18 Sep 2026
Your more-efficient AI probably didn't cut your carbon
In my last post I wrote about picking the right-sized model as the biggest lever on both cost and carbon. But I mentioned that efficiency gains don't reliably shrink your total footprint. Sometimes they grow it.
This is the Jevons paradox, and it has a long track record. LED bulbs are roughly 90% more efficient than the incandescents they replaced. Lighting's share of energy demand barely moved, because cheaper light meant more of it, in more places, for longer.
This year's AI efficiency headlines have the same shape: a custom inference chip (OpenAI's Jalapeño, Hot Chips 2026) claiming up to 1.9x the performance per watt of Nvidia's latest GPUs; sparse mixture-of-experts models activating a fraction of their parameters per token (GLM-5.3 uses about 40B of its 753B); Anthropic's cheapest cache reads now priced at 2.5% of base input cost; Google cutting agentic video-understanding tokens by up to 88% on a long-video benchmark.
Every one of these is a real, verifiable per-task gain. None of them tells you what happens to total emissions once that efficiency makes the workload cheaper to run more of (and cheaper, more efficient tokens get used more).
You can't see the rebound if the metric you're tracking is a rate. Software Carbon Intensity (SCI) measures emissions per unit of work; it can hold steady or improve while absolute emissions climb, because it was never built to catch volume growth.
The one lever that survives this trap is model right-sizing: routing the easy ~80% of requests to a small model and reserving the frontier model for the hard 20%. It works because it applies the saving to demand that already exists, not to a new use case invented to spend the savings on.
The PM move: treat efficiency gains as a budget decision, not a press release. Decide deliberately whether you're banking the saving or reinvesting it into more capability at flat cost, and don't announce per-query efficiency as a climate win while absolute emissions are still climbing.
Sources: OpenAI/Broadcom Jalapeño at Hot Chips 2026 (via Tom's Hardware); GLM-5.3 model card (Hugging Face); Anthropic Claude Fable 5.1 pricing, platform.claude.com (cache reads); Google blog, agentic video understanding, Sept 2026.
Originally posted on LinkedIn ↗