Explore the development journey of Moshi, the first speech-to-speech model, from its inception to real-time inference. This session will dive into key decisions, such as large-scale training on vast datasets and the shift from traditional frameworks like PyTorch to Rust/Candle. Learn how these choices impacted performance, and discover how 500,000 Moshi sessions were served using optimized L4 GPU clusters, minimizing computational demand while maintaining real-time accuracy.

Master Stage

Moshi’s behind the scenes: From conception to inference

NVIDIA solves the world’s most important computational problems through accelerated computing: a complete rethinking and optimization of the entire problem, from chips to racks to data centers, compilers, libraries, frameworks, algorithms, and applications. In a world where traditional methods of improving computational speed have run out of steam, this full-stack optimization is the way forward to improve computational capability and enable the applications of the future. Bryan will be covering how NVIDIA approaches the accelerated computing problem, with examples from DLSS and Megatron development.

Chips will never be enough

The future of AI relies on more than just powerful algorithms—it’s all about the quality of the data. At Kili Technology, we help companies unlock the full potential of their data to create smarter AI systems. This presentation explains how better, more specialized data can push AI forward, making it more effective in industries like finance, law, and more. We’ll also explore how synthetic data and expert knowledge are shaping the next big breakthroughs in AI.

Creativity Room

Frontier Data: The Key to AI’s Next Leap Forward

This 30-minute workshop, led by a former NLP/AI expert now specializing in quantum computing, introduces AI enthusiasts and promoters to the emerging intersection of quantum technology and artificial intelligence. The session is divided into two parts: a 15-minute presentation followed by Q&A, and a 15-minute interactive demonstration.

In the first half, we'll explore quantum computing principles through an AI-centric lens, emphasizing its developmental stage and potential future impact on AI. We'll briefly compare classical AI paradigms with emerging quantum capabilities, highlighting quantum computing's unique features and current limitations. The presentation will also touch on early-stage hybrid architectures that aim to leverage strengths from both quantum and classical systems.

The second half features a live, on-screen demonstration of a simple hybrid quantum-AI model. While the demonstration will be shared on screen for all to follow, developers in the audience will have the option to run the shared notebooks on their own computers, allowing for a more hands-on experience. This interactive component offers a practical glimpse into how quantum computing might enhance future AI applications.

Join us for this concise yet comprehensive introduction to the quantum-AI frontier. Whether you're an AI promoter looking to understand the next big trend, or an enthusiast eager to explore cutting-edge developments, this workshop will equip you with insights into the potential quantum-enhanced computational paradigms of tomorrow.

AI Meets Quantum: Unlocking New Possibilities in Machine Learning

Experts Samuel Rincé and Gaël Varoquaux delve into the challenges of scaling AI while maintaining sustainability. They’ll explore the trade-offs between large foundation models and smaller, domain-specific models, focusing on their environmental impact and practical applications across various fields.

AI sustainability and specialized models: balancing power and efficiency

In this technical guided session, we’ll guide you through a systems-based approach to mapping inference workflows onto the right GPUs. By examining various open-source models and performance benchmarks, you’ll learn how to choose the optimal GPU based on workload. 

You’ll also get a sneak peek of what’s next in Scaleway’s GPU roadmap, along with practical demonstrations of how to use NVIDIA Nsight Systems to profile and assess GPU performance, ensuring your inference pipelines are fully optimized.

Workshop room

Guided session: mapping inference workflows on the right GPU

Laurent Daudet, Co-founder and Co-CEO of LightOn, will explore how AI agents are transforming the capabilities of LLMs, enabling the automation of complex tasks and workflows. He will demonstrate how LightOn’s Paradigm Platform integrates tool orchestration to create flexible generative AI solutions. Laurent will show how these solutions can meet the needs of dynamic environments and evolving user demands.

Agentic capabilities for enhanced LLMs and GenAI applications

In this talk, we will explore the evolution of AI at Meta and the broader industry, highlighting Meta's open approach to research. We will delve into the Llama stack and discuss how it enables trust and safety in AI applications. Finally, we will share success stories in using Llama to demonstrate its potential in driving innovation and advancing the field of AI.

Join us to learn more about the exciting developments in AI at Meta and how they are shaping the future of research and innovation.

AI at Meta: Llama and Open Research at Scale

Join us for an engaging, hands-on workshop, "Simulating Embedding Models for Structured Data" hosted by Neuralk-AI. In this session, we'll explore the challenges companies face when integrating embeddings models for structured data and guide you through a simulated workflow using a practical use case. 

Participants will follow along on their own devices (or a Jupyter notebook will be projected on screen),

 learning key steps like data preparation, feature extraction, model integration, and results presentation. The workshop concludes with an open discussion, allowing for feedback and insights on how these techniques can be applied to real-world projects. This is an excellent opportunity for those looking to deepen their understanding of embedding models and their applications.

Hands-On Integration: Simulating Embedding Models for Structured Data

Generative AI and Large Language Models (LLMs) are rapidly reshaping industries, bringing transformative potential and new and complex security challenges. In this session, we’ll dive into the most pressing risks these technologies pose and, more importantly, offer actionable strategies to protect your organization. Whether it’s data privacy, model reliability, or cyber threats, you’ll gain practical insights to secure your business as you leverage the full power of AI. Join us for an essential discussion on staying secure while embracing the future of AI.

Securing the Future: Strategies to Safeguard Your Business in the Era of Generative AI

After a quick rundown on agentic patterns and what they unlock, we’ll jump right into demos, tackling key skills like generating consistent structured outputs using JSON Schema. You’ll learn practical tricks for leveraging function calling to integrate LLMs smoothly into your tool ecosystem. Plus, we’ll go beyond text, adding Vision for richer, more versatile applications. Join us to see how you can build smarter, more adaptable AI solutions with open-weight LLMs at the core.

Agentic AI powered by Generative APIs: Building high-precision function calling on open-weight LLMs

poolside builds frontier AI models specifically for coding, with everything from pre-training to IDE extensions, chat assistants, and APIs built from the ground up. Join Vitor who will briefly introduce poolside, key design choices, and showcase the current product in a quick demo.

Meet frontier AI for software development

Join Jeff Wittich as he explores how Ampere’s cutting-edge processors are powering AI innovations in the fast-paced worlds of autonomous driving, ride optimization and split-second in-race decision-making. Discover how these industries and others leverage Ampere’s sustainable, high-performance CPUs to deliver real-time AI inference, transforming operations while reducing costs and energy consumption. From millions of daily trips to race-day strategies, Ampere is driving AI in the fast lane.

AI in the Fast Lane: From Autonomous Vehicles to Ride Hailing to Racing

We will explore how the Intel Gaudi 3 AI accelerator, combined with Dell’s high-performance server technology, enables businesses to optimize and scale their generative AI solutions. Participants will gain insights into how the Gaudi 3 architecture enhances model performance, providing a cost-effective, versatile alternative to conventional graphic processing units. Through use case discussions, attendees will learn practical strategies for deploying Gen AI workloads more efficiently and effectively across industries. Whether you're looking to accelerate AI capabilities, manage energy costs, or expand your AI infrastructure, this session offers the knowledge and tools to leverage Intel’s innovative accelerator & software ecosystem for flexible, scalable AI deployment. Join us to see how this cutting-edge hardware empowers choice and performance in Gen AI applications.

Founders café

Bringing Choice to Gen AI with the Intel Gaudi 3 AI Accelerator Powered by Dell Servers

This presentation reveals the architecture of a Retrieval-Augmented Generation (RAG) system specifically designed for legal professionals, addressing critical needs for data sovereignty and response verifiability. Using open-source models on Scaleway’s European infrastructure, the system ensures confidentiality for sensitive legal content while enforcing source citations through structured generation techniques. Explore the technical choices that make this a secure, reliable alternative to proprietary AI solutions for law.

Building a RAG system for lawyers: behind the scenes

The Head of AI at Ampere, Victor Jakubiuk, shares insights on the techniques used by Ampere's AI software engineers in optimizing models such as llama.cpp to deliver excellent LLM inference performance on Ampere's Cloud Native Processors. The team behind Ampere Optimized AI Frameworks (AIO) delivered massive performance improvements over the model's vanilla version, making it possible to deploy solely on Ampere CPUs without a need for dedicated accelerators. 

The workshop will include setting up the demo of an AI agent running on a small parameter version of a Llama 3.2 model.

Leveraging AI agents and efficient LLM inference on COP-ARM

Nacim Rahal, Senior Director of Data & AI at Doctolib, will present the company's AI-powered medical assistant, which helps doctors with real-time note-taking and automated data processing. By reducing administrative burdens, this assistant allows healthcare professionals to focus more on patient care. Nacim will explore how this tool is transforming medical practices and improving efficiency in healthcare systems.

Central Room

AI-Powered Medical Assistant: Transforming Caregivers' Work Life

Join us for an insightful journey into the evolving world of GPU cluster management with open-source technologies. This keynote will delve into the multifaceted process of managing a GPU cluster, spotlighting why bare metal Kubernetes emerges as the most versatile and robust infrastructure approach.

You'll gain a deep understanding of how to optimize GPU utilization through strategic scheduling, ensuring your resources are leveraged to their utmost potential. Discover the essentials of running an observability platform tailored to these dynamic environments, empowering you to maintain transparency and operational excellence.

As we navigate the complexities of day-to-day operations in a rapidly advancing AI landscape, you'll learn practical strategies to sustain and enhance your software ecosystem. Whether you're an AI enthusiast, a cloud architect, or an operations professional, this session promises valuable insights to elevate your GPU cluster management skills with cutting-edge open-source solutions.

Don’t miss this opportunity to explore the intersection of open-source technology and high-performance computing and take your AI infrastructure to the next level. Join us and be part of the future of GPU cluster management!

Managing GPU Clusters with Open Source

The EU AI Act is set to reshape the landscape for AI developers and entrepreneurs, establishing rules for deploying AI systems in the market. This session will break down the regulation into actionable insights, showing you how to navigate compliance while driving innovation. Learn about the risk-based framework, key obligations, and the concrete steps you can take today to future-proof your AI solutions. Whether you're developing machine learning models or growing your AI-driven startup, this talk will give you the key insights to navigate regulations and adapt to the changing AI landscape.

Mastering the AI Act: What every AI Entrepreneur needs to know

This panel will explore how AI is revolutionizing the financial services industry, from enhancing data management and operational efficiency to integrating digital assets. The discussion will focus on the latest innovations and practical challenges in adopting AI-driven infrastructure, offering insights into the future of banking and finance.

AI’s Role in Transforming Financial Services

Industry leaders will explore the full lifecycle of data in AI, from creation and licensing to preparation and transformation, and ensuring high-performance data storage. The panel will address technical challenges and innovations at each stage, offering insights into optimizing AI data pipelines for success.

The Data Pipeline: From Creation to Training

Learn how simple it is to pull down and deploy Hugging Face models on AMD GPUs using ROCm™ software and discover the benefit of end-to-end workflows and optimizations in an open-source development ecosystem.

AMD x Hugging Face: “No Code Change LLM Deployment on AMD GPUs"

Speech AI technologies, including Automatic Speech Recognition (ASR) and Text-to-Speech (TTS), enables personalized, human-level customers experience through applications like Virtual Assistants, Contact Center Agent Assistance, and digital avatars. In this session, we will address the common challenges and demonstrate how to build speech AI pipelines using NVIDIA Riva.

Building Speech AI Applications with NVIDIA Riva

When there are no universal standards for measuring the full environmental footprint of cloud services, anyone can say what they want. In this session, Scaleway will present the methodology they have adopted to offer the most comprehensive and transparent environmental reporting of the cloud sector, covering not only carbon, but also bespoke ACV and water impacts. With AI workloads demanding ever more cloud resources, this session will highlight how even AI operations can be more sustainable, providing valuable insights, and allow attendees to easily identify levers to reduce their own impact.

Accurately measuring your cloud activity’s full impact: wishful thinking?

This workshop introduces NVIDIA BioNeMo, a generative AI platform designed to revolutionize drug discovery. 

Participants will explore how BioNeMo integrates advanced AI models and high-performance computing to streamline critical processes in pharmaceutical research.

Accelerating Drug Discovery with NVIDIA BioNeMo

Join experts from HPE, OCP, and Quandela as they examine the cutting-edge technologies that are redefining AI hardware. This session will cover advancements in quantum computing for AI, the role of open platforms in enhancing flexibility and interoperability, and the modular systems that are enabling more scalable and energy-efficient AI clusters. The discussion will highlight both recent achievements and the challenges ahead, offering insights into what the next 12 to 24 months—and beyond—hold for AI infrastructure, with a focus on sustainability and technological efficiency.

Building sustainable AI infrastructure: quantum, open and modular systems

As AI models, particularly large language models, increase in size and complexity, effectively managing memory bottlenecks becomes imperative. This presentation will delve into the technical architecture and packaging innovations of AMD GPUs, which boast high memory capacity to cater to demanding generative AI workloads. Additionally, this talk covers practical strategies for optimizing performance and highlight the role of open-source tools and collaborations with ecosystem partners in accelerating AI development.

Overcoming memory Hurdles in AI - Performance and choice at scale

Explore the technical challenges of ensuring AI accuracy and explainability in healthcare, defence, and beyond. Experts from Helsing, Owkin, and Giskard will dive into methods for detecting errors like ethical bias, diagnosing cancer cells through computer vision, and ensuring real-time AI systems in defence operate with minimal risk.

Moderator : Sophie Monnier, Business Development Engineer at InstaDeep

Building trustworthy AI: accuracy and explainability across sectors

Top VCs Gabriel de Vinzelles, Alexis Robert, Paul Murphy, and Rana Yared discuss how the distinction between AI builders, like Poolside and Mistral, and AI adopters is reshaping investment strategies. They’ll explore how each approach influences risk, timelines, and growth potential in today’s AI-driven market.

Investing in AI: builders vs. adopters

Experts from Photoroom, Mojo, and Aive discuss the complexities of deploying real-time generative AI for video and photo applications. They’ll tackle issues like heavy content processing, managing hallucinations, and optimizing virtualization/quantization to ensure efficient, scalable solutions for end users.

Real-time generative AI for video and photo: technical challenges uncovered

As the use of generative AI (GenAI) accelerates, so does its environmental impact. In this session, IJO and On Behalf AI will dive into the core principles of carbon balance calculation for AI models, offering a clear view of what we can measure with confidence and the uncertainties that challenge precision. Attendees will learn how different factors—such as data centers, hardware, and model training—contribute to carbon output. This talk will break down known metrics and orders of magnitude, helping developers, engineers, and researchers monitor their AI’s carbon footprint and manage their impact more consciously in their day-to-day work.

Decoding GenAI’s carbon footprint

Experts from Bioptimus, Qubit Pharmaceuticals, and InstaDeep will discuss how they’re tackling real-world challenges like scaling AI models and building specialized bioAI models to accelerate clinical insights.

Discover how quantum computing optimizes drug research, how custom AI models in biology rely on precise data handling, and the techniques behind generative bioAI models for protein and DNA analysis.

Join us for a deep dive into the AI architectures and tools transforming healthcare and biotech today.

Moderator : Chris O'Brien, Founder & Editor of The French Tech Journal

Overcoming AI Scaling Barriers in Healthcare & Biotech

We use cookies to improve your browsing experience, save your preferences and provide us with informations on how you use our website. For more information about cookies, please read our