection of the defect record rather than re-derived. The 58.0% used here is built from NVIDIA's own filed lines (Valuation §3).

(g) AV EARNINGS_ESTIMATES returned {"estimates": []} at HTTP 200 for NVDA — the largest company in the world by market capitalisation and the most covered stock in the market. data/pit/consensus_pit.jsonl carries 0 NVDA records; data/pit/consensus_rejects.jsonl also 0. Recorded INDETERMINATE with the reason; it blocks nothing. The 12-month target base is built from company guidance and labelled as such, not from consensus, and the memo says so in the target section rather than passing guidance off as consensus.

(h) AV transcript quarter labels are FISCAL, as documented. EARNINGS_CALL_TRANSCRIPT quarter=2027Q1 returns the call for the quarter ended 2026-04-26 — NVIDIA's fiscal Q1 FY2027, which is calendar Q2 2026. Any cross-name series keyed on the label would misalign NVDA against a December-FY name by two quarters. The mention-frequency table is labelled fiscal explicitly.

(i) No defect found in the prepared-remarks / Q&A split, because it was checked. DATA_DEFECTS records splitters firing at index 0 on 20 of 40 quarters. Keying on the first speaker whose title contains "Analyst" put the boundary at index 4, 5 or 6 in all thirteen NVDA quarters — never 0 — with 4–6 prepared utterances and 22–38 Q&A utterances per call. The split is sound.

3. Where a documented AV defect did NOT apply — checked, not assumed

sellingGeneralAndAdministrative is correct on NVDA. The documented defect (the field returning G&A only: NOW $369m vs $1,741m, 4.72x; INTU $1,656m vs $7,901m, 4.77x) does not arise, because NVIDIA's income statement carries a single combined "Sales, general and administrative" line. AV's FY2026 value of 4,579 equals the filed 4,579, and Q1 FY2027's 1,300 equals the filed 1,300. This was verified rather than assumed, and it is why the terminal opex bridge closes to zero residual on this name — there is no "other" bucket in which a ~30pp margin error could hide.

Splits: queried explicitly and reported even though there is nothing recent. SPLITS returns 10.0000 on 2024-06-10, 4.0000 on 2021-07-20, and older actions in 2007/2006/2001/2000. No split since 2024-06-10, so the KLA failure mode (a pre-split share count paired with a post-split price — market cap 89% understated, required CAGR wrong by 37pp) cannot arise in the current market cap. All price history was pulled with adjustment=split, which matters for the 10-year multiple series that spans both the 4:1 and the 10:1.

No spin-off or separation in the window, so the WDC/SanDisk failure mode (a spin-off recorded as a 1.3230 split, throwing every pre-separation multiple off by −24.4%) does not apply.

cashAndShortTermInvestments — the new sign-flipping defect, avoided by construction. DATA_DEFECTS was amended during this run to record that AV's cashAndShortTermInvestments returns cash only, silently dropping shortTermInvestments (INSM: $641.3m / 52.4% omitted, flipping +$479.1m of net cash to −$166.2m of net debt). That composite was never used here. Net cash was rebuilt line-by-line from the 10-Q balance sheet — cash $13,237m plus marketable debt securities $37,098m, less short-term debt $1,000m and long-term debt $7,470m — with marketable equity securities and non-marketable securities excluded on judgement rather than by omission. AV's shortTermInvestments for the same date is 67,335, which is marketable debt (37,098) plus marketable equity (30,237) lumped together; taking it at face value would have overstated net cash by $30,237m and understated EV by 0.7%. Neither the AV composite nor the AV component was adopted.

A defect list is a prior, not a fact. Three of the ten records above (§3) are clean results on documented defect classes, reported deliberately. The corpus's amended guidance on sellingGeneralAndAdministrative — a "G&A only" entry generalised from a sample of two that turned out to be byte-identical to the filed line on INSM — is the reason: a prior asserting "this field is always 4.7x light" would have caused this memo to reject NVIDIA's correct $4,579m SG&A and rebuild a terminal margin ~30pp too high. Every defect class was tested against the primary filing on this name, and the negatives were recorded alongside the positives.

Scale cross-check. Net income $58,321m ÷ diluted 24,391m = $2.3911, against a filed diluted EPS of $2.39. It passes — and that check has no power here and is reported as such. CRWD passed the same check at $0.108 vs $0.11 while being 4.0x and $137bn wrong, because a vendor that is internally consistent defeats every internal-consistency test. The check that carried the weight was the nine-line comparison against EDGAR primary documents in Research §0.

4. Field-collapse check

The terminal_margin, company_state, exit_multiple, required_cagr and terminal_margin_basis fields in NVDA_analysis.json are all derived from NVIDIA-specific filed lines and are deliberately unlike the corpus defaults — 58.0% against the universe-wide 14.4%; 13.88x derived from the warranted-multiple identity rather than a peer median; terminal_margin_source naming the 10-K opex table rather than max(own, sector peer median). A value that describes many companies describes none, and none of these should collide with another name's.

5. Structural notes for anyone rebuilding the model

1. Fiscal calendar. FY ends the last Sunday of January. FY2026 = year ended 2026-01-25. Fiscal Q1 FY2027 = quarter ended 2026-04-26 = calendar Q2 2026. 52/53-week filer, so AV's calendar-month-end normalisation will not match EDGAR exactly on any period.
2. EDGAR does not tag fiscal Q4. It sits inside the annual duration. Q4 FY2026 must be synthesised as FY2026 − (Q1+Q2+Q3) = 215,938 − (44,062 + 46,743 + 57,006) = 68,127, which matches the AV quarterly record for 2026-01-31 exactly. Flows subtract; stocks do not — applying the same synthesis to a share count yields nonsense.
3. Two opex lines only. R&D and combined SG&A. Any three- or four-line opex model of NVIDIA is inventing a split the company does not publish.
4. Segment operating income ≠ consolidated operating income. FY2026 segment total $139,297m against consolidated $130,387m; the $8,910m difference is SBC $6,386m + unallocated opex $1,997m + acquisition costs $527m. Using segment margins as company margins overstates by ~4.1pp.
5. Net income is not an operating number. Q1 FY2027 "Other income (expense), net" was $15,929m = 19.5% of revenue = 22.8% of pre-tax income, from marks on a $73.6bn investment book. Model operating income; treat net income and EPS as contaminated.
6. The market-platform series breaks at Q1 FY2027. Pre-Q1-FY2027: Data Center / Compute / Networking / Gaming / ProViz / Automotive / OEM. From Q1 FY2027: Data Center (Hyperscale, ACIE) / Edge Computing. Only one prior period was recast. A multi-period series on either basis longer than that is not constructable from the filings.
7. Working capital is a real drag at this growth rate. Receivables +$17.6bn and inventory +$15.7bn year-on-year. TTM OCF $125,648m sits $33,965m below net income, mostly for this reason plus the non-cash investment gains. Do not model FCF as a fixed share of EBIT.