This article breaks down where that money goes, how the cost compares to the current Blackwell generation, and why companies like Foxconn are more critical than ever to the AI supply chain.

The $47 Billion Breakdown: What a 1-GW Vera Rubin Data Center Costs

The headline figure from Bernstein is broken down into three major components .

• ~3,557 Vera Rubin NVL72 Racks: Each rack is estimated to cost roughly $9.1 million . This figure is notably higher than the earlier market whisper of ~$8 million. The primary reason is a projected increase in memory costs, particularly High-Bandwidth Memory (HBM) .
• Annual Electricity Bill: Running a facility of this scale requires immense power, with an estimated annual electricity cost of about $1.3 billion .
• Other Estimates: For context, a broader analysis by Epoch AI for a "typical" (non-Vera Rubin specific) 1-GW AI data center puts CapEx at $38 billion and annual OpEx at $0.9 billion, with servers accounting for roughly $5 billion per year in annualized costs .

Vera Rubin vs. Blackwell: A Generational Cost Comparison

While a direct per-rack price comparison is difficult to find in published sources, the generational differences are stark.

• Per-Rack Cost: The Vera Rubin NVL72 rack at ~$9.1M is estimated to be 25-50% higher than comparable Blackwell GB200 NVL72 racks, which were estimated in the $6-7 million range .
• Power Consumption: This cost increase comes with a massive leap in performance but also power draw. A Vera Rubin NVL72 rack draws 190–230 kW, a significant jump from the 120–130 kW of the Blackwell GB200 . For comparison, Hopper-era racks drew roughly 40 kW.
• Performance per Dollar: Despite higher absolute costs, Nvidia claims that Vera Rubin delivers 4x better training performance and up to 10x better inference performance per watt . The total cost of ownership per unit of useful compute is therefore lower, with Nvidia citing a one-tenth the token cost of Blackwell .

Key Hardware Cost Drivers: Where the $9.1 Million Per Rack Goes

The increase in per-rack costs is not just about a more expensive GPU. A report from Morgan Stanley provides a detailed look at the changing bill of materials .

• High-Bandwidth Memory (HBM): This is the single largest cost escalator. Morgan Stanley estimates memory cost per rack at roughly $2 million, commanding 25-30% of the total rack BOM . HBM4 supply tightness and higher-density stacks are pushing prices up significantly compared to Blackwell .
• GPU Silicon: The Rubin CPX GPU remains the largest single line item by dollar value, but its share of the total BOM is shrinking from ~65% to ~51% as memory and networking costs grow .
• Cooling: Vera Rubin requires 100% liquid cooling with no air-cooled option . The cooling hardware alone for the larger NVL144 configuration is projected to exceed $55,000 per rack .
• Networking, CPU & Power: Other critical cost drivers include the NVLink 6 fabric, the new 88-core Vera CPU, and the adoption of a new 800V DC power architecture that adds significant facility-level cost .

Foxconn's Pivotal Role in the Vera Rubin Supply Chain

Foxconn is more than just a supplier for Vera Rubin; it is a central, strategic partner in its manufacturing and assembly.

• Primary System Builder: Nvidia lists Foxconn as one of the top system builders in full-scale Vera Rubin production, alongside Dell, HPE, Lenovo, and Supermicro .
• Global Manufacturing Hub: Over 1 million MGX rack components for Vera Rubin are assembled across 25 factory sites in Taiwan, with Foxconn acting as a lead integrator .
• European Expansion: In June 2026, Nvidia initiated European production of the Vera Rubin NVL72 in partnership with Foxconn and Bull (Atos Group) . This covers the design, construction, and validation of systems for the European market.
• Validation of Cost Estimates: Foxconn Chairman Young Liu publicly validated Bernstein's analysis, citing the $47 billion/1 GW cost estimate and $1.3 billion annual power bill .
• Deepening Integration: Nvidia is embedding its proprietary AI software stack into Foxconn's server assembly lines, creating a tighter integration between chip design and manufacturing that increases switching costs .

This article breaks down where that money goes, how the cost compares to the current Blackwell generation, and why companies like Foxconn are more critical than ever to the AI supply chain.

The $47 Billion Breakdown: What a 1-GW Vera Rubin Data Center Costs

The headline figure from Bernstein is broken down into three major components .

• ~3,557 Vera Rubin NVL72 Racks: Each rack is estimated to cost roughly $9.1 million . This figure is notably higher than the earlier market whisper of ~$8 million. The primary reason is a projected increase in memory costs, particularly High-Bandwidth Memory (HBM) .
• Annual Electricity Bill: Running a facility of this scale requires immense power, with an estimated annual electricity cost of about $1.3 billion .
• Other Estimates: For context, a broader analysis by Epoch AI for a "typical" (non-Vera Rubin specific) 1-GW AI data center puts CapEx at $38 billion and annual OpEx at $0.9 billion, with servers accounting for roughly $5 billion per year in annualized costs .

Vera Rubin vs. Blackwell: A Generational Cost Comparison

While a direct per-rack price comparison is difficult to find in published sources, the generational differences are stark.

• Per-Rack Cost: The Vera Rubin NVL72 rack at ~$9.1M is estimated to be 25-50% higher than comparable Blackwell GB200 NVL72 racks, which were estimated in the $6-7 million range .
• Power Consumption: This cost increase comes with a massive leap in performance but also power draw. A Vera Rubin NVL72 rack draws 190–230 kW, a significant jump from the 120–130 kW of the Blackwell GB200 . For comparison, Hopper-era racks drew roughly 40 kW.
• Performance per Dollar: Despite higher absolute costs, Nvidia claims that Vera Rubin delivers 4x better training performance and up to 10x better inference performance per watt . The total cost of ownership per unit of useful compute is therefore lower, with Nvidia citing a one-tenth the token cost of Blackwell .

Key Hardware Cost Drivers: Where the $9.1 Million Per Rack Goes

The increase in per-rack costs is not just about a more expensive GPU. A report from Morgan Stanley provides a detailed look at the changing bill of materials .

• High-Bandwidth Memory (HBM): This is the single largest cost escalator. Morgan Stanley estimates memory cost per rack at roughly $2 million, commanding 25-30% of the total rack BOM . HBM4 supply tightness and higher-density stacks are pushing prices up significantly compared to Blackwell .
• GPU Silicon: The Rubin CPX GPU remains the largest single line item by dollar value, but its share of the total BOM is shrinking from ~65% to ~51% as memory and networking costs grow .
• Cooling: Vera Rubin requires 100% liquid cooling with no air-cooled option . The cooling hardware alone for the larger NVL144 configuration is projected to exceed $55,000 per rack .
• Networking, CPU & Power: Other critical cost drivers include the NVLink 6 fabric, the new 88-core Vera CPU, and the adoption of a new 800V DC power architecture that adds significant facility-level cost .

Foxconn's Pivotal Role in the Vera Rubin Supply Chain

Foxconn is more than just a supplier for Vera Rubin; it is a central, strategic partner in its manufacturing and assembly.

• Primary System Builder: Nvidia lists Foxconn as one of the top system builders in full-scale Vera Rubin production, alongside Dell, HPE, Lenovo, and Supermicro .
• Global Manufacturing Hub: Over 1 million MGX rack components for Vera Rubin are assembled across 25 factory sites in Taiwan, with Foxconn acting as a lead integrator .
• European Expansion: In June 2026, Nvidia initiated European production of the Vera Rubin NVL72 in partnership with Foxconn and Bull (Atos Group) . This covers the design, construction, and validation of systems for the European market.
• Validation of Cost Estimates: Foxconn Chairman Young Liu publicly validated Bernstein's analysis, citing the $47 billion/1 GW cost estimate and $1.3 billion annual power bill .
• Deepening Integration: Nvidia is embedding its proprietary AI software stack into Foxconn's server assembly lines, creating a tighter integration between chip design and manufacturing that increases switching costs .