NVDA › analyze
NVDA · Analyze
Price at analysis: $227.98 (post-earnings, +8.7% on 2026-08-27 print day, -1.0% premarket 2026-08-28) · MktCap $5.51T · EV $5.03T
0. Knowledge check
python .mcp/kb.py find NVDA returned no live notes — this is the first NVDA file in the repo. No prior verdict to reconcile against.
Two live Playbook entries govern this analysis and both are directly tested below:
- [[pattern-ai-build-inflates-earnings-while-destroying-fcf]] — the standing 2026 pattern that AI-capex names show non-operating marks inflating GAAP earnings while capex/COGS destroys FCF. NVDA fires the earnings-inflation half this quarter (see §1) — this is the first name recorded where the tripwire fires on a half-year aggregate rather than a single quarter, and it should be added to that note's ticker list at the next kb-sweep (flagging for the orchestrator; not done here per scope lock).
- [[pitfall-vendor-forward-eps-is-the-wrong-fiscal-year]] — fires exactly as documented: Yahoo's headline forwardPE 15.03x prices forwardEps $15.17, which is FY2028 (two fiscal years out), not the current FY2027. Never quote 15x as "NVDA's forward multiple" — see §3.
- Also relevant, not yet fired in a verdict-changing way here: [[pitfall-yahoo-snapshot-fcf-field-diverges-from-quarter-sum]] — the snapshot freeCashflow field ($46.3B trailing) undercounts against a quarter-sum reconciliation (§1); and the general "just reported earnings, don't trust any vendor field" rule (amendment 5 of the forward-EPS pitfall) — confirmed here too: the vendor's quarterly statement feed had not yet ingested the Q2 FY27 print (reported 2026-08-27) at the time of this pull, so trailingEps (7.09) undercounts the true reconciled TTM (7.91). All figures below are rebuilt from NVIDIA's own press release and CFO Commentary (SEC 8-K, 2026-08-26/27), not the vendor aggregate.
1. Fundamentals
The reconciled TTM (through Q2 FY2027, ended 2026-07-26) — do not use the vendor snapshot
| Quarter | Revenue | GAAP Net Income | Diluted EPS | Data Center rev |
|---|---|---|---|---|
| Q3 FY26 (Oct-25) | $57.01B | $31.91B | $1.30 | — |
| Q4 FY26 (Jan-26) | $68.13B | $42.96B | $1.76 | — |
| Q1 FY27 (Apr-26) | $81.62B | $58.32B | $2.39 | $75.25B |
| Q2 FY27 (Jul-26) | $96.22B | $59.69B | $2.46 | $89.02B |
| TTM (reconciled) | $302.97B | $192.88B | $7.91 | — |
| Vendor snapshot (stale) | $253.49B | $159.61B | $7.09 | — |
The vendor feed is running roughly one quarter behind actual — it has not yet folded in the 2026-08-27 print. Use $7.91, not $7.09, for any TTM multiple. TTM P/E on the corrected figure is 28.8x, not the snapshot's 32.2x.
5-8yr trend (annual, fiscal year ends Jan 31)
| FY | Revenue | Net Income | Diluted EPS | FCF | Shares(dil) |
|---|---|---|---|---|---|
| 2023 | $26.97B | $4.37B | $0.17 | $3.81B | 25.07B |
| 2024 | $60.92B | $29.76B | $1.19 | $27.02B | 24.94B |
| 2025 | $130.50B | $72.88B | $2.94 | $60.85B | 24.80B |
| 2026 | $215.94B | $120.07B | $4.90 | $96.68B | 24.51B |
| 3yr CAGR | 100.0% | 201.8% | — | 193.9% | -0.7% |
Only 4 clean annual years are available (roic.ai's free tier caps history at two years — [[pitfall-roicai-free-plan-caps-history-at-two-years]] — and fin.py's FY2022 column is blank), so the 5-8yr window the framework asks for isn't fully populated. What's available is unambiguous directionally: triple-digit CAGRs on revenue, net income and FCF, funded entirely by the AI infrastructure buildout, with a shrinking share count throughout (buybacks have outpaced dilution every year, -0.7% 3yr CAGR on diluted shares despite heavy SBC).
Capital allocation (FY2026, $ actual)
| Use | Amount | % of FCF ($96.7B) |
|---|---|---|
| Buybacks | $40.1B | 41% |
| Acquisitions | $14.5B | 15% |
| Dividends | $1.0B | 1% |
| Debt repaid | $0 | 0% |
| Cash retained/reinvested | ~$41B | ~43% |
Q2 FY27 alone: $26.0B returned to shareholders (record quarter, buybacks + dividends), against a $99.0B remaining buyback authorization. Capital return is real, growing, and not funded by leverage — until this quarter: NVIDIA issued $25.0B of senior unsecured notes in Q2 FY27 "for general corporate purposes," the first large debt raise from a company that didn't need one. Total debt was $12.8B pre-raise against $56.6B cash+investments — leverage remains low in absolute terms, but the trend just inflected upward for the first time, worth watching alongside the guarantee book (§1.3).
Debt/Assets & balance sheet
Debt/Assets 5.3% (FY2026 year-end, pre the Q2 FY27 note issuance). D/E 6.6%. Current ratio 3.4x. This remains one of the least-levered balance sheets in the mega-cap set — the risk in this name is not leverage, it's contingent exposure (below).
🚩 The earnings-quality flag: [[pattern-ai-build-inflates-earnings-while-destroying-fcf]] fires
| Q2 FY27 | H1 FY27 (6mo) | |
|---|---|---|
| GAAP operating margin | 66.24% | 65.94% |
| GAAP net margin | 62.03% | 66.36% |
| Gains from equity securities, net | $7.77B | $23.71B |
| GAAP NI | $59.69B | $118.01B |
| Non-GAAP NI (ex-equity gains, SBC-inclusive) | $53.95B | $99.50B |
| Equity-mark share of GAAP NI | 13.0% | 15.7% |
[[pattern-net-margin-above-operating-margin-is-a-tripwire]] fires on the H1 aggregate — net margin (66.36%) exceeds operating margin (65.94%) despite NVDA paying real cash tax at 16.5%, because $23.7B of non-operating, non-cash equity-securities gains (marks on strategic AI-lab stakes) more than offset the tax drag. ~16 cents of every reported H1 net-income dollar was a mark on an equity stake, not operating profit. This is not disclosed as a red flag by NVIDIA — it's cleanly broken out in their own non-GAAP reconciliation — but it means the non-GAAP EPS ($4.09 H1) is the right number for a run-rate valuation, not GAAP EPS ($4.85 H1).
🚩 The cash-quality flag: DSO extension + inventory build
| Q1 FY27 | Q2 FY27 | Change | |
|---|---|---|---|
| Operating cash flow | $50.34B | $24.08B | -52% QoQ |
| Free cash flow (company-reported) | $48.55B | $21.34B | -56% QoQ |
| Days sales outstanding | 45 | 60 | +15 days |
| Inventory | $25.8B | $31.6B | +22% QoQ |
| Accounts receivable | — | $63.1B | — |
NVIDIA's own CFO commentary: DSO extended to 60 days "due to extended payment terms on large, multi-quarter agreements with certain investment-grade customers." Inventory built ahead of the Vera Rubin launch (normal, capacity-driven). The DSO extension is the more important of the two — NVIDIA is now financing part of its own sales cycle to customers, which is the mechanism [[pattern-ai-build-inflates-earnings-while-destroying-fcf]] describes generally, showing up here as a working-capital drag rather than a capex line: revenue and GAAP earnings recognize on schedule, cash lags. TTM reconciled FCF is $126.9B against $192.9B TTM GAAP NI — a 66% cash-conversion ratio, the gap explained partly by the equity marks (non-cash) and partly by this real receivables stretch.
This is not yet a crisis — FCF is still enormous and grew 76.6% YoY on a six-month basis ($69.9B H1 FY27 vs $39.6B H1 FY26), which is the opposite direction from GOOGL/AMZN/MU in the parent pattern. But the mechanism — easing terms for large customers to keep the order book moving — is the same family of risk as the vendor-financing story around the whole AI-capex complex, and it is now visible in NVDA's own numbers for the first time this cycle. Watch DSO at the Q3 FY27 print; a further move above 60 days is the tell that this is becoming structural rather than a one-quarter mix effect.
🚩 The circular-financing / guarantee book — the risk the user asked to isolate
NVIDIA's own words, from the Q2 FY27 CFO Commentary: "AI clouds and model makers are seeing extraordinary demand for AI infrastructure, yet many are growing faster than their balance sheets and long-term credit profiles can support. In response, we have entered into arrangements that help select customers secure the land, power and data center capacity needed to support their growth."
| Commitment / guarantee | Amount | Notes |
|---|---|---|
| Supply & capacity (memory procurement) | $279B | up from $119B last quarter — 2.3x in one quarter |
| Cloud service agreements | $29B | R&D infrastructure |
| Data center leases (not yet commenced) | $25B | own use, up to 20yr terms |
| Equity investments (AI labs) | $25B | "AI model makers, infrastructure financiers" |
| AI-cloud revenue-share agreements | $36B | earns revenue on upfront infra sale + rev share |
| SB Energy / OpenAI lease guarantee | up to $105.0B | signed Aug 2026 — NVIDIA guarantees land/power/shell buildout for a 4.25GW site (OpenAI-exclusive, 20yr lease); option on 3.8GW more |
| Total guarantee exposure | $108.5B | |
| Total forward commitments (all categories) | ~$366B + $56B AI-cloud |
This is the concrete shape of the circularity: NVIDIA is now a credit backstop for its own customer's infrastructure buildout — guaranteeing up to $105B so that OpenAI's data-center landlord can finance a site that will be filled with NVIDIA GPUs. NVIDIA's own disclosure frames this explicitly as compensating for counterparties whose "balance sheets and long-term credit profiles" don't yet support the buildout on their own. The guarantee is phased (triggers on data centers becoming "ready for service," first expected FY2029) and declines as OpenAI pays down lease obligations — so it is not a Day-1 liability — but it is a real, sized, multi-year contingent claim on NVIDIA's balance sheet that exists because the underlying demand isn't fully self-funding. Combined with customer concentration — a single "Customer A" ran 20-23% of revenue through FY2026, and two unnamed customers together were reported near 40% of a recent quarter's revenue (Motley Fool/DCD, FY2026 10-Qs) — the picture is a small number of capital-constrained, NVDA-dependent counterparties whose ability to pay is partly underwritten by NVDA itself. This is the single risk factor most likely to matter over a 2-3 year horizon and the one this report weights most heavily against an otherwise excellent quarter.
2. Moat & Competitive Advantage
ROIC / gross-margin trend. ROIC (roic.ai, return on invested capital) fell from 97.6% (FY2025) to 84.7% (FY2026) — still exceptional by any standard, but the direction is down, driven by the denominator (invested capital) growing faster than operating income as capex, the $14.5B FY2026 acquisition spend, and the equity-stake book all add to the capital base. Annual gross margin also dipped FY2025→FY2026 (74.99%→71.07%), but that annual figure is distorted by a $4.5B one-time H20 China export-control charge in Q1 FY26 — the trailing two quarters (Q1-Q2 FY27) both printed 75.0% gross margin, a record, on Blackwell Ultra mix. Read the annual dip as a one-off charge, not moat erosion; read the quarterly run-rate as the real trend, and it says the moat, priced in gross margin, is intact and at a high.
Adversarial stress-test. "You are a well-funded rival — AMD, a hyperscaler's custom silicon (TPU, Trainium, MTIA), or a Chinese domestic champion. How do you take share?" The honest answer: not easily, and not soon, on frontier training — CUDA's software moat (a decade of libraries, frameworks, and developer muscle memory) and NVLink/InfiniBand's networking lock-in mean a rival chip has to be meaningfully better and clear a software-migration cost most large labs won't pay while demand is unconstrained. The more credible attack vector is inference, where custom ASICs (Google TPU, Amazon Trainium) are cheaper per-token for well-defined, static workloads and don't need CUDA's generality. NVDA's own numbers already show this bifurcation: Hyperscale revenue (+13% QoQ, +102% YoY) is decelerating relative to ACIE/AI-native demand (+25% QoQ, +138% YoY) — the hyperscalers, who have the most in-house-silicon incentive and capability, are growing their NVDA spend more slowly than the AI-native labs who have neither the balance sheet nor the time to build alternatives. That's the moat working as intended, but it also means the growth mix is skewing toward the least creditworthy segment of the customer base — which ties directly back to §1's guarantee-book concern.
Disruption forecast, 5-10yr arc. Near-term (1-3yr): no credible near-term disruption to training-workload share; Blackwell Ultra → Vera Rubin roadmap keeps NVDA a full generation ahead on total system performance, not just chip performance. Medium-term (3-7yr): the real disruption vector is customers building enough in-house capability (or enough scale with a second-source ASIC vendor) that they stop needing NVDA for the majority of a workload, compressing NVDA to the highest-value slice rather than the whole stack — margin-dilutive even without unit-share loss. Long-term (7-10yr): a genuine architectural shift away from the transformer/GPU-parallel paradigm (a new compute primitive) is the only scenario that unseats the platform outright, and nothing in the current roadmap suggests it's imminent.
Evergreen assessment. NVDA is as close to a "forever business" as a single-generation-risk hardware company can be, because the moat is increasingly the software and systems stack (CUDA, NVLink, the reference-architecture ecosystem), not the silicon alone — that transfers across chip generations in a way a pure ASIC vendor's advantage does not. The qualifier: the current growth rate is not evergreen — 100%+ revenue CAGR is a buildout-phase number, and the framework's own [[pattern-margin-expansion-is-a-finite-growth-lever]] and the broader AI-capex literature both argue this rate mean-reverts. The business is evergreen; the current growth multiple the market is paying for is not, and that distinction is the crux of §3.
3. Valuation
Graham's Intrinsic Value — √(22.5 × EPS × BVPS) = √(22.5 × 7.91 × 8.07) ≈ $37.9, vs. a $227.98 price. This model is functionally useless here and is not weighted — NVDA's value is almost entirely in future growth and an intangible software moat, not in a book value of $8.07/share. Cited for completeness only, per the framework's instruction to state every model and its applicability, not because it informs the verdict.
The corrected forward multiple. Yahoo's headline forwardPE (15.03x) prices forwardEps ($15.17) — this is FY2028 (fiscal year ending Jan-2028), a full extra fiscal year past NVDA's current FY2027. [[pitfall-vendor-forward-eps-is-the-wrong-fiscal-year]] names this exact failure mode. Reconstructing from the company's own numbers:
| Basis | EPS | Implied fwd P/E | Source |
|---|---|---|---|
Vendor forwardPE (headline, wrong year — FY2028) |
$15.17 | 15.0x | Yahoo forwardEps — do not quote as "NVDA's forward multiple" |
Vendor priceEpsCurrentYear (current FY2027, but pre-print, likely stale) |
$9.05 | 25.2x | Yahoo epsCurrentYear |
| This report's current-FY (FY2027) estimate, non-GAAP, ex-equity-marks | ~$9.2-9.6 | ~24-25x | H1 actual ($4.09 non-GAAP) + Q3 guide ($108B rev, 74% GM, $9.0B non-GAAP opex → ~$2.35-2.45) + Q4 estimate (Vera Rubin ramp continuing, ~$2.6-2.7) |
The real current-year forward multiple is ~24-25x, not the 15x a screen would show. That is still a full multiple for a mega-cap, but it sits in a defensible range against 100%+ revenue growth and a 75% gross margin — the PEG-style logic (vendor's own trailingPegRatio 0.58) roughly survives the correction because both the numerator (P/E) and the denominator (growth) move together once the fiscal year is fixed.
Bogle's Expected Return — dividend yield (0.47%) + earnings growth (normalizing from the current triple-digit rate toward a more sustainable 25-40% over the next 1-2 years as the AI buildout matures) ± P/E change (assume some compression from ~29x TTM toward a more mature 20-24x as growth decelerates, i.e. a negative 15-25 percentage-point drag). Net: roughly flat-to-low-teens expected annual return over a 2-3yr horizon — a name where the growth is real but a meaningful share of it is already in the price, and multiple compression is a real offsetting force, not a tailwind. This is a more sober number than the framework's naive "buy anything growing 100%" instinct would suggest, and is the main reason this isn't a higher-conviction BUY.
DDM / DYT — N/A. Payout ratio ~1%, dividend yield 0.47%, this is not a dividend-return story today. Noted for the record: recent financial press (Yahoo Finance, 2026-08-27) is floating "dividends as NVDA's next Apple-style catalyst" given the $99B remaining buyback authorization and $56.6B cash pile — worth a fresh look if the payout policy changes materially, not before.
Fair value range: $210-290. Built from the corrected current-FY EPS estimate ($9.2-9.6) at a 23-30x band — the low end reflecting the DSO/guarantee-book/customer-concentration risk discount, the high end reflecting the genuine moat and growth. At $227.98 the stock sits inside the band, near its lower third — full, not obviously overpriced, and not the bargain the raw 15x headline multiple implies. Trim set as 32x fwd (current-FY, non-GAAP) rather than a fixed dollar — at the midpoint EPS estimate that's roughly $300, consistent with the analyst mean target ($305.79); it will rise on its own as FY2027/FY2028 earnings come in, per the framework's multiple-trim convention.
4. Sentiment
The Q2 FY27 print (2026-08-26 after close, 2026-08-27 trading day) was unambiguously well received: the stock closed +8.7% on the day ($209.66 → $227.98) against a $96.2B revenue beat (Street was modeling something closer to $92-93B), a Q3 guide of $108B that came in well above the buy-side whisper, and Jensen Huang's characterization of AI having "reached its inflection point" with 70% FY2028 revenue growth commentary reported by CNBC — a figure "far above estimates." Headlines the same week ("How Nvidia silenced the bears," "Nvidia posted a blowout quarter. China remains a quagmire.") capture the split: the print itself was clean, China remains structurally near-zero (Hopper shipments to China were <1% of Data Center revenue this quarter, and the Q3 guide assumes zero China Data Center compute revenue — a conservative, already-de-risked assumption rather than a new negative).
Insider activity shows a long, remarkably regular cadence of CEO Jensen Huang sales (225,000-share tranches at consistent intervals going back over a year, textbook of a pre-arranged 10b5-1 selling plan) plus routine director/officer sales and stock gifts — no insider purchases in the disclosed window, but also no acceleration or unusual timing clustering around the earnings beat. Read this as background liquidity/diversification, not a signal either way — [[pitfall-yahoo-insider-purchases-counts-rsu-grants]] is a reminder not to misread the $0-price stock-award grants littering this list as buying.
Broader tape context (2026-08-28): Nasdaq/S&P futures were reported softer premarket the same morning NVDA is in focus, alongside Fed Governor Warsh's Jackson Hole debut — a macro, not company-specific, crosscurrent. Not weighted into the verdict beyond noting the stock's -1.0% premarket pullback the day after its post-earnings pop is ordinary give-back, not a reversal signal.
5. Synthesis & Verdict
Is this a good business? Yes, emphatically — a 75% gross margin, 85% ROIC platform business growing Data Center revenue 117% YoY with a real, multi-generation software moat (CUDA/NVLink) that the hyperscalers' own custom-silicon programs haven't dented at the frontier. The framework's evergreen test passes.
Has the market already priced that in? Mostly, once the vendor's wrong-fiscal-year multiple trap is corrected. The headline "15x forward" that would scream BUY is a FY2028 number two years out; the real current-FY multiple is ~24-25x, which is full for a mega-cap even against 100%+ growth, and sits inside — not below — this report's own $210-290 fair-value band.
The named tension, weighted honestly: Fundamentals and Moat argue for real quality (near-BUY conviction on business durability alone); Valuation argues for a fair, not cheap, entry (WATCH-territory on price); and the fresh, concrete evidence this quarter of (a) net margin exceeding operating margin on non-cash equity marks, (b) a 15-day DSO extension NVIDIA itself attributes to financing customer growth, and (c) a $105B guarantee explicitly built to backstop a customer whose "balance sheet... can't support" its own buildout, together constitute a real and growing, if not yet acute, circular-financing risk that the market's +8.7% reaction the same day did not appear to price at all. Per §5 of the framework's conflict-resolution rules, this is a case where Fundamentals' quality read and Valuation's full-price read pull in opposite directions and neither should be allowed to silently win — the honest verdict is ACCUMULATE, not BUY: a name worth owning and adding to on any pullback toward the entry zone, sized with the guarantee-book/DSO risk explicitly in mind, not a name to chase at the post-earnings high.
Conviction: 6.5/10. This is a close call because the business quality alone would justify higher conviction, but the combination of a full (if not absurd) corrected multiple and a genuinely new, disclosed-but-underpriced financing-circularity risk caps it below the top of the scale. Recheck at the Q3 FY2027 print (~2026-11-18, to be confirmed closer to the date) — the specific things to re-test: DSO trend (a further move past 60 days would upgrade this from "flag" to "thesis-relevant"), the size of the guarantee book (whether $108.5B grows further), gross margin (should hold ~74-75% per guidance), and whether any China Data Center compute revenue re-enters guidance (currently assumed at zero, so any re-entry is a modest upside surprise, not a risk).
Data: .mcp/fin.py NVDA --news (2026-08-28); Yahoo Finance MCP (get_stock_info, get_financial_statement quarterly income/cashflow, get_holder_info insider/institutional); roic.ai MCP (income_statement, cash_flow, profitability_ratios, annual, NASDAQ:NVDA — capped at 2yr history per [[pitfall-roicai-free-plan-caps-history-at-two-years]]); NVIDIA Q2 FY2027 press release (SEC 8-K, 2026-08-26/27) and CFO Commentary PDF, retrieved directly to correct for vendor-feed lag. Knowledge base checked via python .mcp/kb.py find NVDA (no prior notes) and the Playbook pitfalls named inline above.