Foxconn's Liu Young-way: 1GW Vera Rubin Data Center Capex Reaches $47 Billion
0xBroomberg
Foxconn chairman Liu Young-way disclosed that building a 1 GW AI data center on Nvidia's Vera Rubin architecture costs $47 billion — a figure that locks the supply side of compute into the hands of a very small number of players.
Where does the $47 billion go?
A 1 GW facility requires roughly 3,557 Vera Rubin racks at $9.1 million each, totaling $47 billion in capital expenditure.
Operating costs add up fast: annual power runs about $1.3 billion, while hardware depreciation is six times that — roughly $7.8 billion a year.
This means → building the center is just the starting line. Ongoing operating costs dwarf the one-time construction spend — this is a business you must be able to feed, not just build.
How large is the global compute gap by 2030?
Liu cited data projecting $1.6 trillion in total global data-center investment through 2030.
Compute load will climb from roughly 68 GW in 2024 to 174 GW, adding 106 GW in six years — nearly 18 GW of new power capacity per year.
In plain terms = that is roughly 18 new mega-scale 1 GW data centers every year, a pace far beyond current global construction rates.
Who is buying the compute?
Liu grouped today's largest compute buyers into four categories: model developers, cloud service providers (CSPs), governments, and enterprises.
Model developers and CSPs have clear business models and represent the strongest current demand; governments are exploring but hold enormous potential; enterprises are seen as the next blue ocean.
This reflects a broadening of compute demand from a handful of tech giants toward society at large — yet actual orders remain concentrated in the first two groups.
What is the difference between "AI-enabled" and "AI-native"?
Liu drew a line between two stages of enterprise AI adoption: most companies today are "AI-enabled" — embedding AI tools into existing workflows to boost efficiency.
The end goal is to become "AI-native": every process runs on AI at its core, with humans setting objectives and providing governance oversight only.
This means → the leap from "enabled" to "native" demands an order-of-magnitude jump in compute needs, and the $47 billion per-GW threshold ensures that the vast majority of enterprises can only be buyers, never builders — the supply side of compute will be highly concentrated.
Content is for reference only, not financial advice.