Source: Bloomberg (paywall)
Vals found a critical efficiency cliff in generative AI: single queries consume minimal resources, but multi-stage reasoning tasks like building a web application compound inference costs exponentially, making them orders of magnitude more environmentally expensive than previously measured. This challenges the narrative that open-weight model adoption is greener than closed systems. Environmental footprint depends on task complexity and inference stages, not just model availability. Companies deploying these models for agentic or multi-step workflows face a concrete trade-off: architectural choices around task decomposition and inference depth matter more to environmental impact than switching to open-source alternatives.