27 October 2025 Global Technology AI Value Chain: How much does a GW of data center capacity actually cost, and what goes into it? Stacy A. Rasgon, Ph.D. +1 213 559 5917 stacy.rasgon@bernsteinsg.com Daniel Zhu +1 917 344 8309 daniel.zhu@bernsteinsg.com Alex Wang, CFA +852 2123 2613 alex.wang@bernsteinsg.com Mark Li +852 2123 2645 mark.li@bernsteinsg.com David Dai, CFA +852 2918 5704 david.dai@bernsteinsg.com Mark C. Newman +1 212 845 7822 mark.newman@bernsteinsg.com Arpad von Nemes +1 917 344 8461 arpad.von ...

Summary of Key Points from the Conference Call Industry Overview - The analysis focuses on the AI data center industry, specifically the economics of GB200/NVL72 AI data centers [2][24]. Core Insights and Arguments - **Cost Estimates**: A typical GB200/NVL72 rack costs approximately $3.4 million, with physical infrastructure costs around $2.5 million per rack, leading to total AI data center capital expenditure (capex) of $5.9 million per rack or $35 billion per gigawatt (GW) [3][25]. - **Comparison with Nvidia**: This estimate is significantly lower than Nvidia's projected $50-60 billion per GW, suggesting Nvidia may be anticipating future product cycles [3][28]. - **Cost Composition**: - GPUs account for 39% of total costs, while Nvidia's gross profit dollars represent 29% of total costs, indicating that Nvidia's gross profit dollars constitute about 30% of total AI data center spending [4][26]. - Networking expenses are around 13% of total spending, with switches being the largest component at approximately 3% [5][27]. - Storage costs are minimal, representing about 1.4% of total spending [29]. Additional Important Insights - **Foundry and Supplier Economics**: Foundries capture 2.5-3% of data center capex through GPUs, with additional contributions from CPUs and memory suppliers [6][30]. - **Mechanical and Electrical Equipment**: Major costs include diesel and gas generators (6%), uninterruptible power supplies (4%), and transformers (5%), with thermal management costs at around 4% [7][32]. - **Operational Costs**: The annual electricity cost to run a GW of data center capacity is approximately $1.3 billion, with personnel costs being relatively low [8][33]. - **Market Dynamics**: Companies that serve as bottlenecks in the supply chain are likely to capture a larger share of economic value as demand increases [35]. - **Future Trends**: The power content in data centers is expected to increase significantly, with projections for the Vera Rubin Ultra design indicating a potential increase to 7-8 times the current power content by 2027 [38]. Investment Implications - **Nvidia (NVDA)**: Rated Outperform with a price target of $225, highlighting the significant and early-stage datacenter opportunity [12]. - **Broadcom (AVGO)**: Also rated Outperform with a target of $400, expecting strong growth driven by software and cash deployment [12]. - **AMD (AMD)**: Market-Perform rating with a target of $200, noting high AI expectations and potential growth from a new deal with OpenAI [12]. - **Intel (INTC)**: Market-Perform rating with a target of $35, indicating ongoing challenges [12]. - **Qualcomm (QCOM)**: Outperform rating with a target of $185, despite headwinds from Apple [13]. This summary encapsulates the key points from the conference call, providing insights into the AI data center industry's economics, cost structures, and investment implications.