Somewhere in the high desert of Texas, a parcel of land the size of a small town is being transformed into something that would have seemed like science fiction a decade ago: a single AI datacenter complex drawing more electricity than the entire city of Seattle. This isn’t a distant proposal. Permitted, funded, and actively under construction, gigawatt-class AI campuses have become the defining infrastructure project of the era.
To understand what a gigawatt actually means in this context, consider the trajectory. The hyperscale datacenters that powered the cloud era of the 2010s topped out around 50 to 100 megawatts. Then frontier AI training arrived, and the numbers started climbing. GPT-4’s training cluster operated at roughly 25,000 A100s. Today’s training runs routinely involve hundreds of thousands of H100s and B200s, and the clusters being planned and built now — Microsoft’s Wisconsin campus, xAI’s Memphis Colossus expansion, the CoreWeave buildouts across the Midwest — are targeting power draws measured not in megawatts but in fractions of a gigawatt, with some roadmaps aiming past that threshold entirely.
The engineering challenge this creates is genuinely staggering. At gigawatt scale, power delivery itself becomes a systems problem as hard as anything in the compute stack. You can’t simply run a bigger wire. Substations have to be purpose-built. Transmission infrastructure must be upgraded or constructed from scratch. Some developers are now negotiating directly with grid operators to essentially co-locate generation — gas peakers, dedicated solar farms, in a few ambitious cases small modular nuclear reactors — so that the facility is as close to energy-independent as possible. The latency of electrons across a transmission grid becomes a real concern when you’re moving a gigawatt around.
Cooling is its own frontier. Liquid cooling, once exotic, is now table stakes for high-density GPU racks. The most advanced deployments are moving toward direct-to-chip liquid loops, where coolant flows within millimeters of the die itself. Nvidia’s GB200 NVL72 rack, which packs 72 Blackwell GPUs into a single unit, essentially requires liquid cooling by design — air simply can’t remove heat fast enough. Some facilities are experimenting with immersion cooling at scale, submerging entire servers in dielectric fluid. The thermal engineering happening inside these buildings is pushing the boundaries of industrial heat transfer in ways that will probably have applications far outside AI.
What makes the current moment particularly remarkable is the speed. Traditional datacenter development runs on a five-to-seven-year timeline: permits, environmental review, grid interconnection queues, construction. The AI buildout is attempting to compress that to eighteen months or less, which requires a level of coordination between construction firms, utilities, chip suppliers, and land developers that has no real precedent. xAI reportedly brought its first Memphis phase online in under a year. That pace is almost unreasonable. It’s happening anyway.
The economic logic driving all of this is simple: the models that will be trained in these facilities don’t exist yet, and the companies building them are betting that the capabilities unlocked by another order of magnitude of compute will be worth the capital. Colossal infrastructure has always preceded transformative technology. The interstate highway system preceded the suburban economy. The fiber buildout of the late 1990s, overbuilt and then forgotten by the companies that built it, quietly became the backbone of the modern internet.
The gigawatt datacenter is the infrastructure layer for whatever comes next. The models that will run inside these buildings, trained on compute clusters we’re only now learning to build, are waiting to exist. The concrete is being poured right now.