The Gigawatt Scale Era: Why 100MW is the new 10MW
As LLM training clusters grow exponentially, the definition of hyperscale is shifting. We analyze the infrastructure requirements for 1GW campuses and what this means for grid stability.
Global Scale Research
The nomenclature of the data center industry is struggling to keep pace with reality. Five years ago, a 10MW lease was a headline-making transaction, sufficient to support a major cloud region or a Fortune 500 enterprise. Today, for the hyperscalers and frontier labs training Foundation Models, 10MW is a rounding error. We have entered the Gigawatt Scale Era.
The Physics of LLM Training
To understand this shift, one must look at the hardware physics. An NVIDIA H100 SXM5 server rack draws approximately 40kW to 50kW depending on configuration. A "SuperPOD" architecture connects these racks in tight, low-latency groups.
A cluster of 25,000 GPUs — now considered a baseline for serious training runs of models like Llama 3 or GPT-4 class — requires roughly 30-40MW of power just for compute. When you account for cooling (PUE 1.2-1.3) and ancillary networking gear (Infiniband switches, storage), the site requirement pushes past 50MW.
However, the roadmap for 2026/2027 projects cluster sizes exceeding 100,000 GPUs. This pushes the power envelope of a single training run into the 150MW to 200MW range. To support multiple such clusters and ensure redundancy, campuses are no longer being master-planned for 100MW; they are being designed for 1GW+ total site capacity.
The Grid Crisis
Building the shell is "easy" (relatively speaking). Finding the electrons is the crisis. The US power grid is congested. In major markets like PJM (Virginia/Ohio) and ERCOT (Texas), interconnection queues for new large loads stretch 4 to 7 years.
Utilities are pushing back. They are demanding upfront capital for substation upgrades and signing "take-or-pay" contracts years in advance. This has bifurcated the market:
- The Haves: Hyperscalers who secured land and power agreements 3 years ago.
- The Have-Nots: Enterprises trying to enter the market today using traditional colocation strategies.
The Rise of the "Mega-Campus"
We are seeing a shift in site selection strategy. Proximity to internet exchanges (IXs) is less relevant for training clusters where the data goes in once, and the model comes out months later. The priority is raw, cheap, abundant power.
This is driving development to rural areas with access to high-voltage transmission lines, often bypassing local distribution entirely. We are qualifying sites in Wyoming, North Dakota, and rural Texas where 500+ acres can be secured adjacent to 345kV lines.
Strategic Advice: If you are planning for AI infrastructure in 2027, you cannot rely on looking for "available inventory" in 2026. It won't exist. You must be securing land and power agreements today, essentially acting as your own utility developer.