Foxconn pegs Nvidia Vera Rubin data center at $47B per gigawatt

Foxconn Pegs Nvidia Vera Rubin Datacenter at $47 Billion Per Gigawatt

Foxconn Chairman Young Liu disclosed that a one-gigawatt AI datacenter built around Nvidia’s upcoming Vera Rubin architecture would cost approximately $47 billion, with annual electricity bills of roughly $1.3 billion — figures that illustrate the staggering capital required to power the next wave of artificial intelligence infrastructure.x

The Cost Breakdown

Liu’s remarks, made on June 19, drew on estimates from a Bernstein Research report that priced a single Vera Rubin NVL72 rack at about $9.1 million — up from the $8 million benchmark associated with the current Blackwell generation. At a rack power draw of 220 kilowatts, a gigawatt-scale facility would house roughly 3,557 racks, bringing hardware-related costs alone to approximately $32 billion. Adding physical infrastructure — mechanical and electrical systems, land, and buildings — accounts for the remaining $15 billion.moomoo

Memory represents a key cost driver. High-Bandwidth Memory, CPU DRAM, and directly attached storage together account for roughly $3.2 million per rack, while GPUs contribute about $4 million and networking adds another $1.2 million. The total per-gigawatt figure marks a 17 percent increase from the roughly $40.5 billion estimated for a Blackwell-era datacenter of similar scale.investing

Performance Gains Offset Rising Costs

Despite the higher price tag, analysts argue the investment remains economically justified. Each Vera Rubin NVL72 rack is expected to deliver 2,520 FP8 petaflops, compared with 720 petaflops for Blackwell — a 3.5-fold improvement in compute performance that improves capital efficiency per unit of useful work.moomoo

Hardware depreciation, not electricity, increasingly dominates datacenter economics. At $0.15 per kilowatt-hour, the $1.3 billion annual power cost pales beside roughly $7.9 billion in yearly depreciation under a six-year hardware lifecycle.investing

Foxconn’s Expanding Role

Foxconn is deeply embedded in the Vera Rubin supply chain. The company recently announced that shipments of Vera Rubin servers remain on track for 2026 and unveiled a partnership with Bull to produce Vera Rubin NVL72 platforms for European AI infrastructure. Liu’s cost disclosures underscore how Foxconn, as Nvidia’s primary manufacturing partner for AI server hardware, is positioning itself at the center of what it calls the “AI factory” buildout — a capital cycle that hyperscalers including Microsoft, Meta, and Alphabet are racing to fund.digitimes