NVIDIA

Phase A — Understand the Business

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

Lens 1 · Company Overview (refreshed — now research-layer-grounded)

Business model, unchanged in shape, sharper in the numbers. NVIDIA sells the picks and shovels of the AI build-out and has moved from selling a chip to selling a building's worth of compute — integrated rack-scale AI systems (GPU + CPU + NVLink fabric + NICs/DPUs + Ethernet switching + the CUDA software layer), fabless, manufactured by TSMC/ODMs, sold to a handful of buyers spending tens of billions each.

FY2026 (year ended Jan 25 2026) — the audited full-year baseline the prior dossier lacked: revenue $215.9B, +65% Y/Y; net income $120,067M; gross margin 71.1% (down from ~75% FY2025, pressured by the Hopper-HGX → Blackwell full-system transition and a $4.5B H20/China charge). Prior-year net income $72,880M (FY2025) and $29,760M (FY2024).

Latest quarter (FY27-Q1, ended Apr 26 2026) — now hard:

• Revenue $81,615M, +85% Y/Y (vs $44,062M), +~20% Q/Q off the ~$68.1B FY26-Q4
• Gross profit $61,157M → GM 74.9%; operating income $53,536M → op margin 65.6%
• GAAP net income $58,321M (+211%), GAAP diluted EPS $2.39 — but $15,929M of that pre-tax is non-operating equity-mark gains; operating EPS ex-marks ≈ $1.85, vs prior non-GAAP ~$1.76-1.87
• Diluted shares 24,391M, falling Y/Y (24,611M)

Customers — concentration confirmed at the filing level. FY27-Q1: three direct customers = 21%, 17%, 16% of revenue (= 54%), all Compute & Networking; three customers = 30/18/16% of AR. These "direct customers" are ODM/CSP/integrator names of record (Foxconn/Quanta/Wistron/Supermicro/Dell) building for hyperscalers + neoclouds + AI labs.

Contract book & the financing loop — the defining structural feature, now filing-quantified:

• Manufacturing/supply/capacity commitments $119B ($95B due in the rest of FY2027) — "in certain instances… cancellable, able to be rescheduled, or adjustable" (partial answer to the prior dossier's cancellability question)
• Multi-year cloud-service commitments $30B (NVIDIA renting back capacity, "primarily to support R&D")
• Investment commitments $27B; facility-lease guarantees up to $3.5B (was $860M last read); $32.4B of future data-center leases to commence FY27-FY33
• ~$18.6B of private-company stakes purchased in the quarter alone

Analyst flag — the circularity is now the dominant read. NVIDIA takes equity in OpenAI/Anthropic/CoreWeave/neoclouds → they buy NVIDIA systems → revenue books at ~75% margin → gains on those stakes mark up and flow through the P&L as 23% of GAAP EPS. The OpenAI $100B→$30B right-sizing didn't shrink the loop; it diversified it (>$40B across a portfolio in 4 months). Highest-priority stress-test item — see Lens 10/13.

Lens 2 · Supply Chain (refreshed — chokepoint reshuffled)

The named-link map is unchanged in structure (see deep-dive-2026-06-09.md for the full ASCII chain: ASML EUV → TSMC N3/A16 → TSMC CoWoS → NVIDIA → Foxconn/Quanta/Wistron/Supermicro/Dell → hyperscalers/neoclouds/labs; HBM from SK Hynix > Samsung > Micron). What moved in the delta:

Net: the binding constraint rotated from CoWoS to HBM4, and while multi-sourcing de-risks the single-supplier tail, the practical 2026 consequence is a 25% cut to Rubin unit availability — a real supply-side negative even as demand guidance rose. China revenue $4,550M in FY27-Q1 (−53% Y/Y vs $9,659M), 5.6% of revenue, essentially all non-DC-compute; DC-compute-to-China modeled at zero, so H200 resumption is upside not baseline.

Lens 3 · Competitive Advantages (Moats) (refreshed — the challenge sharpened)

The moat ranking is unchanged (1. full-stack rack-scale co-design; 2. CUDA lock-in; 3. annual cadence; 4. CoWoS/HBM priority allocation), and the financial proof of the widening moat is now filing-grade: Compute & Networking booked $53,335M operating income on $74,550M revenue — a 71.5% segment operating margin; rivals sell a chip, NVIDIA sells the room at 70%+.

What the delta sharpens on the challenge side:

• The window is now contested. AMD MI450 (432GB HBM4 — more memory than Rubin's 288GB — ~40 PFLOPS FP4, TSMC N2, CoWoS-L) arrives H2 2026, the same window as Rubin, marketed as beating "every NVIDIA product including Rubin Ultra" on memory density + TCO. Claim ≠ share, and CUDA still governs training, but this is a credible in-window merchant alternative for memory-bound inference.
• Custom silicon crossed from threat to run-rate. Google TPU v7 "Ironwood" (GA late 2025, JAX-captive), AWS Trainium3 (3nm, shipped Dec 2025, +40% efficiency; AWS chip business ~$20B run-rate), Microsoft Maia, Meta MTIA. The captive-to-one-cloud limitation still caps their addressable market, but the $20B AWS figure is the first hard evidence the escape hatch is being used at scale.
• The cadence edge is eroding by imitation — "AMD, NVIDIA, and hyperscalers commit to annual releases through the decade". Annual is becoming table stakes.

Honest vulnerability, unchanged: SemiAnalysis' 90%+ → 20-30% inference-share-by-2028 call still frames the bear case; the delta adds evidence (MI450 in-window, AWS $20B). The bull case remains "hold the premium training + frontier-inference tier where system integration commands the margin," not "90% of everything."

Bargaining power: over customers, strong-but-peaking (allocation-rationed, 54% in three buyers who are building their own silicon and taking NVIDIA's equity checks); over suppliers, weaker than it looks — the $119B of pre-committed supply is the balance-sheet admission that ASML/TSMC-CoWoS/SK-Hynix hold the physical chokepoints and must be paid years ahead.

Lens 4 · Segments (refreshed — upgraded → )

Every FY27-Q1 segment number the prior dossier carried as is now:

By reportable segment (FY27-Q1):

By market platform (recast): Data Center $75,246M (Hyperscale $37,869M + AI-Clouds/Industrial/Enterprise $37,377M); Edge Computing $6,369M. By geography: United States $63,769M (78%), Taiwan $12,006M (ODM build location), China $4,550M (5.6%), Other $1,290M — outside-US just 22%, down from 42% a year ago.

Segment bottom line (unchanged): a ~92%-Data-Center company in the costume of a diversified semi. The investable question is entirely (a) how long hyperscaler + sovereign + lab capex compounds (delta: it rose to $725B/2026, >$1T/2027), (b) whether the 54%/top-3 concentration + the now-doubled equity-financing loop hold, (c) whether HBM4 supply + mid-70s margins survive the Rubin unit cut and MI450 into 2027-2028.

---

Phase B — Measure Performance

What do the numbers and the tape say?

Lens 5 · Earnings Result (re-run — same quarter, now filing-grounded + a correction)

No new print has landed in the delta window (FY27-Q1 remains the latest; FY27-Q2 reports ~late Aug 2026). But grounding the same quarter in the actual 10-Q corrects the prior web read in one important place — the quality of the beat.

The print (FY27-Q1, ended Apr 26 2026): Revenue $81.62B (+85% Y/Y), consensus ~$78.8-79.2B → ~+3.5% beat; GAAP diluted EPS $2.39; GM 74.9%, right on the mid-70s target. Drivers: Data Center $75.2B (+92%), Networking the standout inside it.

The correction — the beat is lower-quality than the headline. Of $69,903M pre-tax income, $15,929M (22.8%) is non-operating mark-to-market gains on equity stakes, not compute sold. So:

• Reported GAAP EPS $2.39 vs operating EPS ex-marks ≈ $1.85
• The operating business still grew ~65-70%+ and carried a 65.6% operating margin — genuinely excellent — but the reported +211% net-income growth is inflated by the equity-mark tailwind.

Margins: GAAP GM 74.9%, operating margin 65.6% — margin recovery from the H20-charge year complete. Inventory provisions only $0.8B this quarter (vs $2.3B year-ago) — cleaner cost of revenue.

Guidance & tone: Q2 FY27 guide ~$91.0B ±2% with GM ~75%, assuming zero China DC compute — the next print (Aug) is the first real test of the delta.

Balance-sheet flags (now hard): inventory $25,797M (+$4.4B/qtr; raw materials +75% seq to $6,647M — building ahead of Rubin); AR net $40,710M; FCF ≈ $48.6B (OCF $50,344M − $1,757M capex).

Capital allocation — the signal, now filing-sourced: dividend raised 25x ($0.01 → $0.25) and $80.0B added to buyback (May 18 2026) on top of $38.5B remaining = $118.5B authorized; $20.2B repurchased in the quarter. Management is telling you the cash flows are structural.

Market reaction (the delta tell): the stock is down ~12% in 30 days and −14% off the May 14 ATH ($235.47) into this refresh — a beat-and-raise franchise being sold on macro AI-capex-durability fear, not on its own numbers. "Beats get faded" is now "the whole complex gets re-rated."

Lens 6 · Earnings Calls (sentiment trend) (carried — no new call in window)

No earnings call occurred in the delta window (last was FY27-Q1 in May; next ~late Aug). The FY27-Q1 sentiment arc from the prior dossier stands: tone has moved promotional → structural/declarative ("AI factories… the largest infrastructure expansion in human history is accelerating"), anchored on ~$500B Blackwell+Rubin visibility. Interim signals in the window: Jensen confirmed Vera Rubin in full production at CES 2026, shipping Q3 2026, and NVIDIA certified all three HBM4 suppliers — consistent with the "declarative confidence" read. The standing risk unchanged: the more management anchors on $500B/$1T visibility, the more a single soft data point (a capex trim, the Rubin unit cut, an MI450 win) becomes a sentiment shock. See deep-dive-2026-06-09.md for the full multi-call arc.

Lens 7 · Comps (re-run — NVDA line moved; peers carried)

The prior peer table (AMD/AVGO/TSM/INTC/MU, multiples ) is carried by reference — those names did not re-rate materially in a month. NVDA's own line moved and is re-struck:

Where NVDA screens now: the prior dossier's "21x forward" blended NVIDIA's own (optimistic) $9.85 NTM estimate; against consensus the honest figure is ~25x FY27 / ~20x FY28. Still undemanding for ~65-85% operating growth and 100%+ ROE, and now ~12% cheaper than the prior read — but not the "cheapest growth name in the group at 21x" headline. It sits mid-pack-to-cheap versus AVGO (~25x fwd) and a fraction of AMD (~56x fwd). EV/EBIT per-peer and 5-yr-avg ROE remain n/a (do not fabricate). The consensus-vs-my-base gap is itself the signal: the Street's FY27 $8.06 / FY28 ~$10 sit below the prior dossier's base ($8.85 / $11.50) — i.e. consensus already models more deceleration than the base case (see Lens 11).

Lens 8 · Stock-Price Catalysts (re-run — a new >5% move + the driving narrative)

New catalyst in the delta window: NVDA from the $235.47 ATH close (May 14 2026) to ~$202.78 (Jul 9) — −14% off the high, −~12% in 30 days. No single NVDA-specific event; the driver is the AI complex re-rating as "AI spending is surging faster than revenue and markets are starting to notice". This extends the five-year pattern rather than breaking it:

• Narrative/regime shocks > reported numbers (2023 +24% guide shock; 2025-01 DeepSeek −17%/−$600B; now a capex-durability re-rate) — the tape reacts to the AI-capex regime, not the quarter.
• China is a bidirectional catalyst — H200 framework/15%-25%-USG is live optionality.
• Beats no longer move it up; the complex moving down takes it down — NVDA is the index's AI proxy on the way down as much as up.

Implication (unchanged, reinforced): the variant-perception edge is in the narrative inputs — hyperscaler-capex durability, Rubin execution + the unit cut, MI450 traction, the equity-book reflexivity, China optionality — not in modeling the next beat, which is expected and faded.

---

Phase C — Judge People & Books

Can I trust the operators and the accounting?

Lens 9 · Management (carried + one legal delta)

The prior assessment carries in full (see deep-dive-2026-06-09.md): Jensen Huang, co-founder/CEO ~33 years, 3.3-3.8% owner ($150-190B stake), best-in-class capital compounder (ROE ~100%+); CFO 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. Buyback discipline question stands ($20.2B repurchased near ATHs; $118.5B authorized).

Legal delta (material): the 2017-18 crypto-mining channel-inventory securities class action advanced — on March 25 2026 the district court GRANTED class certification (class = purchasers Aug 10 2017 – Nov 15 2018); on April 8 2026 NVIDIA petitioned the Ninth Circuit for leave to appeal (Rule 23(f)). No liability accrued ("reasonably possible, not probable"). The rhyme remains the interesting part: execs accused of downplaying a cyclical demand driver — the precedent the market's "is AI capex durable or a cycle?" question echoes. A new datum worth logging: a $3,957M "accrued purchase consideration related to the Groq, Inc. non-exclusive license agreement" — NVIDIA taking a ~$4B licensing position tied to an inference-chip rival; track, don't over-read.

Lens 10 · Forensic Red Flags (re-run — the headline flag is now ~5x larger)

Earnings remain high-quality on a cash basis (OCF $50.3B > operating net income; the $15.9B equity gains are correctly backed out of OCF ). But the risk migration to the balance sheet — the prior dossier's central forensic thesis — accelerated hard. Ranked by severity:

1. Circular financing / mark-to-market income — the flag is now dominant, not emerging. The equity book doubled to ~$73.6B in one quarter (marketable equity $30.2B + non-marketable $43.4B), plus ~$9.2B warrants/convertibles and $8.9B locked-up publicly-held equity in Other assets — ~$90B of AI-ecosystem investment exposure. $15,929M of gains flowed through the P&L this quarter = 22.8% of pre-tax income; $13.4B of it is unrealized on publicly-held names, and $27.4B of the publicly-held stack is under lock-up (can't be sold). NVIDIA spent $18.6B buying new private stakes in the quarter — nearly as much as the $19.3B it spent on its own buyback. These marks reverse violently if AI-private valuations crack — the same event that would hit the core business. This is no longer a footnote; it is a quarter of GAAP EPS.
2. Customer concentration — extreme, filing-confirmed. Top-3 = 54% of revenue / 30-18-16% of AR, all Compute & Networking. A capex pause at two accounts = a >30% revenue event.
3. Off-balance-sheet + investment commitments stacked higher. $119B supply ($95B within FY27) + $30B cloud + $27B investment commitments + $3.5B facility-lease guarantees (up from $860M) + $32.4B future DC leases — against ~$50.3B liquid (cash $13.2B + marketable debt $37.1B). Much is matched to booked shipments and some supply is cancellable/reschedulable per Note 10 — but the $27B investment + $3.5B guarantees are pure "financing our own customers" exposure.
4. Inventory build. $21.4B → $25.8B (+$4.4B/qtr), raw materials +75% sequentially — pre-building for Rubin; writedown risk is not hypothetical (the H20 precedent), though provisions this quarter were modest ($0.8B).
5. SBC — quantified, still neutralized. SBC $1,928M/qtr, $20.8B unearned over ~2.6yr — but buybacks ($20.2B/qtr) >> SBC, so diluted shares are falling (24,391M vs 24,611M). Unlike most tech, SBC is not silently diluting holders.
6. Effective tax crept up to 16.6% (from 14.3%) as SBC tax benefits shrank relative to income; IRS examining FY2023-24. Minor.
7. GPU-depreciation thesis (industry-level, indirect). Carried: the bear argument that hyperscalers under-depreciate GPUs (5-6yr vs ~2-3yr real life), overstating customer AI profits and thus the durability of their capex (= NVIDIA's revenue). NVIDIA's own D&A is small/clean.

Forensic bottom line (sharpened): still NOT accrual manipulation — cash generation is real and high-quality. But the periphery the prior dossier flagged has become the story: a ~$90B book of stakes in its own customers now contributes ~23% of GAAP EPS via unrealized, largely locked-up, AI-cycle-correlated marks. Every risk — concentration, commitments, inventory, equity marks — points at the same trigger (an AI-capex slowdown), so they compound, not diversify. Confirm fair-value leveling of the private stakes against the FY2026 10-K's critical-accounting-estimates on non-marketable equity.

Regulatory findings (required sub-section). No SEC enforcement history — SEC EDGAR EFTS returns 0 Litigation Releases and 0 AAERs naming NVIDIA over 2021-07-10 → 2026-07-10. 10-K/10-Q Item — Legal Proceedings: the only material litigation is the 2017-18 crypto-mining securities class action (class certified 2026-03-25, NVIDIA seeking interlocutory appeal) + stayed derivative suits; no accrued liability. Non-SEC (web): the material regulatory theme is export control, not enforcement — the USG H20/H200 China licensing regime with a 15% (H20) / 25% (H200) revenue-or-tariff take to the US government; antitrust scrutiny of NVIDIA's AI-market position is a watch-item but no consent decree/fine found. Conclusion: no material regulatory or accounting-enforcement findings — verified via SEC EDGAR EFTS (LR, AAER), 10-K/10-Q Item Legal Proceedings, and web as of 2026-07-10; the live exposures are the securities class action and export-control policy, both disclosed.

---

Phase D — Project & Stress-Test

What's it worth, and what kills it?

Lens 11 · Forward Projection (EPS, next 3 fiscal years) (re-run — re-anchored to consensus + Rubin cut)

forecast.ts is a Brier tracker, not an EPS model — built bottom-up from actuals, labeled ``. NVIDIA FY ends late Jan (FY27 = Feb 2026-Jan 2027). No forecast.ts create this pass (unattended breadth mode).

Anchors: FY27-Q1 actual GAAP EPS $2.39 / operating ex-marks ~$1.85; Q2 FY27 guide ~$91.0B ±2%, GM ~75%. Consensus: FY27 EPS ~$8.06, FY28 ~$10 (dispersed $6.24-9.98). Shares ~24.2-24.4B, falling on the $118.5B buyback. Projection is on an operating/non-GAAP basis — it deliberately excludes the volatile equity-mark line, which is un-forecastable and would flatter GAAP.

Triangulation: at ~$202.78 the market pays ~25x consensus FY27 ($8.06) / 20x FY28 ($10). Consensus now sits below the prior dossier's base and roughly at this pass's (trimmed) base — i.e. the Street has moved to price more deceleration, and the ~12% de-rate did the work. The investable question is still which FY28-29 path you underwrite: base/bull repays ~20-25x; the bear (margin compression into an earnings decline) makes today's multiple ~26x trough earnings and the stock halves.

Lens 12 · Bull vs Bear (re-run — both cases updated for the delta)

Bull case. NVIDIA is the toll booth on a build-out whose toll estimate just went up: hyperscaler 2026 capex raised to ~$725B (+77%), 2027 >$1T; the operating business printed +85% revenue at a 74.9% gross margin and a 65.6% operating margin, with Compute & Networking alone at a 71.5% segment operating margin — proof the moat has widened from "the chip" to "the rack." Management holds ~$500B Blackwell+Rubin visibility, is shipping Rubin from Q3 2026, and just told you the cash is structural ($118.5B buyback + 25x dividend on ~$48.6B/qtr FCF). And the stock is ~12% cheaper than a month ago at ~25x consensus FY27 for a franchise consensus itself rates "Strong Buy" with a ~$300 target — with China H200 relief as un-modeled optionality.

Bear case (permanent-impairment vectors):

1. The equity book is now a second, reflexive business. ~$90B of AI-ecosystem stakes, 23% of GAAP EPS from unrealized marks, $18.6B of fresh private-stake buying in a single quarter, $27B more committed. If AI-private valuations crack, the marks reverse, the neoclouds NVIDIA funds lose their cheap capital, and the demand they represent evaporates — all correlated, all at once. This is bigger and more reflexive than the prior dossier could see.
2. Concentration is a single point of failure. Three buyers = 54% of revenue / 30-18-16% of AR — and they're building Trainium ($20B run-rate)/MTIA/Maia/TPU to escape.
3. The supply-and-competition window turned adverse. Rubin's 2026 units cut 2.0M→1.5M on HBM4; AMD MI450 lands in the same window with a memory-density/TCO claim; inference-share erosion (90%→20-30% by 2028) is the structural end-state.

Pre-mortem (late 2027, the thesis broke). A hyperscaler signals 2027 capex is "digested" not grown, citing utilization below plan and a pivot to its own silicon for inference. The $91B/quarter run-rate — built on a China-zero guide, three buyers, and $500B of "visibility" — loses its marginal buyer. $50B+ of pre-committed CoWoS/HBM4 converts to inventory writedowns; the ~$90B equity book marks down 30-40%, flipping "other income" deeply negative and revealing GAAP EPS was ~23% air. The stock doesn't need a fundamental collapse — just deceleration + one private-AI re-rate — to lose the multiple and de-rate to ~12x on falling estimates.

Are multiples too high? No — ~20-25x forward is below AI-logic peers and undemanding for the growth. The risk isn't the multiple; it's the earnings base and its quality. If FY28-29 follows the bear path (an earnings decline), today's ~25x forward is ~26x trough earnings and the stock halves; and the reported earnings are ~23% mark-to-market, so a downturn hits the "E" twice — operating deceleration and mark reversal.

Contrarian view — what the market is refusing to see. The consensus is still debating whether training demand saturates; the real question the FY27-Q1 filing answers is that NVIDIA's revenue quality is degrading even as the headline rips — it increasingly must finance its own demand (now ~$90B of stakes, $3.5B of lease guarantees, $30B of cloud buy-backs) and book the financing's gains as a quarter of its own EPS. When the cost — and the accounting flatter — of manufacturing demand rises faster than the demand it manufactures, franchise quality has already peaked, regardless of the next print. The edge is in the footnotes, and the footnotes just got louder.

Lens 13 · Devil's Advocate (short-seller) (re-run)

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 and who are building to displace NVIDIA. If customer AI economics don't pencil (and Burry's under-depreciation argument says reported customer profits are ~$176B too high 2026-28), their capex — NVIDIA's revenue — corrects with the accounting.
• The most dangerous competitor is still the customer, and now it has a run-rate. AWS's chip business is at ~$20B; Google TPU/Ironwood is GA; MI450 gives even non-captive buyers a merchant alternative in Rubin's own window. Bulls underestimate that NVIDIA's biggest customers are its most motivated — and now most funded — competitors.
• The accounting tell got worse, not better. A quarter where $15.9B / 23% of GAAP EPS is mark-to-market gains on stakes NVIDIA bought $18.6B more of in the same quarter is vendor-financing-as-earnings at industrial scale. $27B more committed. $3.5B of lease guarantees dressed as credit derivatives. This is what manufactured demand looks like on the way up.
• Assumptions that must ALL hold for $202: (1) hyperscaler capex compounds toward $1T without a digestion quarter; (2) NVIDIA holds premium share vs MI450 + custom; (3) operating margin stays ~52-56% through the HBM4 cost step-up; (4) the ~$90B AI-private book doesn't mark down; (5) no DeepSeek-style efficiency shock. Each has wobbled once; the Rubin unit cut and the −12% tape say two are wobbling now.
• Valuation if growth disappoints 20-30%: bear path → FY28-29 operating EPS ~$7.75-8.25 (a decline), GAAP far worse if marks reverse; on a de-rated 12-14x that's a 40-55% drawdown from $202. The DeepSeek day (−$600B in a session) proved the gamma is real.
• Single permanent-impairment scenario + plausibility: a hyperscaler capex pause coinciding with an AI-private-valuation re-rate — the loop runs in reverse (marks down → neoclouds can't fund → marginal buyer vanishes → $50B+ supply commitments become writedowns). Plausibility: moderate, not tail — it needs only deceleration + one private-AI re-rate, both overdue after a 3-year vertical run, and the equity book that reverses is now ~$90B, not ~$8B.

Lens 14 · Management Questions (carried + 2 updated for the delta)

The prior 15 (ranked by information value) carry in full — see deep-dive-2026-06-09.md. Two are re-pointed by the delta and move to the top:

• [NEW #1] This quarter, $15.9B (23% of GAAP EPS) was unrealized gains on your ~$90B book of stakes in customers, $27.4B of which is locked up. What is your framework for the point at which financing your own demand becomes the demand — and how should investors value earnings you can't sell and that reverse with your customers' valuations?
• [NEW #2] You cut the 2026 Rubin unit target from ~2.0M to ~1.5M on HBM4 verification. Which supplier is the binding constraint into FY27, what is the qualified Samsung/Micron ramp, and does the MI450's 432GB memory advantage change your inference positioning in the same window?
• (Then the prior list: circular-financing revenue share; top-3 backlog-vs-speculative; GPU useful-life vs hyperscaler 5-6yr depreciation; supply-commitment writedown thresholds + cancellability; inference-share defense; $91B backlog-vs-book-and-ship; buyback-discipline trigger; China 2027 realistic revenue; networking attach durability; succession/key-man; Rubin inference-deflation self-cannibalization; equity-stake-as-manufactured-demand; gross-margin floor vs AMD; recurring/software revenue to de-cyclicalize.)

---