The Data Centers That Train A.I. and Drain the Electrical Grid (www.newyorker.com)

Since the ChatGPT era, data centers have become the new factories of AI — sprawling, highly secured warehouses packed with Nvidia GPUs that consume enormous amounts of power, water and capital. The story traces CoreWeave’s rise from crypto-mining start‑up to a public company owning hundreds of thousands of GPUs and running “hero” training jobs for OpenAI, Meta and others. Modern training runs can tie up whole facilities for weeks, need tens of thousands of GPUs, perform on the order of 10^31 operations for the largest models, and rely on water‑cooled GB300-style racks where each rack can use more electricity than a hundred homes. For the AI/ML community this matters because model capability is now constrained by physical infrastructure and utility coordination as much as by algorithms. Training is dominated by massive matrix multiplications and produces a compact weights file that’s inexpensive to copy but extremely costly to produce; inference distributes those weights worldwide and multiplies ongoing energy demand (e.g., ~5,000 tokens ≈ three minutes of a microwave). The implications: rapid data‑center buildout, stress on local grids and cooling/water systems, concentration of compute and IP behind a few operators (and security risks), and a real risk of underbuilding capacity as demand for richer multimodal and real‑time AI grows.