NVIDIA May Be the Most Expensive Cyclical Stock in History

NVIDIA Corporation (NASDAQ: NVDA) In-Depth Stock Research Report

Chapter 1: Executive Summary

1.1 Thesis in a Nutshell

NVIDIA is the most profitable semiconductor company in history (A-Score 8.1, ROE 101.5%), but its $4.31T market cap prices it as a "perpetual AI platform," a valuation applied to a company still fundamentally reliant on CapEx cycles. The analyst's own forecast projects a revenue decline by FY2030, yet the 36x P/E ratio assumes perpetual growth. A great company, but not a great price.

1.2 Core Conflict

The market prices NVDA as a "perpetual infrastructure platform" (P/E 36x), but the inherent cyclicality of AI CapEx suggests it may be "the largest infrastructure build-out cyclical stock in history." This conflict—Perpetual Platform vs. Cyclical Giant—is a central theme throughout the 34 chapters of this report and remains ultimately unresolved.

1.3 Conditional Rating

Worldview	Probability	Valuation Range	Rating
AI as a general-purpose technology, CUDA moat >10 years	15%	$6.0-7.5T	Neutral Watch
Cloud computing path, CUDA moat 5-8 years	35%	$3.5-4.5T	Neutral Watch
Wireless path, CUDA erosion	25%	$2.2-3.2T	Cautious Watch
Fiber optics path, cycle collapse	15%	$1.2-2.0T	Cautious Watch
Black swan / multiple crises	10%	$0.8-1.5T	Cautious Watch
Probability-Weighted	100%	$3.43T	Cautious-to-Neutral (52:48)

1.4 Why "Cautious-to-Neutral" and Not a Stronger Conviction?

Not "Cautious Watch (Strongly Bearish)": Because NVIDIA is, indeed, an exceptionally high-quality company. With an A-Score of 8.1, ROE of 101%, Z-Score of 57, and FY2026 growth of +66%, it is top-tier by any corporate quality metric. The issue is the price, not the company.

Not "Neutral Watch" or "Watch": Because the $4.31T valuation requires scenario B-4 (sustained CapEx growth) to hold true, a single point of failure (SPOF) that our own FY2030E revenue forecast (-2.8%) already suggests is unsustainable. The probability-weighted valuation ($3.43T) is 20% below the current price.

The Meaning of 52:48: The probabilities of a neutral outcome (no decline) and a cautious one (a decline is warranted) are nearly even. This is not an evasion ("we are not sure") but an honest answer to CQ-2 (platform vs. cycle), a question that is fundamentally unpredictable. Whether AI is the third general-purpose technology revolution in human history, following electricity and the internet, is not a question an investment report can answer.

1.5 Core Questions (CQ) List

The following eight core questions are central to our analysis and are answered in a closed loop in Chapter 30.

CQ-1: Is AI CapEx an infrastructure build-out or an inflating bubble? (Weight: 15%)

Final Verdict: Leaning Bearish (56%) — The 4:1 gap between $600B in CapEx and $100B in AI revenue is historically unprecedented, and our own forecast anticipates a revenue decline by FY2030 (-2.8%). The CapEx cycle will eventually peak, though a breakout in AI Agents could extend it.
Key Uncertainty: Whether demand elasticity is >3 (which would cause the gap to self-correct). Tracking Signal: Hyperscaler CapEx growth in FY2027 Q2-Q3.

CQ-2: Is NVDA a perpetual platform or a cyclical supplier? (Weight: 15%)

Final Verdict: Uncertain (50:50) — This cannot be determined, as it depends on the AI penetration rate, a fundamentally unpredictable variable. The business model shows signs of platformization (Software + NVLink), but over 90% of revenue comes from hardware.
Key Uncertainty: Whether software ARR can exceed $3B (a confirmation signal of platformization). Tracking Signal: Subscription revenue as a percentage of total, SaaS pricing.

CQ-3: Will the Rubin transition period create an "air pocket"? (Weight: 12%)

Final Judgment: Moderately Optimistic (70%) — Q1 guidance $78B (+15% QoQ), supply commitment $95.2B doubled, FY2027 H1 will not have an air pocket, but H2 requires attention.
Key Uncertainty: Whether Rubin shipment timing will be delayed (15-20% probability). Tracking Signals: FY2027 Q2 guidance, sequential growth rate.

CQ-4: Is the CUDA moat widening or eroding? (Weight 12%)

Final Judgment: Risk Increasing (65%) — Direction is clear (it is eroding), but slower than expected. Performance gap narrowed to 10-30%, Triton rising, half-life 5-8 years.
Key Uncertainty: Will Triton become the inference standard? Will OpenAI decouple from CUDA? Tracking Signals: GPT-5 training hardware composition, ROCm maturity.

CQ-5: Can in-house chips replace >20% of NVDA within 3 years? (Weight 12%)

Final Judgment: Moderately Pessimistic (61%) — In-house inference share will increase from <10% to 20-25% (in 3 years), training remains secure. Big 5 customers are expected to reduce NVIDIA purchases by 13% in FY2028.
Key Uncertainty: In-house chip yield/software stack bottlenecks. Tracking Signals: Actual deployment volume of AWS Trainium, Google TPU proportion.

CQ-6: Will gross margin structurally decline to 71%? (Weight 10%)

Final Judgment: Moderately Pessimistic (57%) — Product mix change is structural (rising inference share + system-level products), FY2028 gross margin may decline to 69-71%.
Key Uncertainty: Can inference ASP be increased through software value-add? Tracking Signals: FY2027 gross margin trend, networking revenue share.

CQ-7: Is the China market permanently lost? (Weight 8%)

Final Judgment: Pessimistic (67%) — China's AI decoupling from NVIDIA has passed a critical point (share decreased from 66% to 8%), even if restrictions are lifted, the upper limit would be 25-30%. Sovereign AI and growth in other markets have fully offset this.
Key Uncertainty: Will there be a major turning point in US-China relations? Tracking Signals: BIS policy updates, Huawei Ascend 920 mass production data.

CQ-8: What conditions are required for a $4.3T valuation to be justified? (Weight 6%)

Final Judgment: Moderately Overvalued (63%) — Probability-weighted valuation $3.43-3.6T, current $4.31T requires AI Agent explosion + sustained CapEx increase + CUDA moat >8 years to justify.
Key Uncertainty: The ultimate penetration rate of AI as a general-purpose technology. Tracking Signals: P/E natural regression, AI Enterprise ARR.

---

Chapter 2: NVIDIA — From Gaming Graphics Cards to the Architect of AI Infrastructure

2.1 From Gaming Graphics Cards to the Architect of Computing Paradigms

In 1993, when Jensen Huang, Chris Malachowsky, and Curtis Priem founded NVIDIA in San Jose, California, the company's mission was to make graphics on personal computers better. Thirty years later, this company is valued at $4.31T, surpassing Japan's GDP, becoming the most valuable single enterprise in human history.

This leap from being a "graphics card maker" to "defining computing paradigms" was not achieved in one go but underwent three identity revolutions:

First Leap (1999-2006): Gaming GPU Specialist

The GeForce 256, launched in 1999, was dubbed the "world's first GPU." NVIDIA defined a new product category—the graphics processing unit—and established dominance in the PC gaming market through DirectX compatibility and close collaboration with game developers. The business model at this stage was simple: design chips → TSMC foundry → sell to OEM/retail.

Second Leap (2006-2016): General-Purpose Computing Platform

The launch of CUDA in 2006 was an inconspicuous but decisive moment. CUDA allowed scientists and engineers to run non-graphics computing tasks on GPUs using C-style code. At the time, almost no one—including Wall Street—believed this would be significant. But CUDA did something fundamental: it transformed the GPU from a fixed-function hardware into a programmable computing platform.

The cost of this decision was considerable. CUDA's development and maintenance consumed significant R&D resources, while early commercial returns were meager. From 2006 to 2012, CUDA primarily served niche markets such as scientific computing and oil exploration, with negligible direct contribution to revenue. However, it built something crucial: a developer ecosystem.

As of FY2026, the CUDA ecosystem boasts over 4 million registered developers, 1,800+ GPU-accelerated libraries, and 600+ AI/ML framework optimizations. This 20-year accumulation of software assets cannot be quickly replicated and forms NVIDIA's deepest moat.

Third Leap (2016-Present): The Benchmark of AI Infrastructure

In 2012, AlexNet achieved groundbreaking results in the ImageNet competition using NVIDIA GPUs, marking the beginning of the deep learning era. However, the true inflection point was the launch of ChatGPT at the end of 2022. From that moment, every tech company—and an increasing number of non-tech companies—required significant GPU compute power.

The revenue trajectory from FY2024 to FY2026 tells this story:

Fiscal Year	Revenue	YoY	Net Income	Net Margin	Data Center % of Revenue
FY2023	$27.0B	0%	$4.4B	16.2%	~56%
FY2024	$60.9B	+126%	$29.8B	48.8%	~78%
FY2025	$130.5B	+114%	$72.9B	55.8%	~87%
FY2026	$215.9B	+65%	$120.1B	55.6%	~90%

Over three years, revenue grew eightfold, from $27B to $216B. Data Center revenue share rose from 56% to approximately 90%. NVIDIA transformed from a chip company with gaming and data centers each accounting for half of its business into a giant enterprise almost entirely driven by AI infrastructure.

2.2 Four-Layer Business Model

NVIDIA's business model can be understood through four concentric circles:

First Layer: GPU Chip Design (Core Gravity)

NVIDIA is a pure design (fabless) company. All chips are manufactured by TSMC, with HBM memory provided by SK Hynix/Samsung/Micron. NVIDIA's job is to design the best GPU architecture and then outsource manufacturing. This means:

• Extremely High Gross Margin: 71.1% for the full year FY2026, Q4 recovered to 75.0%
• Extremely Low Capital Intensity: CapEx of $6.0B accounts for only 2.8% of revenue
• Extremely High ROIC: 174.4%
• But it also means production capacity is entirely dependent on suppliers (especially TSMC CoWoS packaging).

Second Layer: CUDA Software Platform (The Moat)

CUDA is not just a programming framework, but a complete development toolchain:

• cuDNN (Deep Learning Primitives)
• TensorRT (Inference Optimization)
• NCCL (Multi-GPU Communication)
• Triton Inference Server (Inference Serving)
• NeMo (Large Model Training)

These software layers are provided free of charge but only run on NVIDIA GPUs. This creates the classic "razor-and-blade" model: free software → attracts developers → applications locked to NVIDIA hardware → hardware premium pricing.

Key Figures: Over 95% of global AI research papers use CUDA for GPU programming. Major AI frameworks (PyTorch, TensorFlow, JAX) have the deepest optimizations for CUDA. This is more than just technological leadership—it's 20 years of accumulated network effects.

Third Layer: System-Level Solutions (Expanding)

Starting in FY2025, NVIDIA accelerated its shift from "selling chips" to "selling systems":

• DGX/HGX: GPU server systems
• NVL72/NVL144/NVL576: Rack-scale GPU clusters
• SuperPOD: Data center-scale AI factories
• Spectrum-X: Ethernet network switches (FY2026 Q4 networking revenue $11B, +267% YoY)

This transition has two sides:

• Positive: Increases ASP (from single card ~$25K to system ~$3M), enhances customer stickiness.
• Negative: System-level products include more non-GPU components (networking, cooling, racks), leading to structurally lower gross margins than pure chips.

Fourth Layer: Software Subscriptions (Emerging)

NVIDIA AI Enterprise is a software subscription service (approx. $4,500/GPU/year). FY2026 software ARR has not yet exceeded $1B—this is a key inflection point signal identified by 100baggers.club. If NVIDIA can convert millions of installed GPUs into recurring software revenue, its valuation logic will fundamentally change: from hardware multiples to software multiples.

2.3 Jensen Huang: Founder-CEO's 35-Year Vision

Jensen Huang is one of the very few founder-CEOs in tech history to have led a company from its inception to a $4T market capitalization. Several defining characteristics:

Vision: Launched CUDA in 2006 when no one saw its potential; launched Volta (the first GPU designed specifically for deep learning) in 2017 when AI was still an academic topic; driving the transition from chips to systems in the AI frenzy of 2024. Each time, he positioned the company 3-5 years ahead of the market.

Execution: The annual iteration pace from Blackwell to Rubin has transformed from "promises" to predictable "execution." Q1 FY27 guidance of $78B exceeded expectations by over $5B.

Concentration: Jensen holds approximately 3.4% of NVDA shares (~$150B), completely aligning his personal wealth with the company's destiny. However—insider trading data shows 529 sell transactions vs. only 2 buy transactions for the full year 2025. Is this a signal or noise? For a CEO with a net worth of $150B, a $5M reduction is statistically insignificant, but the pattern itself is noteworthy.

Key Risks: NVIDIA's strategic coherence is highly dependent on Jensen Huang personally. There is currently no clear succession plan. If Jensen retires or an unforeseen event occurs within the next 5 years, NVIDIA's strategic direction, execution, and industry influence could experience unpredictable changes.

2.4 Organizational Structure and Cultural DNA

NVIDIA's organizational structure is a direct embodiment of Jensen Huang's management philosophy:

Extremely Flat Structure: Jensen claims to have 60+ direct reports. There is no traditional hierarchical management – this is extremely rare for a company with 42,000 employees. The advantages are rapid decision-making and unfiltered information; the disadvantage is that Jensen personally becomes the bottleneck for all important decisions.

"Top 5 Things" Culture: Each employee regularly updates Jensen on their top 5 most important tasks. Jensen uses this method to bypass middle management and directly understand frontline situations. This management model can also be seen at Apple (Steve Jobs) and Tesla (Elon Musk) – it is suitable for founder-CEOs with extraordinary energy but lacks institutionalized transferability.

R&D Dominant: Among 42,000 employees, an estimated 25,000+ are engineers. R&D of $18.5B equals approximately $440,000 per person in R&D. This is a purely R&D-driven organization – SG&A of only $4.6B (2.1% of revenue) means sales/marketing/administrative functions are minimized.

Employee Growth Control: From 18,975 in FY2021 to 42,000 in FY2026, the number of employees grew 2.2 times, but revenue grew 12.9 times. This "revenue growth significantly outpacing employee growth" pattern is characteristic of a platform-based business model – the CUDA ecosystem enables a small number of engineers to support a massive revenue base.

Year	Employees	YoY	Revenue/Employee	Net Profit/Employee	Revenue Growth/Employee Growth
FY2022	22,473	+18%	$1.20M	$0.44M	3.4x
FY2023	26,196	+17%	$1.03M	$0.17M	0x (Revenue Flat)
FY2024	29,600	+13%	$2.06M	$1.01M	9.7x
FY2025	36,000	+22%	$3.63M	$2.03M	5.2x
FY2026	42,000	+17%	$5.14M	$2.86M	3.8x

Anomalous Signal: The employee growth rate in FY2025 (+22%) was the highest in the past 5 years, with a ratio of 5.2x compared to FY2025 revenue growth (+114%) – Jensen accelerated hiring during the AI boom. However, the employee growth rate slowed to +17% in FY2026 while revenue growth was +65%, suggesting NVIDIA is controlling the hiring pace to ensure per-capita efficiency does not decline.

2.5 Corporate Governance and Incentive Alignment

Shareholding Structure: Jensen Huang holds approximately 3.4% of NVDA shares (~$150B). While the percentage is not high, the absolute amount ($150B) means Jensen's interests are fully aligned with shareholders – this is one of the largest CEO shareholdings by value on Earth.

SBC (Stock-Based Compensation): FY2026 SBC of $6.4B = 3.0% of revenue. For a tech giant, this ratio is very reasonable (Meta ~12%, Google ~8%). More importantly, NVIDIA's share buybacks ($40.1B) are 6.3 times the SBC – annual SBC dilution is fully covered by buybacks, and net shares outstanding decreased by 0.73%.

Incentive Alignment Matrix:

Stakeholder	Alignment	Basis
CEO (Jensen)	★★★★★	$150B shareholding, founder status
Employees	★★★★☆	Sufficient SBC, stock price appreciation = significant wealth effect
Shareholders	★★★★☆	Buybacks > SBC, net shares outstanding decreased, but no meaningful dividends
Customers	★★★☆☆	Excellent products but potentially excessive pricing power

Insider Trading Signals Re-evaluation: 529 sells vs 2 buys throughout 2025 full year might look concerning, but context is needed:

• The vast majority of sales are preset 10b5-1 automatic trading plans (non-timing decisions)
• For Jensen, who holds $150B in stock, selling $50M represents 0.03% of his assets
• More importantly: Have any executives abnormally increased their selling plans before earnings announcements?
• Current evidence does not support the conclusion that "insiders knew bad news and sold in advance"

---

2.6 Jensen Huang's Management DNA — Founder's Premium and Succession Risk

Jensen's Strategic Decision Tree

Several key decisions made by Jensen Huang over the past 35 years have defined NVIDIA:

Year	Decision	Consensus at the time	Actual Outcome
1999	Defined the "GPU" category	No one used this term	Created a $500B+ market
2006	Launched CUDA	Waste of money, GPUs are only for graphics	Became the strongest moat 20 years later
2016	All-in AI	AI was still academic research	Seized a $4T opportunity
2020	Acquired Mellanox ($7B)	Too expensive	Networking business reached $30B+ in FY2026
2022	ARM acquisition failed → In-house CPU development	Strategic failure	Grace/Vera CPU successfully developed in-house
2024	Annual iteration cadence	Too aggressive	Established a pace of always being one generation ahead
2024	System-level products (NVL72)	Gross margin would decline	Profit contribution per customer actually increased

Pattern Recognition: Jensen's strategic characteristics are "3-5 year ahead planning + extreme execution + unafraid of short-term controversy." CUDA had almost no commercial return for 6 years after its launch, but Jensen persevered. The Mellanox acquisition was considered a 40% premium at the time, but a $7B investment now generates $30B+/year in revenue.

2.7 Management Incentive Alignment Analysis

Executive	Title	Stock + Options Holdings	Annual Compensation	Alignment
Jensen Huang	CEO	~3.4% (~$147B)	~$30M	Very High
Colette Kress	CFO	~0.03% (~$1.3B)	~$20M	High
Debora Shoquist	EVP Operations	~0.01% (~$430M)	~$15M	High

Jensen's $147B holding aligns his interests highly with shareholders. Even with divestments, the absolute value of his holdings remains astonishing.

Note, however: Stock-Based Compensation (SBC) constitutes an extremely high proportion of management compensation. FY2026 SBC could reach $8-10B (approximately 7-8% of operating profit) – this is an implicit shareholder dilution cost. A lesson from our SMCI report: SBC is a real cost and should not be ignored in valuations.

2.8 Quantification of Succession Risk

Jensen is 62 this year, healthy and energetic. However, as the linchpin of a $4.3T company, the absence of a succession plan is a non-zero probability systemic risk:

Analogous Analysis:

Company	Founder Departure Year	Market Cap at Departure	Consequence	Recovery Time
Apple	2011 (Jobs passed away)	$380B	Short-term -7%, long-term sustained growth	<1 year
Microsoft	2000 (Gates handover)	$500B	14 years of market cap stagnation	14 years (Nadella)
Oracle	2014 (Ellison handover)	$180B	Growth slowed	Currently recovering
Berkshire	TBD (Buffett)	$1T	—	—

If Jensen retires/leaves within the next 5 years:

• Short term: Stock price could fall 10-20% (confidence shock)
• Mid term: Depends on the successor's strategic direction
• Risk: NVIDIA could become more "bureaucratic" → lose its aggressive annual iteration pace → competitors catch up
• Mitigation: NVIDIA's product roadmap is planned until 2028 (Feynman), even if Jensen leaves, inertia could sustain for 2-3 years

Successor Speculation: NVIDIA has no publicly known "#2 person". This is in stark contrast to Apple (Tim Cook clearly designated as successor). This is a governance risk worth including in Phase 2 key monitoring metrics.

---