For years, the AI industry assumed that the largest energy cost in vision-language models (VLMs) came from visual processing. This belief drove strategies like visual token pruning, hoping to drastically reduce consumption on edge devices. However, a systematic energy profiling of on-device VLM inference —covering models from 1 billion to 8 billion parameters, multiple resolutions, and platforms like NVIDIA RTX 3070 and Jetson Orin NX— has revealed an uncomfortable truth: seeing is nearly free, but speaking costs a fortune.
The study shows that average inference power is a model-intrinsic constant, invariant to input resolution, image complexity, or prompt type. This power varies less than 5% across all conditions, meaning that all energy variation between different inferences stems solely from differences in execution time, not instantaneous power draw. And what drives that time? Fundamentally, the number of tokens the model generates as output.
The asymmetry between the prefill and decode phases is the culprit. Each output token costs 11 to 39 times more wall-clock time than each input token. Decoding is compute-bound and memory-bound, while prefill —which processes the image and prompt— is relatively fast. This imbalance makes the number of generated tokens the true bottleneck for both latency and energy.
Even image complexity, which one might think increases visual cost, affects energy only indirectly: an image with more objects induces the model to produce longer answers, and that longer output length is what drives consumption. Energy differences of up to 4.1x have been measured between images of the same resolution but with different object counts, entirely attributable to the model's verbosity.
What does this mean for companies deploying VLMs in edge environments? That focusing on visual token pruning is a marginal effort: even removing all visual tokens saves at most 10% of total energy. In contrast, controlling output length can save up to 97% of energy. For larger models, the energy dominance of decoding only grows: the bigger the model, the more expensive it is to speak.
The lesson is clear: in edge VLM inference, the real bottleneck is not what the model sees, but how much it says. This forces a rethink of optimization strategies. Instead of costly investments in vision pruning or quantizing convolutional layers, engineering teams should focus on techniques that limit model verbosity: prompt engineering for concise answers, hard caps on output tokens, using smaller models for simple tasks, and architectures that accelerate decoding.
At Q2BSTUDIO, as a software development and technology company, we tackle these challenges with a comprehensive perspective. Our team combines expertise in AI, cybersecurity, and cloud computing to deliver solutions that maximize energy efficiency without sacrificing functionality. For example, we develop custom software applications that integrate optimized VLMs for edge, employing output control strategies and hybrid local-cloud processing.
The cloud plays a complementary role. Using services like AWS or Azure, we can offload complex reasoning tasks —which require long responses— and leave only quick inferences with short answers on the edge device. This architecture drastically reduces local energy consumption and extends battery life. At Q2BSTUDIO we offer cloud services on AWS and Azure designed to integrate seamlessly with embedded AI systems.
However, energy efficiency cannot come at the cost of security. In edge environments where data can be sensitive, we implement advanced cybersecurity measures: end-to-end encryption, model authentication, and protection against adversarial attacks. Our cybersecurity and pentesting service ensures AI solutions are robust against threats.
We also integrate business intelligence with Power BI to monitor real-time energy consumption and model efficiency. This information allows dynamic adjustment of inference parameters and anomaly detection. At Q2BSTUDIO we develop BI and Power BI solutions that provide clear dashboards for decision-making.
AI agents are another key piece. We design agents that, upon receiving a request, automatically decide whether to process it locally with a short answer or delegate it to a larger model in the cloud. This intelligent orchestration, based on energy and latency thresholds, is possible thanks to our experience in process automation and AI agent development. For more information, see our process automation page.
In summary, the finding that the main energy cost in edge VLMs lies in text generation opens new optimization avenues. The industry must abandon the myth of the visual bottleneck and adopt an output-centric approach. For companies, this represents an opportunity to drastically reduce energy consumption and operational costs, especially in massive deployments of connected devices. At Q2BSTUDIO we are ready to guide that transition, combining cutting-edge artificial intelligence with deep software engineering expertise. Because sometimes, the most efficient thing is to say less.
