We toured CoreSite’s NY3 facility in Secaucus, New Jersey, to understand how the infrastructure behind AI workloads actually works. CEO Juan Font compares his data centers to shopping malls: multiple tenants share power, cooling, and space rather than building their own. That model is scaling fast as tech giants—Amazon, Alphabet, Meta, Microsoft—collectively spend over $600 billion this year on capacity.

Power is the gating factor. Three years ago, clients asked for 10-30 megawatts. Now they ask for hundreds, sometimes gigawatts. The grid can’t keep up, so operators turn to behind-the-meter generation: natural gas, fuel cells, solar. CoreSite uses fuel cell tech from GE Vernova, whose equipment orders surged 16% YoY to $9.3B last quarter. Eaton supplies power distribution units and remote power panels, with its Electrical Americas backlog hitting record highs.

Inside the server rooms, Nvidia and Broadcom dominate. Nvidia’s networking business alone hit $31B last year, driven by NVLink and InfiniBand. Broadcom’s custom chip deals with Google and Meta are testing Nvidia’s lead. Fiber from Corning carries data between racks; its optical communications segment grew 36% YoY, with Meta committing up to $6B for fiber through 2030.

Cooling is the hidden bottleneck. Traditional air conditioning can’t handle racks producing heat equivalent to hundreds of hair dryers. Liquid cooling is now standard, with Eaton’s $9.5B acquisition of Boyd Thermal and Dover’s thermal connectors expected to generate over $1B in AI-related revenue this year.

From power to chips to fiber to cooling, every layer of the stack is under pressure. The engineering trade-offs are clear: grid constraints force alternative energy sourcing; thermal limits drive liquid cooling adoption; and fiber replaces copper for speed and efficiency. This isn’t just a boom—it’s a full-stack infrastructure rebuild.

Source: CNBC