Overview
Bernstein’s research report projects that deploying one gigawatt (GW) of AI data‑center capacity using Nvidia’s forthcoming Vera Rubin architecture will require approximately $47 billion in capital expenditure. The estimate reflects a detailed cost breakdown per rack and the aggregate infrastructure needed for a full gigawatt.
Rack‑Level Cost Structure
A standard Vera Rubin NVL72 rack is priced at $9.1 million. The cost composition is driven primarily by three components: GPUs account for about $4 million per rack, memory and storage—dominated by high‑bandwidth memory (HBM) expectations—contribute roughly $3.2 million, and networking infrastructure adds $1.2 million. Cooling and power delivery are each estimated at $150,000. Analysts note that the $9.1 million figure exceeds the previously cited $8 million estimate mainly because of anticipated higher HBM prices when Rubin systems scale in 2027.
Gigawatt‑Scale Aggregation
Each NVL72 rack is rated at 220 kilowatts (kW). To achieve a full gigawatt, the design calls for 3,557 racks, resulting in rack‑related capital costs of about $32 billion. Adding an estimated $15 billion for broader physical‑infrastructure elements—such as site preparation, power distribution, and ancillary systems—brings the total AI data‑center capex to roughly $47 billion per GW.
Performance and Operating Economics
The Vera Rubin NVL72 architecture is expected to deliver 2,520 FP8 petaflops per rack, a four‑fold increase over the 720 petaflops per rack offered by Nvidia’s earlier Blackwell generation, indicating a substantial boost in compute capacity per dollar invested. Operating costs are projected at $0.15 per kilowatt‑hour; at full gigawatt utilization this translates to an annual electricity expense of about $1.3 billion. In contrast, hardware depreciation—based on a six‑year useful life—accounts for roughly $7.9 billion of annualized cost, suggesting that depreciation, rather than ongoing power consumption, dominates the economics of such facilities.
Outlook and Supply‑Chain Implications
Analysts anticipate that the cost per gigawatt will continue to rise as future AI systems demand more memory, higher‑capacity power infrastructure, and advanced components. They identify growing investment opportunities across the AI supply chain, particularly in power‑system technologies, memory modules, networking equipment, and substrate manufacturing.
Analyst Ratings
Bernstein maintains an Outperform rating 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). Conversely, it retains an Underperform rating on Quanta Computer Inc (TW:2382) and CoreWeave Inc (NASDAQ:CRWV).