StockNews 中文版 →
NVDA 22 min read

NVDA — Investment Tree v1

Stories lie, structure doesn't.

Date: 2026-04-30 Anchor: $209.25 / $5.05T market cap / forward P/E 25.10× Archetype: Dominant Incumbent Under Forced Dual Transition H-0 confidence: ~75% supported

I. The contradiction

The market prices NVIDIA at 25× forward earnings — a multiple consistent with a compounding platform company — while simultaneously stress-testing the position as if NVIDIA's revenue could mean-revert thirty to forty percent on the next AI capex cycle. These two views are logically inconsistent. Either NVIDIA has durable platform moats — CUDA lock-in, sovereign AI demand, the inference-time scaling tailwind — that justify a twenty-five-times multiple, or it doesn't, and it should trade at fifteen to eighteen times like AMD. The current price requires both narratives to be simultaneously true, which forces the analyst to take a position on which description is correct.

The data presented here argues that the platform interpretation is empirically more consistent with the simultaneous facts — but that the market has only partially re-rated NVIDIA, and the partial re-rating is fragile because it has not been named. The mispricing magnitude is modest — perhaps fourteen percent — but the mechanism of mispricing is structurally important: the market is anchored on hardware-cyclicality more than the lifecycle-stage evidence supports, while ignoring the unrealized option of platform monetization that would force a second re-rating event.

This essay reconstructs the thesis from first principles, tests it against twenty leaves of falsifiable evidence, and produces a probability-weighted twelve-month target of $222 — a modest +6% from current levels. The asymmetry is intact, but it is no longer the high-conviction trade it was at the DeepSeek shock low of $110. NVIDIA at $5.05 trillion is a coherent-thesis hold for existing positions; new money has materially reduced asymmetry.

II. The simultaneous facts

Six facts must be reconciled before any thesis can be defended:

F1 — Margin profile. NVIDIA's FY2026 non-GAAP gross margin landed at 71.3% — below the 74-78% range that the scaffold (and most sell-side models) had assumed, but still well above any commodity hardware vendor. This is the Blackwell-ramp dilution showing up in the income statement: the transition from chip-only sales (HGX systems) to full-rack datacenter solutions carries higher capex and lower margin per dollar of revenue. Critically, the 71.3% is uncomfortably close to the 72% line that the falsification framework calls FF2 — gross margin below 72% for two consecutive quarters would constitute partial falsification of the platform thesis.

F2 — Architecture cadence. NVIDIA committed at GTC 2026 to a one-year roadmap: Hopper → Blackwell → Rubin (2027) → Feynman (2028). This is not a hardware commodity cycle. It is a software-subscription trigger that forces existing cluster owners to face a perpetual "buy now or wait one year for two-times performance" decision. The cadence sustains demand through internal NVIDIA upgrade rather than external NVIDIA-to-AMD switching — exactly the behavior of a platform company whose customers face switching costs.

F3 — Post-DeepSeek capex resilience. When DeepSeek R1 demonstrated reasoning at a fraction of frontier-lab compute cost in January 2026, NVIDIA shed roughly $600 billion of market cap in a single session — the largest single-day market-cap loss in U.S. equity history. Within ninety days, all four major hyperscalers — Microsoft, Alphabet, Amazon, Meta — had raised, not cut, their FY2026 AI capex guidance. Combined hyperscaler capex now sits in the $260-320 billion range. This is Jevons paradox in action: efficiency improvements expand total compute consumed faster than they reduce per-task compute requirements. The bear case that "AI capex will normalize because efficiency wins compress demand" has no current data signal.

F4 — Custom silicon at fifteen to twenty-five percent. Hyperscaler-captive ASICs — Google TPU v7, AWS Trainium3, Microsoft Maia 200, Meta MTIA — collectively hold an estimated 15-25% of total hyperscaler AI compute workloads, growing at a 44.6% CAGR. The number is significant but mostly captive-inference: Google now runs about 75% of Gemini computation on TPUs internally. NVIDIA's training franchise — where CUDA lock-in is strongest — remains largely intact. The threat is real but mechanically slow.

F5 — Forward multiple. NVIDIA's forward P/E of 25.10× on FY2027 consensus EPS implies the market is paying for roughly 32% revenue CAGR through FY2029. Consensus midpoint already tracks +46% Q1 FY2027 YoY; the trajectory only needs to decelerate to ~28-32% to justify the multiple. The price is fair-to-modest against consensus, not aggressive. This is the largest delta from the scaffold-time view, when the assumed price was roughly $110 and the multiple was assumed materially more aggressive.

F6 — Sovereign AI as a new demand category. Aggregate sovereign commitments — Japan via SoftBank/NTT, UAE via G42, Saudi Arabia, France, India — total $5-15 billion across FY25-FY26. At 2.6-7.7% of Compute & Networking, this is below the 10% threshold where it would dominate the cyclicality argument. But the category exists, where it didn't in any prior NVIDIA cycle, and it is on a multi-year procurement budget with national-security motivations that do not correlate with hyperscaler capex cycles.

The hardware-cycle interpretation cannot reconcile F1 and F2 simultaneously without invoking platform economics. The platform interpretation reconciles all six.

III. The H-0 thesis

H-0: NVIDIA is priced as a late-cycle semiconductor hardware incumbent by a market anchored on historical chip cyclicality, when the evidence simultaneously supports that NVIDIA is in the early stage of a software-platform transition — where the CUDA/NIM ecosystem is becoming the durable moat and the GPU hardware is the delivery vehicle for compounding platform rents.

The mechanism of mispricing is cognitive bias × lifecycle stage, with structural blindness on software option value as a secondary mechanism. The market anchors on NVIDIA's semiconductor-cyclical history and maps it onto a company that has structurally different economics from its prior hardware cycles. The structural blindness on software means that the unrealized platform option is both unquantified and therefore unstressed-tested in either direction — the bear cannot price its absence and the bull cannot price its presence, so the multiple sits in an intermediate range that satisfies neither narrative fully.

H-0 decomposes into five Level-1 branches:

Together these five branches cover both sides of H-0. The L1A through L1C branches test the bull interpretation; L1D tests the bear interpretation; L1E tests the upside option.

IV. Profit Pool Defense (L1A)

The first question is whether NVIDIA's share of the AI-compute profit pool is expanding, stable, or contracting — and whether sovereign AI plus inference-time scaling expansion is large enough to offset any share losses to custom silicon.

The profit-pool math in FY2026 favors NVIDIA decisively. Compute & Networking revenue of $193.5 billion at 71.3% gross margin produces approximately $138 billion of gross profit. AMD's MI-series and the entire custom-silicon stack — TPU v7, Trainium3, MTIA, Maia 200 — collectively capture far less attributable AI-accelerator gross profit because captive ASICs do not generate merchant economics. NVIDIA's share of the AI-accelerator profit pool, distinct from the AI-accelerator workload pool, comfortably exceeds 75%.

The pace of custom-silicon absorption is the live concern. Workload share moved from near-zero in 2022 to 15-25% by 2026, growing at a 44.6% CAGR. Mechanical extrapolation puts custom silicon at roughly 30% of workloads by FY2027 and 40% by FY2028 — approaching what Andy Grove would call a 10× competitive-dynamic change by FY2028. However, the absorption is captive-inference-skewed. Google running 75% of Gemini computation on TPUs internally is the canonical data point, and inference now represents about two-thirds of total AI compute by the Introl analysis. The training franchise, where CUDA lock-in adds 6-18 months of migration friction per workload, is not the primary loss surface.

Sovereign AI as a structurally non-correlated demand category remains modest in dollar terms — $5-15 billion aggregate against $193.5 billion of segment revenue, or 2.6-7.7%. Below the 10% threshold where it would meaningfully buffer cyclicality, but the category exists, which it didn't in any prior NVIDIA hardware cycle. National-security motivation creates multi-year procurement budgets that do not synchronize with hyperscaler capex cycles.

Inference-time scaling — the o1/o3-class reasoning-chain compute pattern — should structurally expand NVIDIA's TAM as more reasoning queries consume more compute per query. Direct verification is not possible because NVIDIA does not separately disclose inference versus training GPU-hours. Indirect evidence supports the thesis: hyperscaler capex was raised post-DeepSeek, consistent with Jevons-paradox dynamics, and Q1 FY2027 consensus revenue at $78.8 billion annualizes to roughly +46% YoY — antithetical to compression.

L1A verdict — partially supported, leaning supportive. The profit pool is intact in FY2026 and will mechanically erode through FY2028 as custom silicon scales, but the erosion is moderate, not collapse. The thesis names the mechanism: NVIDIA cedes captive-inference share progressively while the training franchise holds; sovereign and inference-TAM expansion partially offset; and platform optionality (L1E) is the unhedged upside that, if it materializes, more than offsets the share-loss math.

V. Valuation Expectations (L1B)

The second question is what revenue trajectory and margin profile the current $209.25 price requires, and whether that scenario is internally consistent with the L1A profit-pool evidence.

The mechanical reverse-DCF works as follows. At 25.10× forward P/E on FY2027 consensus EPS, the market pays roughly 25× for ~$315 billion of forward run-rate revenue (annualizing the $78.8 billion Q1). For a 20× terminal multiple by FY2029, the implied FY2029 EPS must support a $5 trillion equity value at 20× — implying $250 billion FY2029 net income. Working backward at the FY2026 net margin of approximately 56% suggests FY2029 revenue around $500 billion. From $215.9 billion FY2026 to $500 billion FY2029 is 32% CAGR.

This is at consensus midpoint, not above it. Consensus already prices roughly +46% Q1 FY2027 YoY; the trajectory through FY2029 needs only to decelerate to a 28-32% range to justify the multiple. The required scenario is plausible, not heroic. The price is fair-to-modest against consensus, not aggressive — this is the largest delta from scaffold-time analysis, when the price was assumed near $110.

The margin sensitivity is the live falsification mechanism. Consensus appears to embed margin recovery from the 71.3% Blackwell-ramp trough back toward 74-75%. A reasonable bet — Blackwell maturation historically delivers ASP uplift, and Rubin in FY2027 should re-anchor the high end. But 71.3% is at the edge of the 72% line that defines FF2: gross margin below 72% for two consecutive quarters would constitute partial falsification. Q1 FY2027 (May 20) is the next test. AMD MI400 pricing leverage and Trainium3's 3nm process at 144GB HBM3E create plausible pressure for further compression. Each 100bps of margin compression at $315 billion forward run-rate translates to roughly $2.5 billion of net income and 1% to EPS — material at multiples north of 25×.

The relative-valuation question reaches for structural analogues, not point-in-time comparables. AMD's 2018-2020 ramp peaked near 50× forward P/E on a sub-$10 billion revenue base and crashed when crypto ASIC cycles reversed — not analogous because NVIDIA's customer base is enterprise, sovereign, and hyperscaler, not retail. NVIDIA's own 2018 crypto peak ran to 25× then collapsed below 10× — also not analogous because the current driver is enterprise capex with multi-year procurement, not retail demand. The Microsoft-Azure transition in 2013-2015 is the most structurally analogous: a dominant incumbent (Windows / Office) layering a recurring infrastructure platform (Azure) on top of legacy hardware and license revenue. Microsoft traded 14-18× through that period, then re-rated to 25-30× as Azure scaled and proved durable. NVIDIA at 25× sits squarely within the Microsoft-analogue band, which means the current price is consistent with platform-emergence pricing — not pre-pricing it heroically.

The mean-reversion downside scenario is mathematically constructable. A 25% Data Center revenue contraction from FY2026 peak, combined with multiple compression to 15× trough P/E, produces a $1.24 trillion mcap and roughly $51 per share — a 76% drawdown from current levels. The DeepSeek shock proved that this kind of move can happen in a single session. But the trough scenario requires a sustained Data Center contraction that has no present indication in the data. Hyperscaler capex was raised, not cut, post-DeepSeek. Q1 FY2027 consensus encodes continuation, not normalization.

L1B verdict — strongly supported, with margin caveat. At $209.25 with forward P/E 25.10×, the price is fair against consensus, not aggressive. The 32% CAGR-to-FY2029 requirement sits at the consensus midpoint. The Microsoft-Azure structural analogue justifies the 25× multiple. The single live falsification mechanism is gross margin compression below 72% for two consecutive quarters — Q1 FY2027 is the next test.

VI. S-Curve Position (L1C)

The third question is where NVIDIA sits on the AI-compute S-curve — pre-knee, at-knee, or post-knee — and how rents are distributing across the Value Net of TSMC, SK Hynix, and the supporting supply chain.

The penetration test goes first. Total enterprise IT spending globally is roughly $5 trillion. AI infrastructure spending sits at about $300-400 billion run-rate — NVIDIA Compute & Networking alone is $193.5 billion, plus custom-silicon at $50-80 billion, plus networking, power, and real-estate buildout. That places AI infrastructure at approximately 6-8% of total enterprise IT — well below saturation. Most Fortune-500 enterprises remain in pilot or early-deployment phase for production AI; only hyperscalers and a small set of AI-native companies are at scale. Penetration of the addressable enterprise AI workload universe is plausibly below 25%, putting NVIDIA Data Center pre-knee on the S-curve, not at it.

The Blackwell utilization pattern provides the cleanest near-term diagnostic. Q3 FY2026 Data Center revenue was $51.2 billion, up 66% year-over-year, during the Blackwell ramp. If Blackwell were primarily displacing Hopper rather than expanding workloads, total revenue growth would be capped at roughly the ASP-and-performance-uplift multiple times stable cluster count — implying perhaps 25-35% growth, not +66%. The empirical 66% growth rate is consistent with net-new workload expansion alongside displacement. Sovereign AI clusters, new hyperscaler deployments, and reasoning-chain inference all add net-new compute consumption beyond replacement.

The Value Net rent question is more nuanced. TSMC's CoWoS packaging and SK Hynix's HBM3e supply have been bottlenecks throughout FY2026, allowing both to capture incremental rent. SK Hynix's HBM gross margin spiked into the 30-40% range against historical 15-20%. NVIDIA's gross margin compressed from above 75% in FY2025 to 71.3% in FY2026 — partly mix shift to full-system rack solutions, partly the supply chain capturing rent. The compression is real but partly supply-constraint-driven and partly reversible. Rubin's TSMC N2 allocation is multi-year; HBM4 capacity is being added by Micron and Samsung, not just SK Hynix. The Value Net rent migration is more likely partial and transitory than structural and permanent — though the leading-edge node (N2) likely retains durable pricing power.

The architecture cadence test is qualitative but important. NVIDIA's published one-year cadence — Hopper to Blackwell to Rubin to Feynman — creates a "buy now or wait one year for two-times performance" decision that mechanically forces continuous upgrade. AMD's 18-month cadence and ROCm maturation give customers a credible "wait" option, but installed-base CUDA lock-in keeps the upgrade decision NVIDIA-internal rather than NVIDIA-to-AMD. There is no public evidence of customer upgrade fatigue: hyperscaler capex was raised post-DeepSeek; sovereign orders are accelerating; Rubin pre-orders are reportedly oversubscribed. Historical Blackwell ramp delay (CoWoS packaging) shows cadence can slip — but Rubin's N2 allocation is reportedly secured.

L1C verdict — strongly supported. NVIDIA Data Center is pre-knee on the AI-compute S-curve. Enterprise penetration is below 30%; Blackwell is driving net-new workload expansion alongside displacement; the one-year architecture cadence sustains the demand curve through internal upgrade economics; the only caveat is partial Value-Net rent migration to TSMC and SK Hynix which is supply-driven and partially reversible. The lifecycle stage is early compounding, not late maturity.

VII. Strategic Inflection Test (L1D) — the bear case rejected

The fourth question tests the bear's strongest case directly. Andy Grove defined a strategic inflection point as a 10× change in competitive dynamics — the point at which "what got you here won't get you there." The bear case for NVIDIA names three candidate 10× changes: Huawei Ascend reaching parity, AMD ROCm displacing CUDA at scale, and algorithmic efficiency improvements compressing absolute compute demand below the architecture FLOP improvement pace.

Three of these four leaves test bearishly and resolve against the bear. Only one resolves ambiguously.

The Huawei parity test fails empirically and structurally. Ascend 910C reportedly delivers 60-70% of H100 throughput at memory-bandwidth parity; the gap to H200 and Blackwell is wider, perhaps 30-40%. The U.S. global market is already on Hopper-to-Blackwell, will be on Rubin in 2027, and on Feynman in 2028. Even if Ascend reaches H100-class parity by FY2028, the global market will be on Feynman by then — Huawei is a one-to-two-generation behind moving target, not a parity threat in the planning horizon. The BIS January 2026 final rule that loosened export-license review (case-by-case approval for H200 to China) further narrows even the China captive-advantage window. The bear's Huawei case is rejected.

The algorithmic efficiency case fails on the data. DeepSeek R1 demonstrated approximately 10× cost reduction on reasoning at similar quality. Subsequent algorithmic wins — MoE scaling laws, RL-based test-time compute, distillation — show 2-5× per generation. NVIDIA architecture roadmap delivers 2-3× FLOP improvement per generation. The two paces are roughly matched in isolation, but Jevons paradox dominates empirically: post-DeepSeek hyperscaler capex was raised, not cut, because efficiency improvements expand total compute consumed faster than they reduce per-task requirements. The bear's algorithmic-efficiency case is rejected.

The capex normalization signal has not fired. As of Q1 FY2027 print preview (May 20 pending), no hyperscaler has guided FY2027 AI capex flat-to-down. Microsoft, Google, Amazon, and Meta Q4 2025 calls all maintained or raised AI infrastructure capex guidance. Q1 FY2027 consensus revenue at $78.8 billion annualizes to ~$315 billion, +46% YoY — the antithesis of normalization. The bear's capex-normalization case is rejected.

The ROCm displacement test is the only one that resolves ambiguously, and it resolves not yet. AMD's ROCm covers 80-85% of PyTorch operation compatibility but still lacks production-grade profiling, debugging, and compiler toolchain parity. Migration friction sits at 6-18 months of engineering effort per major model. No major hyperscaler has publicly deployed ROCm-only training clusters at scale. Meta has internal contributions and AMD MI-series in mixed clusters, not ROCm-only. The threat exists but has not crossed the inflection threshold; AMD MI400 in FY2027 with ROCm 7+ could change this rapidly.

L1D verdict — not supported. Three of four 10×-change tests fail empirically, and the fourth is forward-looking and inconclusive. The bear's strategic-inflection thesis is not supported by current evidence. This is structurally important for H-0: rejecting the bear's strongest case strengthens the bull case by elimination, even before the platform-optionality (L1E) discussion.

VIII. Software Platform Emergence (L1E) — the unrealized option

The fifth question tests the bull's strongest case: is the CUDA/NIM software layer generating compounding rents that constitute a separate business model from hardware?

The headline finding is that the load-bearing test cannot be currently satisfied because NVIDIA has not separately disclosed software-platform revenue in its FY2026 10-K. NIM microservices were announced and shipping at GTC 2026; NVIDIA AI Enterprise is sold through hyperscalers; CUDA-Enterprise licensing exists. But the discrete revenue line is not broken out. Absence of disclosure does not mean absence of revenue, but it means the test threshold ($500 million ARR run-rate) cannot be verified from public filings.

The cross-side network effect question is empirically observable but causally ambiguous. NVIDIA hardware deployments and CUDA-native software development have grown in lockstep across 2020-2026. Frameworks — PyTorch, JAX, TensorRT-LLM — optimize CUDA paths first; ISVs build CUDA-first; enterprise AI startups assume CUDA. The 95%-of-training-on-CUDA statistic is the network-effect outcome. The interpretive question is whether software is creating hardware demand (platform pull) or following hardware deployment (hardware pull). Distinguishing the two from public data is hard. Meta's PyTorch ROCm contributions show that the ecosystem can become hardware-agnostic if a major participant invests, suggesting the network effect may be path-dependent rather than structural.

The developer community growth rate test, however, resolves clearly in favor of the platform thesis. CUDA developer count grew from approximately 3 million in 2019 to approximately 25 million by 2025-2026 — roughly 50% CAGR over five-to-six years, well above the 20% pre-inflection threshold. The growth rate does not appear to be slowing materially. The platform community is exhibiting pre-inflection-platform dynamics on the developer-side metric.

The historical-analogue test points at Microsoft 2013-2015 (Azure scaling) as the most structurally relevant comparison. Microsoft did succeed in transitioning from a Windows / Office license business to an Azure-centric subscription business. The leading indicators of success were three: discrete Azure revenue disclosure starting in FY2014; developer-community migration to Azure-native services; and hyperscaler-class capex commitment. NVIDIA in 2025-2027 has the second and third — developer ecosystem on CUDA, hyperscaler-class infrastructure spend — but lacks the first. The Microsoft analogue suggests that the platform transition can succeed but that disclosure is the gating event for re-rating. Until NVIDIA discloses, the market cannot fully re-rate.

L1E verdict — partially supported; the platform thesis is plausible but unverifiable. Developer-community dynamics and ecosystem growth are consistent with platform emergence; cross-side network effects are present but causally ambiguous; the historical analogue supports plausibility. The single missing piece is software-revenue disclosure. This is the load-bearing leaf for H-0's upside re-rating: if NVIDIA discloses ≥$500 million software ARR at Q1 FY2027 or a subsequent investor day, the platform thesis activates and the multiple has room to expand from 25× toward Microsoft-Azure-2015 multiples in the 30× range.

IX. Three scenarios

The thesis decomposes into three coherent scenarios, each constructed as a narrative with associated price target.

Bull scenario — "Platform thesis activates" (target $310, +48%). L1C pre-knee S-curve compounds; L1A inference TAM expansion empirically materializes; L1E software ARR disclosure happens at Q1 FY2027 (May 20) or a subsequent investor day. Custom silicon at 25-30% workload share by FY2028 is absorbed by total TAM expansion. Market re-rates from "AI hardware leader" to "AI infrastructure platform" at Microsoft-Azure-2015 multiples. The required FY2027 trajectory is +46% revenue (at consensus midpoint) with margin recovery to 74-75%, FY2028 at +30% with software disclosed at $1.5 billion ARR run-rate. Twelve-month equity value at 30× FY2028 EPS of approximately $210 billion produces a $7.5 trillion market cap, or roughly $310 per share.

Base scenario — "Continuation, no re-rating" (target $240, +15%). L1B verdict (current price fair against consensus) holds. NVIDIA delivers consensus-class revenue and EPS through FY2027-FY2028 without dramatic surprises in either direction. Margin oscillates in the 71.5-73% band as Blackwell matures and Rubin ramp begins. Custom silicon erodes captive-inference share but training profit pool intact. No software disclosure; multiple holds at 25× as the platform thesis remains "implied but unnamed." FY2028 net income of $180 billion at 25× produces a $5.8 trillion mcap, or roughly $240 per share. This is the highest-probability path because it requires no narrative change — only continuation of the current data trajectory.

Bear scenario — "Capex normalization plus margin compression" (target $130, -38%). L1D's bear thesis activates despite current evidence: at least one major hyperscaler guides FY2027 AI capex flat-to-down in Q3-Q4 2026 commentary; FF2 fires (gross margin below 72% for two consecutive quarters); custom-silicon adoption accelerates beyond 44.6% CAGR. Market re-prices NVIDIA as a hardware cyclical at 15× trough P/E. FY2028 net income of $110 billion at 15× produces a $1.65 trillion mcap, or roughly $130 per share. This requires multiple unfavorable resolutions in coordination — lowest-probability path given current data, but historically precedented.

The implied-probability reverse engineering is instructive. At the current $209.25, holding Base at 50% probability, the price implies Bull at 13.5% and Bear at 36.5% — a substantively bearish posture, surprising given the +90% YoY rally from the DeepSeek shock low. The L1-verdict-consistent probability vector, by contrast, would be Bull 30% / Base 50% / Bear 20%, implying a fair value of $239 — about +14% above current.

We adopt a compromise probability vector of 25/50/25: slightly more bullish than the price implies, slightly less bullish than pure verdict-consistent. This produces an expected twelve-month price of $222 — approximately +6% from current. The reward-to-risk asymmetry (+48% / -38%) is approximately 1.26×, favorable but materially less so than at scaffold-time entry near $110.

The mispricing magnitude implied by H-0 is on the order of 14%. The market is partially correct — it has re-rated NVIDIA from semiconductor-cyclical to platform-emerging — but it is anchored on hardware-cyclicality more than the lifecycle-stage evidence supports. The mispricing is modest, not severe.

X. Triggers and red flags

The thesis is structured to update on observable events. Three triggers and three red flags constitute the live monitoring framework.

Triggers — bullish events that would update H-0 toward higher confidence.

T1 — Software ARR disclosure (high impact, ~25-30% probability). NVIDIA discloses NIM, NVIDIA AI Enterprise, or CUDA-Enterprise license revenue at ≥$500 million run-rate, separate from C&N segment. Resolution window: Q1 FY2027 earnings (May 20) or any subsequent investor day. On fire: L1E 1.1 ⊗ → ✅; H-0 confidence 75% → 85%; probability vector shifts to 35/50/15; expected price $245-260. This is the single largest re-rating catalyst because it forces analysts to apply a SaaS-class multiple to a discrete revenue segment.

T2 — Hyperscaler FY2027 capex raised (medium-high impact, ~50% probability). At least three of four (MSFT, GOOGL, AMZN, META) raise FY2027 AI capex guidance vs. prior commentary in Q3 or Q4 2026 earnings calls. Resolution window: July-November 2026. On fire: reinforces L1A 1.4 and L1D 1.4 rejection of bear; H-0 confidence 75% → 80%.

T3 — Sovereign AI ≥$25B aggregate (medium impact, ~40% probability). Cumulative sovereign commitments cross $25 billion, or sovereign AI is disclosed as ≥10% of Data Center revenue. Resolution: rolling. On fire: L1A 1.3 ⚠️ → ✅; adds geopolitical-diversification narrative to thesis.

Red flags — bearish events that fire falsification conditions.

RF1 — Margin compression below 72% for two consecutive quarters (high impact, ~30% probability). Q1 FY2027 (May 20) and Q2 FY2027 (~August) both print non-GAAP gross margin below 72%. This is the only currently-live mechanical falsification trigger (FF2). On fire: L1B 1.2 ⚠️ → ✗; H-0 confidence 75% → 60%; probability vector shifts to 15/45/40. Playbook: re-test whether the compression is supply-driven (Value Net rent migration to TSMC / SK Hynix) or demand-driven (ASP pressure from custom silicon competition). Supply-driven compression is partially reversible through capacity normalization; demand-driven compression is structural.

RF2 — Hyperscaler FY2027 capex flat-to-down (high impact, ~20% probability). Any major hyperscaler explicitly guides FY2027 AI infrastructure capex flat-to-down vs. FY2026 in Q3 2026 earnings commentary. Resolution: July-August 2026. On fire: L1D 1.4 reverses ✗ → ✅ (capex normalization signal); H-0 confidence 75% → 55%; probability vector 10/40/50.

RF3 — Custom silicon CAGR exceeds 60% (medium-high impact, ~15% probability). Independent analysis shows custom-silicon workload share growth above 60% YoY (vs. current 44.6% CAGR), or any hyperscaler exceeds 50% internal-workload share on custom silicon. Resolution: rolling quarterly trade-press. On fire: L1A 1.2 ⚠️ → ✗; Grove inflection thesis activates earlier than FY2028.

The most-watched number in the entire framework is the Q1 FY2027 non-GAAP gross margin print on May 20. A sequential improvement back to 72% or above clears FF2 and reinforces the bull case. A continued sub-72% print partially fires FF2 and forces a probability-vector update on the next quarterly result.

XI. Position guidance and conclusion

The recommendation framework is conditional on which scenarios resolve.

Current state — H-0 supported at ~75%, base case in play. Probability vector 25/50/25; expected twelve-month price $222 (+6%). Existing positions: hold; favorable asymmetry retained from earlier entry. New money: marginal; await Q1 FY2027 margin print before adding meaningfully. Hedge: not currently needed; no triggers fired in either direction.

If H-0 strengthens (T1 software disclosure fires). Probability vector → 35/50/15; expected price $245-260. Add on weakness; the platform thesis activates and the multiple has room to expand. Microsoft-Azure analogue suggests durable re-rating once software disclosure begins.

If H-0 weakens (RF1 margin or RF2 capex normalization fires). Probability vector → 15/45/40 or 10/40/50. Reduce on margin-compression confirmation. Do not average down until the trend reverses. Preserve the option to re-enter on falsification clearance — historical NVIDIA drawdowns (2018 crypto, 2022 inventory, 2026 DeepSeek) have all been V-shaped where the structural thesis remained intact.

The single most important fact to track is the Q1 FY2027 gross margin on May 20, 2026. Everything else is secondary because FF2 is the only currently-live mechanical falsification trigger.

The twelve-month base case is +6%. The twelve-month bull case is +48% conditional on disclosure. The twelve-month bear case is -38% conditional on margin and capex deterioration. The asymmetry is intact but materially reduced from scaffold-time entry. NVIDIA at $5.05 trillion is a coherent-thesis hold, not a high-conviction add. The trade is to retain existing exposure, monitor the live falsification metric, and update the probability vector on the May 20 print and subsequent quarterly cadence.

The structural insight remains: the market has partially named the thesis but cannot fully price the option of platform monetization without disclosure. NVIDIA's management holds the disclosure timing in their hands. The probability that they choose to disclose in the next twelve months — at Q1 FY2027 or a subsequent investor day — is perhaps 25-30%. The probability-weighted upside from disclosure alone is roughly 0.27 × ($310 - $222) = approximately $24 per share, or +11% expected return contribution from the platform-disclosure option specifically.

The trade is partially priced. The mispricing is modest. The thesis is alive. The next data point is in twenty-one days.

Tree v1 complete. Linked artifacts: evidence_2026-04-30.jsonl (R2-upgraded evidence), leaves.md (20-leaf hypothesis test), scenarios.md (Bull/Base/Bear), implied_prob.md (probability-vector reverse-engineering), triggers_redflags.md (live monitoring), dashboard.md (current state). Update cadence: at every material event (next: Q1 FY2027, 2026-05-20).