Thematic Exposure -- 10/10

NVIDIA is the defining infrastructure platform for the AI revolution -- >92% data center GPU market share, extreme CUDA ecosystem lock-in, and absolute pricing power across the most important technology theme of the decade. Weight: 25%

Segment exposure

Segment	% of Revenue	Market Share	TAM	Theme Growth
Data Center AI/Compute	91.5%	~92-97%	$400B+ by 2030	>40% CAGR
Gaming	5.5%	~80% discrete GPU	~$45B	~5-8% CAGR
Professional Visualization	1.9%	~90% workstation GPU	~$10B	~10-15% CAGR
Automotive	0.9%	~20-30% AV compute	~$25B by 2030	>25% CAGR

Q4 CY2025 segment revenue breakdown. TAM estimates from management commentary and industry research.

Oligopoly assessment

Verdict

Effective monopoly with one aspiring competitor. Classic oligopoly with 1 player controlling >90% share.

1. How many competitors >15% share?
Zero in data center AI GPUs. NVIDIA holds ~92%+. AMD has ~8%. Intel is <1%. No competitor exceeds even 10% share.

2. Can customers replace within 12 months?
No. CUDA ecosystem lock-in with 5.9M+ developers, 4,400+ accelerated applications, and a deep software stack (TensorRT, Triton, NeMo, NIM). Switching costs are enormous. AMD ROCm is improving but remains years behind in ecosystem maturity.

3. Price setter or taker?
Absolute price setter. NVIDIA commands premium ASPs ($30K-$40K per GPU) with customers willing to pay for performance-per-watt leadership and lowest TCO. GB300 NVL72 delivers 50x performance per watt vs. prior gen.

Data Center Share

~92%+

AMD ~8%, Intel <1%

CUDA Developers

5.9M+

Ecosystem lock-in

Accelerated Apps

4,400+

Deep software moat

GPU ASP

$30-40K

Absolute price setter

AI revolution infrastructure platform

Defining theme

Management cites $3-4 trillion in cumulative AI infrastructure spend by end of decade.

Three scaling laws intact: Pre-training, post-training, and inference-time compute scaling are all expanding simultaneously, driving sustained demand across the entire AI stack
Hyperscaler CapEx explosion: Aggregate 2026 CapEx estimates now sit at ~$600B, up $200B+ from start of year. Every major CSP (Microsoft, Google, Amazon, Meta, Oracle) is deploying NVIDIA at massive scale
Blackwell/Rubin visibility: The company has visibility to $500B+ in Blackwell/Rubin revenue, with Rubin platform entering volume production in CY2026
Agentic AI inflection: Token generation is growing exponentially -- Microsoft processed 100 trillion tokens in a single quarter, up 5x YoY. This drives inference GPU demand directly
Physical AI next vector: Jensen Huang has repeatedly emphasized physical AI, robotics, and autonomous systems as the next $1T+ opportunity after agentic AI

Cumulative AI Spend

$3-4T

Management estimate through 2030

2026E Hyperscaler CapEx

~$600B

Up $200B+ from start of year

Blackwell/Rubin Pipeline

$500B+

Revenue visibility

Data Center TAM

$400B+

Annual by 2030, >40% CAGR

Competitive moat

CUDA ecosystem lock-in. With 5.9M+ developers and 4,400+ accelerated applications, CUDA represents the deepest software moat in semiconductors. The full stack -- TensorRT, Triton, NeMo, NIM -- creates compounding switching costs that grow with each new developer and application. AMD ROCm is improving but remains years behind in breadth and maturity.

Annual product cadence. NVIDIA has shifted from a biennial to annual architecture cadence: Hopper → Blackwell → Blackwell Ultra → Rubin. Each generation delivers substantial performance-per-watt improvements (GB300 NVL72: 50x vs. prior gen), making it impossible for competitors to close the gap.

Switching costs. Customers face enormous friction in moving workloads off NVIDIA hardware. Enterprise AI deployments are built on CUDA, optimized for NVIDIA architectures, and deeply integrated into production inference pipelines. The cost of retraining, reoptimizing, and revalidating on alternative hardware typically exceeds any potential savings.

Competitive dynamics and secular trends

The competitive landscape in AI accelerators is effectively a one-player market with aspiring challengers. AMD is the most credible competitor with the MI450 architecture seeing deployment at Oracle, but AMD remains at ~8% share and lacks the software ecosystem depth to win large-scale inference deployments. Custom ASICs from hyperscalers (Google TPU, Amazon Trainium, Microsoft Maia) address internal workloads but do not compete for the broader enterprise and sovereign AI market.

The secular trends are overwhelmingly favorable. AI model complexity continues to scale on all three dimensions -- pre-training data, post-training reinforcement, and inference-time compute. Each new generation of frontier models requires more compute, not less. The emergence of agentic AI is particularly bullish: autonomous AI agents generate tokens continuously rather than in response to human prompts, creating a structural increase in inference demand that is growing exponentially.

Sovereign AI represents a nascent but meaningful demand vector. Governments worldwide are investing in domestic AI compute infrastructure for national security and economic competitiveness. NVIDIA is the default platform for these deployments given its ecosystem maturity and support infrastructure.

The key risk to the thesis is not competition but demand sustainability. If hyperscaler CapEx moderates or if AI scaling laws hit diminishing returns, the entire demand curve shifts. However, current evidence suggests we are still in the early innings of the AI infrastructure buildout, with cumulative spend tracking well below the $3-4T management estimate through 2030.

Score rationale

10/10 — >92% share in the defining theme of the decade, classic oligopoly with one player controlling >90% of the market, theme growing at >40% CAGR, extreme switching costs via the CUDA ecosystem (5.9M developers, 4,400+ apps), and absolute pricing power. NVIDIA is not merely exposed to the AI theme -- it is the AI infrastructure platform. No other company in the coverage universe scores higher on thematic alignment, market dominance, and secular tailwind strength.

Data sourced from Daloopa, earnings transcripts, and industry research.