NVIDIA

Refresh delta (since 2026-07-10)

What changed — the six material moves:

1. NVIDIA escalated from equity in its customers to guaranteeing their debt — and the market repriced it violently. Reports on 2026-07-27 that NVIDIA is in talks to provide ~$250B of financing guarantees for OpenAI to lease capacity at a 10GW campus in Piketon, Pike County, southern Ohio (SoftBank's energy arm developing, on a former uranium-enrichment site; total project >$500B), plus a separate ~$350B of financing for GPU purchases under discussion. The backstop covers lease and construction debt, not the chips. Michael Burry: the arrangement has NVIDIA guaranteeing OpenAI's spending on NVIDIA's own chips — "around and around we go". NVDA fell ~5% on 2026-07-27, and — the sharpest tell — the cost of protecting NVIDIA's debt against default for five years rose by the most on record that day.
2. A $1.3T semiconductor drawdown, and NVIDIA lost the largest-company title. More than $1.3T of market value left the largest chipmakers in a few sessions; NVIDIA alone erased $238B, SK Hynix $176B, Samsung $173B, Micron $113B. On 2026-07-28 Apple passed NVIDIA as the world's most valuable company — Apple $4.95T vs NVIDIA $4.76T, NVDA at $196.03 — explicitly on the narrative that Apple "avoided the capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. pitfalls". Three named catalysts stacked: Meta Compute (2026-07-01, Meta building a cloud unit to sell surplus AI capacity — a hyperscaler turning into a neocloud competitor); SK Hynix diverting planned HBM4 capacity to DDR5 (~90% DDR5 margins); and the OpenAI backstop.
3. The tape then recovered — and through the prior dossier's level. NVDA closed $223.96 on Fri 2026-08-07 (+2.27% on the day), i.e. +10.4% vs the prior pass's $202.78, +14.3% off the 2026-07-28 low of $196.03, and only ~4.9% below the $235.47 ATH. Market capMarket capitalisationThe share price multiplied by the number of shares. What the market says the equity is worth. ~$5.4T.
4. The multiple did not expand — earnings estimates did. Consensus FY27 EPS moved $8.06 → $8.79 (+9.1%) while the price rose 10.4%, leaving the forward multiple ~25.5x, essentially unchanged from the prior pass's ~25.2x; an independent NTM read puts forward P/E at 24.46x as of 2026-08-07. Consensus PTConsensus price targetThe average of what published analysts think the share price should be. An opinion poll, not a forecast. ~$302.83 across 61 analysts, 58 of 61 at buy/strong-buy, range $180–$500; KeyBanc raised $310 → $330 on 2026-07-14.
5. A new product line the prior dossiers did not have: the Vera CPU. On the FY27-Q1 call management put ~$20B of standalone CPU revenue visibility this year against a claimed $200B TAM NVIDIA has never addressed — Jensen: "my sense is that we will be supply constrained throughout the entire life of Vera Rubin". Also newly grounded from the call: Anthropic went from ~zero NVIDIA coverage to a named strategic partner, and total supply including inventory purchase commitments and prepaids reached $145B — which reconciles cleanly to the filing as $119B of commitments + $25.8B of on-hand inventory.
6. The customer base is levering up to buy. Hyperscaler incremental annual debt rose from 9% of capex in FY24 to 32% LTM by mid-2026, and equity issuance returned — Alphabet priced an $84.75B equity raise in June 2026. 2026 capex still ~$725B across the big four with 2027 modelled >$1T by Evercore and BofA, but Alphabet's Q2 report was panned and sparked the sell-off on 2026-07-28. Demand is still rising; the funding mix behind it has shifted decisively from cash flow to debt.

What held. The structural thesis stands and the operating moat widened. No operating datum deteriorated in the window — there was no print. Gross margin guidance held at 74.9–75% ±50bp for Q2 and mid-70s for the full year; the $500B Blackwell+Rubin booking figure was reiterated, and management's GTC framing is now a $1T visibility number across Rubin + Blackwell platform revenue. Vera Rubin is in full production with the first rack live at Microsoft Azure, shipping H2 2026. Micron is in high-volume HBM4 production for Vera Rubin. China remains a zero-revenue line — licences exist (~$10B approved) but a Commerce official told Congress on 2026-07-14 shipments have been "very few", and Kress confirmed no China DC-compute revenue. The prior pass's core call — revenue quality is migrating to the balance sheet — is more true again, and this time it migrated one rung further out the risk curve: from equity marks to credit guarantees.

---

Phase A — Understand the Business

How does it actually make money, and is the moat real?

Company Overview (new product line + new reporting framework)

The shape is unchanged; two things are genuinely new. NVIDIA still sells rack-scale AI systems (GPU + CPU + NVLink fabric + NICs/DPUs + Ethernet switching + CUDA), fabless via TSMC/ODMs, to a handful of buyers spending tens of billions each. The FY2026 audited baseline (revenue $215.9B, +65%; net income $120,067M; GM 71.1%) and the FY27-Q1 print (revenue $81,615M, +85%; GM 74.9%; operating margin 65.6%; GAAP diluted EPS $2.39, operating EPS ex-marks ≈ $1.85) carry unchanged from the prior pass.

New #1 — the Vera CPU is a third leg, not a peripheral. Management disclosed visibility to nearly $20B of total CPU revenue this year, standalone (explicitly not counted inside Vera Rubin rack revenue), against a claimed $200B TAM NVIDIA has never addressed. The strategic logic is agentic AI: Jensen's argument is that agents run tool-use, browsing and search on CPUs, so a world going from "a few hundred thousand" agents to "a few billion" needs CPU capacity per agent. Versus x86, Vera is claimed at 1.5x performance/core, 2x performance/watt, 4x density/rack. Treat the $200B TAM as a marketing frame, but the $20B in-year visibility is a management revenue statement, and it is incremental to every prior model.

New #2 — the reporting framework changed. NVIDIA moved to platform reporting (Hyperscale / ACIE / Edge Computing) and posted 9 quarters of recast history. Any model built on the old Data-Center-vs-Gaming split needs re-basing — see Lens 4.

Customers — concentration unchanged and extreme. FY27-Q1: three direct customers = 21%, 17%, 16% (= 54%) of revenue, all Compute & Networking; three = 30/18/16% of AR. One quality improvement from the call: DSO fell to 45 days on favourable collection timing, with management guiding back to the mid-50s in Q2 — a rare instance of management pre-announcing a working-capital deterioration, which is a credibility marker.

The contract book, and the loop — now escalated:

• $119B manufacturing/supply/capacity commitments ($95B due in the rest of FY2027), "in certain instances… cancellable, able to be rescheduled, or adjustable"
• $30B multi-year cloud-service commitments (NVIDIA renting capacity back, "primarily to support R&D"); $6B other vendor commitments
• $145B total supply including on-hand inventory and prepaids (management framing)
• $27B investment commitments; $3.5B facility-lease guarantees; $32.4B of future data-centre leases commencing FY27–FY33
• NEW: a reported ~$250B guarantee of third-party lease/construction debt, plus ~$350B of GPU-purchase financing under discussion — neither on the balance sheet, neither in any filing yet

Analyst flag — the loop changed category this quarter. Through the prior two passes, "circular financing" meant NVIDIA held equity in customers whose gains flowed through the P&L. The reported OpenAI structure is different in kind: NVIDIA would be guaranteeing the debt of a data-centre lessee whose only economic purpose is to buy NVIDIA compute. Equity stakes cap loss at the investment; a guarantee is unbounded until called and converts a demand problem directly into a credit problem on NVIDIA's own balance sheet. That is precisely why the CDS move — not the equity move — is the signal worth carrying forward. Highest-priority stress-test item; see Lens 10 and 13.

Supply Chain (carried by reference — three delta rows)

The named-link map carries in full from the previous dossier (ASML EUV → TSMC N3/A16 → TSMC CoWoSCoWoSTSMC’s method of packing a processor and its memory onto one carrier so they sit close together. Supply of it has been a hard limit on how many AI chips can be built. → NVIDIA → Foxconn/Quanta/Wistron/Supermicro/Dell → hyperscalers/neoclouds/labs; HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips. from SK Hynix > Samsung > Micron). Three rows moved materially enough to record, because one of them was a named catalyst of the $1.3T selloff:

Net: the constraint is no longer NVIDIA-specific execution — it is an industry DRAM allocation fight NVIDIA can lose on price. The Rubin Ultra datum is the one to carry: DRAM scarcity is now writing the specification of the chip, not merely its cost. A memory-density retreat (288GB → 192GB) in the same window AMD ships MI450 at 432GB is a competitively awkward place to be (Lens 3). Offsetting, the SK Group partnership (2026-07-24, LOIs, headline ">$500B over a number of years") includes a long-term SK Hynix next-gen memory supply agreement and a 2GW SK Telecom AI data centre on Vera Rubin + HBM4, first facility 2027. Read the $500B as an LOI-stage aspiration, not a contract — the parties signed letters of intent, and the figure is an estimate of multi-year potential, not a booking.

Competitive Advantages (Moats) (challenge sharpened materially)

Moat ranking unchanged (1. full-stack rack-scale co-design; 2. CUDA lock-in; 3. annual cadence; 4. memory/packaging priority allocation), with filing-grade proof intact: Compute & Networking booked $53,335M operating income on $74,550M revenue — a 71.5% segment operating margin. Share is still overwhelming: NVIDIA holds upward of 95% of the data-centre GPU market vs AMD ~4.5%.

Two moat extensions from the call:

• Anthropic went from ~zero to a named strategic partner. Jensen: "our coverage of Anthropic has been largely zero until just recently… we are gaining share tremendously fast in inference." NVIDIA is securing Anthropic capacity across Azure, AWS, CoreWeave and xAI. This is a genuine share gain at a frontier lab that was previously TPU/Trainium-weighted.
• Networking is the quiet compounder. Data-centre networking revenue $15B, nearly tripled Y/Y, vs data-centre computing $60B, +77% Y/Y; InfiniBand grew >4x Y/Y on XDR. Networking growing ~2x faster than compute is the clearest evidence the unit of sale really has moved from chip to room.

What the delta sharpens on the challenge side — AMD converted claims into named gigawatts:

• AMD MI450 / Helios now has anchor customers: a 6GW deal with Meta, up to 2GW with Anthropic (first GW deploying H1 2027), and Microsoft deploying Helios on Azure for frontier-model inference. MI450 ships H2 2026 at 432GB HBM4 against Rubin's 288GB — and possibly against a 192GB Rubin Ultra (Lens 2).
• The Anthropic datum cuts both ways, and this is the sharpest single fact in the pass. NVIDIA holds a ~$10B equity stake in Anthropic and just named it a strategic partner — while Anthropic simultaneously committed up to 2GW to AMD. Taking NVIDIA's capital does not buy exclusivity. That is direct evidence against the strongest form of the circular-financing bull argument ("the equity locks in the customer").
• Meta Compute turns a customer into a competitor's supplier. Meta building a cloud unit to resell surplus capacity compresses neocloud economics — and NVIDIA's equity book is concentrated in neoclouds (CoreWeave, IREN).

Honest vulnerability, unchanged: the SemiAnalysis 90%+ → 20-30% inference-share-by-2028 call still frames the bear case, and this window added evidence (MI450 anchor customers, Meta Compute) without yet moving reported share. Bargaining power over customers is strong but visibly peaking; over suppliers it weakened this window — SK Hynix reallocating to DDR5 is a supplier optimising against NVIDIA's roadmap.

Segments (reporting framework changed; re-base required)

By reportable segment (FY27-Q1), unchanged:

By market platform — NEW framework, now transcript-grounded:

• Hyperscale $38B — ~50% of data centre, +12% Q/Q
• ACIE (AI Clouds / Industrial / Enterprise) $37B — +31% Q/Q, with AI-cloud revenue more than tripling Y/Y
• Edge Computing $6.4B — +10% Q/Q, +29% Y/Y (Blackwell workstations strong; consumer fell on higher memory and system prices)
• Cross-cut: data-centre computing $60B (+77%) vs networking $15B (~3x)

The mix signal that matters: ACIE is growing ~2.6x faster sequentially than Hyperscale (+31% vs +12%). NVIDIA's marginal dollar is increasingly coming from AI clouds and labs — the exact cohort NVIDIA finances — not from the cash-rich hyperscalers. That is simultaneously the bull case (demand broadening beyond five buyers) and the bear case (the fastest-growing cohort is the least self-funding). Consumer Edge falling on memory prices is also the first place the DRAM squeeze shows up in NVIDIA's own P&L.

By geography (FY27-Q1): United States $63,769M (78%), Taiwan $12,006M (ODM build location), China $4,550M (5.6%, −53% Y/Y), Other $1,290M. Segment bottom line, unchanged: a ~92%-Data-Center company in the costume of a diversified semi — with the investable questions now (a) whether debt-funded hyperscaler capex compounds toward $1T, (b) whether the 54%/top-3 concentration plus the financing loop hold, (c) whether HBM4 allocation and mid-70s margins survive 2027 DRAM scarcity and MI450.

---

Phase B — Measure Performance

What do the numbers and the tape say?

Earnings Result (no new print; guidance and the Q2 setup are the content)

No print landed in the delta window. FY27-Q1 (ended 2026-04-26) remains the latest reported quarter; FY27-Q2 reports after the close on 2026-08-26, and its quarter ended ~2026-07-26.

The Q2 setup, now precise:

• Company guide: ~$91.0B ±2%, GM 74.9% GAAP / 75.0% non-GAAP ±50bp, GAAP opex ~$8.5B, and — restated on the call — zero China data-centre compute revenue in the outlook
• Consensus: revenue ~$91.8B (range $90.3–96.7B), EPS ~$2.08 (range $2.03–2.20); one source carries $93.61B / $2.13
• Full-year FY27 consensus EPS $8.79, +92% Y/Y
• Management also guided full-year opex growth in the "upper forties" percent and reiterated returning ~50% of FCF to shareholders this year

A timing point worth more than it looks. FY27-Q2 ended ~2026-07-26; the circular-financing selloff ran 2026-07-27 to 07-29. The equity marks in the 26 August print are therefore struck before the drawdown — so the ~23%-of-GAAP-EPS mark tailwind most likely persists in Q2, and any reversal lands in Q3 (quarter ending late Oct). Anyone reading the Q2 GAAP number as evidence the reflexivity risk is benign will be reading a pre-selloff mark. This is the single most useful non-consensus observation in the pass.

The FY27-Q1 quality correction carries unchanged and remains the anchor: of $69,903M pre-tax income, $15,929M (22.8%) was non-operating mark-to-market gains on equity stakes — GAAP diluted EPS $2.39 vs operating EPS ex-marks ≈ $1.85.

Balance-sheet flags (now fully in our figures — backfilled this pass): inventory $25,797M (+$4.4B Q/Q, raw materials +75% seq); AR net $40,710M; PP&E net $12,403M; total assets $259,474M; payables $13,097M; total debt $8,470M against cash $13,237M + marketable debt securities $37,098M (net cash ~$41.9B); total equity $195,474M; D&A only $997M/qtr; interest expense $102M. Record FCF ~$48.6B (OCF $50,344M − $1,757M capex), which management stated as $49B, up from $35B in Q4.

Capital allocation — one conflict resolved. Dividend raised $0.01 → $0.25 per share (approved 2026-05-18, record 2026-06-04) and $80.0B added to the $38.5B remaining authorisation = $118.5B; $20.2B repurchased in the quarter.

Market reaction (the delta tell — and it inverted). The prior pass caught the stock down ~12% in 30 days; this pass catches the round trip: $202.78 (Jul 9) → $196.03 (Jul 28, the Apple overtake) → $223.96 (Aug 7), i.e. +10.4% over the window despite the worst circular-financing scare to date. "Beats get faded" has become "the complex sells the narrative and then buys the earnings revisions".

Earnings Calls (sentiment trend) (UPGRADED from web-reconstructed to primary transcript)

This lens materially improves this pass. Both prior dossiers flagged that no transcript was compiled and Lens 6 was web/filing-reconstructed. transcripts/2026-q2.md (call of 2026-05-20, Jen-Hsun Huang / Colette Kress / Toshiya Hari) is now on the shelf. No new call occurred in the delta window — the next is 2026-08-26 — so the sentiment arc is re-read at higher resolution rather than extended.

The tone, read directly: the prior passes characterised the arc as promotional → structural/declarative. The primary text confirms and sharpens that. Kress opens on records ("revenue, operating income, and free cash flowFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. exceeding our prior records… third consecutive quarter of year-over-year acceleration and the fourteenth straight quarter of sequential growth"). Jensen's close is a ranked list — "NVIDIA is the only platform that runs every frontier AI model… Second, we are in every hyperscale cloud" — the rhetoric of an incumbent defending a position, not a challenger selling one.

Three call passages that change how the risk lenses read:

1. Management pre-empted the GPU-depreciation bear thesis directly: "customers are generating profitable revenue beyond the depreciable life of their GPUs". This is the most-cited structural bear argument (Burry's under-depreciation claim), and management is now answering it on the record rather than ignoring it. It is an assertion, not evidence — but it is now a falsifiable management claim, which is more than the bears had before.
2. Management frames the financing loop as a feature, explicitly: "The vast and trusted marketplace for NVIDIA compute is a critical foundation on which billions in AI infrastructure spending is being financed by the ecosystem." Read plainly: NVIDIA is arguing that its compute is good collateral. That is exactly the argument a $250B lease-debt guarantee rests on — and exactly the argument the CDS market repriced ten weeks later.
3. Supply confidence is qualified, not absolute: "we are not immune to supply challenges, we remain confident", alongside total supply raised to $145B — booked ahead of demand, and a hostage to the DRAM allocation fight in Lens 2.

The standing risk, reinforced by primary text: management has anchored on ever-larger visibility numbers — $500B bookings, "$1 trillion" of Rubin+Blackwell platform visibility discussed at GTC, a $200B CPU TAM, ~$20B in-year CPU revenue. Each anchor raises the sentiment beta to a single soft data point. The July selloff is the proof: nothing in NVIDIA's numbers changed, and $238B left the stock in days.

Comps (NVDA line re-struck; peers carried)

The peer table (AMD/AVGO/TSM/INTC/MU) is carried by reference from the previous dossier — but note the peer set moved with NVIDIA this window, not independently: Micron −$113B, SK Hynix −$176B, Samsung −$173B in the same sessions. Relative multiples are unlikely to have re-ranked; absolute levels all fell and partially recovered.

Where NVDA screens now — the finding is the absence of movement. The stock rose 10.4% and the forward multiple went from ~25.2x to ~25.5x, because consensus EPS rose 9.1% in the same window. The prior pass's headline ("~12% cheaper than a month ago") has fully reversed, but not via multiple expansion — the market paid for revisions, not for re-rating. On the MarketCall's own language, NVDA is still "a market multiple": ~25.5x forward for ~65-85% operating growth, mid-pack against AVGO (~25x fwd) and a fraction of AMD (~56x fwd). EV/EBIT per-peer and 5-yr-avg ROE remain n/a; do not substitute silently.

Stock-Price Catalysts (two >5% moves and a credit-market first)

Move 1 — down. 2026-07-27: NVDA −~5% on the WSJ report of the ~$250B OpenAI backstop, inside a complex-wide $1.3T drawdown; NVIDIA alone erased $238B and lost the largest-company title to Apple on 2026-07-28 at $196.03. The novel datum is not the equity move — it is the credit move: five-year protection on NVIDIA's debt widened by the most on record in a single session. NVIDIA carries only $8.47B of total debt against ~$50B of liquid assets, so this was not a solvency judgment — it was the market pricing the contingent exposure of guarantees that appear in no filing.

Move 2 — up. Recovery to $223.96 by 2026-08-07 (+14.3% off the low), helped by the $500B bookings figure being re-circulated and continued hyperscaler capex commitment.

The pattern this extends (unchanged in shape, sharper in mechanism):

• Narrative/regime shocks >> reported numbers. 2023 guide shock +24%; 2025-01 DeepSeek −17%/−$600B; 2026-07 circular-financing −$238B. Not one was an earnings event.
• New this pass — the transmission channel widened from equity to credit. A financing structure, not a product or a print, was the catalyst. Future catalysts should be watched in CDS and hyperscaler debt issuance, not only in the tape.
• China stays a bidirectional catalyst with a zero baseline — ~$10B of licences approved, shipments "very few" as of 2026-07-14, no revenue booked.

Implication: the variant-perception edge remains in the narrative inputs — capex funding mix, the guarantee structure, Rubin/Rubin-Ultra memory specs, MI450 traction, China conversion — not in modelling the 26 August beat, which is expected, well-guided, and historically faded.

---

Phase C — Judge People & Books

Can I trust the operators and the accounting?

Management (a succession event both prior dossiers missed)

The core assessment carries (see the previous dossier): Jensen Huang, co-founder/CEO ~33 years, ~3.3-3.8% owner, best-in-class capital compounder; Colette Kress the financial architect; genuine risks are key-man/succession (Huang is sole CODM reviewing segment data ) and promotional-narrative dependency, not comp or self-dealing.

Management delta — the first real executive succession in the go-to-market seat. Ajay K. Puri, EVP Worldwide Field Operations and a 21-year NVIDIA executive, notified the company of his intention to retire, staying on in a senior advisory role. Nicholas Parker, 55, was appointed EVP Worldwide Field Operations effective ~2026-08-24, joining after 26 years at Microsoft, most recently EVP and Chief Business Officer of Microsoft's Worldwide Sales & Solutions organisation. Package: $1.0M base, $5M sign-on, $1.5M target variable, $35M new-hire RSUs (4yr), $5M multi-year PSUs vesting on TSR vs the S&P 500 over three years.

Both prior dossiers carried Lens 9 without this — it was disclosed 8 days before the 2026-07-10 pass and is only now on the shelf. Read: the person who sold NVIDIA's chips for two decades is being replaced, at the exact moment the sale is shifting from chips to allocated customers toward financed, multi-year infrastructure programmes with sovereigns and enterprises. Hiring Microsoft's enterprise/commercial chief rather than a semiconductor sales executive is a deliberate signal about what NVIDIA thinks it now sells. It is also a key-person transition in the seat that owns the top-3 customer relationships (54% of revenue) — low-probability disruption, high consequence. The TSR-linked PSU is a modest governance positive.

Legal, carried: the 2017-18 crypto-mining channel-inventory securities class action — class certified 2026-03-25 (purchasers 2017-08-10 to 2018-11-15), NVIDIA petitioned the Ninth Circuit 2026-04-08 for leave to appeal; no liability accrued. The rhyme still matters: executives accused of downplaying a cyclical demand driver. Also carried: the $3,957M accrued purchase consideration for the Groq, Inc. non-exclusive licence — a ~$4B position tied to an inference-chip rival; track, don't over-read. FY2027 executive variable-comp plan (adopted 2026-03-02) is keyed to revenue achievement with threshold/base/stretch tiers; Huang's target $4.0M (200% of salary). A revenue-only performance goal, with no margin, ROIC or capital-efficiency gate, is a mild governance negative in a year the company is guaranteeing customer debt to move product — it pays for volume, and volume is exactly what the financing loop manufactures.

Forensic Red Flags (flag #1 changed category; a new regulatory item)

Earnings remain high-quality on a cash basis — OCF $50.3B exceeded operating net income and the $15.9B of equity gains are correctly backed out of OCF. Ranked by severity:

1. Circular financing — escalated from equity marks to CREDIT GUARANTEES. The prior pass ranked this #1 at ~$90B of AI-ecosystem exposure (marketable equity $30.2B + non-marketable $43.4B + ~$9.2B warrants/convertibles + $8.9B locked-up equity in Other assets), with $15,929M (22.8% of pre-tax income) flowing through the P&L, $13.4B of it unrealised and $27.4B of the publicly-held stack locked up. The delta adds a reported ~$250B guarantee of OpenAI's data-centre lease and construction debt, plus ~$350B of GPU-purchase financing under discussion. This is a different instrument class. An equity stake caps loss at cost and marks through OCI/P&L; a guarantee is a contingent liability that is invisible until it is called, and it converts customer demand risk directly into NVIDIA balance-sheet credit risk. The record one-day widening in NVIDIA's 5-year CDS is the credit market saying exactly that. Caveat, held firmly: this is press-reported and in "talks" — it appears in no filing. The first hard test is the FY27-Q2 10-Q's guarantees and commitments notes (Notes 8/10/14) after 2026-08-26. If a guarantee of this scale is real and omitted or under-disclosed, that is a disclosure problem on top of a credit one.
2. Customer concentration — extreme, unchanged. Top-3 = 54% of revenue, 30/18/16% of AR. Now compounded by Lens 4: the fastest-growing cohort (ACIE, +31% Q/Q) is the least self-funding one.
3. Customer-side leverage is now the demand's foundation. Hyperscaler incremental debt went 9% of capex (FY24) → 32% LTM (mid-2026), with Alphabet's $84.75B equity raise in June 2026. NVIDIA's revenue is increasingly debt-financed at the buyer, on top of being NVIDIA-financed at the margin. Two layers of leverage under one revenue line.
4. Off-balance-sheet and commitment stack. $119B supply ($95B within FY27) + $30B cloud + $6B other vendor + $27B investment commitments + $3.5B facility-lease guarantees + $32.4B future DC leases — against ~$50.3B liquid. Some supply is cancellable/reschedulable per Note 10.
5. Inventory build into a memory-price squeeze. $21.4B → $25.8B (+$4.4B Q/Q), raw materials +75% sequentially; provisions modest at $0.8B. The H20 precedent means writedown risk is not hypothetical, and Lens 2's DRAM scarcity raises the cost of what is being pre-bought.
6. SBC — quantified, still neutralised. $1,928M/qtr, $20.8B unearned over ~2.6yr, but buybacks ($20.2B/qtr) far exceed it, so diluted shares are falling (24,391M vs 24,611M).
7. Effective tax 16.6% (from 14.3%); IRS examining FY2023-24. Minor.
8. Receivables quality — a self-flagged deterioration. DSO 45 days on favourable collection timing, guided back to mid-50s in Q2. Disclosed in advance, which is the good version of this.

Regulatory findings (required sub-section). No SEC enforcement history — EDGAR EFTS returns 0 Litigation Releases and 0 AAERs naming NVIDIA over 2021-07-10 → 2026-07-10. Legal Proceedings: the 2017-18 securities class action (class certified 2026-03-25) plus stayed derivative suits; no accrued liability. NEW non-SEC item: the French Competition Authority's antitrust investigation into NVIDIA is winding down — the general rapporteur stated in early July 2026 "we are nearing the end of the investigation"; French law allows penalties up to 10% of global annual revenue. On FY2026 revenue of $215.9B that is a theoretical ceiling of ~$21.6B — actual EU-style fines land far below statutory maxima, and no charge has been issued; this is a first European checkpoint, not a finding. NVIDIA also challenged EU regulators over an antitrust review. Export control remains the dominant regulatory theme (H200 to China at a 25% USG revenue share, ~$10B of licences, effectively zero shipments). Conclusion: still no material regulatory or accounting-enforcement findings — but the French decision moves from "watch-item" to near-term binary for the first time.

Forensic bottom line (sharpened again). Still not accrual manipulation — cash generation is real, high-quality, and net-cash-funded. But the periphery has now escalated twice in two passes: from equity stakes in customers (Jun), to equity marks as 23% of GAAP EPS (Jul), to guaranteeing customers' third-party debt (Aug). Every risk — concentration, commitments, inventory, marks, guarantees, and now customer leverage — points at the same trigger (an AI-capex slowdown), so they compound, not diversify. The novel feature this pass is that the compounding now has a credit transmission channel, which is faster and less forgiving than a mark-to-market one.

MarketCall review finding (assessment only — no call row touched). The ACTIVE MEDIUM call reads "NVIDIA: paying a market multiple for a franchise financing its own demand." Both legs are SUPPORTED, and the second is materially STRENGTHENED. Multiple leg: forward P/E is ~25.5x vs ~25.2x at the prior pass — the de-rate reversed, but via +9.1% EPS revisions, not multiple expansion, so "a market multiple" is still an accurate description; the call is not being invalidated by the stock's 10.4% rise. Financing leg: the window produced the strongest evidence yet — a reported $250B debt guarantee (a category beyond the equity stakes the call was written against), ~$350B more in GPU-purchase financing under discussion, and a record one-day CDS widening that priced it. One datum cuts against the strong form of the thesis and should be held honestly: Anthropic took NVIDIA's ~$10B and a strategic partnership, then committed up to 2GW to AMD — financing demand does not appear to buy exclusivity, which is a worse return on the financing than the bull case assumes but also means the loop is less closed than the pure-circularity bear case claims. Net: premise supported, not broken; conviction unchanged at MEDIUM. Route any change to our position log (human-gated).

---

Phase D — Project & Stress-Test

What's it worth, and what kills it?

Forward Projection (EPS, next 3 fiscal years) (re-anchored to raised consensus)

Model integrity — read before citing anything below. model.xlsx was built this pass (our model), and the balance-sheet backfill means it now reports "Opening balance sheet sourced: yes" — an improvement over the prior passes, where it did not exist. But it produces NO usable output and nothing below is taken from it. our model reports: "No computed values. The workbook has 51 formulas and no cached results" — the formulas require Excel to evaluate. It also carries two seed gaps: only 2 quarters of financials on the shelf (LTM figures are partial) and, consequently, revenue growth seeded at a default 8%, which is not an estimate of anything. Do not cite the workbook. Every figure in this lens is `` with arithmetic shown. our model is a Brier tracker, not an EPS model; no our model create this pass (unattended). NVIDIA FY ends late Jan (FY27 = Feb 2026 – Jan 2027).

Anchors: FY27-Q1 actual GAAP EPS $2.39 / operating ex-marks ~$1.85; Q2 guide ~$91.0B ±2%, GM 74.9-75%; consensus Q2 EPS ~$2.08; consensus FY27 EPS $8.79 (up from $8.06 last pass); shares ~24.2-24.4B and falling on the $118.5B authorisation. Projection is on an operating/non-GAAP basis — it deliberately excludes the volatile equity-mark line, which is un-forecastable and would flatter GAAP. Operating net margin is anchored at the FY27-Q1 actual: ~55%.

Triangulation: at $223.96 the market pays ~25.5x consensus FY27 ($8.79) and 20x the base FY28 ($11.25). The key change from the prior pass: consensus is no longer below my base — it has caught up to it. Last pass the Street's FY27 $8.06 sat below the dossier's $8.30 base, and that gap was itself the signal; this pass consensus ($8.79) sits at the base (~$8.80). The market has stopped pricing extra deceleration into FY27 — which removes the cushion the prior pass was implicitly leaning on. The investable question is still which FY28-29 path you underwrite: base/bull repays ~20-25x comfortably; the bear — margin compression into an earnings decline — makes today's ~25.5x roughly 26x trough earnings, with the stock halving.

Bull vs Bear (both cases updated)

Bull case. NVIDIA is the toll booth on a build-out whose toll estimate keeps rising: ~$725B of 2026 hyperscaler capex, >$1T modelled for 2027 by Evercore and BofA. The operating business printed +85% revenue at 74.9% GM and 65.6% operating margin, with Compute & Networking at a 71.5% segment operating margin — the moat has widened from the chip to the rack, and networking (~3x Y/Y) is now growing twice as fast as compute, proving it. The window added two genuine expansions: Anthropic converted from ~zero coverage to a strategic partner, and the Vera CPU opened a new line with ~$20B of in-year visibility against a claimed $200B TAM. Management holds ~$500B of bookings and ~$1T of platform visibility, is shipping Rubin from H2 2026 with the first rack live at Azure, and is returning ~50% of FCF ($118.5B authorised, dividend up 25x) off ~$49B/quarter of FCF on a net-cash balance sheet. And the market just stress-tested the scariest version of the bear story — a $1.3T selloff on circular financing — and the stock closed the window UP 10.4% at an unchanged ~25.5x, with 58 of 61 analysts at buy and a ~$303 target.

Bear case (permanent-impairment vectors):

1. The financing loop escalated to credit, and the credit market noticed first. A reported ~$250B guarantee of OpenAI's lease/construction debt plus ~$350B of GPU-purchase financing under discussion sits on top of a ~$90B equity book that already supplies 23% of GAAP EPS in unrealised, partly locked-up marks. NVIDIA's 5-year CDS widened the most on record in a single session. Equity stakes cap loss; guarantees do not.
2. Two layers of leverage under one revenue line. Hyperscaler capex funding went from 9% incremental debt (FY24) to 32% LTM, plus Alphabet's $84.75B equity raise — the customers are borrowing to buy, while NVIDIA guarantees a further slice. And the fastest-growing cohort (ACIE +31% Q/Q) is the least self-funding.
3. Concentration plus a competitor with anchor tenants. Three buyers = 54% of revenue; AMD's MI450/Helios now has Meta (6GW), Anthropic (up to 2GW), and Microsoft on Azure — and Anthropic took NVIDIA's ~$10B and bought AMD anyway. Meanwhile Meta Compute turns a top customer into a supplier of competing capacity.
4. The memory squeeze is now writing the product spec. SK Hynix is diverting HBM4 capacity to ~90%-margin DDR5, and NVIDIA is evaluating cutting Rubin Ultra to 192GB (8-Hi HBM4) on a 2027 DRAM shortage — in the same window MI450 ships at 432GB.

Pre-mortem (late 2027, the thesis broke). A hyperscaler signals 2027 capex is "digested", citing utilisation below plan and a pivot to internal silicon for inference — and does so while carrying the debt it raised in 2026. The $91B/quarter run-rate, built on a China-zero guide, three buyers, and ever-larger "visibility" numbers, loses its marginal buyer. The AI-private complex re-rates; the ~$90B equity book marks down 30-40%, flipping other income deeply negative and revealing GAAP EPS was ~23% air. Then the new part: OpenAI's Ohio lessee cannot service the debt NVIDIA guaranteed, the guarantee is called, and NVIDIA — a company with $8.5B of actual debt — takes a multi-billion contingent hit at the same moment its revenue decelerates and its marks reverse. $50B+ of pre-committed supply becomes writedowns into a DRAM market that has moved to DDR5. The stock does not need a collapse in demand — deceleration plus one private-AI re-rate does it — and the credit channel makes it faster than 2025's DeepSeek day.

Are multiples too high? No — ~25.5x forward is not the problem; it is below AI-logic peers and undemanding for the growth. The risk is the earnings base and its quality, and it now has three layers: (a) operating deceleration, (b) ~23% of reported EPS is marks that reverse with the same trigger, and (c) a contingent credit exposure that appears in no filing. A downturn hits the "E" twice and the balance sheet once.

Contrarian view — what the market is refusing to see. The July selloff was widely read as "the market finally noticed circular financing" — and then the market bought it back within eight sessions, at an unchanged multiple, on earnings revisions. Both reactions missed the actual change. The point is not that NVIDIA finances its customers; the prior two dossiers already established that at ~$90B. The point is that in this window NVIDIA moved up the capital structure of its own demand — from owning equity in customers to guaranteeing their senior obligations — and that a firm which does this is telling you its customers cannot raise that capital unaided at an acceptable price. The tell is not the equity tape, which recovered; it is the CDS, which repriced NVIDIA's own credit for the first time on record over a liability it has not booked. When the cost of manufacturing demand migrates from the P&L to the balance sheet to the credit markets, the franchise is being asked to underwrite the cycle it profits from. The edge is in the footnotes — and the footnote to watch does not exist yet. It is due on 26 August, in Notes 8/10/14.

Devil's Advocate (short-seller) **

You are a skeptical short-seller dismantling the bull case.

• What structurally breaks the machine? 92% Data Center, 54% in three buyers whose AI ROI is unproven, who are now funding purchases with debt (32% of capex, from 9%), and who are building silicon to displace NVIDIA. If customer AI economics don't pencil, their capex — NVIDIA's revenue — corrects, and this time it corrects with leverage on both sides of the trade.
• The most dangerous competitor is still the customer — and it now has anchor contracts. AMD MI450/Helios: Meta 6GW, Anthropic up to 2GW, Microsoft on Azure. AWS's chip business at ~$20B run-rate; Google Ironwood GA. The killer detail: Anthropic is simultaneously an NVIDIA equity holding (~$10B), a newly announced NVIDIA strategic partner, and a 2GW AMD customer. If ~$10B of NVIDIA capital does not secure a frontier lab's roadmap, what exactly is the ~$90B book buying?
• The accounting tell escalated from aggressive to structural. Last quarter, $15.9B / 23% of GAAP EPS was marks on stakes NVIDIA bought $18.6B more of in the same quarter. Now it is reportedly guaranteeing $250B of a customer's debt and considering $350B more for GPU purchases — obligations that appear in no filing. Burry's "around and around we go" is the polite version. And note the timing: FY27-Q2 ended ~26 July, one day before the selloff — so the 26 August print's marks are struck at the pre-crash peak. Bulls will read a clean GAAP number that is already stale.
• Assumptions that must ALL hold for $224: (1) hyperscaler capex compounds toward $1T while funded increasingly by debt; (2) NVIDIA holds premium share against MI450 with anchor tenants, possibly at 192GB vs 432GB on Rubin Ultra; (3) operating margin survives 2027 DRAM scarcity and SK Hynix's DDR5 pivot; (4) the ~$90B mark book doesn't reverse; (5) the guarantees are never called; (6) no DeepSeek-style efficiency shock; (7) the French antitrust decision is immaterial. Seven, up from five last pass — and #5 is new and unbounded.
• Valuation if growth disappoints 20-30%: bear path → FY28-29 operating EPS ~$7.75-8.50 (a decline), GAAP far worse on mark reversal; at a de-rated 12-14x that is $95-120, a 47-58% drawdown from $223.96. The DeepSeek session (−$600B) and this July (−$238B in days) both prove the gamma.
• Single permanent-impairment scenario + plausibility: a hyperscaler capex pause coinciding with an AI-private re-rate, running the loop in reverse — marks down → neoclouds and labs lose cheap capital → guarantees get called → the marginal buyer vanishes → $50B+ of supply commitments become writedowns. Plausibility: moderate, not tail — and higher than last pass. It needs only deceleration plus one private-AI re-rate, and the July CDS move demonstrates the credit channel is live and reflexive rather than theoretical.

Management Questions (carried + 3 new for the 26 Aug call)

The prior 15 (ranked by information value) carry in full — see the previous dossier and the two re-pointed questions in the previous dossier. Three new, ranked, for the FY27-Q2 call:

• [NEW #1] Reports describe ~$250B of guarantees for OpenAI's Ohio lease and construction debt and ~$350B of GPU-purchase financing. Which of these are binding obligations as of quarter-end, where do they appear in Notes 8, 10 and 14, and what is the maximum contingent exposure? Your 5-year CDS had its largest one-day widening on record on this news — what did the credit market get wrong?
• [NEW #2] Anthropic became a strategic partner this year and NVIDIA holds roughly $10B of its equity — and Anthropic has committed up to 2GW to AMD's MI450. What does NVIDIA's ecosystem investment buy if not roadmap share, and how do you measure return on the ~$90B book?
• [NEW #3] With SK Hynix converting planned HBM4 capacity to DDR5 and Rubin Ultra reportedly under evaluation at 8-Hi/192GB, what is the committed HBM4 volume by supplier for FY28, and does a 192GB Rubin Ultra remain competitive against a 432GB MI450 in memory-bound inference?
• (Then the prior list: circular-financing revenue share; top-3 backlog-vs-speculative; GPU useful-life vs hyperscaler 5-6yr depreciation — now that you've claimed customers profit beyond depreciable life, show the evidence; supply-commitment writedown thresholds and cancellability; inference-share defence; $91B backlog-vs-book-and-ship; buyback discipline; China 2027 realistic revenue; networking attach durability; succession/key-man — now including the Puri→Parker handover of the top-3 relationships; Rubin inference-deflation self-cannibalisation; gross-margin floor vs AMD; recurring/software revenue to de-cyclicalise.)

---