Vals analysis: open-weight models performing multistage tasks, like building a web app, can have an environmental impact 10K times greater than simple queries
Benchmarker Vals AI finds that multi-stage tasks can require models to use exponentially more energy or water than simple queries do.
Context & Ripple Effects
Vals AI’s finding extends a 2025 study of 14 open-source models that linked reasoning workloads to sharply higher energy use without better answer accuracy. It also adds an environmental constraint to the shift toward test-time compute as a way to improve model performance.
The relevant unit is no longer the individual prompt: a multistage application can impose a radically different resource burden from a simple query. That matters as data-center electricity demand has already been a focus of coverage of AI’s expanding power use.
First-order effects
- Vals AI gives developers deploying open-weight models a task-level environmental benchmark: treating a web-app-building workflow like a single query can understate its energy or water impact by up to 10,000 times.
- Application operators using multistage model workflows have a concrete reason to measure and optimize full execution paths, rather than relying on simple-query estimates.
Second-order effects
- Model hosts and AI application teams face pressure to route, cap, or redesign high-step workflows, because inference economics and environmental accounting vary with task structure rather than model choice alone.
- Benchmarks that report only answer quality or single-prompt efficiency become less informative for buyers comparing systems intended to execute extended tasks.
Third-order effects
- If task-level reporting becomes standard, AI deployment decisions will increasingly weigh capability against the marginal energy and water cost of orchestration, reasoning, and repeated tool use.
- The industry’s efficiency debate shifts from model-level claims toward system-level measurement, where an application’s workflow design determines a material share of its resource footprint.
The trend: AI inference is moving toward workload-level economics, in which multistep execution—not merely the selected model—sets both operating cost and environmental impact.