“The GPU has done to AI what the microprocessor did to personal computing it is the enabling hardware layer, and whoever controls its supply chain controls the frontier.” A graphics processing unit (GPU) is a specialized processor originally designed for rendering images but uniquely suited to the parallel matrix computations that underlie neural network training and inference, making it the critical hardware substrate for modern AI.
Executive Summary
GPUs perform thousands of simple mathematical operations simultaneously, in contrast to CPUs that perform fewer operations sequentially. This parallel architecture maps almost perfectly onto the matrix multiplication at the heart of neural network computation, allowing GPUs to train AI models orders of magnitude faster than general-purpose processors. Nvidia dominates the AI GPU market with approximately 70-80% share, making its H100 and upcoming Blackwell GPU architectures the most strategically contested hardware in the world. The US government’s decision to make advanced AI GPU exports to China the primary lever of technology competition has transformed an electronic component into a geopolitical focal point.
The Strategic Mechanism
- Parallel processing advantage: A modern GPU contains thousands of cores operating simultaneously, enabling the matrix algebra of neural networks to run in hours rather than months compared to CPU-based computation.
- Memory bandwidth: AI training requires moving enormous amounts of data between processor and memory at high speed. The H100’s 3.35 TB/s of memory bandwidth is purpose-designed for this bottleneck.
- NVLink interconnects: Training large models requires multiple GPUs to communicate at high speed. Nvidia’s NVLink allows up to 900 GB/s GPU-to-GPU transfer, enabling the large GPU clusters that train frontier models.
- CUDA software ecosystem: Nvidia’s proprietary CUDA programming platform has become the standard development environment for AI, creating software lock-in that competitors AMD, Intel, and Huawei must overcome to gain market share.
- Supply chain concentration: Advanced GPU manufacturing requires TSMC’s most advanced semiconductor processes (4nm and below), concentrating the entire AI compute supply chain in a single geographic chokepoint Taiwan.
Market & Policy Impact
- Nvidia’s market capitalization exceeded $3 trillion in 2024, briefly making it the world’s most valuable company, driven almost entirely by AI GPU demand from hyperscalers and AI labs.
- H100 GPUs were selling for $25,000-$40,000 on secondary markets in 2023 more than 4x their list price as supply constraints created the first hardware bottleneck in AI development history.
- US export controls banned H100 and A100 GPU sales to China beginning October 2022, forcing Chinese labs to seek domestic alternatives and prompting Huawei to accelerate Ascend 910B development.
- Saudi Arabia’s Project Transcendence and UAE’s G42 both centered their sovereign AI programs on acquiring large H100 GPU clusters, with the US conditioning sales on security agreements limiting Chinese access.
- TSMC’s 3nm process (used in Nvidia’s next-generation Blackwell GPUs) requires extreme ultraviolet lithography machines from ASML a single Dutch company creating a two-node chokepoint in the entire global AI hardware supply chain.
Modern Case Study: Nvidia’s H100 and the AI Hardware Arms Race, 2023-2025
When Nvidia shipped its H100 GPU in volume from late 2022, it created a hardware scarcity crisis that restructured the entire AI industry. Microsoft, Google, Meta, and Amazon each committed to purchasing H100 clusters of 16,000-100,000 units, totaling hundreds of billions of dollars in orders that pushed Nvidia’s data center revenue to $47.5 billion in fiscal 2024. Chinese buyers, blocked by US export controls from purchasing H100s, drove a gray-market supply chain through Singapore and third-country intermediaries, prompting the US to expand entity-list restrictions in October 2024. Meanwhile, Huawei’s Ascend 910B estimated at 60-80% of H100 performance entered production, and Chinese labs began optimizing architectures to run efficiently on inferior hardware. The H100 episode established that advanced GPU access is a direct capability constraint for frontier AI development, validating the US export control strategy while simultaneously accelerating China’s domestic semiconductor development program.