Foxconn has put a brutal price tag on the next wave of AI infrastructure: a 1 GW data center packed with Nvidia Vera Rubin systems could cost about $47 billion to build, plus another $1.3 billion a year for electricity. That is before depreciation, which Foxconn says would run six times higher than the power bill. In other words, the chips are only the beginning of the invoice.
The company’s estimate helps explain why AI data centers are turning into industrial megaprojects rather than ordinary server farms. A 1 GW site is not fantasy territory either – some companies are already building facilities at that scale, albeit with older-generation hardware. The real leap here is that Nvidia’s Vera Rubin systems are being positioned as a step beyond Blackwell for agentic AI workloads, which means the race is not just for more racks, but for enough capital and electricity to keep them fed.
What Foxconn’s 1 GW calculation includes
Foxconn says such a facility could hold up to 3,557 Vera Rubin racks. That makes the numbers feel less abstract: this is not a single expensive box, but an entire industrial ecosystem built around power delivery, cooling, networking, and the constant churn of hardware replacement.
- Estimated build cost: $47 billion
- Annual electricity cost: $1.3 billion
- Depreciation: six times the electricity bill
- Capacity: 1 GW
- Rack count: up to 3,557 Vera Rubin racks
Why Vera Rubin changes the spending math
Foxconn’s pitch is that Vera Rubin will unlock a level of agentic AI performance that Blackwell could not reach. That is the kind of claim that nudges buyers toward even larger deployments, because once the industry decides the next model tier needs more silicon, the bill climbs in a straight line and then keeps going.
The bigger story is that scale itself has become a moat. Companies with access to cheap power, land, and financing can still play; everyone else gets to admire the benchmark charts from a safe distance.
The new bottleneck is not just GPUs
The Foxconn estimate underlines a familiar pattern in AI: the hardware gets the headlines, but the long-term winner is usually whoever can stomach the infrastructure bill. If Vera Rubin systems deliver what Nvidia is promising, the next question is not whether customers want them. It is which customers can afford to keep them running for more than a quarter.