← All companies

NVIDIA

NVDANASDAQUSDAccelerated computing and AI infrastructureReviewed Nov 15, 2026

AI platform operating cash and capital

8 of 11 research checks complete. Still to complete: Security and evidence base; Opportunities and funded scenarios; Variant view and catalysts · Trailing four quarters to 25 Oct 2026 (Q3 FY2027 at guidance); July 2026 balances; August/October funding reserves

GPU compute, networking and software serve the AI infrastructure buildout. Long-run returns depend on customer economics, platform durability and the funding needed to convert demand into cash.

  • Customer GPU utilization, paid AI demand and cash funding
  • Gross versus operating margin; supplier advances, inventory and extended terms
  • New chip/custom ASIC competition; software/platform switching costs
  • Ecosystem investments, partner capacity and contingent guarantees
Inspect live quotes, sensitivity grid & cash-flow bridge

Scenario assumptions

AI platform operating cash and capital · Trailing four quarters to 25 Oct 2026 (Q3 FY2027 at guidance); July 2026 balances; August/October funding reserves

DriverDownsideBaseUpside
Initial growth from estimated run-rate sales-15.0%28.0%58.0%
Revenue growth from year 3 onward3.0%5.0%8.0%
Year 5 GAAP operating margin30.0%50.0%58.0%
Capital spending / revenue6.0%4.0%3.0%
Operating working capital / revenue30.0%22.0%18.0%
Normalized operating tax rate23.0%20.0%17.0%
Exit enterprise value / unlevered NOPAT13.00×18.00×22.00×
Annual share dilution0.0%0.0%0.0%
Retained equity marks/cost recovered40.0%80.0%100.0%
Loss stress on partner capacity and guarantees25.0%0.0%0.0%
Revenue growth
Year 1: -15.0%
Years 2–5: 3.0%
Year 1: 28.0%
Years 2–5: 5.0%
Year 1: 58.0%
Years 2–5: 8.0%
Margin
Year 1: 53.8%
Year 2: 41.9%
Years 3–5: 30.0%
Year 1: 60.4%
Year 2: 55.2%
Years 3–5: 50.0%
Year 1: 63.1%
Year 2: 60.5%
Years 3–5: 58.0%
  • Downside case. Digestion: revenue falls 15% below the trailing four quarters in year 1 (about 30% below the Q3 run rate, the base-rate range of past capex-cycle busts), then recovers at 3%. Margin reverts to 30% within three years, working capital 30%, capex 6%. Equity stakes recover 40%; 25% of guarantees and partner capacity paid in year 3. Exit 13x trough-normalized NOPAT.
  • Base case. Customer-funded growth: year 1 +28% (hyperscalers at 20% of their cash-plus-financing budget, ACIE about 6 GW), then 5% as custom silicon takes share. Margin normalizes to 50% within three years. Capex 4%, working capital 22%. Exit 18x normalized NOPAT; stakes at 80%.
  • Upside case. Near the demand ceiling: year 1 +58% (559bn against a 576bn ceiling), then 8% as NVIDIA holds share. Margin 58%, capex 3%, working capital 18%. Exit 22x, capped because margins stay near peak; stakes at 100%.

Research notebook

Reviewed Nov 15, 2026

8 of 11 research checks complete. Still to complete: Security and evidence base; Opportunities and funded scenarios; Variant view and catalysts
Business model

NVIDIA sells accelerated computing systems: GPUs, CPUs, NVLink and network fabrics, reference rack designs and the CUDA software stack that ties them together. Data Center produced USD89.0bn of USD96.2bn Q2 FY2027 sales (nvda-q2-fy2027-release). The rest ("Edge": gaming, professional visualization, automotive) is a mature franchise. Manufacturing is outsourced to TSMC, memory makers, packaging houses and ODMs. External manufacturing does not remove capital needs: supply commitments, inventory, receivables, ecosystem equity stakes and customer credit support all use shareholder capital.

Revised 3 Nov 2026 for the research pipeline. Fiscal years end on the last Sunday of January, so FY2026 is roughly calendar 2025. Money is in USD millions unless stated. Fact IDs are in financials/facts.json.

Value-creation mechanism

NVIDIA creates value by turning a fixed, front-loaded R&D budget (USD23.6bn trailing, researchFY2026 etc.) into a new full AI computer system every year. It then secures enough of the world's scarce leading-edge wafers, advanced packaging and high-bandwidth memory to deliver that system before any alternative can. Customers pay for useful compute delivered soon, and judge the price against their cost per token and the revenue that idle capacity forgoes. The integration step creates most of the value: chip, interconnect, networking and CUDA libraries tuned together, sold into every cloud. Management sizes its revenue opportunity per gigawatt of AI capacity at about USD18bn for Hopper, USD25bn for Grace Blackwell and USD40bn for Vera Rubin (revenuePerGwRubin, nvda-q2-fy2027-call pp.4 and 10). Each generation captures more of the customer's capacity budget.

Supply orchestration is the second value step. NVIDIA commits ahead of demand: USD279bn of supply and capacity commitments at July 2026, mainly memory, up from USD95.2bn in January (supplyCommitments, supplyCommitmentsFY2026). This makes NVIDIA the bottleneck owner in a capacity-constrained market, which lets a 75% gross margin coexist with rising component costs. The third step is newer, and its value is unproven: NVIDIA lends its balance sheet to customers without investment-grade credit. That takes the form of equity stakes in AI labs (nearly USD50bn, aiLabInvestments), guarantees (USD108.5bn maximum) and extended payment terms (DSO 60 days, dsoQ2). This step creates demand and also absorbs capital and risk.

Business-model profile

Comparable firms are chosen by these traits, not by the industry label.

Growth vehicle: Organic. Revenue rose from 26,914 in FY2022 to 215,938 in FY2026, driven by data-center demand for its own platforms. Acquisition cash was 0.6% of revenue over FY2024–FY2026. Growth is organic: it comes from customers' capital spending on accelerated computing and is funded by the company's own operating investment. The model forecasts organic units and capacity, with no deal budget.

Capital intensity: 2.5% of sales, asset-light. Capital spending is about 2.5% of revenue because manufacturing is outsourced to TSMC, memory makers, packaging houses and ODMs. The capital growth uses sits elsewhere: inventory, receivables, supply prepayments and purchase commitments (tangible operating capital of 85.7bn at July 2026), plus ecosystem equity stakes and customer credit support.

Comparables. A comparable firm is a fabless platform supplier into a customer capital-spending cycle: revenue depends on a few large buyers' capex budgets, growth is organic and internally funded, and margins depend on a technology lead and a software ecosystem. Base rates must treat it as a very large supplier at the top of a spending boom. Not the label: "Semiconductors" mixes NVIDIA with capital-intensive foundries and commodity memory; it is a fabless systems and software platform exposed to its customers' capex cycle.

Declared in valuation/model.json (businessProfile). npm run check measures deal and capital intensity from the reported facts, and the outside view grades the forecast only against firms that share them.

Economic snapshot

Definitions:

  • Trailing (TTM). FY2026 + H1 FY2027 − H1 FY2026.
  • NOPAT. GAAP operating income × (1 − 17%), using the FY2027 guided tax midpoint, with stock compensation charged. Investment gains and interest are excluded.
  • Balances. At 26 July 2026.
MeasureValueBasis
TTM revenue302,970revenueTtm
TTM GAAP operating income197,579 (65.2%)operatingProfitTtm
TTM NOPAT163,991197,579 × 0.83
Q2 gross margin75.0%FY2026 71.1%, FY2025 75.0%
Incremental operating margin, FY2025→FY202657.3%Δ operating income 48,934 / Δ sales 85,441
Receivable days60Up from 51 in January
Inventory days on Q2 cost of sales12031,575 / (24,079 × 4) × 365
Payable days5715,059 / annualized cost of sales

Capital base at July 2026:

ComponentAmount
Working capital excluding cash (receivables 63,059 + inventory 31,575 + prepaids 3,409 − payables 15,059 − accrued liabilities 26,960)56,024
PP&E14,285
Right-of-use assets5,390
Other operating noncurrent assets (supply prepayments and similar, ex equity and warrants)9,965
Tangible operating capital85,664
Goodwill and intangibles (mostly Groq, 2026)24,123
Ecosystem equity stakes at carrying value98,897
Cash and debt securities56,586
Debt33,500
Lease liabilities5,494

Returns on that capital:

  • Return on tangible capital, capitalizing leases: 191%.
  • Return on capital employed, expensing leases (excluding the ROU asset): 204%. Including goodwill and intangibles: 157%.
  • Including the ecosystem stakes as capital that now supports demand: 79%.

These are ending balances rather than averages, which flatters nothing because capital is growing.

Normalization: TTM capex is 7,474 against depreciation of 3,687. That is about twice D&A, which reflects growth capacity rather than under-maintenance, so no add-back is made. There are no minorities, and recent acquisitions are licenses and teams with no separately disclosed profit. The real normalization question is cyclical. The Q2 75% gross margin and 65% operating margin sit far above NVIDIA's own pre-AI record:

  • FY2022: 64.9% gross, 37.3% operating.
  • FY2023 downturn: 56.9% gross, 15.7% operating, including the Arm write-off and inventory provisions.

Stage 7 and the scenarios treat today's margin as a shortage-era level, not a through-cycle norm.

Economic segmentation

Reported segments are evidence, not the answer. In FY2026, Compute & Networking earned 130,141 of 139,297 segment operating income, or 93%, on a 67% segment margin. Graphics earned 9,156, a 41% margin (segmentComputeNetworkingOpFY2026, segmentGraphicsOpFY2026). Cut by how value is created and who decides the purchase:

Economic unitSizeApproximate profit share (estimate)Demand driverCharacter
Hyperscale AI systems (Microsoft, Google, Amazon, Meta, Oracle)48.7bn of Q2 Data Center, about 55% (hyperscaleRevenueQ2)~50%The buyers' own capex budgets. Partly for rented capacity, where their cloud customers ask for CUDA; partly for internal workloads, where their own chips compete.Growing, but this is where buyer power and substitutes are strongest.
ACIE: neoclouds, sovereigns, enterprises and AI labsAbout 40bn of Q2, +138% year on year~40%AI labs' and start-ups' compute needs, funded by venture capital (over USD400bn in H1 2026, per management), debt and NVIDIA's own equity and credit. About a quarter of next year's business is with AI labs NVIDIA supports with its balance sheet (aiLabFinancedShare).Fastest growing. Weaker credit and highly cyclical with capital markets.
Networking (NVLink, InfiniBand, Spectrum-X)31,376 in FY2026, +142% (dataCenterNetworkingFY2026)Inside the two units aboveAttached to each compute deploymentThe Mellanox franchise. It raises content per gigawatt and switching cost.
Edge: gaming, workstations, automotive22,459 Graphics segment revenue in FY2026~7%PC and replacement cycles; memory prices now weigh on PCsMature and cyclical; a small share of value.

The profit split between hyperscale and ACIE is an estimate. NVIDIA does not disclose margins by customer type. I assume similar gross margins, because hyperscalers' volume discounts and ACIE's extended terms and financing roughly offset.

The implication: about 90% of profit and nearly all of the value depend on one product cycle sold to two kinds of buyer:

  • Five hyperscalers, all of which build their own accelerators.
  • A set of capital-hungry AI labs and neoclouds whose purchases NVIDIA partly finances.

A sector label ("semiconductors") hides this. The demand is a capital-spending cycle, and the cycle's financing is partly NVIDIA's own.

Unit economics

Unit: one gigawatt of Vera Rubin AI capacity. Customers add capacity in gigawatt-scale AI factories, and management prices its opportunity per gigawatt. Revenue comes from management's sizing. The other lines are this repository's estimates, built from guided ratios.

Line, per GWAmount (USD bn)Basis
NVIDIA revenue40.0Management sizing (revenuePerGwRubin)
Cost of goods at the 72.5% FY2028 margin guide11.0Memory-heavy (grossMarginGuideFY2028Low, p.7)
Gross profit29.0
Allocated operating expense3.4GAAP opex at 8.5% of sales (Q3 guide 9.2bn on 108bn). Mostly fixed R&D; truly variable selling and support are under 1 point.
Unit operating profit25.664%
NOPAT at 17%21.2
Working capital at 22% of sales8.8
Capex at 1.5% of sales0.6
Tangible capital9.4
Return on tangible capital~225%

Adding NVIDIA's financing changes the picture. OpenAI's ~12 GW is supported by part of the ~USD50bn in lab equity, about USD4bn per GW (nvda-q2-fy2027-call p.6). With that included, capital becomes about USD13bn per GW and the return about 160%. The NVIDIA-side unit is therefore not capital-constrained: its value depends on price per gigawatt, not on capital intensity.

The consensus question is the customer's unit, because their returns fund NVIDIA's price. Management puts a full gigawatt at about USD60bn of customer spend today (p.9). Servers are depreciated over five to six years at Amazon, Meta, Alphabet and Oracle (amzn-q3-2026, meta-q3-2026, googl-q3-2026, orcl-q1-fy2027). On that basis the customer carries roughly 8–10bn of annual depreciation per GW. Power at an assumed USD0.08/kWh adds about 0.7bn. Earning 10% pre-tax on USD60bn therefore needs about USD15–17bn of annual revenue per GW. If annual product cadence shortens the economic life to three years, the requirement rises to about USD22bn+.

Customer revenue is responding:

  • Google Cloud +103% to 30.7bn in Q3.
  • AWS +39% to 46.0bn.

But customers' capex now exceeds their operating cash flow:

  • Alphabet free cash flow −9.1bn in Q3.
  • Amazon trailing free cash flow −42.4bn.
  • Meta −2.1bn in Q3.
  • Oracle −5.4bn, while selling USD20bn of stock.

The customer unit is not yet self-funding at the margin.

Failure case: H20 for China. NVIDIA designed a product to fit export rules, then a regulator closed the market. The result was a 4.5bn inventory and purchase-obligation charge, about USD60m of licensed revenue, and no H200 revenue under the February 2026 license (inventoryProvisionsFY2026, nvda-fy2026-10k). The levers that work with customers (performance, supply, software) fail when the counterparty with the outside option is a government.

An earlier failure was FY2023. Inventory doubled (2,605 to 5,159, inventoryYeFY2023), the gross margin fell 8 points and operating income fell 58% while revenue was flat. Shipped capacity outran end demand.

Independent checks

  1. Segment accounts. Compute & Networking earned a 67.3% segment margin in FY2026, before unallocated stock compensation and acquisition costs. Group GAAP operating margin was 60.4%. The per-GW model's 64% sits between them, consistent with allocated fixed R&D.
  2. Customer budgets. The five hyperscalers' own 2026 guidance sums to about 810–830bn:
    • Microsoft about 170bn.
    • Alphabet 195–205bn.
    • Amazon about 220bn.
    • Meta 135–145bn.
    • Oracle FY2027 about 90–95bn. This figure comes from a call summary and is unverified. NVIDIA puts it at "nearly 800bn" (p.3). NVIDIA's hyperscale revenue annualizes to about 195bn, roughly 24% of that capex. That is consistent with 40bn of NVIDIA content in about 60bn per GW, once land, buildings, power and these customers' own TPUs, Trainium and MTIA are excluded.
  3. Competitors' audited revenue. In the quarter to early August, Broadcom's AI semiconductor revenue was 16.7bn (+221%) and AMD's Data Center revenue was 6.7bn (+107%) (avgo-q3-fy2026, amd-q2-2026). NVIDIA's Data Center revenue was 89.0bn (+117%). NVIDIA still takes about 79% of these three firms' AI data-center revenue. Custom silicon is growing about twice as fast from a smaller base.
  4. Adverse allocation. Charge all trailing R&D (23.6bn) and all stock compensation (7.2bn) to the hyperscale unit alone. Its operating margin still exceeds 55%. No unit is unprofitable today, so the question is durability of price, not current profitability.

Where the consensus disagrees: one camp calls this a supply-constrained secular buildout; the other, circular financing. Three observations would settle it:

  • Receivable days and receivables growth against revenue.
  • The share of revenue from customers NVIDIA finances.
  • Whether hyperscaler free cash flow turns positive while capex keeps rising.

Stage 11 freezes these as prospective checks.

History, demand and counterparties

NVIDIA's GPU business grew from graphics into general accelerated computing. Its software ecosystem then made it the default data-center platform for training and inference. The financial question today is how much of the AI infrastructure buildout becomes lasting owner cash, after suppliers, employees and customers receive the capital they need.

The record shows how fast the business moved:

  • Revenue went from 16,675 in FY2021 to 215,938 in FY2026.
  • GAAP operating income went from 4,532 to 130,387.
  • Trailing revenue at July 2026 was 302,970.

Fiscal labels are not calendar years: FY2026 ended 25 January 2026 (nvda-fy2026-10k, nvda-xbrl-companyfacts). Prior H20 restrictions and inventory provisions distort year-on-year comparisons. The financing structure changed in 2026:

  • USD25bn of notes issued in June.
  • About USD99bn of equity stakes.
  • A USD105bn SB Energy guarantee (nvda-q2-fy2027-10q, notes 5–10; nvda-sb-energy).

No external NVIDIA investment thesis or published shareholder IRR was supplied. Scenario schedules are this repository's estimates.

Demand drivers and cycle

Who uses, who pays, who decides. The end user is a business or consumer using an AI service. The payer is whoever owns the capacity:

  • A hyperscaler, for its own services or to rent out.
  • A neocloud.
  • A sovereign program.
  • An AI lab, often through a cloud contract.

The decision-maker is the payer's infrastructure team, but cloud renters' software choices constrain it. Hyperscalers buy NVIDIA for rented capacity because their customers' code runs on CUDA. For internal workloads they can choose their own silicon.

Split of demand. Using management's own categories and this repository's estimates, current demand is:

Kind of demandEvidence
Mostly growth capacityNeoclouds expected to go from about 3 GW installed at end-2025 to 8 GW by end-2026 (nvda-q2-fy2027-call p.3).
Increasingly replacementAnnual product cadence; prior-generation Trainium and GPU capacity still rents at a premium (amzn-q3-2026), so replacement is driven by economics rather than wear.
Partly regulation-drivenSovereign "AI for data residency" programs grew 35% sequentially (p.5).
A financing-driven partAI labs supported by NVIDIA's balance sheet are about a quarter of next year's business (aiLabFinancedShare).

Little is discretionary consumer demand.

The cycle, measured. NVIDIA's own record has two inventory cycles in the last eight years. The second, FY2023, had:

  • Flat revenue (26,914 to 26,974).
  • Inventory doubling (2,605 to 5,159).
  • Gross margin down from 64.9% to 56.9%.
  • Operating income down 58%.

Today's customer side shows the ingredients of a late capital cycle (googl-q3-2026, amzn-q3-2026, meta-q3-2026, orcl-q1-fy2027). Capex roughly doubled year on year in Q3 2026:

CompanyQ3 2026 capexQ3 2025 capex
Alphabet51.4bn24.0bn
Amazon77.3bn35.1bn
Microsoft39.4bn19.4bn

It now exceeds internal cash generation:

  • Alphabet Q3 free cash flow −9.1bn.
  • Amazon trailing free cash flow −42.4bn, plus 48.1bn of new long-term debt in Q3.
  • Oracle −5.4bn, plus a 20bn equity sale.

Revenue is following the capex:

  • Google Cloud +103%.
  • AWS +39%, backlog 578bn.
  • Microsoft commercial RPO 677bn.
  • Oracle RPO 664bn.

Much of that backlog is AI-lab commitments that are themselves externally financed. Microsoft lowered its calendar-2026 capex estimate from about 190bn to about 170bn, citing delayed third-party deliveries and lease accounting (msft-fy2027-q1).

Market misperception. The common bull view is a "supply-constrained secular" buildout. The common bear view is "circular financing". The series that tests both is customer capex against customer operating cash flow and external financing, alongside NVIDIA's receivable days (45 → 60) and its financed share of revenue. The current data support strong 2027 demand: Alphabet and Amazon both guide 2027 capex higher. They also show that marginal demand is now funded by debt, equity and NVIDIA's own balance sheet rather than by hyperscaler surplus cash. That makes revenue more cyclical than an installed-base replacement model would suggest.

Competition

Competition happens in three places:

  1. Inside each hyperscaler's own fleet. NVIDIA competes with:
    • Google's TPU: generation 8t/8i, nearly 100 customers, external system sales ramping in 2027, Anthropic as a large user.
    • Amazon's Trainium: commitments up more than 20x, to be sold outside AWS from late 2027.
    • Microsoft's Maia 200/300: Microsoft claims 40% better tokens per watt than leading GPU systems.
    • Meta's MTIA, built with Broadcom. Sources: googl-q3-2026, amzn-q3-2026, msft-fy2027-q1. Here the buyer is also the competitor, and the battleground is internal inference cost.
  2. Merchant accelerators for rented and lab capacity. AMD's MI400/Helios has commitments from Anthropic (up to 2 GW), Meta (up to 6 GW), OpenAI, Oracle and Microsoft (amd-q2-2026). OpenAI's own Broadcom XPU is, on Broadcom's call, on track for 1.3 GW in 2027. That call comes from an unverified third-party transcript.
  3. Networking. Broadcom and Ethernet switch vendors compete against NVLink, InfiniBand and Spectrum-X.

Share. In the latest comparable quarter NVIDIA had about 79% of NVIDIA + AMD Data Center + Broadcom AI revenue (89.0 / 112.4bn). Broadcom grew 221% and AMD Data Center 107%, against NVIDIA's 117%. The number of serious competitors is rising, not falling: Broadcom alone reports six XPU customers.

The advantage in one sentence. NVIDIA delivers the most useful compute soonest to any customer in any cloud, because it ships a fully integrated system every year, on supply it secured first, with the software most AI developers already use. The durability evidence is mixed:

  • For durability: share held through Hopper, Blackwell and Rubin; a 75% gross margin despite record memory costs; purchase orders for Vera Rubin from every major hyperscaler (p.4).
  • Against durability: the largest buyers, and NVIDIA's own investees OpenAI and Anthropic, are building alternatives (p.11).

The competitor that could hurt NVIDIA is Google. It owns the whole stack: chip, compiler, cloud and frontier model. It is now selling TPU systems outside its own fleet, and Anthropic's commitments give it scale. It has not hurt NVIDIA yet for three reasons:

  • TPUs were internal-only until recently.
  • Their supply depends on the same TSMC and HBM capacity that NVIDIA reserves first.
  • Most outside developers' code and tooling target CUDA.

Counterparty incentives

Hyperscalers (Microsoft, Alphabet, Amazon, Meta, Oracle; about 55% of Data Center revenue).

  • What they want: lower cost per token, supply security and freedom from a single vendor's margin.
  • What NVIDIA gives them: the capacity their cloud customers ask for, delivered first.
  • Their outside options: proven in-house chips, AMD.
  • Rational behaviour: move internal inference to their own silicon while buying NVIDIA for rented capacity and frontier training.
  • What changes the deal: their own chips becoming good enough for external renters (Google's external TPU sales and Amazon's 2027 Trainium sales are the first moves).
  • Dependence runs both ways. NVIDIA's response is to grow ACIE, the customers without their own silicon, to "roughly half" of Data Center (p.5). That deliberately reduces buyer power.

AI labs (OpenAI with about 12 GW of NVIDIA commitments, Anthropic, others).

  • What they want: the most compute at the lowest cost, plus supplier diversity. Both are building or buying alternatives: OpenAI's XPU, Anthropic's TPU, Trainium and AMD.
  • What NVIDIA gives them: equity (nearly 50bn), credit enhancement for about 2 GW at another lab, and guarantees for OpenAI's 4.25 GW Ohio site (p.6).
  • Their outside options: many for chips, few for credit. NVIDIA's balance sheet is therefore its bargaining chip.
  • Risk to NVIDIA: if lab revenue disappoints, NVIDIA loses sales, receivables, equity value and guarantee exposure at the same time.

Suppliers (TSMC, SK hynix, Samsung, Micron, packaging firms).

  • What they want: long commitments that pay back their capacity, and pricing power while capacity is short.
  • What NVIDIA gives them: scale and up to 279bn of commitments.
  • Their outside options: Google/Broadcom, AMD, Amazon.
  • Current state: memory makers are taking value. Memory pricing "exceeded expectations", and NVIDIA lowered its margin expectations to 71–72% in Q4 and 72–73% in FY2028 (p.7).

The U.S. government wants control over AI capability exports. China Data Center compute is out of the outlook. Regulation can remove a market regardless of customer demand, as the H20 charge showed.

ODMs and distributors (the 22% and 14% direct customers) pass the systems through. Their power is small.

Management and judgment

Jensen Huang co-founded NVIDIA in 1993. His resignation as president, CEO and director was accepted on 11 November and disclosed on 12 November. Colette Kress is interim CEO and continues as CFO; Harvey C. Jones chairs the board. The permanent successor is not yet named. See the filed 8-K, Item 5.02. The investment case depends on architectural execution, supply coordination and, increasingly, capital allocation across customers, financiers and suppliers. Management's product and demand statements are evidence to examine, not achieved cash flow.

13 November source review. Three original Stevens Forms4 (filed2,8,22September) report4,236,740 shares sold31August–18September for approximatelyUSD946.63m at the disclosed weighted prices. All three10b5-1 flags are0. These are secondary sales from trusts, not issuer cash, dilution or evidence of motive. The latest form retains11,544,612 direct shares,15,017,750 Envy Trust shares and970,531 Third Millennium Trust shares. This corrects the limits of a media headline without implying misconduct. Reuters' Kress profile (AOL syndication, updated13November02:16UTC) describes attributed disagreements over financing; it is not a cancellation notice or a new authorized cash commitment. MHA's12November original statement distinguishes domestic false-declaration evidence from espionage: it reports an inaccurate-information warning/bar, insufficient evidence against Huang on the server declarations and an ongoing Megaspeed investigation. No all-clear or quantified NVIDIA loss follows.

8 November 2026 governance update. The board announced an independent investigation into reports of foreign intelligence targeting Huang. Independent directors and outside counsel oversee it. The board retains confidence in current leadership and says it has no reason to believe company secrets were compromised; those are the board's positions while the inquiry is open. The CEO's issuer-published response supports the inquiry. Those 8 November statements preceded the confirmed 11 November leadership decision; they no longer describe the current succession position. Monitor decision-making continuity, information security and regulatory/customer consequences; personal conduct alone does not quantify owner-cash impairment.

On the August call, management gave a preliminary, supply-constrained outlook of about 70% FY2028 revenue growth (fy2028GrowthOutlook). That informs near-term cases only. It does not establish a decade of growth or customers' returns.

Starting in FY2027, NVIDIA's non-GAAP measures retain stock compensation, and history was recast (nvda-q2-fy2027-release). The model uses GAAP operating expenses and never adds stock compensation back.

Ownership and incentives

Sources: nvda-proxy-2026 pp.35–36 (ownership at 23 Mar 2026), the summary compensation table and CD&A pp.44–46; nvda-huang-form4.

PersonRole at the historical proxy dateShares (m)Value at USD217.94FY2026 reported payNotes
Jen-Hsun HuangFounder, CEO870.6 (3.58%)~USD190bnUSD36.3mSalary 1.5m, PSUs 24.8m, cash incentive 6.0m, security and other 4.0m. 10b5-1 plan of 20 Mar 2025 for up to 6m shares (about 0.7% of holdings), expired 31 Dec 2025. No open-market sales in 2026, only tax withholding and gifts (438k shares to his foundation and a donor-advised fund in Sep 2026). In Dec 2025, 39.85m shares were distributed into an independently managed LLC that he no longer beneficially owns.
Colette KressCFO4.4~USD1.0bnUSD14.3m10b5-1 plans adopted and terminated in 2026
Jay PuriEVP Worldwide Field Operations3.3~USD0.7bnn/a10b5-1 plan for 1m shares (Sep 2025)
Debora ShoquistEVP Operations1.6~USD0.3bnn/a10b5-1 plan for 0.31m shares
Tim TeterGeneral Counsel2.7~USD0.6bnn/a10b5-1 plan (May 2026)
All 14 directors and officers957.3 (3.94%)~USD209bnLargest outside holders: BlackRock 7.43%, Vanguard 7.31%

Historical founder alignment. The proxy holdings and USD217.94 valuation below predate the transition. They are not a current beneficial-ownership update, an incentive schedule for Kress, or evidence that Huang retains executive authority. No new severance or award-forfeiture amount is disclosed in the transition filing.

What the stake meant under the former CEO. If the stock compounds at 10% a year for five years, Huang's stake gains about USD116bn. Five years of pay at the FY2026 level is about USD180m, roughly 650 times smaller. Ownership, not pay, sets his interests. His sales have been small relative to his holdings and pre-scheduled. There is no evidence of exercising early to sell near a peak.

What pay rewards. FY2026 pay had three parts:

  • Cash plan: paid on revenue. The base goal of USD130bn equalled FY2025's actual USD130.5bn, so zero growth earned a 100% payout. Actual revenue of USD215.9bn paid 200%.
  • One-year PSUs: paid on non-GAAP operating income. The base was USD71bn, below FY2025's level. The stretch goal was revised down from USD120bn to USD96bn after the H20 export controls. Actual was USD137.3bn (cashPlanBaseRevenueFY2026, psuBaseOperatingIncomeFY2026).
  • Multi-year PSUs: paid on three-year TSR against the S&P 500. FY2024–FY2026 finished at the 100th percentile (1,055% TSR).

Backing out the targets: base operating income of USD71bn is about USD59bn of NOPAT at 17%. Against a market value of about USD5.3tn, that is roughly 90 times earnings, so the incentive goals say nothing about value at today's price.

No metric charges management for capital used. Not return on capital, nor the roughly USD100bn of ecosystem stakes, nor the USD108.5bn of guarantees. Revenue-weighted pay rewards exactly the customer financing that inflates revenue. Pay is too small relative to ownership for this to dominate Huang's behaviour. It is a weaker check on the wider executive team.

Track record and references

Record under the pre-transition team.

  • Returns on capital: FY2022 operating margin 37.3%, FY2023 15.7%, FY2026 60.4%. GAAP operating NOPAT went from 4,413 in FY2023 to 110,676 in FY2026 (operatingNopatFY2023, operatingNopatFY2026).
  • Market won: about 79% of the three leading AI-accelerator vendors' data-center revenue (research/02-history.md).
  • Total shareholder return: 1,055% over FY2024–FY2026 (nvda-proxy-2026).

Decisions that worked.

  • The Mellanox acquisition (completed FY2021): networking reached 31.4bn in FY2026, more than ten times FY2021 (dataCenterNetworkingFY2026).
  • Committing to CUDA and data-center systems a decade before the demand arrived.
  • Buying back USD10.0bn of stock in FY2023, during the downturn (buybacksFY2023).

Decisions that failed.

  • Arm: terminated for regulatory reasons, with a 1.35bn write-off of the prepayment, though NVIDIA kept the 20-year license (nvda-arm-termination).
  • FY2023 inventory glut: inventory doubled and gross margin fell 8 points.
  • H20 for China: a 4.5bn charge.
  • Cryptomining disclosure: an SEC penalty over FY2018 disclosure of cryptomining effects (sec-nvda-crypto-2022).

References. None were collected. This repository has done no interviews with former employees, customers or suppliers. Outside views read here are counterparties' own statements:

  • Broadcom's and AMD's claims about competing chips.
  • OpenAI's public claims about its own chip, which an analyst raised on NVIDIA's call (p.11).

Those are competitors' claims, not references. The proxy-reported bench is long-tenured (Kress, Puri, Shoquist, Teter). Still, product architecture and the ecosystem-financing strategy are visibly Huang's, so key-person risk is real and unquantified.

Capital allocation principles

  1. Zero-based allocation. Partly met. R&D has followed data-center opportunity: Compute & Networking is 90% of revenue, and the networking, CPU and Groq inference technology all aim at raising content per gigawatt. Gaming was not starved but was not favored either.
  2. Funding strategies, not projects. Met, and the strategy is explicit. Stakes in AI labs, guarantees for neocloud and lab sites, and six infrastructure-finance partnerships all aim to build a customer base that does not design its own chips (p.6). The strategy is coherent. It also makes NVIDIA a lender to its own demand.
  3. Capital is accessible but has a cost. Mixed:
    • In favour: including stock compensation in non-GAAP from FY2027.
    • Against: the CFO justified the nearly USD50bn of lab investments as "a small fraction" of expected free cash flow (p.6), a test of affordability, not of return against a hurdle.
    • Against: buybacks run at about USD20bn a quarter regardless of price.
  4. Zero tolerance for bad growth. Partly met. NVIDIA walked away from Arm when regulators blocked it, wrote off H20 rather than chase China, and exited cryptomining products. But guarantee and equity exposure has expanded quickly without a stated limit.
  5. Price versus value. Weak evidence either way:
    • FY2023's buyback was bought cheaply.
    • FY2025–H1 FY2027 repurchases of about USD114bn (33.7bn + 40.1bn + 39.8bn) were made near record prices, with no stated valuation discipline.
    • The equity stakes are marked up (23.7bn of H1 gains) but not yet realized.

Governing objective before the transition. Platform share and the AI buildout dominated the former team's public allocation decisions. The founder's historical stake does not establish the new interim team's authority or future discipline. Assess Kress's actual commitment decisions and technical-management structure rather than assign her either Huang's results or an assumed capital-allocation improvement.

Integrity and governance

Governance structure.

  • One share class with one vote per share; no dual-class control.
  • Nine of ten directors are independent; an eleventh joined in July 2026.
  • Shareholders passed a proposal to remove supermajority voting.
  • Say-on-pay passed with about 93.6% of for-plus-against votes (15.7bn for, 1.1bn against; nvda-agm-2026-votes).

Related-person transactions (nvda-proxy-2026 p.34):

  • Huang's two adult children are employees, paid about USD1.2m and USD1.3m.
  • A director's son is an employee.
  • The Huang foundation buys GPU time from CoreWeave to donate to researchers, and NVIDIA plans free engineering support for some recipients. CoreWeave is also an NVIDIA customer and investee. The amounts are small, but the overlap should be watched.

Integrity evidence.

  • Bonus goals were revised automatically, and downward, when export rules changed. That was a formula, but it protected payouts.
  • The 2022 SEC settlement concerned under-disclosure of a cyclical demand source. That is relevant today, when financed AI-lab demand is about a quarter of next year's business.

Agency costs.

  • Control of resources. The material agency cost. Pay keyed to revenue and operating income, together with a strategy of financing customers, can grow revenue while shifting risk onto the balance sheet: about USD99bn of stakes, USD25bn of further commitments and USD108.5bn of maximum guarantees.
  • Risk tolerance. The founder's concentrated stake is aligned with owners, though his wealth makes him tolerant of large bets.
  • Time horizon. One-year and three-year PSUs are shorter than this five-to-ten-year horizon.
  • Buybacks. They do not merely offset stock compensation. USD79.9bn was repurchased in FY2026–H1 FY2027, against USD10.3bn of stock compensation expense, and diluted shares fell from 24,804m to 24,285m.

How the valuation treats agency costs. No separate discount. The model charges GAAP stock compensation, credits no buyback accretion, recovers only 40%/80%/100% of equity stakes, and pays a 25% loss on contingent exposure in the downside.

The dated review records the filed-suit notice and sources distinguishing allegations, requests and findings. The complaint/docket was not directly reviewed. No new restriction, leadership decision or quantified company loss is established. Keep the existing operating paths and explicit unquantified tail; reassess on discriminating cash/contract or governance evidence.

12 November succession assessment

The dated review reassesses the complete distribution. Founder-led architecture, talent coordination, supplier commitments and customer financing may be harder to coordinate during an interim period. Long-tenured operating leaders, installed CUDA and existing funded orders remain assets; a departure does not erase them. Annual cash paths and terminal multiples stay unchanged. A small, low-confidence transfer of five percentage points from upside to downside makes both horizons 45%/45%/10%. This is a judgment about joint execution outcomes, not a measured probability of misconduct or a cash-loss estimate.

Capital allocation and cash definitions

This section was revised on 3 Nov 2026. Money is in USD millions, and fiscal years end in late January. The annual figures come from each 10-K's statements, extracted through nvda-xbrl-companyfacts. Each measurement is a named step in valuation/calculations.json.

Sources and uses of cash

FYOperating cashCapexAcquisitions and GroqEquity stakes, netBuybacksRSU tax withholdingDividendsStock-plan proceedsDebt, netOther investingUnreconciledChange in cash and securities
20229,108−976−26300−1,904−3992813,977−24−1539,647
20235,641−1,833−49−77−10,039−1,475−39835500−37−7,912
202428,090−1,069−83−861−9,533−2,783−395403−1,250−12429312,688
202564,089−3,236−1,007−1,315−33,706−6,930−834490−1,2502290317,226
2026102,718−6,042−14,535−17,418−40,086−7,948−974644002,98719,346
Total209,646−13,156−15,937−19,671−93,364−21,040−3,0002,1731,477−1263,99350,995

Cash and marketable securities rose from 11,561 to 62,556, which matches the total change (cashSecuritiesYeFY2021, cashSecuritiesYeFY2026).

The unreconciled 3,993 is about 1.9% of operating cash. It consists mainly of changes in securities marks and FX, PP&E principal payments (101 in FY2026) and other small financing lines. The FY2026 residual is larger and has not been traced. I record it rather than plug it.

Inside operating cash:

  • Stock compensation added back: 19,385 over five years.
  • Working capital absorbed: 30,317 (balance-sheet basis).
  • Operating cash before working capital and stock compensation: about 220,600.

Priorities revealed. About half of five-year operating cash (114,404) went back to shareholders as buybacks plus RSU tax withholding, which is economically a repurchase. Dividends were trivial until the FY2027 increase to USD0.25 a quarter.

The pattern changed sharply in FY2026 and H1 FY2027, with the same team in charge (Huang as CEO, Kress as CFO):

  • Acquisitions and licenses rose to 14.5bn in FY2026, mostly Groq.
  • Ecosystem stakes rose from 1.3bn to 17.4bn in FY2026, then 42.4bn in H1 FY2027 alone, about 57% of that half's 74.4bn operating cash.
  • NVIDIA issued 25bn of debt in June 2026.

Investment rates per sales change

Over the five years, sales rose 199,263, from 16,675 in FY2021 to 215,938 in FY2026 (salesChangeFy21Fy26).

RateAmountPer unit of ΔsalesStep
Incremental M&A (acquisitions + Groq; no material divestitures)15,9378.0%acquisitionRate
Internal intangibles, all R&D52,69326.4%researchRate
Internal intangibles, R&D step-up only14,5737.3%researchGrowthRate
Fixed capital (capex 13,156 − D&A 8,933)4,2232.1%fixedCapitalRate
Working capital30,31715.2%workingCapitalRate
Ecosystem equity stakes (financial, shown separately)19,93510.0%equityStakeRate

NVIDIA invests through expensed R&D and working capital, not fixed assets. Each dollar of new sales needed about 2 cents of net fixed capital and 15 cents of working capital.

The inflections:

  • Working capital as the main use. It turned from a small use into the main one after FY2024. DSO rose from 51 to 60 days and inventory reached 31.6bn at July 2026 (dsoQ2).
  • Ecosystem stakes. They now exceed every other operating use.
  • Capex. It rose to about 2.5% of sales (TTM 7,474 against 3,687 of D&A) as NVIDIA builds its own AI and R&D infrastructure. Contracted capex is 8bn.

Reinvestment and returns

MeasureValueStep or basis
Owner earning power, FY2021 GAAP operating NOPAT4,453operatingNopatFY2021
Owner earning power, FY2026110,676operatingNopatFY2026
Owner earning power, TTM at 17%163,991
Five-year reinvestment rate22.5%Growth investment 50,477 (net capex + working capital + acquisitions) / matching NOPAT 224,603 (reinvestmentRateFy22Fy26). Including equity stakes, 31%.
Lagged five-year ROIIC on tangible capital472%ΔNOPAT 106,223 / Δcapital FY2021→FY2025 of 22,521 (roiicLaggedFy26)
Rolling three-year ROIIC from the FY2023 trough530%ΔNOPAT 106,263 / Δcapital FY2022→FY2025 of 20,052 (roiicLaggedFy23Fy26)
Operating leverage63.2%ΔOperating income / Δsales, FY2021→FY2026 (operatingLeverageFy21Fy26). Operating margin went from 27% to 60%.

The identity, growth ≈ reinvestment rate × ROIIC, does not hold here, and the reason is informative. NOPAT grew at a 90% compound rate. 22.5% × 472% gives 106%, roughly the right order. But the ROIIC is not a capital-productivity measure: tangible capital did not create the earnings. Expensed R&D did, along with price and mix on a product in shortage and operating leverage on largely fixed R&D.

Treat all five years of R&D as unamortized capital, an upper bound on the capital base. After tax, that adds 43,735 of investment, which lifts the reinvestment rate to about 40% and cuts the incremental return to about 140% (106,223 / (22,521 + 52,693)). Those are still extraordinary, but they are a rent on scarcity and integration, not a reinvestment engine that compounds at a stable rate.

So the forecast does not use a Harvia-style reinvestment × ROIIC grower. It funds growth capital explicitly and judges margins separately.

The mix by use:

  • Organic tangible investment (net capex and working capital): 34,540, earning the returns above.
  • Acquisitions: 15,937. Their earnings cannot be measured separately (see below).
  • Equity stakes: 19,935 over five years plus 42,404 in H1 FY2027. Their return so far is unrealized marks: 23.7bn of H1 FY2027 gains, and July carrying value of about 99bn against about 62bn of purchases since FY2023. Marks are not owner cash.
  • Payouts: 117,404 including RSU withholding and dividends.

The forward diagnostics are generated from the shared engine in research/10-capital-efficiency.md.

Acquisition record

DealDatePriceWhat was boughtResult
MellanoxCompleted FY2021Within 8,524 of FY2021 acquisition cashInfiniBand and Ethernet networkingNetworking revenue 31,376 in FY2026, more than 10× FY2021 (dataCenterNetworkingFY2026). Profit not disclosed separately. Strategically the most successful deal: it raised content per gigawatt.
ArmSigned Sep 2020, terminated Feb 20221.25bn prepayment forfeited; 1.35bn chargeWould have been the CPU architecture ownerRegulators blocked it. NVIDIA kept a 20-year license (nvda-arm-termination).
Small dealsFY2022–FY20251,402 in totalSoftware and teamsNot separately measurable
Groq (non-exclusive license and related consideration)Dec 2025 (FY2026)13,000 paid in FY2026 + 3,921 accrued, about 16.9bnInference IP and staffGoodwill rose from 5,188 to 20,832. Management credits Groq with raising revenue per gigawatt. No separable profit is disclosed.
Other FY2026 acquisitionsFY20261,535Not itemized in the excerpts readNot measurable
Hugging Face (pending)Agreed 2 Sep 2026; close expected H1 2027About 11,900 + up to 1,000 equity retentionOpen AI-model platformCommits to supporting other silicon vendors. No revenue disclosed (nvda-hugging-face-8k).

In aggregate, about 30bn of 2025–2026 deal and license spend buys ecosystem position and talent, not disclosed profit streams. A deal yield cannot be measured with matched definitions. Nothing indicates NVIDIA is a serial acquirer whose value comes from deal arithmetic.

The underwriting treats Groq's remaining 986 and Hugging Face's 11,900 as year-1 cash claims with no separately credited earnings. Any benefit has to show up through the scenario margins.

Current cash definitions and funding bridge

These definitions feed the forecast.

Trailing free cash flow. Reported trailing operating cash is 134,360. Subtracting 7,354 of capital purchases and 120 of related principal reproduces the issuer's 126,886 of free cash flow. That measure adds back 7,241 of stock compensation. A separately labeled owner-cash proxy deducts it, giving 119,645. This proxy and the forward operating-profit model are different measures, and the forecast does not deduct stock compensation twice.

Starting point. Re-underwritten 3 Nov 2026. The baseline is the trailing four quarters to the 27 October anchor: reported Q4 FY2026, Q1 and Q2 FY2027 plus the Q3 guidance midpoint (108bn of sales, 74% gross margin, 9.2bn of opex). That gives 353,964 of revenue and 232,289 of GAAP operating income, a 65.6% margin.

The earlier vintage annualized the Q3 guide to 432bn and then grew it 50%, counting the near-term ramp twice. Operating profit still excludes the large unrealized equity gains that boosted net income (nvda-q2-fy2027-release, nvda-q4-fy2026-release).

Opening liquidity. July cash plus debt securities was 56,586. The pro forma bridge reserves:

  • 6,048 for the October dividend.
  • 1,500 for the SB Energy equity stake.
  • 1,100 for the KYEC advance.

That leaves 47,938 of liquid funds against 33,500 of debt principal. No working-capital catch-up is reserved, because working capital starts at the reported July balance (75,002) and grows with forecast revenue. Year 1 also pays the Groq balance (986) and the Hugging Face consideration (11,900). Unreported later cash generation, buybacks and new financing are not invented (nvda-supply-oct2026, nvda-hugging-face-8k).

Equity commitments. The full 25bn of reported conditional equity commitments is reserved in year 1, an accelerated funding assumption. The cost paid becomes a separately retained investment, subject to the scenario recovery haircut. Locked public equity, private equity and equity-method interests are not cash available to pay operating bills.

Supply and capacity commitments. These are 279bn and mostly buy inputs for future sales, which margins, capex and working capital must fund. Subtracting every gross contract again as debt would count ordinary costs twice. Unused partner capacity and guarantees are different exposures and are tested in the downside as a cash loss.

Terminal value. The terminal denominator is unlevered GAAP NOPAT. The valuation deducts net debt, adds separately retained investments and deducts other claims once. Issued debt keeps paying its estimated financing cost in annual cash flow. Dividends stay at USD1 per share a year; dividends already paid are excluded.

Post-July financing bridge —13 November2026

The full2bn September Nscale subscription now has a separate first-year reserve, sourced from the counterparty's amended registration rather than an issuer headline. July reported cash and the25bn pool are not silently rolled to November. Funded cost joins retained financial assets once; no operating earnings or appreciation is credited. Exact overlap0–2bn remains a sensitivity to reconcile at17November accounts, when actual cash, holdings and remaining commitments must be rolled together. Nscale's preliminary losses are not measured mature project returns and its debt is not NVIDIA debt.

Investment thesis

Current assessment,15 November2026: atUSD162.85, five-year base IRR is4.87%, weighted−0.71%; ten-year weighted3.72%. Operating paths, fixed exits and45%/45%/10% weights remain unchanged. The post-July Nscale reserve is paid once and retains only the existing case recovery. Customer funding and security/legal outcomes remain unresolved. Current evidence; 13 November funding revision and fixed-price effect; prior probability revision. Historical tables below retain their original quote and weights.

NVIDIA is an exceptional business at a cyclical peak. Its unit economics are extraordinary: about 190–225% after-tax return on tangible capital (research/01-business.md). The risk is not capital intensity. It is three things:

  • The price per unit of compute once the largest buyers' own chips mature.
  • A digestion once the buildout must be funded from customers' own cash rather than debt, equity and NVIDIA's balance sheet.
  • Margins that sit far above NVIDIA's own history.

The 3 November re-underwrite starts from a guidance-inclusive four-quarter estimate to the 27 October anchor (three reported quarters plus unreported Q3 guidance):

  • Revenue: 354bn, using reported Q4 FY2026, Q1 and Q2 FY2027 plus the Q3 guidance midpoint.
  • GAAP operating income: 232bn, a 65.6% margin.

This replaces the annualized Q3 run rate (432bn), to which the earlier vintage then added another 50% of growth. A sourced demand ceiling, built from customers' cash and financing, bounds every scenario. Mechanisms, the ceiling and the outside view are in research/08-underwriting-plan.md.

InputDownsideBaseUpside
Year-1 growth against trailing 354bn−15%+28%+58%
Annual growth from year 23%5%8%
GAAP operating margin, reached in year 330%50%58%
Capital spending / sales6%4%3%
Operating capital / sales30%22%18%
Normalized tax rate23%20%17%
Exit enterprise value / unlevered NOPAT13x18x22x
Retained investment recovery40%80%100%
Loss stress on partner capacity and guarantees25%0%0%

By year 5 these produce:

DownsideBaseUpside
Revenue (demand ceiling 844bn)339bn551bn761bn
NOPAT (trailing about 193bn)78bn220bn366bn

At the USD217.94 quote of 27 October and the fixed 27 October 2031 exit:

DownsideBaseUpsideWeighted
5-year XIRR−24.5%−1.2%12.9%−4.4% (40/45/15)
10-year XIRR−10.5%3.2%11.5%1.8% (40/45/15)

The weighted figure is the IRR of probability-weighted dated cash flows. The before/after comparison is in updates/2026-11-03-reunderwrite.md. These paths are not confidence intervals.

Pre-mortem

  • One-foot or three-foot hurdle? Three-foot. The product economics are understandable. The value depends on post-2027 AI capital spending, custom-silicon substitution and AI-lab credit, none of which a filing can settle.
  • Cyclical or secular? Both. AI compute is a secular need, but NVIDIA's revenue is a capex cycle. NVIDIA had downturns in FY2009 (revenue −16%), FY2020 (−7%, operating income −25%) and FY2023 (operating income −58%). Today the customers' capex exceeds their own cash (research/02-history.md).
  • How could it be a value trap? By valuing shortage-era earnings as permanent. Those earnings could fall at the same moment the ecosystem stakes and guarantees lose value.
  • Cigar butt or productive asset? A productive, appreciating asset: R&D, CUDA and supply relationships, with about 79% of merchant and custom AI-accelerator revenue.
  • Agency cost at the original underwriting date? Pay was small beside the founder's then ~USD190bn stake. That historical alignment no longer describes the interim CEO's authority or incentives. Material in control of resources: revenue-weighted incentives alongside about USD99bn of stakes, USD25bn of commitments and USD108.5bn of guarantees (research/03-management.md).
  • Good or bad execution? Assume good product execution. Assume average-to-weak capital allocation: buybacks regardless of price, and lab investments justified by affordability rather than return.
  • Turnaround? No.
  • Big promises with little record? Partly. The 70% FY2028 outlook exceeds what the demand ceiling allows. Financed demand of about a quarter of the business has no record through a downturn.
  • IP a larger competitor could squash? No competitor is larger in AI compute. But each of the five largest customers is larger as a company, and Google has a complete alternative stack.
  • Single product or fad? Concentrated: about 90% of profit is one product cycle. Past network buildouts ended in digestions (Cisco −15%, Applied Materials −31% after 2001).

Exit multiple

Own history. A verified historical price and multiple series has not been reconstructed, and none is invented. Today:

  • Market value is about USD5.29tn (24,285m diluted shares at USD217.94).
  • Net of about 23bn of net liquidity and about 99bn of ecosystem stakes, enterprise value is about 5.17tn.
  • That is about 27x the trailing four-quarter NOPAT of 192.8bn. The estimated trailing figure includes the unreported Q3 guidance midpoint; it is not reported earnings.

Peers. Peer multiples are not reconstructed; that is left open rather than guessed.

Koller check. Value = NOPAT × (1 − g/RONIC)/(r − g). The required return r is 10% because the earnings are cyclical and concentrated. No case uses a peak multiple on peak earnings.

CaseInputsImplied multipleApplied to
Downsideg 3%, RONIC 40%13xTrough-normalized earnings (30% margin). At zero growth, 1/r would give 10x.
Baseg 5%, RONIC 50%18xNormalized 50%-margin earnings
UpsideLong-run g 6%, RONIC 60%22.5x, capped at 22xEarnings that stay near peak margin

Sensitivities, for information:

  • Base at 22x: 2.1%.
  • Base with a 55% margin: 0.7%.
  • Downside at 11x: −26.4%.

Bear case

The downside is "the price we could sell at after five years if we weren't forced sellers". Its mechanism:

  1. Customers' AI capex already outruns their cash (research/02-history.md). After the 2025–2026 build, financing tightens and buyers absorb capacity.
  2. Revenue falls 15% below the trailing four quarters in year 1, about 30% below the Q3 run rate. That is the base-rate range for capex-cycle suppliers: Cisco −15%, Applied Materials −31%, Micron −49%, NVIDIA FY2009 −16%.
  3. Revenue then recovers at 3% a year.
  4. Hyperscalers move internal inference to their own chips. Financed labs stretch payments. NVIDIA defends volume with price.
  5. Operating margin reverts within three years to 30%, NVIDIA's pre-boom average.
  6. Working capital rises to 30% of sales.
  7. Equity stakes recover 40% of carrying value.
  8. 25% of the 164.5bn of guarantees and partner capacity is paid in year 3.
  9. The exit is 13x trough-normalized NOPAT.

Five-year XIRR: −24.5%.

Tails it omits:

  • A Micron-scale −49% collapse. A −30% first year gives −27.4% (information only).
  • Equity stakes going to zero.
  • Guarantee calls above 25%.
  • A Taiwan supply disruption.
  • Revenue failing to recover at all.

Why the opportunity exists

Evidence on attention.

  • Size: NVIDIA is among the largest listed companies (about USD5.3tn) and a member of the S&P 500.
  • Ownership: BlackRock (7.43%) and Vanguard (7.31%) are the largest holders; insiders own 3.94% (nvda-proxy-2026).
  • Coverage: essentially every large broker covers it.
  • Not measured: daily trading value, short interest and retail ownership were not retrieved.

Neglect is impossible.

Recent events that drove sellers.

  • The August gross-margin reset.
  • The debate about "circular financing".
  • Custom-silicon announcements.

These are real risks, not noise.

Variant view at the original USD217.94 quote

No mispricing in our favour; if anything, the price assumes more than customers can fund. At USD217.94:

  • The base earns about −1% a year for five years.
  • The probability-weighted return is about −4%.
  • A 12% base return would need a price near USD118, and a 12% weighted return about USD100.

The market appears to price NVIDIA's current margin and growth as durable, or to expect customers' AI spending to keep exceeding their own cash for years. This research concludes otherwise:

  • Customers' cash plus observed financing caps year-1 revenue near 576bn (stage 8).
  • Past capex-cycle suppliers, NVIDIA included, have had digestions (outside view).
  • NVIDIA's own pre-boom margins were 27% operating and 62% gross (stage 7).

At that historical quote the business quality did not provide a sufficient margin of safety. The current ranked comparison and probabilities are in the 12 November review; this historical comparison is not a current quote or return.

Catalyst. None hard. The Q3 FY2027 report on 17 November 2026 is an evidence date.

What would prove us wrong

ObservationRefutesWhere and when
Hyperscaler free cash flow turning positive while capex keeps rising more than 30% a yearThe demand ceiling, which would be too lowAlphabet, Amazon, Meta and Microsoft quarterly releases from Q4 2026
FY2028 revenue above about 576bn with operating margin above 60%The base's customer-funded bound; would support the upsideFY2028 results, about February 2028
DSO back below 50 days with revenue still growing over 50%The financing-dependence concernQ3 FY2027 (17 Nov 2026) and later CFO commentary
Gross margin above 70% through FY2029 despite custom-silicon growthThe base's margin normalizationQuarterly releases
Hyperscale share of Data Center falling below 45%, external TPU and Trainium sales in the tens of billions, or guarantee calls and stake impairmentsConfirms the downside rather than refuting itAlphabet and Amazon 2027 filings; NVIDIA 10-Q notes 5–10

valuation/prospective-checks.json was frozen on 3 November against the earlier model hash (f5ae5c50024e). It stays unedited, so its predictions also score the superseded vintage. Permanent-loss paths and unverified inputs are in research/06-risks.md.

Judged view — 5 November 2026

Low in precise forward returns; platform advantage supported, customer-paid demand and durable margins unverified. Retain existing coarse horizon-specific weights after reassessment: scenarios already represent these adverse/favorable operating mechanisms. Numerical odds are subjective, not verified frequencies; unresolved gaps prevent high conviction. No tuning to return.

The current dated scenario returns and the evidence for each changed or retained input are in the independent bull–bear judgment. This review supersedes earlier undated/current return statements above; dated historical/source examples remain calibration. 17 November Q3 actuals replace guidance-inclusive starting estimate; reconcile customer-funded purchases, RPO/cash, receivables, guarantees and utilization.

Risks and disconfirming evidence

15 November update. Original class-action complaint pp.1–3/17 and derivative complaint pp.1–3/33–34 were inspected: the former seeks damages against NVIDIA and Huang; the latter seeks recovery for NVIDIA and governance changes. Allegations and requested remedies are not findings or a quantified net company loss. The Insurer's14November report describes executive Side A cover; the actual policies were not available and corporate reimbursement is not presumed. Megaspeed's emailed statement is reported by newswires, not independently verified clearance. Current odds remain45/45/10. Evidence and limits. Dated entries below preserve earlier assessments.

Demand and customer funding. Large infrastructure orders can precede profitable use. Three direct customers represented 16%, 15% and 13% of H1 sales; direct and end customers are not identical. H1 receivables absorbed USD24.590bn and inventory USD10.204bn of cash. Extended payment terms and financially weaker clouds make conversion and credit quality central to valuation. Growth in token usage is insufficient without paid, collectible demand.

Platform and competitive risk. CUDA and system integration are advantages, but AMD, custom accelerators and software portability can reduce pricing or share, particularly for repeatable inference workloads. AMD's August release describes Helios/MI400, ROCm and named cloud/customer partnerships. Those are competitive claims, not independent proof of superior economics. AMD Q2 issuer results.

Margins and supply. Memory, packaging, networking and faster product transitions create cost, quality and inventory risk. Supplier advances and USD279bn supply/capacity commitments are real economic capital despite external manufacturing. Export restrictions already caused product charges; the Q3 outlook excludes China data-center compute. The scenarios do not assume restoration of a full China market.

Ecosystem correlation. July equity marks were about USD99bn. Supply and deployment are increasingly supported by investments and guarantees. Maximum phased guarantees of USD108.5bn and USD56bn of partner cloud/third-party lease arrangements are not immediate debt, but can become cash losses. The downside pays a chosen USD41.125bn stress in year 3. This is not an estimated default probability or full worst-case exposure. Optional further support is not exercised.

Customer/supplier checks. Microsoft's July discussion reported substantial CPU/GPU capex alongside growing cloud cash flow. TSMC's October report described strong leading-edge demand and further capacity spending. These sources corroborate current activity while also showing the industry's funding needs. They do not prove NVIDIA's future share, margins or return on customer investment. Microsoft, TSMC.

Coverage does not establish actual utilization or profitability for every model maker or cloud. Private contracts, later cash movements and unquantified supply advances remain unresolved. Reported investment marks may also be difficult to realize, particularly when operating demand weakens at the same time.

Permanent-loss paths (reviewed 3 Nov 2026)

  1. Substitution at the largest buyers. Hyperscalers are about 55% of Data Center revenue and each builds its own accelerators. Google now sells TPU systems externally to nearly 100 customers, Amazon will sell Trainium outside AWS from late 2027, and Broadcom's AI revenue grew 221% (googl-q3-2026, amzn-q3-2026, avgo-q3-fy2026). If internal inference moves off NVIDIA, price per gigawatt falls, not just volume. That is the downside's permanent-margin mechanism.
  2. Financed demand. AI labs that NVIDIA supports with equity and credit are about a quarter of next year's business (aiLabFinancedShare). If their revenue or funding disappoints, NVIDIA loses sales, receivables (DSO already 60 days), equity value (about 99bn carrying) and guarantee exposure (108.5bn maximum) at once. These losses are correlated, not diversified.
  3. Customer cash, not supply, as the binding constraint. Alphabet, Amazon, Meta and Oracle had negative free cash flow in their latest quarters while capex roughly doubled. Continued growth needs open debt and equity markets.
  4. Supply commitments in a downturn. 279bn of commitments, mostly memory and concentrated in the next three fiscal years, become onerous if demand pauses. In FY2026, H20 provisions alone were 4.5bn.
  5. Regulation. Export rules removed China data-center compute. A Taiwan supply disruption is an omitted tail.
  6. Capital allocation. Buybacks of about USD20bn a quarter regardless of price, plus deal and license spend of about 30bn (Groq, Hugging Face) with no measurable profit.

Unverified inputs

  • Profit split by economic unit (hyperscale against ACIE). Estimated from revenue; NVIDIA does not disclose it.
  • Customer-side gigawatt economics. USD60bn per GW is a management figure. Power cost and economic life are this repository's assumptions. No neocloud or lab P&L has been reconciled.
  • Oracle's FY2027 capex guide and Broadcom's FY2027/FY2028 AI-revenue and gigawatt remarks. These come from third-party call summaries and automated transcripts.
  • The FY2026 sources-and-uses residual (2,987) is not traced.
  • NVIDIA's own historical valuation multiples, peer multiples, trading value and short interest. Not retrieved.
  • The terms of the lab stakes, including whether NVIDIA's equity carries preferences, and the exact funding dates of the 25bn of commitments.
  • No fieldwork. No interviews or channel checks were done.

Surviving debate attacks — 5 November 2026

Low in precise forward returns; platform advantage supported, customer-paid demand and durable margins unverified. Reconstructed the independent isolated stress at the same quote/date/horizons; do not treat stress size as a probability or a measured current fact.

Discriminating evidence: 17 November Q3 actuals replace guidance-inclusive starting estimate; reconcile customer-funded purchases, RPO/cash, receivables, guarantees and utilization. The complete debate sizes these risks using the shared forecast; a sensitivity is neither a measured fact nor a calibrated probability.

External scrutiny,6November2026

The House committee chairs’5November release confirms a request for Jensen Huang to discuss Megaspeed contacts and information-security controls. Reply requested11November; briefing requested18November. Attendance and the underlying allegations remain unconfirmed in the checked official sources. External official pages and dates are in researchScope.externalDisclosureWatch.

8 November: issuer inquiry confirmed. The board's full statement, paragraphs1-5, establishes a new independent-director/counsel investigation. It also states continued confidence in management and no present reason to believe national security or company secrets were compromised. The CEO's response commits to cooperation. An inquiry is a material governance development, not a finding or an all-clear.

Potential cash mechanisms are leadership disruption, compromised intellectual property, customer/partner loss, wider export restrictions and legal/remediation costs. Neither statement quantifies those effects or announces a new restriction or management replacement. Keep operating inputs and both horizons'40/45/15 weights after explicit reassessment; do not invent a loss amount or a precise additional event probability. China data-centre compute remains excluded, but indirect routing, non-China restrictions and information-security damage are separate risks. The three operating paths do not bound this tail. Revisit the whole distribution when findings, actual contract/cash changes or a leadership decision supply discriminating evidence. Dated assessment.

9 November follow-up

The dated review records the filed-suit notice and sources distinguishing allegations, requests and findings. The complaint/docket was not directly reviewed. No new restriction, leadership decision or quantified company loss is established. Keep the existing operating paths and explicit unquantified tail; reassess on discriminating cash/contract or governance evidence.

11 November evening follow-up

Evening evidence and assessment distinguish reported conditional congressional escalation from an issued demand or investigative finding. Retain the unquantified governance tail and revise the distribution when evidence changes cash, contracts, leadership or restrictions.

11 November late follow-up

Reuters reports that the board met late on 11 November to decide whether to remove the CEO. It attributes the preparations to the WSJ, a second relationship to the NYT, and a replacement search to Bloomberg. Nothing is confirmed by the issuer. Leadership continuity is now an active, near-term risk rather than a remote tail. Its cash channels are execution cadence, talent retention, customer/government relationships and remediation cost; none is yet observed. Dated assessment.

12 November confirmed transition

The 8-K, Item 5.02 resolves the prior uncertainty about a leadership decision. It does not resolve the investigation, securities allegations or operating consequences. Interim continuity and permanent succession remain distinct questions. Both horizons now use 45%/45%/10%; annual drivers and 13x/18x/22x exits remain unchanged to avoid a second arbitrary charge for the same uncertainty. The review records potential coordination, talent, product and customer mechanisms plus unquantified legal/security tails. Reported acceptance of a congressional briefing is not confirmation of the originally requested 18 November date.

13 November customer-funding and governance evidence

Nscale's actual financing needs must be separated from NVIDIA's July pool and from the operating demand ceiling. The extra2bn reserve is paid once and receives only the existing scenario asset recovery. Whether an earlier commitment overlaps is a0–2bn cash-reserve range, not an invented probability. Project funding for future service agreements remains uncommitted; loss-making expansion can create value, but its funding and mature project return need proof.

The Singapore MHA statement and US Senate Finance preservation letter are primary process evidence, not findings of spying, a new export ban, damages or issuer clearance. House acceptance/subpoena status was unconfirmed before its8pm EST13November deadline. Reported Senate December3/10 requests and potential CEO candidates remain unconfirmed in original committee/issuer notices. Recheck those originals; keep requests distinct from issued process and permanent appointments.

Assumption register

Reported facts retain periods, units and source IDs in financials/facts.json. valuation/calculations.json connects those facts and explicit judgments to the model. The arithmetic is shared; the choice of definitions and inputs follows NVIDIA's business. There is no supplied external NVDA investor forecast to reproduce.

9 November review. The dated assessment adds a filed shareholder suit and unpriced investment announcements to the source record. No new financial report, quantified loss, binding restriction or operating interruption sizes a numerical revision. All financial inputs and both horizons’40/45/15 weights are retained with explicit tail limits; assessment evidence/date change. Unspecified investment amounts are neither free nor added twice to the existing reserve.

Input or definitionEvidence and calculationJudgment and change trigger
Trailing revenueFY2026 + H1 FY2027 - H1 FY2026 = USD302,970m.Preserve fiscal dates. Do not relabel a forecast as reported trailing sales.
Reported FCF and owner cashOperating cash 134,360 - capital cash 7,474 = 126,886; less recurring SBC 7,241 = 119,645.Owner-cash proxy differs from issuer FCF. Forward GAAP profit already charges SBC, so no duplicate subtraction.
Starting point (re-underwritten 3 Nov 2026)Trailing four quarters to the 27 Oct anchor: TTM revenue 302,970 − Q3 FY2026 57,006 + Q3 FY2027 guide 108,000 = 353,964. Operating income 197,579 − 36,010 + (108,000 × 74% − 9,200) = 232,289 (revenueTrailingToAnchor, operatingProfitTrailingToAnchor).Before: annualized Q3 guide 432,000 / 282,880. Year-1 growth was then added to an already-annualized peak quarter, counting near-term growth twice. The trailing window ends at the anchor, so year-1 growth applies once. Revise when Q3 actuals replace the guide (17 Nov 2026).
Revenue pathBottom-up year 1: hyperscalers 20% of a 1,065 cash-plus-financing budget (213); ACIE about 6 GW × 35 (210); Edge 30. Total 453 = +28%. Downside sized from capex-cycle declines (Cisco −15%, Applied Materials −31%, Micron −49%, NVIDIA FY2009 −16%).Year 1 −15% / +28% / +58%, then 3% / 5% / 8% (growthFadeYears 2). Before: +10/+50/+70% fading to 0/5/10% over three years from 432,000. Every year must stay under the demand ceiling; the upside year 1 (559) sits just under 576. Management’s +70% FY2028 (about 690) exceeds the ceiling and is not used.
Operating marginPre-boom FY2016–FY2022 operating margin averaged about 27% (range 15–37%). Gross margin averaged 62.1% over FY2018–FY2022. Opex is now about 9–12% of sales against 25–35% pre-boom. Management guides 72–73% gross margin for FY2028.Normalize within three years (marginTransitionYears 3) to 30% / 50% / 58%. Before: 35/55/65% over five years. Base = pre-boom gross margin minus opex at today’s scale. Downside = pre-boom average. Revise on sustained gross margin above 70% or opex intensity shifts.
Operating capitalReported July proxy 75,002 (core receivables, inventory and prepaids less operating accruals, plus residual noncurrent operating assets less long deferred revenue).Starts at the reported balance (before: scaled to the 432,000 run rate with a 9,181 catch-up reserved from cash). Then 30% / 22% / 18% of revenue.
Capex and D&ATTM D&A / sales about 1.22%; historical capital spending about 2.47% of sales. Contracted capex is USD8bn.Spend 6/4/3% of sales on infrastructure and recurring PP&E/intangible capital, above historical intensity. Do not equate accounting D&A with maintenance or subtract the same contracted spending twice.
Taxes and financingFY tax guide 16–18%; June debt issuance and contractual rates are disclosed.Start at 17%, transition to 23/20/17%; estimate 4.5% continuing debt cost, refinancing and no new net borrowing. Carry the historical noncurrent tax liability separately.
Cash availableJuly cash and debt securities 56,586 less October dividend 6,048, SB 1,500 and KYEC 1,100 = 47,938 (proformaOpeningCashTrailing).No catch-up reserve with a trailing start. Not an observed October balance; later cash generation and buybacks are not invented.
Supply and ordinary operationsSupply/capacity 279,000; own R&D cloud 29,000 and uncommenced own leases 25,000.Fund product inputs through margins and capital, not by subtracting all contracts as extra debt. Unused capacity/cancellations can still require revision.
Equity commitmentsReported conditional equity commitments 25,000.Reserve all in year 1 and add funded cost to retained assets. Conditional does not mean already paid. Future follow-on investments are not automatically free.
Post-July Nscale subscription (13 November review)Counterparty S-1:September15 NVIDIA subscription1,000 plus further1,000 expected around16November.Reserve2,000 once in year1 outside theJuly26 pool; funded cost enters the existing40/80/100% retained-value rule. No operating earnings or note income. Pool overlap0–2,000 remains a sensitivity; reconcile current cash, held securities and remaining commitment together at17November accounts.
Anthropic subscription overlap (15 November review)November10 S-1 pp.23/347/F-67: conditional IPO subscription10,000. The November18,2025 original partnership already announced up to10,000 before the July pool.Provisionally treat this as fulfillment within the25,000 pool; composition is unitemized, so inclusion is a judgment. Extra0–10,000 funding sensitivity changes5y weighted IRR by0 to−0.01922pp. Do not add a second liability merely from a later agreement. Reconcile cash/holdings/remaining commitments at17November; no IPO proceeds or valuation gain assumed.
Retained securitiesPublic equity including locked noncurrent holdings plus nonmarketable/equity-method securities = 98,897, plus the separate SB reserve.Recover 40/80/100% of marks and funded cost; no appreciation or income. This is a valuation stress, not a known private investment exit price. Disposal taxes/restrictions can reduce value.
Contingent partner exposureGuarantees 108,500 plus partner cloud/third-party leases 56,000.Base/upside no payment; downside chosen 25% loss paid once in year 3. No claim of zero risk or common default probability.
Shares and distributionsLatest quarter weighted-average diluted shares 24,285m; declared dividend USD0.25 quarterly.Hold net shares flat, leave economic SBC cost charged and give no net buyback accretion. USD1 annual dividend on approximate future dates. Actual grants/settlements/repurchases require later reconciliation.
Terminal valuationKoller (1 − g/RONIC)/(r − g) at r 10%: downside g 3%, RONIC 40% → 13x; base g 5%, RONIC 50% → 18x; upside g 6%, RONIC 60% → 22.5x, capped at 22x.13 / 18 / 22x on unlevered GAAP NOPAT (before: 15 / 22 / 28x). No peak multiple on peak earnings. Net cash and retained investments enter once.
Hugging Face consideration (added 3 Nov 2026)8-K of 3 Sep 2026: about 11,900 to stockholders plus up to 1,000 equity retention (huggingFaceConsideration).Paid in year 1 as a cash claim, alongside Groq's remaining 986. No acquired earnings are credited because no revenue is disclosed. The retention equity is future stock compensation inside GAAP opex. Revise on close, a price adjustment or a failed approval. The 27 Oct vintage omitted this.
Cycle and windfall normalizationQ2 FY2027 margins are 75% gross and 66% operating, against FY2022's 64.9% and 37.3% and FY2023's 56.9% and 15.7%. Management guides gross margin to 71–72% in Q4 and 72–73% in FY2028 on memory costs.The opening run rate is a shortage-era margin. Operating margins of 30%, 50% and 58%, reached in year 3, are the through-cycle judgment (revised 3 Nov from 35/55/65% over five years). The downside's reversal is charged as permanent substitution. Do not raise the base margin on one strong quarter.
Historical capital-allocation measurementsFive-year rates per Δsales: M&A 8.0%, R&D 26.4% (7.3% step-up only), fixed capital 2.1%, working capital 15.2%, equity stakes 10.0%. Reinvestment 22.5%; lagged ROIIC 472% (calculations.json steps).Diagnostics only. They show that tangible capital does not drive the earnings, so no ROIIC-based grower is used. They are not bound to model inputs.
Supply prepayments and extended termsOther noncurrent operating assets about 9,965; DSO 60 days; extended terms with investment-grade customers.Inside the operating-capital proxy (22% of sales in the base, 30% in the downside). A move back to about 50 days DSO would release about 12bn at the run rate. Not credited.
Demand ceiling (added 3 Nov 2026)Five hyperscalers’ operating cash 793,016 (annualized latest quarter; Amazon TTM) + observed financing 272,408 (Amazon debt + Oracle equity, × 4) = 1,065,424. × 25% maximum share (observed 21.0%) = 266,356; + ACIE 7 GW × 40,000 = 280,000; + Edge 30,000. Total 576,356, growing 10% a year (demandCeilingYear1, ceilingGrowth).Bound to model.demandCeiling; the engine rejects any scenario above it. Revise when hyperscaler cash flow or financing pace changes materially, or when NVIDIA’s observed share of their capex moves above 25%.
Outside viewNVIDIA downturns FY2009, FY2020 and FY2023; Cisco FY2002, Applied Materials FY2002, Micron FY2023 (primary filings).Five-year revenue CAGRs −0.9% / +9.2% / +16.6% from the trailing 354bn. Generic large-company base rates are pending (no sourced copy of the Base Rate Book).

The five-year exit is 27 October 2031. The 3 November re-underwrite is recorded in updates/2026-11-03-reunderwrite.md. Every changed estimate requires a dated old/new value, evidence, economic explanation and return effect isolated from stock-price/calendar movement. Scenarios are not probabilities. The decade extension does not constitute ten years of product-cycle research.

11 November evening: governance evidence reassessed. No supported numerical cash, margin, funding or ownership revision; all annual drivers and both horizons' weights retained. The pending process is not a quantified loss estimate.

11 November late: reported board deliberation on CEO removal. Not yet an issuer fact; annual drivers retained under review. Revision trigger: an issuer-announced departure, interim CEO or reduced role. Assessment.

Current probability revision — 12 November 2026

InputBeforeCurrentEvidence and judgment
Five-year downside/base/upside weights40%/45%/15%45%/45%/10%Filed succession decision, Item 5.02; near-ceiling upside needs uninterrupted coordination as well as funded demand. Interim leadership supports retaining the base, while a permanent successor remains unknown.
Ten-year downside/base/upside weights40%/45%/15%45%/45%/10%The same early disruption can have lasting product/share and capital-allocation effects; no automatic decade recovery is credited. Separate long-horizon assessment, not a measured transition frequency.
Annual financial paths, commitments and terminal valuationExisting source-linked paths and 13x/18x/22x endpointsRetainedNo quantified order, margin, compensation or funding change disclosed. No separate arbitrary key-person haircut.

The dated review and change record retain old inputs and same-price/date effects. Reassess on the permanent successor, technical-management structure, actual orders, product cadence and 17 November accounts. Unknown loss amounts remain unknown; three paths do not exhaust possible outcomes.

Underwriting plan

Written 3 Nov 2026 and revised the same day after the owner challenged the first version as extrapolation (updates/2026-11-03-reunderwrite.md). The model vintage stays anchored at 27 Oct 2026, with exits fixed at 27 Oct 2031 and 2036. Money is in USD bn unless stated. Rolling forecast years run October to October; NVIDIA's fiscal years end in late January.

Opportunity ledger

#OpportunityStarting point (evidence)Why NVIDIA wins / customer needReachable scale five years out (estimate)Investment and cash costTiming and limitsEvidence against
1Volume: hyperscaler AI systems (Vera Rubin, then its successor)Hyperscale 48.7 in Q2, about 195 annualized (hyperscaleRevenueQ2). Top-five capex about 810–830 in 2026 (company guidance), "significantly" or "meaningfully" higher in 2027 (googl-q3-2026, amzn-q3-2026).Rented cloud capacity has to run what cloud customers' code targets (CUDA). Rubin has purchase orders from every major hyperscaler (nvda-q2-fy2027-call p.4).The five's own cash plus observed financing is about 1,065 in year 1. At the observed 21% share that is about 225; in the base, 20% (213); ceiling 25% (266). Management's 1.3tn for 2027 exceeds their cash plus financing.Working capital about 20–22% of incremental sales; supply commitments (279, mostly memory) paid through cost of goods. Ceiling: 25% of the hyperscalers' cash-plus-financing budget, about 266 in year 1.Supply-constrained through FY2028 by management's account. Limited by memory and CoWoS capacity and by customers' power and shells.Internal inference moving to TPU, Trainium, Maia and MTIA. Broadcom's AI revenue +221%. Microsoft cut its 2026 capex estimate from about 190 to about 170. Customers' free cash flow is negative.
2New customers: ACIE (neoclouds, sovereigns, enterprises)About 40 in Q2, +138% year on year. Neocloud installed base 3 → 8 GW during 2026 (p.3). Sovereign +35% sequentially (p.5).These customers lack their own chips. NVIDIA supplies reference designs (DSX) and financing partners.Ceiling: at most 7 GW added in year 1 × 40 per GW = 280. Base: about 6 GW at a blended 35 per GW = 210.Extended terms (DSO 60); guarantees of 3.5 for neocloud land, power and shells; equity-method stakes in financiers (3.3).Bound by the availability of debt and equity for neoclouds and by power.Highly sensitive to credit markets. Neocloud returns are unmeasured here.
3Financed AI labs (OpenAI, a second frontier lab, others)About 25% of next year's business (aiLabFinancedShare). OpenAI commitments about 12 GW (p.6).Labs have demand but lack investment-grade balance sheets, and NVIDIA supplies the credit.12 GW × 40 per GW is up to about 480 of content through 2030, if delivered on NVIDIA systems at Rubin pricing. That is an upper bound, not a forecast.Equity: nearly 50 invested plus 25 committed. Guarantees: 105 (SB Energy) plus credit enhancement for about 2 GW. Shareholders bear these costs.Funding phases through 2030; SB Energy phases from FY2029.OpenAI's own XPU (1.3 GW in 2027 per Broadcom's call, unverified) and Anthropic's TPU, Trainium and AMD purchases. Revenue that depends on NVIDIA's own credit.
4Price and content per gigawattManagement sizes content at about 18 (Hopper), 25 (Blackwell) and 40 (Rubin) per GW, including the Vera CPU, NVLink and networking (pp.4 and 10).Each generation delivers more tokens per watt, which justifies price per GW.Already in the Rubin figure. Successors are not assumed to raise content further in the base.Memory costs: gross margin down to 71–72% in Q4 and 72–73% in FY2028 (p.7).The margin reset is already visible.Memory suppliers taking value; competitors' tokens-per-watt claims (Maia 300 +40%).
5Replacement of the installed baseHopper, Blackwell and Rubin installed across hyperscalers and neoclouds. Prior generations still rent at a premium (amzn-q3-2026).Annual cadence plus power limits: replacing old GPUs is how a power-capped site grows output.Supports the mature growth rate from year 2: 3%, 5% and 8% by scenario.Ordinary working capital.Starts to matter from about 2029 as 2025–2026 fleets age (five-to-six-year accounting lives).Customers stretching useful lives; Amazon is "finding ways to extend" them.
6Networking and CPUsNetworking 31.4 in FY2026 (+142%). Management sees about 20 of server-CPU demand for Vera.Attached to systemsIncluded in #1–#4; not added separately, to avoid double counting.Same as aboveEthernet competition (Broadcom)
7China data-center computeExcluded from guidance. H20 and H200 shipments are under 1% of Data Center.Zero in every scenarioDepends on regulationExport rules
8Edge: gaming, workstations, automotive22.5 of Graphics segment revenue in FY2026Mature franchiseLow single-digit growth, inside total revenueMemory prices weigh on PCs
9Acquisitions (Hugging Face, Groq balance)11.9 + 1.0 retention; Groq 0.986 outstandingEcosystem reachNo separate earnings creditedCash paid in year 1Close expected H1 2027Regulatory approval

Management targets as a bound. The roughly 70% FY2028 outlook (February 2027 to January 2028) is a supply-constrained planning number that implies FY2028 revenue of about 690. The demand ceiling built below caps year-1 revenue at 576. Management's plan therefore requires hyperscaler capex beyond their own cash plus the financing pace observed so far, or a larger NVIDIA share than today's. Management cites 1.3tn of top-five capex in 2027, but the companies themselves give no 2027 number. Executives' ranges are overprecise, and here the plan exceeds what customers' budgets fund. So the base is set from the bottom-up customer budget, not from the plan.

12 November opportunity reassessment

The filed succession decision, Item 5.02, changes who coordinates opportunities 1–6, not their measured customer budgets or installed base. In hyperscale design wins and Rubin/system content, slower architecture and supply decisions could concede lasting share. In financed labs, a changed approval process could reduce risky deployment as well as delay sales; an assumed improvement in lending discipline is not credited without actual decisions. Existing CUDA users, product teams and funded orders support continuity. Reachable units, margins, capital needs and ceilings are retained pending operating evidence. The joint uninterrupted upside therefore loses five percentage points to downside, while the customer-funded base keeps 45%. See the dated impact audit. This judgment was recorded before inspecting the revised return.

Demand ceiling

Built in valuation/calculations.json (steps hyperscalerOcf to demandCeilingYear1) and bound to model.demandCeiling. The engine rejects any year above it.

ComponentYear 1Evidence
Hyperscalers' own operating cash, annualized793Microsoft 58.0 (Q1 FY27) × 4; Alphabet 42.3 × 4; Meta 32.8 × 4; Oracle 23.1 × 4 (Jun–Aug); Amazon TTM 168.3. Alphabet, Meta and Oracle are derived as free cash flow plus capex.
External financing at the observed pace272Amazon 48.1 of Q3 debt plus Oracle 20.0 of Q1 equity, × 4. Microsoft repaid debt.
Hyperscaler capex budget1,065Their latest-quarter capex is already 231 (926 annualized), above their own cash
× largest share reaching NVIDIA25%Observed 21%: Q2 hyperscale revenue 48.7 / next-quarter capex 231. Plus up to 4 points for CPU and networking attach. TPU, Trainium, Maia, MTIA, buildings, power and storage take the rest.
Hyperscaler ceiling266
Non-hyperscaler capacity × full Rubin content280At most 7 GW in year 1 (5 GW added in 2026, p.3) × 40 per GW. These buyers do not design their own chips.
Edge30Q2 run rate 7.2 × 4
Ceiling, year 1576Grows 10% a year: customers' operating cash about +15%, with financing flat in absolute terms. Year 5: 844; year 10: 1,359.

The ceiling is an upper bound, not a forecast. It uses the full share in every component at once.

Outside view

Generic large-company base rates (sales growth and margin distributions by size bucket) are pending. A sourced copy of the Base Rate Book could not be obtained, so no remembered percentiles are used. The comparison below uses only primary filings.

NVIDIA's own downturns (10-K values via nvda-xbrl-companyfacts):

DownturnRevenueOperating income
FY20094,098 → 3,425m (−16%), then −3% in FY2010836m → −71m
FY202011,716 → 10,918m (−7%)3,804 → 2,846m (−25%)
FY2023Flat10,041 → 4,224m (−58%); margin 37% → 16%

That is three downturns in about fifteen years. Pre-boom operating margins (FY2016–FY2022) ranged from 15% to 37% and averaged about 27%. Gross margin averaged 62.1% over FY2018–FY2022.

Other suppliers into customer capex peaks:

SupplierPeak yearRevenue declineSource
CiscoFY2001 → FY200222.3bn → 18.9bn (−15%)csco-fy2002-results
Applied MaterialsFY2001 → FY20027.3bn → 5.1bn (−31%)amat-fy2002-10k
MicronFY2022 → FY202330.8bn → 15.5bn (−49%)mu-fy2023-results

This model's paths against that record. Five-year revenue CAGRs run from the trailing 354; margins are GAAP operating.

PathRevenue path5-year revenue CAGRMargin path5-year NOPAT CAGR (from 192.8)
DownsideYear 1 −15% against trailing (about −30% against the Q3 run rate), then +3%−0.9%65.6% → 30% in three years−16.5%
Base+28%, then 5%+9.2%→ 50%+2.7%
Upside+58%, then 8%+16.6%→ 58%+13.7%

The downside's decline sits between Cisco's and Applied Materials'. Its margin is NVIDIA's own pre-boom average.

The base still grows revenue 9% a year from an already enormous base. That is plausible only because year 1 completes orders already placed (Rubin purchase orders from every hyperscaler) and later years grow at half the ceiling's rate. Its margin is far above NVIDIA's history. The named mechanism: scale has cut opex from about 25–35% of sales before the boom to about 9–12%, while gross margin returns to the pre-boom 62%.

The upside's 16.6% revenue CAGR at this scale has no precedent among the suppliers above. It is allowed only because it stays inside the customer-funded ceiling, and it gets 15% weight.

Humility check. How would a disciplined value investor underwrite a supplier at a cyclical peak? This repository's method (docs/method/growth-opportunities.md, stage 7 and stage 9) answers it:

  • Normalize the margin. Do not capitalize the peak.
  • Bound revenue by customers' funds. Not by management plans.
  • Assume that at least one digestion occurs within the horizon. That is what the three NVIDIA downturns and the other suppliers' records show.
  • Price the exit on normalized earnings at a multiple the Koller identity earns.
  • Then ask whether today's price leaves a margin of safety even in the base.

Done that way, the base earns about −1% a year for five years at USD217.94. A 12% base return would need a price near USD118, and a 12% weighted return about USD100. When the critical forecast (post-2027 AI capex) cannot be known, the method says to widen the range or decline to underwrite, not to fit a growth rate to the price. The honest finding is that the price assumes far more than customers can currently fund.

Scenario mechanisms

The three cases are coherent business outcomes. Inputs and their inputNotes are in valuation/model.json; the arithmetic is in valuation/calculations.json. Year 1 carries the cycle position, and from year 2 revenue grows at the mature rate. Margins normalize within three years.

DriverDownsideBaseUpside
MechanismDigestion. Customers' capex already exceeds their cash, so after the 2025–2026 build financing tightens and buyers absorb installed capacity. Hyperscalers move internal inference to their own chips. Financed labs slow purchases. NVIDIA cuts price to defend volume while memory stays expensive.Customer-funded growth. Year 1 completes the Rubin ramp at the pace customers' cash and observed financing allow. Thereafter replacement plus inference demand grow 5% while custom silicon takes share at the hyperscalers. Competition and memory costs normalize margins.Share held near the ceiling. Customers keep spending at the edge of their cash plus financing. NVIDIA keeps its share and most of its pricing.
Year-1 revenue (vs trailing 354; ceiling 576)301 (−15%)453 (+28%)559 (+58%)
Growth from year 23%5%8%
Year-5 revenue (ceiling 844)339551761
Operating margin, reached in year 330%. Pre-boom average; gross margin about 50% with opex about 18% of a smaller base. Permanent, because substitution, not under-use, causes it.50%. Gross margin back to the pre-boom 62%, opex about 12%.58%. Gross margin about 68%, opex about 10%.
Capex / sales6%4%3%
Operating capital / sales30% (stretched terms, excess inventory)22%18%
Tax rate23%20%17%
Retained-investment recovery40%80%100%
Contingent loss25% of 164.5 (about 41) in year 300
Financing and dilutionNo new net borrowing. Debt refinanced at 4.5%. Net shares flat. GAAP stock compensation charged. No buyback accretion.SameSame
Exit (Koller: (1 − g/RONIC)/(r − g), r = 10%)13x (g 3%, RONIC 40%) on trough-normalized NOPAT18x (g 5%, RONIC 50%) on normalized NOPAT22x (long-run g 6%, RONIC 60%), capped because the margin stays near peak

Temporary under-use versus permanent impairment. A pause in hyperscaler orders alone would be temporary under-use, like FY2023: inventory provisions, a few weak quarters, then recovery with the next generation. The downside instead pairs the cyclical decline with permanent substitution. It therefore reverts margins to the pre-boom level and grows only 3% after the trough. Annual granularity places the decline in year 1. In reality it would begin after the visible orders ship, which moves cash timing but not the year-5 exit.

Funding before cash. Growth working capital and capex, the25bn July equity commitment pool plus the2bn post-July Nscale funding reserve and both acquisition claims (Groq 0.986, Hugging Face 11.9) are paid in year 1 before any of their earnings. Neither acquisition is credited with earnings. Supply commitments flow through cost of goods.

Not modeled:

  • China revenue.
  • Future investment gains.
  • Interest on accumulated cash.
  • Buybacks above the stock-compensation offset.
  • New borrowing.

13 November funding update

Nscale's amended S-1 (pp108–110,161–162; F-64) establishes aSeptember subscription after NVIDIA'sJuly balance/commitment anchor. Reserve2bn once; do not give this financing an extra increment of operating revenue or treat Nscale's own debt as NVIDIA's debt. Its Anthropic infrastructure funding is still being arranged, so a signed service contract is not a fully financed GPU delivery. Keep the existing demand ceiling and downside financing stress; the public filing does not isolate a defensible change in aggregate NVIDIA shipments or margins. A negative-FCF project is not automatically bad: the missing question is funded delivery and matched mature cash return. Verify the expected16November closing and roll actual NVIDIA cash/holdings/remaining commitments together on17November.

15 November customer and commitment check

Anthropic's original S-1 describes a multi-chip orchestration foundation that lowers its switching effort. The existing ledger's CUDA advantage is therefore customer-specific, not an immovable industry barrier. This supports the already-modeled substitution/share and margin fade; one advanced lab does not measure every customer's switching time or NVIDIA's future absolute shipments. Lower compute cost can also expand paid demand. The article's exact three-day benchmark was not verified in the primary pages.

The same filing's10bn conditional NVIDIA subscription may fulfill the program announced in2025. The unitemized July pool cannot prove inclusion or a new incremental liability. Keep the explicit overlap judgment and0–10bn funding sensitivity in the 15November review. Customer reported operating cash is positive despite large noncash valuation losses, but future compute commitments and conditional IPO funding remain distinct from collected cash or fully financed delivery. No incremental revenue or terminal investment premium follows from these disclosures.

Capital productivity and reinvestment

Forecast assessment:2026-10-27, operating re-underwrite3November; funding reserve updated13November2026. The model now starts from trailing four-quarter figures, adds a sourced demand ceiling and uses cycle-based scenarios. Money is in USD millions. These calculations use the researched scenarios and the shared forecast ledger. They are forward diagnostics, not independently measured historical ROIIC or realized acquisition IRRs. Local sensitivities may differ.

PeriodScenarioCapital return / proxyPeriod-weighted reinvestmentNew growth capitalMatching earnings
5 yearsdownside-111.3%22.0%102989.674468262.855
5 yearsbase23.7%10.8%115829.2621074853.710
5 yearsupside144.0%7.5%120456.1621615799.053
10 yearsdownside-49.2%23.2%207741.458896022.645
10 yearsbase37.1%10.1%238226.5642352931.976
10 yearsupside127.6%6.9%270691.4933936544.962

Return definition: Operating ROIIC · partial capital proxy. Numerator 27485.803 / denominator 115829.262. Earnings basis: Unlevered operating NOPAT. The forecast runs from 2026-10-27 to 2031-10-27; a ten-year selection uses its own summed ledger through 2036-10-27. The measured historical record (FY2021–FY2026: lagged ROIIC 472% on tangible capital, reinvestment 22.5%, operating leverage 63%) is in research/04-capital-allocation.md; it is mostly price and margin on expensed R&D, so it is not used to extrapolate these forward figures.

Average period reinvestment is total new growth investment / total matching earnings, not the arithmetic mean of annual rates. Old acquisition claims, dividends, debt repayment and financial stakes do not acquire a second increment of operating profit.

Five-year base funding: cash generated before new investment 1068733.260, less distributions 121425.000, less old claims 12886.000, plus opening cash drawn / (cash built) 0.000, plus net borrowing / (deleveraging) -791592.998. This funds growth capital 115829.262 and separately funded financial stakes27000.000 (July pool25000 plus post-July Nscale reserve2000). Reconciliation error is below 0.000001. Rates above 100% are funded, not clipped.

Forecast NOPAT change / recorded net physical and working-capital investment: 23.7%. The base NOPAT rises only modestly because the margin normalizes from 65.6% to 50% while revenue grows. Changes in margins and tax affect this numerator, while expensed growth investment is omitted from this denominator. It is a partial forecast diagnostic, not a measured acquisition-cohort return or an all-in economic ROIIC.

Limits:

  • This operating-capital proxy uses net capex and working capital. Expensed R&D and other intangible growth investment are not capitalized; it is not all-in economic ROIIC. Changes in margins and tax also affect the numerator.
  • Financial stakes have no assigned operating earnings and are excluded from operating ROIIC; their funding is shown separately.
  • These are period-specific forecast diagnostics, not independently measured historical ROIIC or a project XIRR. The matched, lagged historical measurement is in research/04-capital-allocation.md; it is a separate diagnostic, not a calibration of these forecast ratios.

Expensed growth investment

The physical-capital proxy omits expensed R&D. A separate sensitivity treats forecast R&D above a fixed run-rate sustaining amount as after-tax growth investment, amortizing each new cohort over an explicitly assumed life. Both earnings and capital change by the same net amount; shareholder cash flows and XIRRs stay unchanged. The split, constant revenue intensity and lives are assumptions, not issuer useful-life disclosures. Prior R&D capital and other growth expenses are not fully reconstructed, so this remains a partial economic-capital diagnostic.

Assumed lifeROIIC proxyWeighted reinvestmentGrowth capital
3 years22.1%12.6%138288.642
5 years22.5%13.3%146544.061

Definitions and original methodology links are in docs/method/cash-flow-and-reinvestment.md; source locators are in this company’s registry. Probabilities are separate judgments recorded in research/11-scenario-probabilities.md.

Scenario probabilities

Current review15 November2026: both horizons retain45%/45%/10%. The Nscale reserve is a separately funded cash use with existing recovery assumptions. The newly inspected complaints seek unquantified relief; reported insurance terms and Megaspeed's statement do not establish a customer cancellation, permanent successor or enacted export restriction. Fundraising risk is already part of the joint downside. Coarse odds remain low-confidence and do not quantify every legal or financing tail. Current review; funding revision.

Assessment dates:

  • 2026-10-27: initial weights.
  • 2026-11-03 (first review): weights unchanged.
  • 2026-11-03 (re-underwrite): the scenarios were redefined after the owner's challenge, so the weights were reassessed. They now apply to the new paths: digestion, customer-funded growth, and share held near the ceiling.
  • 2026-11-09: filing notice and current public developments reviewed; numerical weights retained with the unquantified tail explicit. Decision.
  • 2026-11-12: confirmed succession; five percentage points move from upside to downside at both horizons. Annual cash paths and terminal endpoints retained. Decision and impact.
  • 2026-11-08: reassessed after the issuer confirmed an independent board investigation. Both horizons retain40/45/15; the reasons and additional unquantified tail limitation are recorded in the model.

These are subjective judgments held with low confidence. Financial sources support the mechanisms, not measured numerical odds, and no external author provided or endorsed them.

HorizonDownsideBaseUpside
5 years45%45%10%
10 years45%45%10%

Previous weights through 11 November: 40/45/15 at both horizons. Initial weights before the 3 November re-underwrite: 5 years 35/40/25; 10 years 40/35/25.

8 November governance assessment. The board statement, paragraphs1-5, confirms a process, while retaining confidence in leadership; the CEO response confirms cooperation. These sources do not establish a business restriction, leadership replacement or measurable change in customer-funded orders, cash margins or claims. The material risk record changes; numerical path weights remain unchanged. This is a low-confidence judgment pending findings, not a probability of exoneration. Leadership disruption, information loss and wider enforcement can create losses beyond the modeled downside; the grid is not a complete quantification of them. Revise paths and odds together if evidence establishes any of those mechanisms. Old weights and source dates remain in dated records and version history.

Historical five-year reasoning (3 November; superseded weights). A digestion within five years is close to as likely as not:

  • NVIDIA itself had revenue or profit downturns in FY2009, FY2020 and FY2023, three in about fifteen years.
  • Every capex-cycle supplier read here had a 15–49% revenue decline after a customer spending peak: Cisco FY2002, Applied Materials FY2002, Micron FY2023.
  • Customers' capex now exceeds their own cash.
  • About a quarter of next year's demand comes from AI labs NVIDIA finances.
  • Custom silicon is growing about twice as fast as NVIDIA's Data Center business.

The base keeps a plurality because year 1 completes orders already placed and the base never exceeds what customers can fund. The upside needs customers to keep spending at the ceiling for five years and NVIDIA to hold share, so it originally received 15%; the succession revision reduces that to 10%.

Historical ten-year reasoning (3 November). At least one digestion over a decade is likely, and the competitive position against in-house silicon is less visible. Every path already contains its recovery growth after year 1, so no weight moves to the upside for the longer horizon.

The three paths:

  • Downside: a digestion after the 2025–2026 build. Revenue falls about 30% below the Q3 run rate and margin returns to the 30% pre-boom level, sized from NVIDIA's own and other suppliers' reported downturns.
  • Base: growth funded by customers' own cash and observed financing, with share lost to custom silicon and margins normalized to 50%.
  • Upside: NVIDIA holds share while customers spend near the demand ceiling; margin stays near 58% and the exit is capped at 22x.

Price movement alone did not change the weights. The change follows the redefinition of the scenarios and the base-rate evidence recorded in research/08-underwriting-plan.md.

Review triggers:

  • Customer AI monetization/capex and concentration.
  • Gross margins, alternative silicon, supply commitments and inventory.
  • Receivables, economic R&D, partner stakes and contingent support.

Each path represents a broad outcome region, not the probability of hitting an exact share price or IRR. Paths are a coarse approximation of a continuous joint distribution; the displayed downside and upside are not bounds. Joint driver dependencies must be considered, rather than multiplying marginal driver guesses. Scenario weights sum to 100% within this approximation. Severe outcomes remain possible outside the modeled paths.

Price movement alone does not change these weights. A changed causal premise, acquisition capacity, financial report or scenario definition does. Local assumption sensitivities suspend the researched weights until the full distribution is reassessed. No weighted-average IRR is presented as an expected shareholder return.

Sources: nvda-q2-fy2027-10q, nvda-q2-fy2027-release, nvda-q2-fy2027-call, nvda-q2-fy2027-cfo, nvda-xbrl-companyfacts, msft-fy2027-q1, googl-q3-2026, amzn-q3-2026, meta-q3-2026, orcl-q1-fy2027, avgo-q3-fy2026, csco-fy2002-results, amat-fy2002-10k, mu-fy2023-results, nvda-sb-energy. Method: Damodaran, chapter 33, printed pp.894-897. See valuation/model.json for machine-readable weights and docs/method/cash-flow-and-reinvestment.md for the capital definitions.

11 November evening reassessment. Dated evidence and decision: retain 40%/45%/15% at both horizons. The new process evidence does not identify a cash-loss amount or a defensible standalone probability. This leaves the existing omitted tail explicit; it is not an exoneration probability.

11 November late reassessment. Dated evidence and decision. Both horizons stay at 40%/45%/15%, now explicitly under review. Reports that the board is weighing removal of the CEO are attributed and unconfirmed, and no transmission into orders, R&D cadence or margins is yet observed. Reweighting on headlines would fit odds to news rather than evidence. On an issuer-confirmed leadership change, reassess the downside weight, paths and terminal multiple together.

12 November current reassessment. The issuer-confirmed departure activates the prior review trigger. Architecture, talent, supply and customer-financing decisions now face an interim coordination period. Transfer a coarse five percentage points from the joint near-ceiling upside to downside, which already contains share, margin and correlated funding stress. Installed CUDA, operating teams and funded customer demand support keeping the base at 45%. Over ten years, lost early design wins and capital decisions can persist, so no automatic reversion restores upside weight. Exact numerical odds are unmeasured and low confidence. No separate cash loss or exit-multiple haircut is added. The same-price/date audit holds every annual row and capital metric fixed. Confirmed economic costs or gains require revising the paths themselves; these weights do not price every legal/security tail.

Bull–bear review — NVIDIA, 5 November 2026

Independent advocates prepared opening cases without seeing the other draft, then exchanged point-by-point cross-examinations. A separate judge who did not author the prior model assessed the competing claims. Source reconstructions retain their stated provenance; inaccessible private reports were not newly read. Quoted sensitivities test arguments and are not automatically adopted forecasts.

Fixed quote USD 200.13, market time 2026-11-04T21:00:01.000Z; valuation date 2026-11-05. Original forecast exits retained.

HorizonBefore bear/base/bullAfter bear/base/bullWeighted before → after
5 years-23.24% / 0.53% / 14.88%-23.24% / 0.53% / 14.88%-2.77% → -2.77%
10 years-9.75% / 4.15% / 12.54%-9.75% / 4.15% / 12.54%2.74% → 2.74%

Bull case

As of 5 November 2026. No opposing advocate draft was read. Existing company research, fact/calculation ledgers and exact registered source locators were examined; original issuer sources additionally checked live: Lumine Q3, CSU Q2, Diploma Q3 transcript, NVIDIA CFO commentary, TerraVest Highland/KBK, WOSG FY26. Harvia Q3 PDF retrieval failed; its registered extraction remains the evidence limit. Private source interpretations below are the repository's documented reconstruction, not a new claim to have read an unavailable private original.

The strongest bear case is that major customers' AI capital spending is externally financed and concentrated, while custom silicon compresses NVIDIA's shortage margins. GPU demand can grow secularly and still suffer a capex digestion, correlated guarantee losses and a poor shareholder return.

NVDA-B1 — The installed software/system advantage buys time. CUDA, networking and full-rack delivery reduce time-to-deployment and total customer engineering cost; sovereign/enterprise inference can broaden demand beyond hyperscalers. Test58% first-year revenue growth versus 28%, holding later 5% growth and capital/credit assumptions. Sources nvda-q2-fy2027-call, pp.3–7,10; nvda-q2-fy2027-cfo, hyperscale/ACIE and inventory discussion. It stays below the repository's funded customer-demand ceiling. Management's70% preliminary FY2028 view is not independent demand evidence. Broadcom's AI growth (avgo-q3-fy2026, release segment/outlook) is simultaneous evidence of expanding demand and substitution.

NVDA-B2 —50% may over-normalise the present advantage. Test58% operating margin versus 50%, keeping revenue and funded investment fixed. R&D software reuse, hardware/software optimisation and platform ecosystem can preserve rents even as memory costs rise; 71–73% gross-margin management expectations support a bounded possibility, not proof (nvda-q2-fy2027-call, p.7). The pre-boom27% operating-margin outside view describes a different mix, but today's65.6% likewise cannot be a permanent baseline. No claim of customer spending-independent growth is made.

NVDA-B3 — A finite competitive lead can support 18×. Test22× versus 18×.18× requires about 14 post-exit years at 8% earnings growth, 50% marginal return and 10% required return before mature no-growth value. The mechanism is CUDA, interconnect and developer adoption; the challenge is customers' own chips and model-efficiency gains. Source nvda-q2-fy2027-10q, MD&A pp.25–32 and commitments note 10. Both14years and 50% are explicit forward judgments, not measured duration. The starting 353.964bn revenue includes the unreported Q3 guidance midpoint: it is a guidance-based trailing estimate, not four reported quarters. Correct that label; replacing it mechanically with an older reported quarter would not improve the forward judgment.

Isolated shared-engine effects

Saved quote 200.13; market timestamp 2026-11-04T21:00:01.000Z. Fixed valuation date 2026-11-05; model 6728a011341f; exits remain the original model vintage. These are sensitivities, not joint forecasts or probability revisions. Baseline base IRR 0.53% / 4.15% (5/10 years).

ClaimExact input/path5-year base10-year baseChange, pp (5/10)
NVDA-B1cases.base.inputs.revenueGrowth4.81%6.35%+4.28 / +2.21
NVDA-B2cases.base.inputs.operatingMargin3.50%5.71%+2.96 / +1.57
NVDA-B3cases.base.inputs.exitMultiple3.91%5.68%+3.38 / +1.54

Bear case

Strongest opposing case

CUDA, networking and a complete rack system lower customer deployment time, while inference expands compute demand beyond the original training cycle. Hyperscalers have placed orders, NVIDIA has abundant cash and high tangible returns, and custom accelerators can coexist with a much larger total market. A near-term credit collapse cannot be asserted merely because customers are investing heavily.

Bear case

The binding risk is who pays for the next unit of compute and who captures its economics. Customers' capex includes buildings, power, custom silicon, CPUs and other equipment; it cannot all be mapped to NVIDIA sales. The demand ceiling in the current opportunity ledger makes these distinctions, but some of its GW/content figures and future customer funding still come from management propositions and third-party transcript summaries. Customer investment announcements are not deployed, utilized, cash-profitable capacity. The ongoing laboratory/neocloud P&L reconciliation is a material gap, not a reason to extrapolate demand. [nvda-q2-fy2027-call p.3–10; research/08-underwriting-plan.md Opportunities and customer sources/uses; googl-q3-2026, amzn-q3-2026, meta-q3-2026, orcl-q1-fy2027 cash-flow tables.]

NVIDIA's largest buyers also have incentives and resources to substitute their own chips for repeatable workloads. Current cloud demand is evidence for compute, not for perpetual merchant-GPU share or price per watt. The base's50% GAAP operating margin already assumes lasting operating leverage far above the pre-boom average near 27%. The unit moat can survive while that margin does not. [nvda-fy2026-10k history and concentration; nvda-q2-fy2027-10q MD&A pp.25–32; googl-q3-2026 TPU discussion; amzn-q3-2026 custom-chip discussion; avgo-q3-fy2026 operating results.]

The outside view includes NVIDIA's own earnings declines and Cisco, Applied Materials and Micron inventory/capex cycles. They establish that digestions are possible; their percentages are not interchangeable forecasts. A first-year15% revenue stress and 40% mature operating margin test customer delay and bargaining without claiming AI disappears. [csco-fy2002-results; amat-fy2002-10k; mu-fy2023-results; research/08-underwriting-plan.md Outside view.]

Lab equity, purchase/capacity commitments and guarantee exposures are correlated with those customers' demand. An80% recovery on investments and zero calls in the base should be challenged together; the model's sensitivities show these claims matter less to the five-year value than operating margin and exit price, so they should not distract from the main earnings risk. An18x exit requires 5% profitable perpetual growth beyond 2031 in the Koller calculation. No finite advantage duration is evidenced; constant secular demand does not prove constant supplier rents. [nvda-q2-fy2027-10q notes 5–10; research/05-thesis.md Exit multiple.]

Contested claims

  • NVDA-B1 —Test year 1 revenue−15% against+28%, leaving subsequent 5% growth fixed. Orders and finance availability can slip without permanent demand destruction.
  • NVDA-B2 —Test40% operating margin versus 50%; a disclosed sensitivity between historical 27% and current peak, not an observed normalized level.
  • NVDA-B3 —Test13x exit against 18x to shorten the premium for durable excess returns.
  • NVDA-B4 —Test40% investment recovery plus 25% claim/capacity stress jointly, preserving demand in this isolated sensitivity to avoid disguising correlated risks as independent probabilities.

The current issuer release was reopened. The pending 17 November quarter must replace the guidance-inclusive baseline when published; Q3 guidance is not reported revenue. This review does not validate private contracts, lab profitability or every funding source.

Shared-engine effects

Fixed saved quote USD200.13; market timestamp 2026-11-04T21:00:01.000Z; valuation date 5 November 2026; fixed original exit calendar. Base before: 0.53% / 4.15% (5/10years). These are one-change sensitivities to the base, except named bundles; existing scenario probabilities are not revised or claimed as validated.

ClaimChange5-year base IRR10-year base IRR5-year effect10-year effect
NVDA-B1Year1 demand decline 15%-7.26%-0.02%-7.79pp-4.17pp
NVDA-B2Operating margin 40% from year 3-3.74%1.83%-4.27pp-2.32pp
NVDA-B3Base exit 13x-4.45%1.88%-4.98pp-2.27pp
NVDA-B4Investments40% recovery and 25% claim/capacity stress0.16%4.01%-0.38pp-0.13pp

Rebuttals

Bull response

5 November 2026. Written after independently frozen openings were exchanged. Every numbered bear claim is addressed below; original bull openings and sensitivities remain unchanged.

  1. Bear first-year−15%: retain as adverse, reject as default. Actual hyperscale/ACIE revenue and orders show broad installed demand (nvda-q2-fy2027-cfo, revenue discussion), and customers possess recurring operating cash plus disclosed funding. They do not prove utilisation or returns on every marginal rack. My58% upside uses the existing funding ceiling, but the ceiling itself includes forward management GW/content assumptions. End-user economics and credit conditions remain surviving tests. No statement that orders equal cash-profitable demand is defended.
  2. Bear40% normalized margin: concede uncertainty, rebut precise historical transplantation. Pre-boom27% operated with a different product mix, scale and platform position; current 75% gross margin is equally unsuitable as permanent proof. Memory pass-through and custom silicon can lower rents while CUDA survives. The base50% and upside 58% are explicit judgments bounded by unit economics, not measured through-cycle returns. Stress40% is valid; a single capex-cycle analogue does not establish it centrally.
  3. Bear13× versus 18×: concede finite duration not established. My14-year diagnostic is deliberately demanding. Product lead, CUDA and networking support a premium, but hyperscalers also fund competing systems. A lower multiple requires a named cash-growth/reinvestment/advantage-duration path, not merely aversion to the quote. Historical supplier boom/bust examples establish possibility, not the exact future drawdown or capitalisation.
  4. Bear correlated 40% stakes recovery/25% calls: concede material joint risk. Preserve the actual contractual perimeter and avoid charging a purchase commitment already funded in cost of sales twice (nvda-q2-fy2027-10q, notes 5–10). The isolated balance-sheet stress's modest IRR effect does not neutralise correlation with demand/margins. The existing adverse case already combines these mechanisms; weights are joint scenario judgments, not independent event odds.
  5. Baseline evidence attack: concede prose correction, rebut mechanical staleness.353.964bn includes 108bn Q3 guidance, not reported sales. SEC CFO Outlook explicitly calls 108bn expected, 74% gross margin and 9.2bn GAAP opex. A guide-based trailing estimate is permissible if separated from reported facts and replaced 17 November. Reverting silently to an older trailing base without rebuilding customer-funded future sales is not a method improvement. Conviction high in present product economics, low in long-horizon price/value confidence.

Bear response

5 November 2026, after independent openings were exchanged.

  1. Bull58% first-year growth: hardware/system lead conceded, cash-demand claim unresolved. Orders, actual hyperscale/ACIE sales and available customer financing support a strong near-term book; they do not establish final-customer utilization or profitable cash demand. The bull concedes the ceiling still uses management GW/content assumptions and private lab accounts remain unavailable (nvda-q2-fy2027-cfo, revenue/outlook; 08-underwriting-plan, customer sources/uses). Retain58% as conditional upside and−15% as digestion stress; neither is selected centrally by analogy.
  2. Bull58% margin: modern mix matters, precise durability unproven. I concede a pre-boom27% average is not a central forecast for today's software/rack/network mix. Equally, present peak gross margin and scale do not establish 58% GAAP operating margins indefinitely.40% was a sensitivity between historical and current conditions, not an observed normal level. Base50% remains a reasoned conditional judgment until memory prices, custom silicon and customer value capture distinguish them (nvda-q2-fy2027-10q, MD&A; callp.7; source-linked own-history record).
  3. Bull18/22x exit: duration uncertainty survives. CUDA and interconnect can persist while the share of customer profit accruing to NVIDIA shrinks. Fourteen post-exit years 8% growth at 50% marginal return is a demanding joint claim, and the bull concedes no observed moat life. My13x stress is not uniquely central either. Disclose both, keep continuation assumptions explicit, and do not let a high multiple on a lower margin hide peak-rent extrapolation.
  4. Correlated investment/guarantee stress: agreed. Investment recovery and guarantee calls can deteriorate with demand, so do not treat their odds as independent. Preserve the contractual perimeter and do not double-charge supply commitments already in cost of sales (nvda-q2-fy2027-10q, notes 5–10). Isolated balance-sheet sensitivity is smaller than margin/price sensitivity, which does not refute severe joint losses.
  5. Guidance-inclusive baseline: factual label correction required, not an automatic cash revision.353.964bn includes 108bn unreported Q3 guidance. It is a guided/estimated trailing bridge, not four reported quarters. I agree with the bull that silently substituting an older quarter without rebuilding future customer demand would be mechanical staleness. Reported facts and estimates must be separated now and the report replaced 17 November. No other central numerical change is established by this exchange.

High conviction in current product economics does not establish a margin of safety atUSD200.13; moderate-to-low conviction in post 2027 cash/terminal forecasts remains appropriate.

Judgment

Existing estimates receive no presumption merely because they are already in the model. Positive company evidence, the relevant outside view and the capital/funding bridge determine each verdict; neither arbitrary conservatism nor an unsupported favorable sensitivity replaces it. The evidence and reconstruction below explain numerical changes or explicit no-change conclusions claim by claim.

Contested claimVerdictEvidence/locatorReason and model action
bull advocate: NVDA-B1: cases.base.inputs.revenueGrowthrejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. Orders, CUDA/networking and funding ceiling support ongoing growth but not adopting management 70% as independent demand.58% remains upside.Reconstructed the isolated proposed input/path from independent opening and shared engine, at frozen saved quote/date/horizons; source mechanism separated from unsupported numerical extrapolation.
bull advocate: NVDA-B2: cases.base.inputs.operatingMarginrejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. Current product/system advantage supports margins above preboom mix; 50% normalizes current 65.6%.58% belongs upside, with recurring R&D still charged.Reconstructed the isolated proposed input/path from independent opening and shared engine, at frozen saved quote/date/horizons; source mechanism separated from unsupported numerical extrapolation.
bull advocate: NVDA-B3: cases.base.inputs.exitMultiplerejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. 18x requires roughly 14 post-exit years at 8%growth/50%marginalreturn/10%requiredreturn. CUDA/networking provides a mechanism; its duration remains unmeasured and 22x is not justified merely by growth.Reconstructed the isolated proposed input/path from independent opening and shared engine, at frozen saved quote/date/horizons; source mechanism separated from unsupported numerical extrapolation.
bear advocate: NVDA-B1: Year1 demand decline 15%rejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. A15% digestion is credible and already adverse; recent funded orders and current customer capacity commitments do not establish an imminent central contraction.Reconstructed the independent isolated stress at the same quote/date/horizons; do not treat stress size as a probability or a measured current fact.
bear advocate: NVDA-B2: Operating margin 40% from year 3rejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. Customer bargaining and customchips are genuine; normalizing 65.6% to 50% already shares that risk.40% stays an adverse path until workload/mix evidence discriminates.Reconstructed the independent isolated stress at the same quote/date/horizons; do not treat stress size as a probability or a measured current fact.
bear advocate: NVDA-B3: Base exit 13xrejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. 13x quantifies shorter advantage, not a uniquely established terminal value. Retain18x as explicitly conditional finite-duration judgment, low conviction.Reconstructed the independent isolated stress at the same quote/date/horizons; do not treat stress size as a probability or a measured current fact.
bear advocate: NVDA-B4: Investments40% recovery and 25% claim/capacity stressrejectnvda-q2-fy2027-call pp.3–10; nvda-q2-fy2027-10q MD&A pp.25–32 and notes 5–10; googl-q3-2026 and amzn-q3-2026 customer cash/customsilicon discussion; nvda-q2-fy2027-cfo. Correlated funding/investment risk is admitted in downside; no verified claim call or recovery observation supports assigning 40%/25% centrally.Reconstructed the independent isolated stress at the same quote/date/horizons; do not treat stress size as a probability or a measured current fact.

Retain existing coarse horizon-specific weights after reassessment: scenarios already represent these adverse/favorable operating mechanisms. Numerical odds are subjective, not verified frequencies; unresolved gaps prevent high conviction. No tuning to return.

Conviction

Low in precise forward returns; platform advantage supported, customer-paid demand and durable margins unverified. These are forward estimates with surviving attacks, not confirmed outcomes. Neither numerical calibration, a completed checklist nor passing software tests establishes investment skill. Reported facts stay separate from forward judgments, and adverse paths are not worst-case bounds.

What would change the view: 17 November Q3 actuals replace guidance-inclusive starting estimate; reconcile customer-funded purchases, RPO/cash, receivables, guarantees and utilization.

Exact inputs, isolated sensitivities, model hashes and same-price/date before/after returns.

Public source library

Primary filings, investor communications, and industry evidence.

primary

NVIDIA Q2 FY2027 Form 10-Q ↗

Quarter ended 26 July 2026. Statements pp. 3–8; notes 5–7 investments and working capital; note 9 debt; note 10 supply, equity/cloud commitments and guarantees; MD&A pp. 25–32 China, margins, receivables and ecosystem funding.

Published Aug 26, 2026 · Reviewed Oct 27, 2026
primary

NVIDIA Q2 FY2027 results and Q3 outlook ↗

GAAP income, balance-sheet and cash-flow tables; Q3 revenue 108,000m ±2%, gross margin 74% ±50bp, GAAP opex 9,200m, FY tax range 16–18%. Beginning FY2027 non-GAAP figures retain SBC. Dividend USD0.25 paid 1 October.

Published Aug 26, 2026 · Reviewed Oct 27, 2026
primary

NVIDIA FY2026 annual report ↗

Fiscal year ended 25 January 2026. GAAP income p. 51 and cash flows p. 55; inventory, compensation, customer concentration and supply commitments. Issuer PDF mirror: https://s201.q4cdn.com/141608511/files/doc_financials/2026/ar/2026-Annual-Report-Web.pdf

Published Feb 25, 2026 · Reviewed Oct 27, 2026
primary

NVIDIA issuer-hosted Q2 FY2027 earnings transcript ↗

CFO pp. 3 and 7: approximately 70% FY2028 revenue growth is a preliminary supply-constrained outlook; Q&A on pricing/capacity; p. 15 confirms 17 November earnings call. Management expectations, not achieved sales. Reviewed 3 Nov 2026: p.3 hyperscale 49bn and ACIE 40bn, top-five hyperscaler capex near 800bn (2026) and 1.3tn (2027) as management cites; p.4 and p.10 revenue opportunity per gigawatt about 18bn (Hopper), 25bn (Grace Blackwell), 40bn (Vera Rubin); p.6 nearly 50bn invested in frontier AI labs, 12 GW of OpenAI commitments, AI labs supported by the balance sheet about a quarter of next year’s business; p.7 gross margin to bottom at 71–72% in Q4 and 72–73% in FY2028 on memory costs; DSO 60 days.

Published Aug 26, 2026 · Reviewed Nov 3, 2026
primary

NVIDIA supply-chain capacity expansion ↗

Substrate, packaging and testing plans, with USD1.1bn KYEC reservation advance. IBIDEN multibillion advance is not separately quantified here; do not invent its amount or assume every capacity plan is operating.

Published Oct 21, 2026 · Reviewed Oct 27, 2026
primary

Microsoft FY2026 Q4 earnings discussion ↗

CFO: USD41bn quarterly capex, around two thirds short-lived CPU/GPU assets; cloud demand and cash generation; lease-accounting effects. Customer evidence, not a direct order to NVIDIA.

Published Jul 29, 2026 · Reviewed Oct 27, 2026
primary

TSMC Q3 2026 earnings discussion ↗

pp. 3–4: USD60–64bn capital budget, strong leading-edge/AI demand, constrained tools and cautious non-AI end markets. Upstream evidence does not establish NVIDIA share or customer returns.

Published Oct 15, 2026 · Reviewed Oct 27, 2026
primary

AMD Q2 2026 issuer results ↗

Helios/MI400, ROCm and named customer partnerships. Competitive product and deployment claims remain issuer claims, not independent performance validation.

Published Aug 4, 2026 · Reviewed Oct 27, 2026
market data

NVDA delayed market quote ↗

Runtime quote keeps the provider timestamp and currency. No current quote is used to reset revenue, margins or terminal valuation.

Published Oct 27, 2026 · Reviewed Oct 27, 2026
primary

NVIDIA FY2025 annual report (Form 10-K) ↗

Fiscal year ended 26 Jan 2025. Income statement, balance sheet and cash-flow statement; FY2024 and FY2023 comparatives restated for the June 2024 ten-for-one split.

Published Feb 26, 2025 · Reviewed Nov 3, 2026
primary

NVIDIA FY2024 annual report (Form 10-K) ↗

Fiscal year ended 28 Jan 2024. Statements; FY2022 comparatives (pre-split share counts). Arm termination cost of USD1.35bn recorded in Q1 FY2023.

Published Feb 21, 2024 · Reviewed Nov 3, 2026
primary

NVIDIA FY2023 annual report (Form 10-K) ↗

Fiscal year ended 29 Jan 2023. Statements; inventory provisions and the Arm write-off depressed FY2023 operating income.

Published Feb 24, 2023 · Reviewed Nov 3, 2026
primary

SEC XBRL company facts for NVIDIA (CIK 1045810) ↗

Machine-readable 10-K values (us-gaap tags) used to extract FY2021–FY2026 annual statements; each value was taken from the 10-K in which it appears, preferring the latest filing for restated comparatives.

Published Aug 26, 2026 · Reviewed Nov 3, 2026
primary

NVIDIA Q4 FY2026 CFO commentary ↗

FY2026 Data Center 193,737m (compute 162,361m; networking 31,376m); hyperscalers slightly over 50% of Q4 Data Center revenue; supply commitments 95.2bn; DSO 51 days; multi-year cloud service agreements 27bn.

Published Feb 25, 2026 · Reviewed Nov 3, 2026
primary

NVIDIA Q2 FY2027 CFO commentary ↗

Hyperscale revenue 48.7bn of 89.0bn Data Center; inventory 31.6bn for Vera Rubin; supply and capacity commitments 279bn mainly memory; DSO 60 days from extended terms with investment-grade customers; cloud service agreements 29bn; China Hopper under 1% of Data Center.

Published Aug 26, 2026 · Reviewed Nov 3, 2026
primary

NVIDIA 2026 proxy statement (DEF 14A) ↗

Beneficial ownership at 23 Mar 2026 pp.35–36; related-person transactions p.34; CD&A pp.44–46 (FY2026 revenue and non-GAAP operating income goals, relative-TSR PSUs, target pay); summary compensation table; board independence p.14.

Published May 12, 2026 · Reviewed Nov 3, 2026
primary

NVIDIA agreement to acquire Hugging Face (Form 8-K item 8.01) ↗

About USD11.9bn to Hugging Face stockholders, subject to adjustment, plus an equity retention program of up to about USD1.0bn; closing expected in the first half of calendar 2027 subject to regulatory approval.

Published Sep 3, 2026 · Reviewed Nov 3, 2026
primary

Jen-Hsun Huang Form 4 filings ↗

Rule 10b5-1 plan adopted 20 Mar 2025 for up to 6m shares (expired 31 Dec 2025); 2026 filings show tax withholding and gifts, no open-market sales; direct holdings 70.1m at 16 Sep 2026.

Published Sep 18, 2026 · Reviewed Nov 3, 2026
counterparty primary

Microsoft FY2027 Q1 results and call ↗

Quarter to 30 Sep 2026: cash PP&E 39.4bn; capex incl. finance leases 45.2bn; calendar-2026 capex about 170bn; Maia 200/300 tokens-per-watt claims; commercial RPO 677bn.

Published Oct 28, 2026 · Reviewed Nov 3, 2026
counterparty primary

Alphabet Q3 2026 results and call ↗

Q3 purchases of property and equipment 51,355m; 2026 capex range 195–205bn; 2027 to increase significantly; TPU 8t/8i, nearly 100 TPU customers, TPU system sales to ramp in 2027; Cloud revenue 30,709m, backlog 561bn.

Published Oct 28, 2026 · Reviewed Nov 3, 2026
counterparty primary

Amazon Q3 2026 results ↗

Q3 purchases of property and equipment 77,259m; TTM free cash flow -42,357m; 2026 cash capex about 220bn; AWS revenue 46.0bn, backlog 578bn; custom chips business growing at triple-digit rates, Trainium commitments up over 20x.

Published Oct 29, 2026 · Reviewed Nov 3, 2026
counterparty primary

Meta Q3 2026 results ↗

Q3 capex incl. finance-lease principal 34.9bn; 2026 capex guidance 135–145bn; Q3 free cash flow -2.12bn.

Published Oct 21, 2026 · Reviewed Nov 3, 2026
counterparty primary

Oracle FY2027 Q1 results ↗

Quarter to 31 Aug 2026: capex 28,499m; free cash flow -5,396m; RPO 664bn; 20bn at-the-market equity sale.

Published Sep 10, 2026 · Reviewed Nov 3, 2026
counterparty primary

Broadcom FY2026 Q3 results ↗

Quarter to 2 Aug 2026: AI semiconductor revenue 16.7bn (+221%); Q4 guide 21.7bn. Call remarks on FY2027 AI revenue (about 115bn) and six XPU customers come from a third-party automated transcript and are treated as unverified.

Published Sep 2, 2026 · Reviewed Nov 3, 2026
primary

NVIDIA Q4 and FY2026 results release ↗

Q4 FY2026 revenue 68,127m and GAAP operating income 44,299m; comparative Q3 FY2026 revenue 57,006m and GAAP operating income 36,010m.

Published Feb 25, 2026 · Reviewed Nov 3, 2026
base-rate primary

Applied Materials FY2002 Form 10-K ↗

Net sales 5,062m in FY2002 against 7,343m in FY2001 (−31%); FY2001 restated for SAB 101.

Published Jan 8, 2003 · Reviewed Nov 3, 2026
counterparty primary

Meta prospectus supplement: forward sale of Class A shares ↗

40m shares (46m with option) at 625; expected net proceeds about 24.9bn (28.7bn) on settlement at Meta’s choice by 19 November 2027; no proceeds at closing; general corporate purposes.

Published Oct 27, 2026 · Reviewed Nov 3, 2026
counterparty primary

Meta Form 8-K: 6.50% mandatory convertible preferred ↗

250m depositary shares, each 1/20 of a 1,000 liquidation-preference share (about 12.5bn liquidation preference; proceeds not confirmed); mandatory conversion about 15 August 2029 (research agent read).

Published Oct 29, 2026 · Reviewed Nov 3, 2026
primary

House committee chairs request Jensen Huang briefing ↗

Official release and stated reply/briefing dates. Request is confirmed; underlying allegations and attendance are not established. Linked letter PDF unavailable through research tool.

Published Nov 5, 2026 · Reviewed Nov 6, 2026
primary

Tench Coxe Form4 filed6November2026 ↗

Transaction4November:500,000 common shares, codeG, no consideration. Plan footnote dated19March2026. Existing share transfer, not issuer financing.

Published Nov 6, 2026 · Reviewed Nov 7, 2026
primary

NVIDIA Board announces independent investigation ↗

Complete board statement, paragraphs1-5: independent directors and outside counsel oversee inquiry into reported foreign intelligence targeting. Board supports current leadership and says it has no reason to believe national security or company secrets were compromised. This is the board’s statement, not an independently established clearance or investigative finding.

Published Nov 8, 2026 · Reviewed Nov 8, 2026
primary

NVIDIA CEO message supporting board inquiry ↗

Complete issuer-published message; investment-relevant final paragraphs: CEO supports and will cooperate with independent inquiry and states continuing commitment to company. No quantified operating or financing revision; personal details do not establish a business cash loss.

Published Nov 8, 2026 · Reviewed Nov 8, 2026
primary

NVIDIA/Rebellions/SK Telecom NVLink Fusion collaboration ↗

Complete release, deployment summary and qualifications. No contracted units, price, revenue, margin or deployment date disclosed. Custom NPUs and NVIDIA infrastructure participation have different economics.

Published Nov 8, 2026 · Reviewed Nov 9, 2026
primary

TSMC October 2026 revenue report ↗

Complete revenue paragraph and official monthly table: October NT$586.89bn, +14.7% month on month and +59.7% year on year; January-October NT$4,485.62bn. Diversified foundry sales do not identify NVIDIA allocations or customer cash returns.

Published Nov 10, 2026 · Reviewed Nov 11, 2026
primary

NVIDIA announces CEO transition ↗

Complete release: Kress interim CEO and continuing CFO, Jones board chair, permanent-successor search, stated continuity and confirmed17November results. No financial impact amount disclosed.

Published Nov 12, 2026 · Reviewed Nov 12, 2026
primary

NVIDIA transition Form8-K, Item5.02 ↗

Full filing. Earliest event11November; resignation accepted and appointments effective then, disclosed12November. No quantified severance/award revision or new earnings figures.

Published Nov 12, 2026 · Reviewed Nov 12, 2026
primary

Moolenaar statement on NVIDIA leadership transition ↗

Complete public statement read through original HTML after reader403. Political call for changed practices/export-control legislation, not an enacted restriction; filing controls the legal description of the departure.

Published Nov 12, 2026 · Reviewed Nov 12, 2026
primary

House Foreign Affairs, China Select and Intelligence chairmen: second request to Huang, NVIDIA general counsel and BIS under secretary ↗

Complete release and all three one-page letters read (scanned PDFs, read as images). Requested closed-door transcribed interviews: Huang 1 December, general counsel Teter 2 December; Kessler hearing 3 December 10 a.m. Written acceptance requested by 8 p.m. EST 13 November, with a stated subpoena otherwise. Requests and a threatened subpoena, not an issued subpoena, finding or enacted restriction.

Published Nov 12, 2026 · Reviewed Nov 12, 2026
primary

Singapore MHA: investigations into Alice Huang and Megaspeed ↗

Original statement paragraphs2–9: domestic false-declaration investigation, insufficient evidence concerning Huang, warning for inaccurate information, ongoing Megaspeed inquiry; no espionage investigation by Singapore.

Published Nov 12, 2026 · Reviewed Nov 13, 2026
primary

Mark Stevens Form4 filed2September2026 ↗

Original ownership XML: transaction dates31Aug–2Sep, share amounts and weighted prices; aff10b5One0; trust and direct holdings.

Published Sep 2, 2026 · Reviewed Nov 13, 2026
primary

Mark Stevens Form4 filed22September2026 ↗

Original XML:18September transactions;1,366,000 shares; aff10b5One0; remaining trust and direct holdings. No inference of motive or unlawful trading.

Published Sep 22, 2026 · Reviewed Nov 13, 2026
primary

Kistenmacher v NVIDIA, filed securities complaint ↗

Filed complaint pp.1–3 and17 inspected directly. Case5:26-cv-13448; requests class certification and damages against NVIDIA and Huang, without a quantified award. Allegations are not findings; no current docket status or insurance recovery is inferred.

Published Nov 9, 2026 · Reviewed Nov 15, 2026
primary

Vanegas derivative complaint on behalf of NVIDIA ↗

Filed complaint pp.1–3 and33–34 inspected directly (court page labels, PDF omits final verification page36). Case3:26-cv-13596; individual defendants and NVIDIA as nominal defendant. Requested recovery goes to NVIDIA; legal fees, company indemnification and any securities-case loss remain distinct. No judgment or quantified net company payment.

Published Nov 11, 2026 · Reviewed Nov 15, 2026
secondary

The Insurer: reported NVIDIA Side A D&O tower ↗

Public article read. Unnamed-source report concerns executive Side A cover; actual policies and company reimbursement rights not inspected. No model insurance offset, court award or loss probability is imported.

Published Nov 14, 2026 · Reviewed Nov 15, 2026
primary

Anthropic amended S-1,10November2026 ↗

Offer summary p.23; MD&A liquidity/commitments pp.174–182; multi-chip compute pp.211–213; concurrent private placement p.347; cash statement F-12 and subsequent events F-67. NVIDIA USD10bn subscription is conditional on the IPO, not closed cash. Relate to the up-to10bn partnership announced18November2025 before adding to the July25bn pool; exact pool composition remains unknown. Public issuer PDF downloaded without credentials; web PDF URL redirects to European landing, so SEC text is the stable reference. No universal three-day migration claim verified.

Published Nov 10, 2026 · Reviewed Nov 15, 2026
primary

Anthropic: Microsoft/NVIDIA strategic partnerships,18November2025 ↗

Original investment/compute partnership announcement: NVIDIA up toUSD10bn, predating the July2026 issuer commitment pool. This explains potential fulfillment/overlap; it does not establish the exact composition of that pool.

Published Nov 18, 2025 · Reviewed Nov 15, 2026