Definition
The accumulation of unnecessary size, context, or compute cost in an AI model or the systems built around it — without a proportional gain in real-world usefulness.
Symptoms: rising inference cost, slower responses, inconsistent quality across sessions, growing operational overhead that outpaces the value it produces.
Field notes
- Global data center electricity use is projected to more than double this decade, growing around 15% a year, with AI accelerators driving most of that growth. — IEA, Energy and AI report
- Data centers accounted for roughly half of all new U.S. electricity demand last year — more growth than housing, industry, and transport combined. — IEA, via Fortune, Apr 2026
- Developer trust in AI-generated code accuracy fell from 40% to 29% in a single year, even as daily usage climbed to 84%. — Stack Overflow Developer Survey
- Unresolved technical debt introduced by AI coding tools grew from a few hundred issues in early 2025 to over 100,000 by February 2026 — and is still climbing. — "Debt Behind the AI Boom," arXiv
- AI now writes 42% of committed code at the average company, on track for 65% by 2027 — yet 96% of developers say they don't fully trust the output. — Sonar, State of Code Developer Survey 2026
Used in a sentence
Our inference bill tripled but the eval scores didn't move — classic model bloat.
That update wasn't a feature. It was model bloat with a changelog.
We need a model-bloat audit before the next training run.