AI boom’s 2025 emissions rival New York City’s CO2 output, study finds

The study highlighted here won’t change model scaling plans overnight, but it crystallises a narrative that will increasingly shadow the race to AGI: frontier AI is not just a math problem, it’s a resource problem. If emissions and water use are already comparable to a mega‑city at today’s deployment levels, the implied trajectory under the kinds of $100 billion+ compute buildouts now being mooted for OpenAI and others is politically hard to ignore. ([theguardian.com](https://www.theguardian.com/technology/2025/dec/18/2025-ai-boom-huge-co2-emissions-use-water-research-finds))

This matters strategically because it reframes AI build‑out as infrastructure policy, not just innovation policy. Power grids, water rights, and local environmental regulation become de‑facto gates on scaling. Regions that can offer cheap, low‑carbon energy and water‑efficient cooling will gain leverage over where the next wave of GPU clusters land. For labs, it adds another axis of competition: not just model quality and price, but embodied emissions per useful unit of reasoning. In the medium term, pressure from regulators, activists and even corporate sustainability teams will nudge the frontier toward more efficient architectures, aggressive reuse and better scheduling—adding engineering overhead, but potentially extending Moore‑like scaling in the face of physical constraints.