Key insights
- A Bernstein report estimates deploying 1GW of Nvidia's Vera Rubin AI infrastructure could cost $47 billion, driven by higher High-Bandwidth Memory (HBM) costs. Despite increased capital intensity, the new architecture promises significant performance gains. The report highlights hardware depreciation as a key driver of data center economics, with annual depreciation costs potentially exceeding power expenses. This indicates a continued, albeit expensive, expansion in AI infrastructure, potentially benefiting semiconductor and hardware providers.
Investing.com -- A Bernstein report estimates that building 1 gigawatt of AI data center capacity using Nvidia’s upcoming Vera Rubin architecture could cost roughly $47 billion, highlighting the growing capital intensity of next-generation AI infrastructure.
The report estimates a typical Vera Rubin NVL72 rack will cost about $9.1 million, above the widely cited $8 million figure.
Analysts attributed the difference primarily to higher expected memory costs, particularly High-Bandwidth Memory (HBM), which they expect to become significantly more expensive by the time Rubin systems are deployed at scale in 2027.
Memory and storage are projected to account for roughly $3.2 million per rack, while GPUs remain the largest cost component at around $4 million. Networking infrastructure is expected to contribute another $1.2 million, with cooling and power delivery costs estimated at roughly $150,000 each.
Based on a rack power rating of 220 kilowatts and an estimated 3,557 racks per gigawatt, the report calculates rack-related costs of about $32 billion per gigawatt.
Adding approximately $15 billion in physical infrastructure expenses brings the total AI data center capital expenditure to around $47 billion per gigawatt.
Despite rising costs, analysts see substantial gains in computing performance.
The Vera Rubin NVL72 architecture is expected to deliver 2,520 FP8 petaflops per rack, compared with 720 petaflops for Nvidia’s Blackwell generation, implying significant improvements in compute capacity per dollar invested.
The report also argues that hardware depreciation, rather than operating expenses, increasingly drives data center economics.
At an electricity cost of $0.15 per kilowatt-hour, running 1 gigawatt of AI data center capacity would cost roughly $1.3 billion annually in power, compared with about $7.9 billion in annual depreciation expenses under a six-year hardware life cycle.
Looking ahead, analysts expect cost per gigawatt to continue rising as AI systems require more memory, power infrastructure, and advanced components. They also see growing opportunities across the AI supply chain, particularly in power systems, memory, networking, and substrate technologies.
The report maintained Outperform ratings on Nvidia, Digital Realty Trust Inc (NYSE:DLR), Equinix Inc (NASDAQ:EQIX), Delta Electronics Inc (TW:2308), Unimicron Technology Corp (TW:3037), and Chroma ATE Inc (TW:2360), while keeping Underperform ratings on Quanta Computer Inc (TW:2382) and CoreWeave Inc (NASDAQ:CRWV).
ProPicks AI evaluates NVDA alongside thousands of other companies every month using 100+ financial metrics. Using powerful AI to generate exciting stock ideas, it looks beyond popularity to assess fundamentals, momentum, and valuation. The AI has no bias—it simply identifies which stocks offer the best risk-reward based on current data with notable past winners that include Super Micro Computer (+185%) and AppLovin (+157%). Want to know if NVDA is currently featured in any ProPicks AI strategies, or if there are better opportunities in the same space?