
Nvidia (NVDA)
Fabless chip design — designs GPUs/CPUs/networking + CUDA software, outsources fab to TSMC, sells silicon, full systems (DGX / GB-NVL racks) and software/services; monetizes a full-stack platform sold mainly to cloud hyperscalers and enterprises.
Sources — 28 figures with citations
- Board statement, Nov 8 2026filed2026-11-08Independent investigation 'in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort'; committee of independent directors with outside counsel; 'no reason to believe that either national security or NVIDIA secrets were compromised'nvidianews.nvidia.com — Statement From NVIDIA Board of Directors, Nvidia newsroom.
- Independent directors' statement, Nov 10 2026filed2026-11-10Retained former acting US Attorney (SDNY) Joon H. Kim and Wachtell, Lipton, Rosen & Katz to conduct the 'thorough and independent investigation'nvidianews.nvidia.com — Statement From the Independent Members of the NVIDIA Board, Nvidia newsroom; no timeline or scope beyond that.
- CEO transitionfiled2026-11-12Nov 11 2026: the board accepted Jensen Huang's resignation as President, CEO and director, effective immediately; Colette M. Kress interim CEO; Harvey C. Jones chairnvidianews.nvidia.com — Nvidia newsroom, 'NVIDIA announces CEO transition' (Nov 12 2026), and Form 8-K Item 5.02 filed Nov 12 2026 (https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000082/nvda-20261111.htm); neither states the investigation's outcome.
- Board meetingfiled2026-11-12Board met late Wednesday Nov 11 to decide whether to oust the CEO (source familiar, Reuters); bylaws allow removal 'with or without cause' by majority vote; ~$600B of market value lost since the investigation was announced; shares -8% in extended tradingfrance24.com — France 24 with Reuters, Nov 12 2026 05:05; the $600B figure is the article's own calculation.
- Analyst actionsmarket2026-11-12Baird downgrade to Neutral, target $500 → $200 (Nov 11, Tristan Gerra); Bernstein to Market-Perform, target $400 → $200, and Oppenheimer to Perform, target removed (Nov 12); Raymond James kept Strong Buy ahead of the Nov 17 results247wallst.com — 24/7 Wall St. (Nov 12 2026) for the Baird and Bernstein notes; Oppenheimer's downgrade and Raymond James, Mizuho and UBS keeping bullish ratings as reported by Investing.com and investinglive.com (Nov 12 2026); Yahoo Finance's live coverage of Nov 11–12 2026 for the Raymond James quote.
- Share price and market capmarket2026-11-13$162.85 at the Nov 13 2026 close, market cap $3.92T (Nasdaq); Nov 12: open $151.01, low $148.80, close $166.19; Nov 11: close $173.73 (+2.87%), $158.76 after hours (-8.62%, by about 8 pm ET), market cap $4.20T, trailing P/E 21.97; 24.15B shares; EPS (TTM) $7.91; prior 52-week range $164.27–$243.37stockanalysis.com — stockanalysis.com price history, matching Nasdaq historical data, for Nov 11–13; Nasdaq's quote summary for the Nov 13 market cap; the stockanalysis.com quote page for the Nov 11 market cap, P/E and after-hours print, and for the Nov 3 close of $206.69.
- Consensus estimatesmarket2026-11-12Strong Buy, 62 analysts (45/10/6/0/1); average target $314.72, high $515, low $180; FY2027 revenue $412.0B (+90.8%) / EPS $9.32 (high $10.10, low $9.01)stockanalysis.com — stockanalysis.com forecast page (S&P Global poll), last updated Nov 12 2026 — an aggregator, not a primary consensus provider. Nov 3 average was $327.14.
- Q3 FY27 results datefiled2026-10-28Tuesday Nov 17 2026, conference call 2 p.m. PT, for the quarter ended Oct 25 2026investor.nvidia.com — NVIDIA investor-relations release, Oct 28 2026.
- Nov 5 2026 sellofffiled2026-11-05Nvidia -9.3%, AMD -17.3%, Micron -17.5%, Broadcom -10.4% on Nov 5 (close); Nebius -15%, Iren -17%, CoreWeave -10%cnbc.com — CNBC Nov 6 2026; Fortune Nov 5 reported the Philadelphia Semiconductor index on track for its worst day since March 2020 and the Nasdaq -5.2% midday.
- Post-midterm policy statementsfiled2026-11-05Sen. Sanders: 'National data center moratorium now'; Rep. Ansari: Day-1 AI pause; Texas governor-elect Hinojosa pledged a moratorium on data-center approvals; Texas = 27% of planned US capacity (Wells Fargo)cnbc.com — CNBC Nov 5 2026 3:16 PM EST; Pennsylvania's August executive order predates the election.
- CoreWeave Q3 2026filed2026-11-10Revenue $3,661M (+168% YoY, computed from the stated figures); backlog ~$132.5B; contracted power >4.6 GW, active >1.7 GW; 2026 revenue guide raised to $13.0–13.3B; capex $9,441M in the quarter; net debt proceeds $10,301Minvestors.coreweave.com — CoreWeave Q3 2026 results release, Nov 10 2026.
- Anthropic placementfiled2026-11-10NVIDIA to buy ~135.1M Class N shares, ~$10.0B at the $74 assumed midpoint, in Anthropic's concurrent private placement; Anthropic's committed compute spend ~$900.9B across Google TPUs, AWS Trainium, NVIDIA GPUs and AMD GPUssec.gov — Anthropic S-1/A, Nov 10 2026.
- Alphabet Q3 2026 capexfiled2026-10-28Capex $51.4B in Q3; FY2026 guide held at $195–205B; free cash flow -$9.1B; future purchase commitments $924B (CFO)sec.gov — Alphabet 8-K exhibit 99.1 for the figures; the CFO's commitment figure is from the Q3 call transcript.
- Q3 FY27 results (quarter ended Oct 25 2026)filed2026-11-17Revenue $115.8B (+20% QoQ, +103% YoY); Data Center $107.6B (+110% YoY); GAAP / non-GAAP gross margin 74.8% / 74.9%; GAAP operating income $77.520B; GAAP / non-GAAP diluted EPS $2.70 / $2.68; free cash flow $43.777Bsec.gov — Recorded Nov 21 2026; figures as the source states them.
- Q4 FY27 outlookfiled2026-11-17Revenue $130.0B ±3% (quarter includes an extra week); GAAP and non-GAAP gross margin 71.0%–72.0%; opex ~$11.0B GAAP / ~$10.7B non-GAAP; no Data Center compute revenue from China assumedsec.gov — Recorded Nov 21 2026; figures as the source states them.
- Commitments and guarantees at Oct 25 2026filed2026-10-25Supply and capacity commitments $390B (from $279B); cloud-service and data-center lease commitments $102B; neocloud agreements $32B; maximum guarantee exposure $118.0B ($105.0B SB Energy, $9.5B power and systems credit support, $3.5B neocloud land, power and shell)sec.gov — Recorded Nov 21 2026; figures as the source states them.
- Balance sheet and working capital at Oct 25 2026filed2026-10-25Cash, equivalents and marketable debt securities $53.5B; long-term debt $32.371B, no short-term debt; receivables $81.0B (DSO 64, from 60); inventory $39.7B (from $31.6B); $28.1B returned in Q3; $227.3B buyback authorization remainingsec.gov — Recorded Nov 21 2026; figures as the source states them.
- Gross-margin path (CFO, Q3 FY27 call)filed2026-11-17Q3 beat on better product mix; Q4 at 71%–72% is the bottom, as orders already received are honoured before new pricing in FY28; 72%–73% in FY28; FY28 revenue outlook raised to about +75% from about +70%earningswhispers.com — REPORTED; recorded Nov 21 2026; from a third-party transcript of the Nov 17 2026 call.
- China data-center sales (CFO commentary and call)filed2026-11-17Data Center Hopper shipments to China $2.4B in Q3 FY27 (CFO commentary); on the Nov 17 call the CFO said licensed Hopper shipments to China have stopped, the Q4 guide includes none, and new Commerce export controls were released that weeksec.gov — Recorded Nov 21 2026; figures as the source states them.
- CEO searchfiled2026-11-21No permanent CEO named in any release or 8-K through Nov 21 2026; on the Nov 17 call the chair said the board does not intend to reinstate the founder, has retained a search firm and is considering internal and external candidates; the Nov 17 8-K is signed by Colette M. Kress as Interim CEO and CFOsec.gov — Recorded Nov 21 2026; figures as the source states them.
- National Security and Export Control Oversight Committeefiled2026-11-17New standing board committee chaired by Suzanne Nora Johnson; the National Security and Global Trade Compliance teams report to it and to Interim CEO Colette Kressnvidianews.nvidia.com — Recorded Nov 21 2026; figures as the source states them.
- Price and valuation at Nov 20 2026 (aggregator)market2026-11-20Close $125.53 (-14.01% on the day); Nov 17 close $155.46 (-0.44%), Nov 18 $150.96 (-2.89%), Nov 19 $145.98 (-3.30%); market cap $3.03T; TTM EPS $9.31; P/E 13.48; 52-week range $124.50–$243.37stockanalysis.com — Recorded Nov 21 2026; figures as the source states them.
- Consensus at Nov 20 2026 (aggregator)market2026-11-20FY2027 revenue $423.75B, EPS $9.67; FY2028 revenue $725.52B, EPS $16.61; average target $244.18 across 63 analysts (low $130, high $425); November ratings 34 Strong Buy, 11 Buy, 16 Hold, 0 Sell, 2 Strong Sellstockanalysis.com — Recorded Nov 21 2026; figures as the source states them.
- SEC probe (REPORTED)filed2026-11-17Wall Street Journal, Nov 17 2026, via InvestmentNews: a civil probe 'examining in part whether Nvidia's disclosures misled investors', questions centred on Asia; neither Nvidia nor the SEC has confirmed itinvestmentnews.com — REPORTED; recorded Nov 21 2026; the Wall Street Journal report of Nov 17 2026, read via InvestmentNews.
- Watched signal: Q3 FY27 print on Nov 17 2026, guided $108.0B and 74.0% ±50bpsfiled2026-11-17Printed $115.8B and 74.8% GAAP / 74.9% non-GAAP, above the top of both rangessec.gov — Recorded Nov 21 2026; figures as the source states them.
- Watched signal: working-capital quality (DSO 60, FCF $21.3B in Q2)filed2026-10-25DSO 64; free cash flow $43.8B; inventory $39.7Bsec.gov — Recorded Nov 21 2026; figures as the source states them.
- Board falsifier check (state-of-nvidia)filed2026-11-21Kill (1), clean non-GAAP GM <60%: not fired (74.9%). Early-warning, clean non-GAAP GM below ~73–74%: not fired on Q3; the Q4 FY27 guide of 71.0%–72.0% sits below it. Kill (6), CEO change without a named successor: fired Nov 11; still no permanent CEO at Nov 21 2026sec.gov — Recorded Nov 21 2026; figures as the source states them.
- Board falsifier check (state-of-ai-compute)filed2026-11-21GM-fade print below ~60%, or Q3/Q4 FY27 GM guided to the low 60s: not fired — Q3 74.9% non-GAAP, Q4 guided 71.0%–72.0%sec.gov — Recorded Nov 21 2026; figures as the source states them.
The thesis on this name
State of AI Compute
Fwd P/E ~22-23x already discounts the inference-commoditization fear and a GM fade — own the still-intact frontier-training + NVLink-system + CoWoS-supply moat that holds through ~2027-28 (UALink not at volume pre-2027-28). Downgraded: the durable margin floor is ~60% not 68% (merchant base rate), the supply-bottleneck is reflexive/self-relieving, and the label overstates the margin of safety — it's fairly-valued/moat-intact with asymmetric downside on the first GM-fade print.
State of AI Compute
Long NVDA over 12-24mo: at ~20-22x fwd EPS the market already discounts inference-commoditization and a gross-margin fade, so re-underwriting at a ~60% through-cycle GM floor (not the current 75%) still clears spot — the edge is structural (NVLink rack-scale + CoWoS supply lock + CUDA-inference inertia holding through 2027-28 before UALink/MI400 reach volume), bought at a price that needs the bear case to be right just to break even. Checked Nov 21 2026: ~13.0x FY27 consensus EPS at the $125.53 Nov 20 close; the Q2 and Q3 FY27 windows passed at 75.0% and 74.9% non-GAAP gross margin (the Q3 print of Nov 17 came under an interim CEO), and Q4 is guided to 71.0%–72.0%, a gradual fade rather than a cliff.
State of AI Compute
At ~20x fwd the price already discounts a GM fade to a ~60% floor; reduces to cheapish-multiple + CoWoS/CUDA inertia after two edges weakened. Only as the long leg of a factor-hedged pair vs an open-accelerator basket short (run structurally, not just print-window); cap 5%, not 7%.
State of Nvidia
NVDA is the highest-quality franchise in semis (90% of revenue from a Data Center business compounding ~92% YoY at ~75% GM, ~$119B TTM FCF, net-cash, $80B buyback) — but it was priced for perfection at the ~$196 June entry, and 'cheap at ~16-20x forward' is an illusion that depends entirely on a peak-cycle ~75% gross margin the company itself has shown can swing ~31pts. The debate is not the numbers (pristine) but the DURABILITY of the ~$150-225B DC gross-profit pool: frontier training + the NVLink/CoWoS/HBM system-moat is defensible through ~2027, while the larger, faster-growing INFERENCE half is the structurally erodible flank that custom ASICs (Broadcom/TPU/Trainium) are repricing now.
State of AI for Healthcare
Genuinely the supplier to every private health-AI leader — the Abridge clinical-conversation foundation model on Nemotron announced Jun 2026, Mayo pathology models on DGX Blackwell, a reported $1B Eli Lilly partnership — and healthcare is not a disclosed revenue segment at any level inside ~$302B of TTM revenue and a ~$3.9T market cap (Nov 13 2026).
State of Frontier AI
Reference-only arm and the value chain's universal toll-collector: every lab on this board — Anthropic, OpenAI, Mistral (13.8k-chip buildout), SSI (Nvidia is a strategic backer), the neoclouds — pays NVDA regardless of who wins the model race. Q3 FY27 data-center revenue $107.6B, +110% YoY (Nov 17 2026 print); ~$3.0T cap (Nov 20 2026). The durable-compounder case is the CUDA + systems moat; the cap on conviction-into-sizing is the circular-financing reflexivity (NVDA→neocloud debt→hyperscaler offtake→NVDA) and, until a permanent CEO is named, the founder-CEO's Nov 11 2026 resignation.
State of Frontier AI
The universal toll on every lab in this board — Anthropic, OpenAI, Mistral, SSI, the neoclouds all pay it — but a ~$3.0T cap (Nov 20 2026), circular-financing reflexivity and a CEO transition cap the asymmetry: the board accepted the founder-CEO's resignation on Nov 11 2026 and named CFO Colette Kress interim CEO, so hold at medium conviction until a permanent CEO is named, then resume accumulate-on-weakness. Checked Nov 21 2026: the Nov 17 results, reported under the interim CEO, beat on revenue ($115.8B) and gross margin (74.9% non-GAAP), with Q4 guided to $130.0B at 71.0%–72.0%; no permanent CEO has been named, so medium holds.
State of Frontier AI
Every lab in this board (incl. Mistral's 13.8k-chip buildout, SSI's Nvidia strategic stake, the neoclouds) pays the same toll. Q2 FY27 data-center rev $89.0B +117% YoY. The cleanest expression of 'whoever wins the model race, the picks-and-shovels get paid' — but sized below GOOGL because a ~$3.9T cap, circular-financing exposure and a CEO transition (the founder-CEO resigned on Nov 11 2026 and an interim CEO leads; adds wait until a permanent CEO is named) cap the asymmetry.
State of Physical AI
The arms dealer that wins regardless of which humanoid/robot OEM wins. GR00T N1.7 (largest open 3B VLA) + Jetson Thor edge compute (2070 FP4 TFLOPS) + Isaac Sim is becoming the default brain+sim stack across the ecosystem (Figure, Agility, Unitree reference designs). Robotics is only ~1% of revenue but growing ~72% YoY — you pay for the AI-datacenter franchise and get the physical-AI platform as a free call option. Entry note (Jun 2026): lagged ~+12% YTD vs +85% for the semi-ETF, the cleanest entry multiple of the megacaps; by the Nov 20 2026 close it was $125.53 (~$3.03T) after a sector shock, the founder-CEO's resignation and a 19% fall in the three sessions after the Nov 17 print, so the multiple is lower still but the near-term driver is the CEO transition, not robotics.
State of Physical AI
Own the robot brain+sim toll-road for free on top of the AI-datacenter franchise; held at medium conviction until a permanent CEO is named after the founder-CEO's Nov 11 2026 resignation — the multiple is lower than at entry, but the near-term driver is governance, not robotics. Checked Nov 21 2026: the Nov 17 results, reported under the interim CEO, beat on revenue and gross margin, and the CFO put FY27 Physical AI revenue on track to exceed $25B (call); no permanent CEO has been named, so medium holds.
State of Physical AI
The one name that wins regardless of which humanoid OEM wins — GR00T N1.7 (3B open VLA) + Jetson Thor + Isaac Sim is the default brain+sim stack. Robotics is ~1% of revenue growing 72% YoY, so you pay for AI datacenter and get the robotics call option free. NVDA was ~$192 / ~$4.7T at entry (Jun 2026, lagging ≈+12% YTD vs +85% for the semi ETF); at the Nov 20 2026 close it was $125.53 / ~$3.03T after a sector shock, the founder-CEO's resignation and a 19% fall in the three sessions after the Nov 17 print, so the entry multiple is lower still but the near-term catalyst is the CEO transition, not robotics; conviction is held at medium until a permanent CEO is named.
State of Quantum Computing
The purest pick-and-shovel: NVIDIA monetizes quantum WITHOUT betting on a qubit modality — CUDA-Q/NVQLink is becoming the control/orchestration layer every QPU plugs into (~75% of public QPUs already integrate). You buy it on AI/datacenter merits and get the quantum tollbooth as free optionality. Reference name (already covered) — do NOT re-pitch as a pure-play; own via the AI thesis and QTUM.
State of Quantum Computing
Purest pick-and-shovel — CUDA-Q/NVQLink is becoming the hybrid quantum-classical 'OS' (~75% of public QPUs integrated); monetizes quantum WITHOUT a modality bet. Own on the AI thesis; quantum is free optionality.
Earnings, margins, COGS & capex
Nvidia is a hyper-growth, hyper-margin franchise: Q3 FY27 (ended Oct 25 2026) revenue was $115.8B, +103% YoY, with Data Center at $107.6B (~92.9% of revenue) growing +110% YoY (fact). It converted that into 74.8% GAAP gross margin and ~67% GAAP operating margin, and free cash flow recovered to $43.8B from $21.3B in Q2 even as receivables days sales outstanding rose to 64 from 60 and inventory rose to $39.7B ahead of the Rubin ramp (fact). The model is fabless and capex-light (~4% of revenue); the weight sits in commitments instead: $390B of supply and capacity purchase commitments, $102B of cloud-service and data-center lease commitments and up to $118.0B of guarantees (fact). Nvidia returned $28.1B to shareholders in Q3 and guided Q4 to $130.0B at a 71.0%–72.0% gross margin (fact). The single concentration risk is structural: one segment (Data Center) and a handful of hyperscaler and frontier-lab customers drive the whole result.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~25¢ is cost of goods and ~8¢ operating expense, leaving ~67¢ of operating profit (~56¢ net).
Revenue trend
Margins
down 0.2pt QoQ on the initial Rubin ramp, up 1.4pt YoY on Blackwell Ultra mix (fact, Q3 FY27); Q4 FY27 guide 71.0%–72.0%, which the CFO called the bottom before 72%–73% in FY28
up (op income +115% YoY) (fact, Q3 FY27)
less flattered than in Q2: equity-securities gains fell to $1.8B from $7.8B (fact, Q3 FY27)
up sequentially (FCF $43.8B vs $21.3B in Q2 and $22.1B a year ago), helped by the timing of federal cash tax payments (fact, Q3 FY27)
COGS structure
COGS is dominated by TSMC wafer/foundry + CoWoS advanced packaging fees, HBM memory (SK Hynix/Samsung/Micron), substrates, test, and as the mix shifts to full GB-NVL rack-scale systems, more third-party content (networking, CPUs, trays). HBM and CoWoS supply are the binding cost/availability constraints; the ~75% gross margin says Nvidia still captures most of the value, but rack-scale system mix and any HBM price spike are the swing factors (analysis).
Capex
Capex is low ($4.4B in Q3 FY27, ~3.8% of revenue; fact, Q3 FY27 release) because Nvidia is fabless — it funds office/datacenter R&D facilities, lab/test equipment, and supercomputers for internal model training, NOT fabs (fact, Q1 FY27). The real 'capex' that drives Nvidia sits on customers' balance sheets — hyperscaler AI capex (~$380-725B across MSFT/AMZN/GOOGL/META in 2026 estimates) (estimate).
Latest earnings
Beat — revenue $115.8B against a guide of $108.0B ±2% ($105.8B–$110.2B); gross margin 74.8% GAAP / 74.9% non-GAAP against 74.0% ±50bps; non-GAAP EPS $2.68 (fact).
Q4 FY27 (includes an extra week): revenue $130.0B ±3%; GAAP and non-GAAP gross margin 71.0%–72.0%; GAAP opex ~$11.0B and non-GAAP ~$10.7B; no Data Center compute revenue from China assumed; full-year FY27 tax rate 16.0%–18.0% (fact, company guidance issued Nov 17 2026). On the call the CFO called Q4 the bottom for gross margin, with 72%–73% in FY28, raised the FY28 revenue outlook to about +75% from about +70%, and said licensed Hopper shipments to China have stopped (call transcript). The Q3 guide of $108.0B ±2% was exceeded by the $115.8B actual (fact).
- Data Center revenue
- $107.6B, +110% YoY / +21% QoQ, ~92.9% of total (fact, Q3 FY27); Hyperscale $60.1B (+71% YoY) and AI clouds, industrial and enterprise $47.5B (+195% YoY) (fact); Rubin brought in nearly $20B in the quarter its shipments began (call transcript)
- DC Networking revenue
- Above $21B in Q3 FY27 (CFO, Nov 17 call); $14.8B in Q1 FY27 (fact)
- Gross margin
- 74.8% GAAP / 74.9% non-GAAP (fact, Q3 FY27), against a 74.0% ±50bps guide; Q4 FY27 guide 71.0%–72.0%
- Operating income
- $77.520B on $115.8B revenue (66.9% GAAP operating margin), GAAP opex $9.156B; non-GAAP operating income $77.883B (fact, Q3 FY27)
- Diluted EPS
- $2.70 GAAP vs $2.68 non-GAAP (fact, Q3 FY27) — GAAP includes $1.8B of net equity-securities gains, after a $1.8B impairment, that non-GAAP strips out; non-GAAP has included stock-based compensation since Q1 FY27, so the $2.68 is not comparable with pre-FY27 non-GAAP EPS as first published.
- Net income (GAAP)
- $65.287B (fact, Q3 FY27) — includes $1.8B of net gains from equity securities (fact); non-GAAP net income was $64.870B.
- Trailing EPS / P-E
- TTM EPS $9.31 (GAAP diluted, four quarters to Oct 25 2026: the $7.91 to Jul 26 2026 less Q3 FY26's $1.30 plus Q3 FY27's $2.70); trailing P/E about 13.5 at the Nov 20 2026 close (fact; stockanalysis.com shows 13.48)
- Free cash flow
- $43.8B, operating cash flow $48.3B less $4.4B of purchases of property, equipment and intangibles and $0.1B of principal payments (fact, Q3 FY27; vs $21.3B in Q2 and $22.1B a year ago); $113.7B over nine months. Receivables rose to $81.0B (days sales outstanding 64 vs 60) on Rubin shipment timing and extended terms with certain investment-grade customers, and inventory rose to $39.7B from $31.6B for the Rubin ramp (fact, CFO commentary).
- Capital returns
- Quarterly dividend of $0.25 per share, next payable Dec 23 2026 to holders of record Dec 2 2026; the board added $150B to the buyback authorization in September 2026, leaving $227.3B at Oct 25 2026, which the company plans to use by the end of fiscal 2028; $28.1B returned in Q3 FY27 through repurchases and dividends (fact)
Growth drivers
- Hyperscaler AI capex supercycle — MSFT/AMZN/GOOGL/META 2026 capex guided to ~$380B+ (some estimates $600-725B), the overwhelming majority into AI datacenters and Nvidia GPUs (estimate)
- Blackwell → Vera Rubin annual cadence — Rubin in full production on TSMC, HBM4 ramp; each generation raises ASP and rack-scale content (fact/estimate)
- Networking attach — Data Center Networking $14.8B, +199% YoY (NVLink/InfiniBand/Spectrum-X) monetizes rack-scale, not just the GPU (fact, Q1 FY27)
- Inference demand inflecting — reasoning/agentic models multiply tokens-per-query, shifting the workload mix toward sustained inference compute (analysis)
- Sovereign AI + enterprise — nation-state and enterprise buildouts diversify beyond US hyperscalers (estimate)
- Software/services optionality — CUDA, AI Enterprise, NIM microservices, Omniverse as a higher-margin recurring layer (analysis)
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2026-02-25. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’26
Bull & bear
The compute layer of the largest technology buildout in history, and a Q2 FY27 spine that reconciles to the filing on every load-bearing line: $96.2B revenue +106% YoY, Data Center $89.0B +117%, 75.0% GAAP and non-GAAP gross margin, a $108.0B ±2% Q3 guide that assumes no China Data Center compute revenue (fact). Aggregate 2026 hyperscaler capex is running ~$690–725B (estimate, secondary source; only Alphabet's print verified), and at Jul 26 2026 Nvidia held $56.6B of cash and marketable debt securities against $33.4B of debt (fact, 10-Q). Rubin is in full production with named cloud deployments, and a METI-backed Japanese buildout gives a demand channel structurally independent of US hyperscaler free cash flow. These operating figures are as of the Aug 26 2026 print; the stock closed at $173.73 on Nov 11 2026 and at $162.85 on Nov 13, after the board accepted the founder-CEO's resignation — the bear summary carries the November events. Checked Nov 21 2026: Q3 FY27 beat its revenue and margin guides ($115.8B against $108.0B ±2%; non-GAAP gross margin 74.9% against 74.0% ±50bps), free cash flow was $43.8B, and Q4 is guided to $130.0B at 71.0%–72.0%, which the CFO called the bottom; the operating case held.
- Correcting our own number, upward: we published ~$380B+ for 2026 hyperscaler capex; the aggregate is running ~$690–725B (ESTIMATE, secondary source — only Alphabet's print was verified). Alphabet ALONE raised FY2026 capex guidance to $195–205B from $180–190B on $44.9B of Q2 capex (~2x YoY) with $811B of contracted future purchase commitments, up ~$500B QoQ, and signalled 2027 'significantly' higher (fact, Jul 22 2026 print); on Oct 28 2026 it held the $195–205B guide and reported $51.4B of Q3 capex (fact). MSFT/AMZN/META Q2 capex actuals are not covered here.
- The financial spine reconciles to the filing — these are the figures a lower read would have to contradict: $96.2B revenue +106% YoY, Data Center $89.0B +117% YoY / +18% QoQ, 75.0% GAAP and non-GAAP gross margin, operating income $63.734B on GAAP opex of $8.408B (~66.2% GAAP operating margin) (fact, Q2 FY27). Free cash flow of $21.3B and capex of $2.7B (~2.8% intensity) come from the filed cash-flow statement (fact).
- Correcting our own number, upward: liquid assets are $13.237B cash and equivalents + $67.335B marketable securities = ~$80.6B, not the ~$50.3B we published (that was operating cash flow, which our own key-metrics line correctly labelled). Alongside: $80.0B additional buyback authorization with no expiration, $20.0B returned in the quarter, and the quarterly dividend raised 25x from $0.01 to $0.25 effective Jun 26 2026 (fact, Q1 FY27, IR press-release balance sheet).
- The annual-cadence claim is shipping rather than promised: Rubin is in full production, Rubin-based partner products available H2 2026, with first cloud deployments announced at AWS, Google Cloud, Microsoft, OCI plus CoreWeave, Lambda, Nebius and Nscale (fact).
- Sovereign AI is a concretely funded, non-hyperscaler demand channel: Japan's METI-backed FRONTia project with Noetra Corp will build a Vera Rubin AI factory of 13,750 Vera CPUs and 27,500 Rubin GPUs delivering 140MW on the NVIDIA DSX platform (fact) — demand structurally independent of US hyperscaler free cash flow, which is where every one of our capex signals keys.
- Rack-scale attach keeps rising: Data Center Networking $14.8B, +199% YoY / +35% QoQ (fact, Q1 FY27) — Nvidia captures more of each datacenter dollar, not less.
- The neocloud canary did not fire: CoreWeave reported Q3 2026 revenue of $3,661M (+168% YoY, computed from the stated figures), a revenue backlog of ~$132.5B, more than 4.6 GW of contracted power, raised its 2026 revenue guidance to $13.0–13.3B and financed $10.3B of net debt in the quarter (fact, Nov 10 2026 release) — the 'neocloud default' leg of our falsifier is further away, not closer, on the numbers.
- Demand at the frontier is being contracted in Nvidia's favour: Anthropic's prospectus shows Nvidia buying ~$10.0B of Class N stock at the IPO price alongside Google ($10.0B), AMD, Broadcom, Micron, Microsoft and TSMC, and discloses ~$900.9B of committed compute spend across Google TPUs, AWS Trainium, NVIDIA GPUs and AMD GPUs (fact, S-1/A Nov 10 2026).
- Consensus still reads Buy, but fewer brokers say Strong Buy: FY2027 consensus revenue $423.8B (+96%) and EPS $9.67, average target $244.18 across 63 analysts (34 Strong Buy, 11 Buy, 16 Hold, 0 Sell, 2 Strong Sell) (ESTIMATE, retail aggregator, Nov 20 2026) — down from $314.72 on Nov 12; EPS estimates rose after the Q3 print while targets and Strong Buy counts fell.
- Relative strength on the tape, stated narrowly and now dated: NVDA went $200.09 (Jun 30) → $200.75 (Jul 31), +0.3%, while the SOX fell -22.7% from its Jun 22 record close (fact). That held until November: from the Nov 3 close the stock fell 15.9% to Nov 11 and a further 8.6% after hours, on a sector shock and then a company-specific governance event.
Two black swans landed in one week on top of the structural bear set. Company-specific: on Nov 8 the board opened an independent investigation into reporting that the founder-CEO was 'the target of a foreign intelligence effort'; on Nov 11, after the Wall Street Journal reported it was preparing to remove him, the board accepted Jensen Huang's resignation as President, CEO and director and named CFO Colette Kress interim CEO and Harvey C. Jones chair (Nov 12 release and Form 8-K); the shares fell 8.6% after hours on Nov 11 and closed at $162.85 on Nov 13, and Baird, Bernstein and Oppenheimer were among the brokers that downgraded. Sector-wide: the Democratic midterm sweep brought calls for data-center moratoria, and the PHLX Semiconductor Index fell 12.2% on Nov 5 (exchange data), its worst day since March 2020 (Fortune). The governing bear question remains how the capex is financed — Nvidia's own $108.5B of guarantees included — and the Q3 FY27 print on Nov 17 was a margin test for a quarter that ended Oct 25, before the change, reported by an interim CEO with no permanent successor named. Checked Nov 21 2026: Q3 passed the margin test (non-GAAP 74.9%) and no permanent CEO is named; the bear moves to the 71.0%–72.0% Q4 margin guide, $118.0B of guarantees, $390B of supply commitments, stopped China Hopper sales and an SEC civil probe the Wall Street Journal reported on Nov 17, which neither Nvidia nor the SEC has confirmed.
- Governance, as the board itself has stated it: on Nov 8 2026 the board announced an independent investigation 'in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort', overseen by a committee of independent directors with outside counsel, while saying it had 'no reason to believe that either national security or NVIDIA secrets were compromised'; on Nov 10 the independent directors retained former acting US Attorney Joon H. Kim and Wachtell, Lipton, Rosen & Katz to run it; on Nov 11 it accepted Huang's resignation as President, CEO and director, effective immediately, naming CFO Colette Kress interim CEO and Harvey C. Jones chair (fact, Nvidia newsroom statements and Form 8-K, Nov 12). As of Nov 13 2026 no findings of the investigation had been published.
- Governance, as reported before the decision: the Wall Street Journal reported on Nov 11, citing people familiar with the matter, that the board was preparing to push him out; Bloomberg reported Nvidia was working with an outside firm to identify replacements; Reuters' source said the board met late that day, and bylaws allow removal 'with or without cause' by a majority of directors (REPORTED — CNBC, Fortune, France 24/Reuters). The board accepted his resignation that day (fact, Form 8-K).
- Succession is interim: CFO Colette Kress became interim CEO on Nov 11 and Harvey C. Jones chair, with no permanent CEO named (fact, Form 8-K); Baird downgraded to Neutral on the uncertainty (Nov 11, target $500 → $200), Bernstein cut to Market-Perform with a $200 target from $400 and Oppenheimer moved to Perform (Nov 12), while Raymond James kept Strong Buy ('we doubt the distraction... alters the fundamentals') (REPORTED, Yahoo Finance; investinglive.com and Investing.com). House Foreign Affairs Committee chairman Brian Mast was reported ready to subpoena Huang over a closed-door briefing request (REPORTED, ABC via Yahoo). Checked Nov 21 2026: no permanent CEO has been named in any Nvidia release or 8-K; on the Nov 17 call the chair said the board does not intend to reinstate the founder and has retained a search firm, with internal and external candidates (call transcript).
- The policy shock: Democrats retook the House and Senate on Nov 3–4 (Fortune); Sen. Sanders posted 'National data center moratorium now', Rep. Ansari called a Day-1 AI pause a priority, Texas governor-elect Gina Hinojosa had pledged a moratorium on data-center approvals (Texas holds 27% of planned US capacity, Wells Fargo), and Pennsylvania's governor had already signed an order slowing development (fact, CNBC Nov 5). The PHLX Semiconductor Index fell 12.2% on Nov 5 (exchange data), Nvidia -9.3% (CNBC); Dan Ives now expects Big Tech capex growth to peak in 2027 rather than 2028/29 and ~10% of 800–1,000 planned US AI data centers to come off the roadmap, with up to 20% facing longer delays (analyst estimate, Fortune); Barclays' Dominique Toublan: 'scarce power favors the biggest balance sheets' (quote, CNBC Nov 6).
- Capex FINANCING, not capex level, is the variable our signal was not calibrated to. Alphabet Q2 2026 (Jul 22): capex $44.9B (~2x YoY), FY26 guide RAISED to $195–205B, $811B of contracted commitments — alongside its first negative FCF quarter since the 2004 IPO (-$5.9B), and the stock fell as much as 7% (fact). Q3 (Oct 28) kept the guide, printed $51.4B of capex and -$9.1B of free cash flow, and the CFO put future purchase commitments at $924B (fact). The demand proxy went UP while the spender was sold.
- Circular financing is now disclosed by the company, at a smaller headline size than the press figure. On Aug 17 2026 Nvidia signed residual-value guarantees for leases of ~4.25 GW at SB Energy's PORTS-Pike campus in Ohio, tenant an OpenAI affiliate, capped at $105B for its initial commitment and effective in phases as sites become ready for service (first expected in fiscal 2029); OpenAI has agreed to reimburse and indemnify any payment, though the 10-Q warns recovery may not be prompt or complete, Nvidia can add ~3.8 GW more at its sole discretion, and the obligations end if OpenAI earns a satisfactory credit rating. Total maximum guarantee exposure was $108.5B at Jul 26 2026, including $3.5B of earlier guarantees for AI-cloud partners' leases, and $118.0B at Oct 25 2026 after a new $9.5B power-and-systems credit support line (fact, 8-K and Q2/Q3 FY27 CFO commentary). The press figure in late July was ~$250B; the $105B rests on one tenant's credit.
- Hyperscaler-as-seller is a genuinely new vector, but the market did not price it as first-order for NVDA. Meta announced 'Meta Compute' on Jul 1 2026, renting out its EXCESS AI GPU capacity; ~$200B of market value was erased across semis/cloud that day, with META +8.8%, MU -10.6%, AMD -6.9% and NVDA -1.25% (REPORTED — Bloomberg via secondary coverage). Counterweight from the same file: renting surplus GPU capacity is the existing business of CoreWeave, Lambda and Nebius, which appear as named Rubin launch partners. Watch whether it becomes a pricing mechanism; do not yet weigh it as one.
- Regulatory exposure is the one real structural gap in our coverage model, now closed: the French competition authority said on Jul 9 2026 its Nvidia probe (open since 2024) is nearing its end with a possible statement of objections over GPU dominance, with separate EU and US DOJ reviews open (fact). The alleged conduct is CUDA/bundling — i.e. the same asset we score as our top strength. Unquantified: no timing, no fine or remedy range is stateable.
- Export control is escalating from a revenue ceiling to an enforcement question: Huang met Commerce Secretary Lutnick on Jul 28 2026 as the administration examines potential violations involving chip exports to China (REPORTED via SCMP; no primary agency document located). Commerce's May 31 2026 guidance already extends license requirements to any company whose ultimate parent is China-headquartered, wherever incorporated (fact). The November reporting ties the intelligence inquiry to the same thread, chip shipments to Chinese firms (REPORTED, France 24/Reuters).
- China H200 is SYMMETRIC, not bearish, and we will not book it as a negative: ~$10B of licences across ~10 firms at up to 75,000 chips each exist, yet Commerce's Jeffrey Kessler told the House Foreign Affairs Committee on Jul 14 2026 that actual shipments remain 'trivial' (fact). Whether the $91.0B Q2 guide assumes any China contribution is unknown — so it is simultaneously an unpriced upside option and an unquantified guide risk.
- Earnings quality is softer than the headline: GAAP diluted EPS $2.46 exceeds non-GAAP $2.22 because GAAP includes $7.8B of net equity-securities gains that non-GAAP strips out, and since Q1 FY27 non-GAAP includes stock-based compensation, so the $2.22 cannot be compared with any pre-FY27 non-GAAP EPS (fact, Q2 FY27). Net gains from equity securities were $23.7B in the first half, and non-marketable securities rose to $51.2B from $22.3B at Jan 25 2026 (fact).
- Balance-sheet strength is real but less clean than 'net cash': at Oct 25 2026 Nvidia held $53.5B of cash, equivalents and marketable debt securities against $32.4B of long-term debt, about $21B net on that measure (analysis). A further $44.2B of marketable equity securities and $65.5B of non-marketable securities sit outside that measure, and supply commitments of $390B (from $279B a quarter earlier) plus up to $118.0B of guarantees stand beside it (fact, Q3 FY27 CFO commentary).
- The Q3 FY27 print on Nov 17 2026 cleared its bar — revenue $115.8B against $108.0B ±2% and gross margin 74.8%/74.9% against 74.0% ±50bps (fact) — and set the next one: Q4 FY27 is guided to $130.0B ±3% at a 71.0%–72.0% gross margin with no China Data Center compute revenue (fact), which the CFO called the bottom before 72%–73% in FY28 (call transcript). Working capital is still the part to watch: receivables days sales outstanding rose to 64 from 60 on Rubin shipment timing and extended terms, and inventory rose to $39.7B from $31.6B, though free cash flow recovered to $43.8B from $21.3B (fact).
- Valuation after the drawdown: at $125.53 / ~$3.03T (Nov 20 2026 close, aggregator), NVDA is about 13.5x trailing EPS of $9.31 (fact) and ~13.0x FY2027 consensus EPS of $9.67 (estimate, aggregator, Nov 20 2026). The 15–18x-trailing bear arithmetic on the new TTM EPS is ~$140–168, and the close sits about 10% below its bottom: the close implies a lower E than consensus or a lower multiple than our bear assumption, and no reported figure yet shows either (analysis). The multiple and the E remain co-dependent — if the cycle digests, both compress.
- Concentration is unchanged and structural: Data Center is $89.0B of $96.2B, ~92.5% of revenue (fact, Q2 FY27). One direct customer was 16% of Q2 revenue and three direct customers were 16%, 15% and 13% of first-half revenue; five direct customers were 22%, 14%, 13%, 11% and 10% of receivables at Jul 26 2026 (fact, 10-Q) — any single capex cut is an air-pocket.
- Watch item, deliberately not a pillar: a reported Shanghai-built domestic immersion-DUV lithography tool in small-batch production (~5 units 2026, ~20 in 2027, destined for SMIC, Hua Hong, CXMT) moved the complex hard (ASML -5.8%, SK Hynix -12%, Samsung -10%, KOSPI -8% circuit breaker) (REPORTED — The Information via CNBC, which flagged 'big caveats'). It is second-order for a fabless Nvidia and first-order only for the premise that the China AI stack stays supply-constrained.
- Regulatory overhang (REPORTED; neither Nvidia nor the SEC has confirmed the probe): the Wall Street Journal reported on Nov 17 2026 an SEC civil probe examining in part whether Nvidia's disclosures misled investors, with questions centred on Asia (via InvestmentNews); California's attorney general said 'red flags have been raised' about possible diversion of chips to China (Bloomberg TV via InvestmentNews). No Nvidia release or 8-K found through Nov 21 2026 addresses the reports; on the Nov 17 call the chair said the company would not take questions on the resignation or the board's investigation (call transcript).
What it is worth
Forward P/E on aggregator consensus, cross-checked against the trailing multiple and against the semiconductor complex's de-rating. PROVENANCE WARNING: no primary consensus provider (FactSet/Visible Alpha/LSEG) was used — every consensus figure below is an ESTIMATE from a retail aggregator (stockanalysis.com), so the forward multiple, the FY27 EPS, the price target and the 'bar above guidance' claim all inherit that weakness.
A single bear number is still not stateable — it would rest on a cut-EPS anchor we cannot ground in anything verified. In place of a target we publish the arithmetic on figures we hold: 15–18x the verified TTM EPS of $9.31 is ~$140–168, and the Nov 20 close of $125.53 sits about 10% below it; 15–18x the aggregator FY27 consensus EPS of $9.67 is ~$145–174. The close sits below both bands; the interim CEO and an SEC civil probe the Wall Street Journal reported on Nov 17, which neither Nvidia nor the SEC has confirmed, are risks no verified figure captures. The 15–18x multiple range is an ASSUMPTION, not a verified input, so treat the bands as illustrative arithmetic, not a price call (analysis).
~$145–174
15–18x the FY2027 consensus EPS of $9.67 (estimate, aggregator, Nov 20 2026): the E rose into the print, and the base case is that the multiple recovers part of the Nov 18–20 fall once a permanent CEO is named and if an SEC civil probe the Wall Street Journal reported on Nov 17, which neither Nvidia nor the SEC has confirmed brings no adverse finding; until then the Nov 20 close of $125.53 (~13.0x) sits below the band, and the Q4 FY27 print (71.0%–72.0% gross margin guided) is the next test (analysis).
~$244.18
the average analyst target across 63 analysts (34 Strong Buy, 11 Buy, 16 Hold, 0 Sell, 2 Strong Sell), consensus Buy (ESTIMATE, retail aggregator, Nov 20 2026), down from $314.72 on Nov 12; the low target is $130, about 4% above the Nov 20 close. Requires the capex cycle to stay financeable at the spenders, gross margin to recover from the guided 71.0%–72.0% Q4 bottom to the 72%–73% the CFO guides for FY28, a permanent CEO to be named without a strategy break, and no adverse finding from an SEC civil probe the Wall Street Journal reported on Nov 17, which neither Nvidia nor the SEC has confirmed; financeability now includes Nvidia's own guarantees of up to $118.0B (fact).
Re-derived Nov 21 2026: at $125.53 / ~$3.03T (Nov 20 2026 close, aggregator; 24.15B shares outstanding), NVDA trades at about 13.5x trailing EPS of $9.31 (fact) and ~13.0x FY2027 consensus EPS of $9.67 on revenue of $423.8B, +96% growth (estimate, aggregator, Nov 20 2026), and ~7.6x the FY2028 consensus of $16.61. The E rose on the Nov 17 print (revenue $115.8B against a $108.0B guide, Q4 guided to $130.0B, FY28 revenue guided up about 75%) while the price fell 19.3% from the Nov 17 close in three sessions, so the whole fall was multiple compression (analysis). The drawdown since our Nov 3 check ($206.69, -39%) came in three parts: the Nov 5 sector shock (-9.3%, CNBC), the Nov 8–12 governance event that ended in the founder-CEO's resignation, and the Nov 18–20 fall after the print. Four of our own published calls were corrected IN PLACE on Nov 3, each in the company's favour: (1) our '~35x trailing' was wrong — it was 30.74x at the Jul 31 2026 close; (2) our forward inputs were low — consensus was $8.99 on $393.60B at Jul 31 2026, not ~$8.06 on ~$367B; (3) our arithmetic did not hold — $201 / $8.06 is 24.9x, not the '~22x' we asserted; (4) our market cap was internally inconsistent. On the sector-relative frame: the SOX fell -22.7% from its Jun 22 record close to Jul 31 while NVDA went +0.3%, so through October NVDA's multiple was the one that had NOT compressed — that cushion is now spent, and the forward multiple rests on an E that embeds +96% growth in FY27 and +71% in FY28 (estimate, analysis).
SWOT
Strengths
- CUDA software lock-in — 15+ years of libraries, frameworks and developer mindshare; the highest switching cost in compute (analysis). Note the tension we now carry explicitly: CUDA/bundling conduct is the alleged abuse in the French probe nearing a decision (fact, Jul 9 2026).
- Full-stack rack-scale platform — GPU + CPU (Grace/Vera) + networking (NVLink/InfiniBand/Spectrum-X) + systems; competitors sell a chip, Nvidia sells the datacenter (analysis). Corroborated by Data Center Networking at $14.8B, +199% YoY / +35% QoQ (fact, Q1 FY27).
- Margin and cash structure intact — 75.0% GAAP and non-GAAP gross margin, ~66.2% GAAP operating margin ($63.734B on $96.2B), $56.6B of cash and marketable debt securities, ~$99.0B remaining under the buyback authorization, $26.0B returned in the quarter (fact, Q2 FY27). Free cash flow was $21.3B in Q2, down from $48.6B in Q1 on working capital (fact).
- Product cadence executing, not just announced — Rubin in full production, partner availability H2 2026, first cloud deployments at AWS, Google Cloud, Microsoft, OCI plus CoreWeave, Lambda, Nebius, Nscale (fact).
- Relative strength through a sector drawdown — +0.3% ($200.09 → $200.75) while the SOX fell -22.7% from its Jun 22 record close and 19 tech/semi names fell >25% (fact, Jun 30 – Jul 31 2026). Tape evidence of relative preference, not of fundamentals.
Weaknesses
- Single-segment dependence — Data Center $89.0B, ~92.5% of revenue (fact, Q2 FY27); everything else is immaterial.
- Extreme customer concentration — a few hyperscalers drive the majority of Data Center revenue (analysis).
- Key-person transition — founder Jensen Huang resigned as CEO on Nov 11 2026 with no permanent successor named; CFO Colette Kress is interim CEO (fact, Nov 12 release and Form 8-K).
- Liquidity headline overstates clean net cash — $53.5B of cash, equivalents and marketable debt securities against $32.4B of long-term debt at Oct 25 2026, with a further $44.2B of marketable equity securities and $65.5B of non-marketable stakes outside that measure; $390B of supply commitments and up to $118.0B of guarantees sit off the balance sheet (fact, Q3 FY27 CFO commentary).
- Reported earnings flattered by non-operating items — GAAP net income $58.321B above $53.536B operating income (fact/analysis) — and non-GAAP EPS is no longer comparable to prior years after SBC stopped being excluded beginning Q1 FY27 (fact). The headline $1.87 needs that definitional warning attached to it.
- Supply-chain choke points outside its control — TSMC CoWoS packaging and HBM are the binding constraints; Nvidia cannot unilaterally add capacity (analysis).
- China is licensed but not flowing — ~$10B of H200 licences across ~10 firms at up to 75,000 chips each, yet shipments described as 'trivial' by Commerce on Jul 14 2026 (fact); whether the $91.0B Q2 guide assumes any China contribution is unknown, so the exposure is unquantified in both directions.
- The de-rating cushion is spent — 26.13x trailing at the Nov 3 close became about 20.6x at $162.85 on Nov 13 and about 13.5x at $125.53 on Nov 20 on the new TTM EPS of $9.31 (fact) — a statement about the multiple's path, not about the business; the E rose on the Nov 17 print while the price fell.
Opportunities
- Sovereign AI, concretely funded — Japan's METI-backed FRONTia with Noetra Corp: 13,750 Vera CPUs + 27,500 Rubin GPUs, 140MW on NVIDIA DSX (fact) — a demand pool structurally independent of US hyperscaler free cash flow.
- Rubin ramp raising rack-scale content per deployment, with H2 2026 partner availability and named first cloud deployments (fact).
- Networking attach compounding — $14.8B, +199% YoY / +35% QoQ (fact, Q1 FY27).
- China as an option the guide does not count — licensed Hopper data-center sales to China were $2.4B in Q3 FY27, and on Nov 17 the CFO said those shipments have stopped, the Q4 guide includes none, and new Commerce export controls were released that week (fact/call transcript); ~$10B of H200 licences had been approved while shipments were 'trivial' (Commerce, Jul 14 2026), so any resumption would be upside outside the guide (analysis).
- Inference + agentic AI inflection — reasoning models multiply token volume, extending demand beyond training (analysis). Management's own version — 'Free AI should be great for hardware' (Huang, Jul 22 2026) — is a management claim, not evidence.
- Enterprise + software monetization — AI Enterprise/NIM/Omniverse as a recurring, higher-margin layer on the installed base (analysis).
Threats
- Founder-CEO governance event (Nov 2026) — an independent board investigation into reporting that Huang was the target of a foreign intelligence effort (fact, Nov 8 statement), then his resignation as CEO and director on Nov 11 with an interim CEO and no permanent successor (fact, Nov 12 release and Form 8-K) — key-person risk at a company whose strategy was closely identified with its founder (analysis).
- US data-center siting politics after the midterms — calls for a national moratorium, a Texas approval-pause pledge, local restrictions passed; Dan Ives expects ~10% of planned US AI data centers off the roadmap and the capex-growth peak in 2027 (estimates); the PHLX Semiconductor Index fell 12.2% on Nov 5 2026 (exchange data).
- Capex financing stress at the spenders — Alphabet raised FY26 capex to $195–205B AND printed its first negative FCF quarter since 2004 (-$5.9B), then -$9.1B in Q3, with $924B of future purchase commitments (fact, Jul 22 and Oct 28 2026). Other hyperscalers' Q2/Q3 capex actuals are not covered here.
- Vendor/circular financing at a now-disclosed scale — $105B of capped guarantees for ~4.25 GW of OpenAI-tenanted leases at one Ohio campus (option for ~3.8 GW more), $108.5B of total maximum guarantee exposure ($118.0B at Oct 25 2026), with OpenAI contractually agreeing to reimburse any payment, though the 10-Q warns recovery may not be prompt or complete (fact, Aug 2026). The mechanism was already in our published bear set; the filing caps the initial commitment at $105B in phases, the first expected to take effect in fiscal 2029, against the ~$250B reported in late July.
- Hyperscalers reselling surplus capacity — Meta Compute, announced Jul 1 2026 (REPORTED, Bloomberg via secondary coverage); ~$200B erased across semis/cloud on the day, but NVDA only -1.25% while META rose 8.8%, and the model mirrors what CoreWeave/Lambda/Nebius already do as Rubin partners.
- Antitrust — French authority nearing a decision on a statement of objections over GPU dominance (Jul 9 2026), with separate EU and US DOJ reviews open (fact); timing and remedy exposure unquantified.
- Export-control escalation from revenue ceiling to enforcement risk — Huang meeting Commerce as the administration examines potential violations involving chip exports to China (REPORTED, Jul 28 2026, no primary agency document); May 31 2026 Commerce guidance extends licensing to China-headquartered ultimate parents (fact); November reporting ties the intelligence inquiry to chip shipments to Chinese firms (REPORTED, Reuters via France 24).
- Custom ASICs — Google TPU (nearly 100 external customers, four systems in 2026), AWS Trainium, Meta MTIA, Microsoft Maia (Broadcom/Marvell-designed) absorbing internal workloads; custom silicon moving toward 20–30% of some buildouts (estimate).
- AMD MI450/Helios ramp — now shipping: MI455X in volume, first Helios racks live at Anthropic, Meta, Microsoft, OpenAI and Oracle, Data Center $8.5B (+97%) in Q3 2026, Q4 guided up ~50% sequentially (fact, AMD Nov 3 2026) — a credible second source at ~55% gross margin, i.e. competing on price.
- China supply-chain sovereignty — reported Shanghai domestic immersion-DUV tool in small-batch production for SMIC, Hua Hong, CXMT (REPORTED, 'big caveats' per CNBC); second-order for a fabless vendor, tracked as a watch item rather than a thesis risk.
Moats, dependencies & bottlenecks
Moats
15+ yrs of libraries/frameworks/training; the highest switching cost in compute. Rivals can match a chip, not the stack (analysis).
NVLink/InfiniBand/Spectrum-X + Grace/Vera + DGX/GB-NVL racks; sells the datacenter, not a part. Networking +199% YoY proves attach (fact, Q1 FY27).
Blackwell→Rubin→Feynman keeps a 1-2 gen lead; ~$48.6B/qtr FCF funds a budget rivals can't match (fact).
Priority allocation of TSMC CoWoS + leading-edge HBM; an advantage, but the same choke points constrain Nvidia too (analysis).
Models, tooling, and ops are built around Nvidia; being the default compounds with each generation (analysis).
Dependencies
Sole leading-edge foundry; CoWoS packaging is the binding capacity constraint on units shipped (analysis).
HBM availability/price gates output and gross margin; SK Hynix is the lead, Samsung/Micron secondary (fact/estimate, 2026).
MSFT, AMZN, GOOGL, META, Oracle, CoreWeave A few buyers drive most Data Center revenue; their capex decisions ARE Nvidia's revenue (analysis).
Entire thesis depends on AI compute demand not digesting; an unproven end-market ROI is the macro swing factor (analysis).
China market capped/lost; further node or country restrictions can cut TAM at any time (analysis).
Advantages
- Software lock-in (CUDA) that turns a hardware sale into a platform commitment (analysis)
- Full-stack integration — performance/TCO at rack scale that point-solution chips can't match (analysis)
- Cash-flow dominance — ~$48.6B quarterly FCF funds an R&D and supply-priority lead rivals can't fund (fact, Q1 FY27)
- Pricing power — ~75% gross margin even mid-Blackwell/Rubin ramp signals genuine differentiation (fact)
- Time-to-market cadence — annual generations reset the competitive bar before rivals ship the prior node (analysis)
- Standard-setter network effects — being the default compounds with the installed base each cycle (analysis)
Weaknesses
- Customer concentration — a handful of hyperscalers; loss/cut from any one is a revenue air-pocket (analysis)
- Single-segment dependence — ~92% Data Center; no diversified ballast (fact, Q1 FY27)
- Customers are also building competitors (TPU/Trainium/MTIA/Maia) — the buyer funds the substitute (estimate)
- Supply-chain dependence on TSMC + HBM oligopoly that Nvidia doesn't control (analysis)
- Valuation leaves no margin for error — a ~$3.9T cap (Nov 13 2026) priced on continued ~90% growth (FY2027 consensus revenue +90.8%, estimate, aggregator, Nov 12 2026; analysis)
- China revenue structurally limited by export controls (analysis)
Bottlenecks
- CoWoS advanced-packaging capacity at TSMC — the literal cap on how many accelerators can ship (analysis)
- HBM supply (HBM3E→HBM4) — quantity and price gate both volume and gross margin (fact/estimate, 2026)
- Leading-edge wafer allocation (TSMC N2/A16) shared with the whole industry (analysis)
- Datacenter power + cooling at the customer — increasingly the physical limiter on deployment, not chip supply (analysis)
- Customer capex budgets — demand is gated by hyperscaler willingness to keep spending, not by Nvidia's ability to design (analysis)
Top signals & trends
Top signals
Revenue and gross margin beat their guides and the E rose (FY28 revenue guided up about 75% on the call; Rubin brought in nearly $20B in the quarter its shipments began), while gross margin is guided to a Q4 bottom of 71.0%–72.0% before 72%–73% in FY28 and no permanent CEO has been named (Form 8-K exhibits 99.1 and 99.2, Nov 17 2026; call transcript). On Nov 17 the Wall Street Journal reported an SEC civil probe of Nvidia's disclosures, which neither Nvidia nor the SEC has confirmed (stockanalysis.com for prices). The analysis text is as of the entry's stated date and predates the print unless dated otherwise.
Nov 8: independent investigation opened 'in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort' (fact, board statement); Nov 10: Joon H. Kim and Wachtell retained (fact); Nov 11: after a WSJ report that it was preparing to remove him (REPORTED), the board accepted Huang's resignation and named Colette Kress interim CEO and Harvey C. Jones chair (fact, Nov 12 release and Form 8-K). Stock -8.6% after hours Nov 11 and $162.85 at the Nov 13 close; Baird to Neutral and Bernstein to Market-Perform (both with $200 targets), Oppenheimer to Perform. Next: a permanent CEO, and any published findings of the investigation.
Democrats took the House and Senate; Sen. Sanders posted 'National data center moratorium now', Texas governor-elect Hinojosa had pledged a moratorium on approvals (Texas = 27% of planned US capacity, Wells Fargo), local measures restricting data centers passed; the PHLX Semiconductor Index fell 12.2% on Nov 5 (exchange data), its worst day since March 2020 (Fortune), NVDA -9.3% (CNBC). Dan Ives moved the Big Tech capex-growth peak to 2027 (analyst estimate). A timing risk to the capex the Q3 guide assumes, not a demand event on the evidence.
Custom-silicon share loss remains the structural erosion case and is unchanged — Google now counts nearly 100 TPU customers and launched four TPU systems in 2026 (Oct 28 call; Nov 9 blog) — but it is not what repriced the complex in July 2026, and the bear set is wider than this signal. Three mechanisms sit outside it, each reported rather than confirmed: (a) Meta announced 'Meta Compute' on Jul 1 2026, a business renting out its EXCESS AI GPU capacity — a hyperscaler turning from buyer into seller, which attacks the perpetual-GPU-scarcity premise itself (Bloomberg via secondary coverage; NVDA -1.25% on the day, ~$200B of value erased across semis/cloud); (b) on Aug 17 2026 Nvidia signed residual-value guarantees for leases of ~4.25 GW at SB Energy's PORTS-Pike campus in Pike County, Ohio, tenant an OpenAI affiliate, capped at $105B and effective as leases commence, with an option to add ~3.8 GW; OpenAI has agreed to reimburse any payment, and total maximum guarantee exposure was $108.5B (fact, 8-K and Q2 FY27 CFO commentary) and $118.0B at Oct 25 2026 (Q3 CFO commentary) — against the ~$250B reported in late July — which converts revenue quality into a credit question and puts Nvidia's own balance sheet behind demand; (c) a reported Shanghai-built domestic immersion-DUV lithography tool (Jul 27–28 2026) is second-order for a fabless Nvidia but first-order for the assumption that the China AI stack stays supply-constrained (unverified report that moved markets; CNBC flagged 'big caveats').
Custom-silicon share loss remains the structural erosion case and is unchanged — Google now counts nearly 100 TPU customers and launched four TPU systems in 2026 (Oct 28 call; Nov 9 blog) — but it is not what repriced the complex in July 2026, and the bear set is wider than this signal. Three mechanisms sit outside it, each reported rather than confirmed: (a) Meta announced 'Meta Compute' on Jul 1 2026, a business renting out its EXCESS AI GPU capacity — a hyperscaler turning from buyer into seller, which attacks the perpetual-GPU-scarcity premise itself (Bloomberg via secondary coverage; NVDA -1.25% on the day, ~$200B of value erased across semis/cloud); (b) on Aug 17 2026 Nvidia signed residual-value guarantees for leases of ~4.25 GW at SB Energy's PORTS-Pike campus in Pike County, Ohio, tenant an OpenAI affiliate, capped at $105B and effective as leases commence, with an option to add ~3.8 GW; OpenAI has agreed to reimburse any payment, and total maximum guarantee exposure is $108.5B (fact, 8-K and Q2 FY27 CFO commentary) — against the ~$250B reported in late July — which converts revenue quality into a credit question and puts Nvidia's own balance sheet behind demand; (c) a reported Shanghai-built domestic immersion-DUV lithography tool (Jul 27–28 2026) is second-order for a fabless Nvidia but first-order for the assumption that the China AI stack stays supply-constrained (unverified report that moved markets; CNBC flagged 'big caveats').
Q3 beat its 74.0% ±50bps guide and slipped 0.1pt QoQ on the initial Rubin ramp (fact). The CFO called Q4 the bottom, with orders already taken honoured before new pricing, and guided FY28 to 72%–73% (call). Holding 72%+ = pricing power intact; a print below the Q4 range signals commoditization.
Buy-side whisper now exceeds the print; sentiment, not fundamentals, is the near-term governor (fact/analysis) — the Q3 FY27 beat on Nov 17, reported under an interim CEO, was followed by a 2.9% fall on Nov 18 (stockanalysis.com).
More supply = more units shippable (bullish volume) but also more competitor enablement (analysis).
Proof that AI spend pays off is what de-risks the capex cycle's durability (analysis); Anthropic's prospectus ($17.0B of Q3 revenue, ~$900.9B of compute commitments, Nvidia a $10.0B placement buyer) and CoreWeave's Q3 (backlog ~$132.5B, guide raised) are the November data points (fact).
Nvidia agreed to acquire Hugging Face for about $11.9B plus an equity retention program of up to about $1.0B, expected to close in the first half of 2027 subject to regulatory approvals (fact, 8-K). Nvidia committed to keep the platform open and to keep supporting other silicon vendors. The filing flags government restrictions on open-source models, including those originating in China, as a risk to the platform and to Nvidia's business.
Trends
Largest infrastructure buildout in tech history; directly funds Nvidia's Data Center revenue (estimate).
Multiplies tokens-per-query and sustained-compute demand, extending the runway (analysis).
Broadcom/Marvell-designed TPU/Trainium/MTIA/Maia absorb internal workloads, capping share (estimate); Google now sells TPU systems to outside customers.
Expands dollar content per GPU; networking +199% YoY (fact, Q1 FY27).
Caps the China TAM and adds recurring regulatory overhang (analysis).
New, capex-funded demand pool less correlated to US hyperscalers (estimate).
Moratorium calls and state approval pauses add timing risk to the buildout the capex guides assume (fact for the statements; analysis for the read-through).
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Sole leading-edge foundry (N2/A16) + CoWoS advanced packaging — the binding capacity constraint.
Lead HBM supplier — ~70% of Rubin HBM4 allocation (2026); Korea-listed.
HBM4 supplier (Vera Rubin) + foundry; Korea-listed.
HBM3E/HBM4 supplier, especially for inference-tier accelerators (Rubin CPX).
Upstream — EUV lithography enabling TSMC's leading-edge nodes.
Power + liquid-cooling for the rack-scale systems Nvidia ships.
Top hyperscaler buyer (Azure); also building Maia in-house.
Major GPU buyer; GPUs ~60% of AWS AI build 2026 even as Trainium ramps.
Buys GPUs for Cloud + runs its own TPU; customer and competitor.
~$70B+ 2026 capex, heavy GPU buyer; also ramps MTIA.
OCI AI capacity buildout; fast-growing GPU customer.
GPU-cloud neocloud — pure-play Nvidia capacity reseller.
Instinct MI450 Series / Helios racks on TSMC N2 + HBM4 — MI455X shipping in volume since Q3 2026 with the first Helios racks live at Anthropic, Meta, Microsoft, OpenAI and Oracle (AMD, Nov 3 2026); the credible #2 GPU at ~55% non-GAAP gross margin, i.e. competing on price (fact/analysis).
Designs the custom ASICs (Google TPU, Meta MTIA, OpenAI/Anthropic Titan); ~$8.4B AI rev Q1 FY26 +106% YoY, $73B backlog — the arms dealer for Nvidia's customers (fact, 2026).
Second custom-ASIC co-design house (with Broadcom ~95% of the market); Trainium/Maia content (estimate).
TPU v7 powers internal + Cloud AI; the most mature in-house alternative to Nvidia GPUs (estimate).
Trainium3 ramping; GPUs still ~60% of AWS AI build in 2026 but custom share rising (estimate).
Gaudi/Falcon Shores has struggled to gain AI-accelerator traction; distant also-ran (analysis).