## Chaîne d'approvisionnement NVIDIA : chaque fournisseur et partenaire

NVIDIA is the most important company in the AI revolution, but it does not operate in isolation. Its supply chain is a web of critical dependencies that investors must understand.

Fabrication : TSMC (100% Dépendance)

Tout GPU NVIDIA est fabriqué par TSMC. Le H100 utilise le processus 4nm de TSMC, tandis que les GPU Blackwell B100/B200 utilisent la couche 3nm. NVIDIA n'a aucune capacité de fabrication interne. Cela fait de la relation TSMC la liaison de chaîne d'approvisionnement la plus importante en technologie.

Memory: The HBM Triad

Each H100 GPU contains 80GB of HBM3 memory, sourced from three suppliers: SK Hynix (primary), Samsung, and Micron. The Blackwell B200 doubles this to 192GB of HBM3E. With HBM supply constrained, memory allocation has become a gating factor for GPU production.

Packaging: CoWoS Constraints

TSMC's Chip-on-Wafer-on-Substrate (CoWoS) advanced packaging is another bottleneck. Each AI GPU requires CoWoS to integrate the compute die with HBM stacks. TSMC has been aggressively expanding CoWoS capacity, but demand continues to outstrip supply.

Networking: Mellanox and the Interconnect Stack

NVIDIA acquired Mellanox in 2020 for $7 billion, gaining control of InfiniBand - the dominant networking technology for AI clusters. This vertical integration gives NVIDIA an end-to-end platform advantage that competitors struggle to match.

Server Partners

Super Micro Computer (SMCI), Dell, HPE, and Lenovo build the server systems that house NVIDIA GPUs. SMCI has been the fastest-growing partner, capturing significant share of AI server builds due to faster time-to-market.

The Customer Concentration

Microsoft Azure is NVIDIA's single largest customer, estimated at $15B+ annually. Google Cloud, Amazon AWS, Meta, and Oracle round out the top five. These hyperscalers collectively represent roughly 60% of NVIDIA's data center revenue, creating significant customer concentration risk.

The Competitive Landscape

AMD's MI300X is the only credible GPU alternative, but with less than 10% market share. The real threat may come from custom silicon: Google's TPU (designed by Broadcom), Amazon's Trainium, and Microsoft's Maia. However, NVIDIA's CUDA software ecosystem remains a formidable moat.