Bernstein estimates that building a 1GW AI data center requires between $34.6 billion and $39.5 billion in capital expenditure, depending on the accelerator architecture deployed. NVIDIA's Vera Rubin architecture commands the highest construction costs, while OpenAI's self-developed Jalapeno ASIC architecture is positioned at the lower end of the range. The firm revised its cost estimate for NVIDIA's Rubin NVL72 single rack down 17% from $9.1 million to $7.52 million, citing adjusted expectations for HBM pricing and NAND storage capacity. Bernstein identifies massive capital expenditure and resulting depreciation as the primary economic burden for AI data centers, rather than electricity costs.