Core Thesis

I maintain a neutral stance on NVIDIA at $208.93 despite the consistent earnings beats. The company's data center revenue growth of 427% year-over-year in Q4 FY2024 reaching $47.5 billion masks underlying margin pressure as hyperscalers optimize for inference workloads over training. The H100 dominance window is narrowing as competitors close the performance gap and custom silicon proliferates.

Data Center Revenue Analysis

NVIDIA's data center segment generated $60.9 billion in FY2024, representing 78.4% of total revenue. The sequential growth decelerated from 206% in Q1 to 427% in Q4, indicating demand normalization. More critically, the revenue per GPU has declined 23% since peak pricing in Q2 FY2024 as enterprise customers negotiate volume discounts.

The H100 80GB commands approximately $25,000-$30,000 per unit versus the previous $40,000 peak. This 25% price erosion directly impacts gross margins, which compressed from 78.4% in Q1 to 73.0% in Q4 FY2024.

Compute Architecture Economics

Training workloads, which drove the initial AI boom, require different compute characteristics than inference. Training demands maximum throughput per GPU, justifying premium H100 pricing. Inference workloads optimize for cost per token, driving demand for lower-cost alternatives.

Meta's custom MTIA chips handle 50% of their inference workloads. Google's TPU v5 delivers 2.8x better performance per watt for LLM inference versus H100. Amazon's Trainium2 offers 30% better price-performance for training workloads under 70B parameters. These developments compress NVIDIA's addressable market by approximately $15-20 billion annually.

Competitive Positioning Metrics

AMD's MI300X delivers 1.3x memory bandwidth versus H100 at 60% of the price point. Intel's Gaudi3 achieves 40% better inference throughput per dollar for transformer models. While NVIDIA maintains architectural advantages in software ecosystem and interconnect technology, the hardware performance gap has narrowed from 5x in 2022 to 1.8x in 2024.

CUDA's moat remains substantial with 4.1 million registered developers, but PyTorch 2.0's compilation improvements reduce CUDA dependency by enabling 80% performance portability across accelerators.

Financial Trajectory Analysis

FY2025 guidance implies $112-115 billion revenue, requiring 84% year-over-year growth. This assumes data center revenue reaches $90-95 billion, demanding 50% sequential growth in H1 FY2025. Historical precedent suggests this growth rate is unsustainable given enterprise budget cycles and infrastructure deployment timelines.

Operating expenses increased 35% year-over-year to $7.9 billion in Q4, primarily driven by R&D investments in Blackwell architecture and software stack enhancements. The OpEx growth trajectory indicates margin pressure will persist through FY2025 as competition intensifies.

Blackwell Architecture Assessment

The B100 and B200 GPUs deliver 2.5x training performance and 5x inference performance versus H100. However, manufacturing complexity using TSMC's 4NP process creates supply constraints limiting initial production to 200,000 units in H1 FY2025.

Blackwell's 208 billion transistors require advanced packaging techniques, increasing manufacturing costs by 40% versus H100. This cost structure challenges NVIDIA's ability to maintain gross margins above 70% while competing on price.

Market Demand Quantification

Hyperscaler capital expenditure totaled $200 billion in 2024, with 35% allocated to AI infrastructure. Microsoft's $50 billion AI spending, Google's $32 billion, and Amazon's $28 billion create a $110 billion addressable market for AI accelerators.

However, inference workload growth will comprise 70% of compute demand by 2026 versus 30% training. This shift favors cost-optimized solutions over performance-maximized H100s, compressing NVIDIA's revenue per workload by an estimated 40%.

Risk Assessment

China export restrictions eliminate 20% of potential market, equivalent to $22-25 billion annual revenue impact. Geopolitical tensions could expand restrictions to additional semiconductor categories, further constraining growth.

Customer concentration risk persists with four hyperscalers representing 65% of data center revenue. Any reduction in capital allocation from these customers creates immediate revenue volatility.

Bottom Line

NVIDIA trades at 31x forward earnings despite decelerating revenue growth and margin compression. The inference workload transition and competitive encroachment justify neutral positioning until architectural advantages translate to sustained margin expansion or market share defense becomes quantifiable.