NVDANVIDIA Corporation

NVDA — dossier, 2026-09-21

Slot names-ai-capex-1, name pass. Drafted unattended by the cloud research routine following .claude/skills/stock-sourcing-pipeline/SKILL.md. Nothing here is human-verified. Every figure carries the document it was read from; anything from background knowledge is marked [background]. Never edited after today — corrections go in scorecard.md's log.

Price reference: $222.27 close 2026-09-18 (ROIC.ai). Shares outstanding 24,100,000,000 as of 2026-08-21 (EDGAR dei:EntityCommonStockSharesOutstanding, Q2 FY27 10-Q cover). Implied market capitalisation ≈ $5.36tn.

---

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), cluster ai_semis, coverage: sec_domestic, cik 1045810. No Tier 0 promotion: queue.json promoted was empty at the start of this slot.

---

S2 — Kill test

All figures from EDGAR XBRL companyconcept (free, unmetered) unless noted. Fiscal years end late January; FY2026 ended 2026-01-25.

Share-count note: the FY2025 jump is the 10-for-1 split, not issuance. On a like-for-like basis diluted shares went 24.80bn → 24.51bn → 24.29bn (Q2 FY27) — a falling count, funded by buybacks. Dilution is not a risk here; it is the opposite of AMD's position.

Net debt / EBITDA: not a meaningful constraint. FY26 operating income alone is $130bn against $7.5bn of long-term debt. Even after the June-2026 issuance (below), gross debt of $33.4bn is roughly one quarter's operating income.

Analyst coverage count: not captured this pass — no source on the ladder returns it without spending an Equibles call better used elsewhere. Marked thin.

Verdict — survives or dies: NVIDIA does not die on its operating numbers. Revenue grew 65% on a $130bn base, gross margin is 71%, operating margin 60%, and free cash flow was ~$97bn in FY26 against $6bn of capex. Nothing in the income statement is close to a kill. The kill test has moved to the balance sheet, and the balance sheet changed character in two quarters:

Liquid position: $56.6bn of cash and marketable debt securities against $33.4bn of debt → net cash ≈ $23.2bn, before crediting $93.9bn of marketable equity and non-marketable securities. That investment book is now 29% of total assets and did not exist at this scale a year ago.

---

S3 — Filings, at gate-minimum depth

Read this pass: FY26 10-K (Item 1A risk factors and Item 7 MD&A only), Q2 FY27 10-Q (Item 2 MD&A delta), DEF 14A 2026-05-12 (ownership and 5% holders), the June-2026 424(b)(5), and the Q2 FY27 earnings call. Thin against the skill's own bar: one transcript, not two — the Equibles budget for this run was 25 calls across three names and 22 were spent; Q1 FY27's call was the item dropped. Not read: the second 10-K, older 10-Qs, comment letters, critical audit matters, auditor tenure and fees.

The three biggest risks management itself discloses

All three are management's own words, located by exact substring in the filing text and copied with their locator. They are not this run's ranking of what matters — they are what the company puts first, quantified.

1. Customer concentration — two customers are 36% of revenue. 10-K FY26, Item 1A, "We receive a significant amount of our revenue from a limited number of partners and distributors…":

For fiscal year 2026, sales to one direct customer represented 22% of total revenue and sales to another direct customer represented 14% of total revenue, all of which were primarily attributable to the Compute & Networking segment.

The cluster pass read the Q1 FY27 10-Q and found four direct customers at 60% of H1 FY27 revenue and five at 70% of receivables. The FY26 10-K figure is the audited annual version of the same fact.

2. Export control and the loss of China. 10-K FY26, Item 7 MD&A, "Recent Developments, Future Objectives and Challenges":

In April 2025, the USG informed us that a license is required for exports of our H20 product into the China market. As a result of these requirements, we incurred a $4.5 billion charge in the first quarter of fiscal year 2026 associated with H20 for excess inventory and purchase obligations, as the demand for H20 diminished.

Updated in the Q2 FY27 10-Q: a further $0.4bn charge on H200 in H1 FY27, shipments under the H200 licences "account for less than 1% of Data Center revenue in our most recent quarter," and a 25% US import tariff that the company states it has been unable to pass on. On the Q2 FY27 call the CFO said there is no China data center compute revenue in the forward outlook at all. China has been written to zero in the guide — which removes it as a downside and as an upside.

3. The customers cannot fund what they want to buy. 10-Q Q2 FY27, Item 2 MD&A:

We believe AI clouds and AI model makers have significant demand for training and inference compute and currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity to secure the AI infrastructure necessary to grow.

This is the company stating, in a filing, the precise condition the cluster pass identified from the outside — and then describing what it is doing about it, which is the next section.

Proxy skim (DEF 14A filed 2026-05-12, table as of 2026-03-23)

• Directors and executive officers as a group (14 persons): 957,313,517 shares, 3.94%.
• 5% holders: BlackRock 7.43%, Vanguard Capital Management 7.31%.
• Individual director lines are small by comparison (Mark A. Stevens 34.1m, A. Brooke Seawell 2.5m, both marked "*" = less than 1%), so the 3.94% is overwhelmingly the founder-CEO.
• Compensation structure and what the bonus is paid on: not extracted this pass — thin.

Insider pattern (12 months to 2026-09-21)

Zero open-market purchases by any insider (Equibles GetInsiderTransactions NVDA, transactionType: Buy, fromDate: 2025-09-21 → "No insider transactions found"). The Form 4 feed is busy — 556 Form 4s on the current EDGAR submissions page, newest 2026-09-18 — but none of it is conviction buying. Sales and award mechanics were not itemised this pass.

Red-flag sweep

• Going concern: no flag. Newest examined filing is the 10-Q filed 2026-08-26 (Equibles GetGoingConcernStatus); coverage there starts 2026-05-20, so earlier filings are unexamined.
• Executive changes (8-K Item 5.02, via Equibles): Chief Accounting Officer handover — Donald Robertson retired effective 2026-05-04, Scott Gawel appointed from the same date. EVP Worldwide Field Operations Ajay K. Puri retiring, Nicholas Parker starting ~2026-08-24. Director Persis Drell resigned mid-term, effective immediately, 2026-01-20, including from the Compensation Committee; the 8-K gives no reason. Suzanne Nora Johnson added to the Board and Audit Committee 2026-07-13, board size ten → eleven. None of this is an accusation. A CAO change and a mid-term director resignation inside the same year that the balance sheet transformed are simply worth knowing.
• Auditor change / restatement / NT 10-K: not run this pass — the efts.sec.gov full-text sweep was dropped for time. Thin; flagged for the next pass.
• Customer concentration: carried from the 10-K (above) and the cluster file.

---

The finding this pass exists to report: NVIDIA is now financing its own demand

The cluster pass of 2026-09-20 split the complex into sellers of the shortage, which self-fund, and debt-funded buyers, which do not. That split no longer holds cleanly at the top. Four things, all from NVIDIA's own filings or its own call, dated within the last four months:

1. $25.0bn of notes, June 2026. Seven tranches maturing 2028 through 2056 — $3.5bn 4.250% 2028, $3.5bn 4.350% 2029, $4.0bn 4.500% 2031, $3.5bn 4.750% 2033, $4.0bn 4.950% 2036, $3.0bn 5.550% 2046, $3.5bn 5.625% 2056 — off the S-3ASR shelf filed 2025-05-28.

We estimate that the net proceeds from this offering will be approximately $24.9 billion, after deducting underwriting

…for "general corporate purposes, including the repayment and refinancing of outstanding notes." Long-term debt went $7.5bn → $32.4bn. And:

The indenture does not limit the amount of debt we may incur.

Baker's stated bubble sequence is "issue debt → cancel dividends → stop buybacks." NVIDIA has now done the first while doing the opposite of the third — $39.0bn of buybacks in H1 FY27 with $99.3bn of authorisation remaining as of 2026-07-26, plus a $0.25 quarterly dividend. That is a different shape from Oracle's, and the cluster file's Oracle framing should not be transplanted onto it. It is also not the shape of a company funding everything from cash flow.

2. A $105bn guarantee cap behind one customer. From the Q2 FY27 10-Q:

In August 2026, we entered into guarantees with SB Energy Corp. to provide credit support on the land, power, and shell buildout at SB Energy's PORTS Technology Campus in Pike County, Ohio, covering leases for approximately 4.25 gigawatts of IT load.

The campus will exclusively host our compute under 20-year leases to OpenAI, subject to limited exceptions, with our obligation capped at $105 billion in the aggregate

Obligations begin as each of nine data centers is placed in service, expected from fiscal year 2029, and terminate if OpenAI reaches a satisfactory credit rating. NVIDIA holds an option on roughly 3.8 GW more. The exposure is capped, staged and conditional — and it is NVIDIA carrying OpenAI's lease credit.

3. A $36bn take-or-pay back to the AI clouds.

Our commitments, which are typically six years in duration, totaled $36 billion as of July 26, 2026, and decrease as capacity is used by third-party customers or by us for our research and development efforts.

The clouds buy NVIDIA hardware; NVIDIA commits to buy cloud services back. The 10-Q states the clouds "can unilaterally stop providing to us and sell to third-party customers at more advantageous rates" — the optionality runs against NVIDIA.

4. >$500bn of third-party capital being organised.

In August 2026, we entered into memorandums of understanding with several large capital providers to establish independent financing platforms designed to mobilize more than $500 billion of third-party capital over time to support the deployment of AI infrastructure.

The call named six: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR. The 10-Q names none and says these "may not lead to definitive agreements."

Management's own size for the circularity. From the Q2 FY27 call, the CFO said NVIDIA has invested nearly $50bn in the frontier AI labs, addressed the objection directly — "we know some will call this circular financing. We see it differently" — and then quantified it: demand from the AI labs for which NVIDIA expects to leverage its balance sheet contributes toward roughly a quarter of the business next year. (Machine transcript; the wording above is what the transcript renders, and that transcript mangles proper nouns elsewhere — "SB Energy" as "SoftBank Energy", "Groq" as "Grok", "Jensen Huang" as "Jensen Wong". The quantum — a quarter of next year's business — is the load-bearing fact and it is not corroborated by a filing read this pass. Treat it as management commentary, not as a filed figure.)

---

S5 — Ownership

Who is on the other side: essentially nobody. Short interest is ~1.2% of the float with 2 days to cover, and there is no visible activist or concentrated skeptic. The register is index money and the founder. The absence of a short case in the tape is not evidence the thesis is safe — it means the market is not offering a second opinion at this price.

---

S8 — Valuation and sizing (model; both decisions are ⏳ OPEN (user))

Out-year model. Revenue and margin inputs are management's own guidance where it exists, so the exercise is "what is priced against what management says", not an independent forecast.

EV: market cap ~$5.36tn less ~$23.2bn net cash → EV ≈ $5.33tn, ≈ 13.2x FY2027E revenue and ≈ 7.7x FY2028E revenue. The $93.9bn investment book is excluded from EV; crediting it would take EV to ~$5.24tn.

• Bear: FY28 growth comes in at 30% rather than 70% (supply eases but demand does not follow, or one of the two 20%+ customers slows), gross margin holds at the 71–72% trough rather than recovering. Revenue ~$525bn, EPS ~$11.5, and the multiple de-rates because the growth story is what supports it. A 40–50% drawdown is arithmetically ordinary from here.
• Base: management's guide is met. ~$16.4 FY28 EPS, and the stock is at 13.6x that today.
• Bull: the ~100% demand growth management says it sees converts as supply unlocks. Revenue ~$800bn+, EPS ~$19–20.

The honest summary of the valuation: on management's own numbers this is not an expensive stock — it is a cheap one. That is precisely the tell worth thinking about. The market is pricing meaningful probability that the FY28 guide does not happen, and the reason is visible in S3: a rising share of the demand is being underwritten by the seller.

Liquidity: average daily volume 139,118,326 shares (FINRA, 2026-08-31 settlement) at ~$222 ≈ $31bn traded per day. No sizing constraint at any plausible personal-account size.

⏳ Buy price: OPEN (user). Size: OPEN (user). Both are owner-only under the skill.

---

S9 — Monitoring

• Next earnings: Q3 FY2027 ends ~2026-10-25; the results call has historically landed in the second half of November. Derived from the reporting cadence, not from a sourced calendar — ROIC.ai's earnings calendar requires a paid plan and returned an upgrade notice. Thin.
• Filing feed: the Tier 0 EDGAR sweep (watch_edgar.py, 06:40 UTC) covers cik 1045810 and will surface the 10-Q, any 8-K and any further 424B5 takedown without a model call.
• Leading indicator between quarters: the supply and capacity commitments line in Note 10 of each 10-Q — $119bn at Q1 FY27, $279bn at Q2 FY27. It moves a quarter ahead of revenue and it is a filed number rather than commentary. Second: days sales outstanding, now 60 days on extended terms.
• Trigger that converts this to a buy: the owner's, not this routine's. What this routine can say is the condition that would remove the central objection — two consecutive quarters in which revenue grows at or above guide while receivables, the investment book and the guarantee cap stop growing faster than revenue. That would establish the demand is being paid for rather than financed. Until then the FY28 multiple is cheap for a reason that is disclosed.

---

Access and connector notes for this run

• SEC EDGAR: reachable, with the required User-Agent: ddboard-research gorkem@celebico.com. data.sec.gov XBRL, data.sec.gov/submissions and www.sec.gov/Archives all returned 200 via plain curl, so filings were read directly rather than through WebFetch.
• Equibles: reachable. 22 calls across the three names of this slot, against the 25-call cap.
• ROIC.ai: reachable for prices; list_earnings_calendar is gated behind a paid plan.
• EDINET DB / eu_data: not needed for three sec_domestic names; not called.
• Not called, per the skill: Alpha Vantage, Bigdata.com, FMP.

Process support, not investment advice.