# Vision-Language Models' Hidden Energy Cost: It's Not What They See, But What They Say > Source: > Published: 2026-07-13 10:26:15+00:00 # Vision-Language Models' Hidden Energy Cost: It's Not What They See, But What They Say A new study reveals that the true energy burden of Vision-Language Models lies not in visual processing but in the generation of output tokens. Controlling output length could save up to 97% of energy across various models. Vision-Language Models (VLMs) have become indispensable in the space of embodied AI, touted for their ability to process complex visuals and language in tandem. Yet, their energy consumption on edge devices remains a largely uncharted territory. This latest study sheds light on a key aspect: the very assumptions about where these models expend the most energy might be fundamentally flawed. ## Challenging Prevailing Assumptions Previous efficiency efforts predominantly zeroed in on reducing visual tokens, working on the premise that visual processing was the dominant energy consumer. However, this study conducted a systematic energy profiling across five VLMs, spanning three architecture families, four input resolutions, and two hardware platforms, the [NVIDIA](/glossary/nvidia) RTX 3070 and the Jetson Orin NX. The findings are rather striking. First, the study shows that average [inference](/glossary/inference) power is essentially a model-intrinsic constant. The [parameter](/glossary/parameter) count and hardware platform remain invariant to input resolution, image complexity, or prompt type, with less than a 5% variation across these conditions. What does this mean? The energy variation we observe isn't about how much power the models draw but how long they operate. ## Output Tokens: The Real Energy Culprit Crucially, the study unveils that each output token consumes 11 to 39 times more wall-clock time than each input token. This results from the asymmetry between [compute](/glossary/compute)-bound prefill and memory-bound decode processes. Thus, the length of output becomes the main driver of latency and energy usage, not the complexity or the number of visual tokens. Interestingly, even the number of objects in an image can lead to as much as a 4.1x increase in energy consumption, driven not by increased visual processing, but by longer output lengths. This underlines a fundamental flaw in strategies focusing solely on pruning visual tokens, which at most save 10% of total energy for fixed-token models. ## The Larger Implications For models ranging from 1 billion to 8 billion parameters, controlling the output length can save up to 97% of total energy. The data shows that the energy dominance of decoding intensifies as the model scales up. This raises a pertinent question: Are we focusing too much on visual token pruning when real gains lie elsewhere? The takeaway is clear. The energy bottleneck in edge VLM inference isn't primarily about what the model processes visually. Instead, it's about what it articulates. Shouldn't energy efficiency efforts pivot towards optimizing output lengths rather than visual token reductions? For developers and researchers, this offers a new direction. By controlling output length, energy efficiency in VLMs can be vastly improved, paving the way for more sustainable AI applications on edge devices. Western coverage has largely overlooked this. But the [benchmark](/glossary/benchmark) results speak for themselves. Get AI news in your inbox Daily digest of what matters in AI.