Bernstein Breaks Down AI Infrastructure Costs: Up to $39.5 Billion Investment per Gigawatt

Bernstein research shows that building a 1GW data center using different AI accelerator architectures requires capital expenditure of approximately $34.6 billion to $39.5 billion. Among these, Nvidia's Vera Rubin architecture has the highest construction cost, while OpenAI's self-developed ASIC architecture Jalapeno is relatively lower. Bernstein significantly lowered its cost estimate for Nvidia's Rubin NVL72 single rack from the previous $9.1 million to $7.52 million, a reduction of about 17%, mainly reflecting adjustments to expectations for HBM prices and NAND storage capacity. The research report also points out that the main economic burden of AI data centers is not electricity costs, but the massive capital expenditure and the depreciation it generates.