Optimizing AI Inference: How to Cut Costs, Latency & Energy

Running production-level AI (especially LLMs and generative models) is facing a wall: cost, performance and hardware availability. Cloud bills are skyrocketing, latency is impacting user experience, and reliance on high-end GPUs is creating a bottleneck. On top of that, new regulations such as the EU AI Act are increasingly requiring transparency on energy consumption.

Green AI is an emerging field dedicated to making AI environmentally sustainable without compromising model performance. This discipline goes hand in hand with engineering efficiency: using fewer resources to achieve equivalent results. By optimizing models to reduce energy consumption, we address two critical goals at once: we significantly improve operational margins and scalability, and we minimize the carbon emissions.

In this session, we will demonstrate how to address these challenges using GAISSA, a UPC spin-off project that automates AI model optimization. We will show a Live Demo covering:

• A walk-through of how a standard model is compressed and adapted for efficiency
• A Return on Investment (ROI) comparison of a standard vs. an optimized model
• A visualization of cost reduction (€), latency (ms), and emissions (CO2)

Who Are We
We are the team behind GAISSA, researchers and engineers from Universitat Politècnica de Catalunya (UPC). We specialize in AI efficiency and model optimization, and our mission is to make advanced AI accessible, affordable, and compliant with future standards.
More on: https://gessi.upc.edu/en/research-areas/sustainability-of-artificial-intelligence-systems

Location and sponsor
This event is sponsered by basetis, a technology consultancy that stands out not just for its IT services, but for its radical approach to corporate structure. The event will take place in their office, which is located in La Pedrera (Casa Milà). The access to the building will be by Carrer Provença entrance.

Event schedule
Doors open from 18:30 to 19:00. Please join us after the presentation for pizza and drinks to network with other attendees.