A one-gigawatt AI datacenter built around NVIDIA's Vera Rubin architecture will require capital expenditure of up to $47 billion, according to Foxconn Chairman Young Liu. The figure, reported by Commercial Times Taiwan, sets the price of entry for the next generation of agentic AI infrastructure.

The cost breakdown is specific. A 1GW installation needs roughly 3,557 server racks, with each Vera Rubin rack priced at $9.1 million. Annual electricity costs reach $1.3 billion. Hardware depreciation, Liu stated, runs six times the power bill — implying depreciation costs on the order of $7.8 billion per year for a single gigawatt-class site.

What does the hardware cost?

Liu's rack pricing aligns closely with independent estimates. A recent Morgan Stanley Research bill-of-materials breakdown put the cost of VR200 NVL72 servers at approximately $8 million per rack, with memory alone — HBM4 and LPDDR5X — accounting for around $2 million of that total.

Quoted in Commercial Times Taiwan, Liu said: "The construction cost is also very high. Building a 1GW AIDC with Vera Rubin as its core would require a capital expenditure of up to US$47 billion and about 3,557 racks, while a single Vera Rubin rack costs US$9.1 million; the annual electricity cost of a 1GW AIDC is US$1.3 billion, and the hardware depreciation cost is six times the electricity cost."

The pricing lands as Vera Rubin enters full volume production. NVIDIA is already shipping first systems to major cloud providers, which are validating and testing the hardware before scaling deployments. The company expects the platform to exceed the commercial success of Blackwell.

How large is the power problem?

The market context compounds the per-site costs. By 2030, the global datacenter market is projected to reach $1.6 trillion, with global compute consuming 174GW — more than double the 68GW required in 2024. Meeting that trajectory demands roughly 18GW of new electricity capacity per year through 2025–2030.

Demand comes from four primary customer categories:

• AI model developers
• Cloud service providers
• Governments
• Enterprises

Most of these customers remain in early stages of AI adoption. Their stated goal, however, is the AI-native organization: processes running with AI at the core, with humans setting objectives, managing goals, and supervising workflows and results.

Where will these facilities be built?

Liu has proposed establishing "Taiwan-style" science and technology parks in the United States, primarily in Arizona and Texas. Efforts to make these parks operational are already underway, with a target of taking shape by the end of this year.

Multi-gigawatt AI datacenters are no longer theoretical; several are already in planning or construction. At $47 billion per gigawatt, each additional gigawatt multiplies both the capital commitment and the operational burden — from annual power bills to depreciation schedules that Foxconn now quantifies at six times the electricity cost.

The Vera Rubin generation delivers compute density that Blackwell-era systems do not approach. But the numbers Liu presented frame the industry's near-term test: whether power capacity, capital, and construction can scale at the pace the 2030 projections demand. The economics, not the silicon, may determine who builds first.