Analysis: The AI Data Center Revenue Stack
A new analysis breaks down the revenue flow from AI data centers, showing Nvidia captures the lion's share at 73%. The rest is split between memory providers like SK Hynix and Micron (12%), foundry/packaging from TSMC (8%), and other components (7%). The data highlights TSMC's critical but relatively small piece of the value chain, a key dynamic for Apple's own silicon manufacturing.
Nvidia's data center revenue surged 75% year-over-year to $62.3 billion in its latest quarter, now accounting for over 91% of the company's total sales. This growth is fueled by massive investments in AI infrastructure from major cloud providers like AWS, Google Cloud, and Microsoft Azure. The demand has been so intense that Nvidia has scaled its data center business by nearly 13 times since the emergence of ChatGPT. The High-Bandwidth Memory (HBM) market, critical for AI accelerators, is dominated by SK Hynix, which held a 62% market share in the second quarter of 2025. Micron followed with a 21% share, while Samsung's share dropped to 17%. This demand has led to SK Hynix's entire 2026 HBM stock being sold out. TSMC's role, while representing a smaller revenue slice, is foundational, particularly its advanced packaging technology known as CoWoS (Chip-on-Wafer-on-Substrate). Advanced packaging is expected to account for over 10% of TSMC's revenue in 2025, with gross margins higher than the company average. Nvidia has already secured over 70% of TSMC's CoWoS-L capacity for 2025 to support its Blackwell architecture GPUs. For Apple, this supply chain is paramount. Its most advanced silicon, including the upcoming M5 chip, relies on TSMC's cutting-edge manufacturing processes in Taiwan. While TSMC is expanding its presence in Arizona, where it produces older chips like the A16 Bionic, the latest process nodes and crucial advanced packaging steps remain offshore, highlighting a strategic dependency for its highest-performance products. The sheer cost of this ecosystem is staggering. A single AI server rack can cost over $500,000, and building a new AI-ready data center can range from $10 million to over $20 million per megawatt. The specialized GPUs are the most significant expense, with a high-end Nvidia H100 costing tens of thousands of dollars. Looking ahead, the AI chip market is projected to experience explosive growth, with some forecasts predicting it will approach $500 billion in revenue in 2026, making up roughly half of all global chip sales. This growth is concentrated in high-value AI chips, which, while driving half the revenue, represent less than 0.2% of total chip unit volume.
