How Data Centers Actually Work (www.wired.com)

WIRED’s Uncanny Valley podcast episode breaks down the AI data center boom—what these facilities do, why Big Tech is pouring hundreds of billions into them, and why critics worry the expansion may be unsustainable. Hosts and guest Molly Taft explain the end-to-end flow: a user query is tokenized, routed through authentication/moderation and load balancers, then runs on specialized hardware (mostly GPUs like NVIDIA H100s) where inference predicts next tokens and returns a response in seconds. Recent industry moves (big hyperscaler builds, OpenAI’s high-profile AMD chip deals) signal expectations of near‑limitless demand and rapid capacity scaling. The episode highlights the core technical and policy issues for the AI/ML community: data centers are extremely energy- and water-intensive (cooling, power, and network gear), their emissions depend heavily on grid mix and operational cycles, and capacity projections are massive—Meta’s cited Hyperion build is ~5 GW and some regions already face strain (Ireland’s centers use >20% of national electricity). Transparency problems—proprietary reporting, omitted supply‑chain emissions, and contested per‑query energy figures (e.g., Sam Altman’s 0.34 Wh claim)—make true efficiency and climate impact hard to assess. That mix of technical scaling, local infrastructure stress, and opaque metrics raises engineering, regulatory, and environmental challenges the AI community must confront.