Accelerated Networking for AI Infrastructure
Gain a holistic understanding of AI factory networking, exploring how key communication methods between GPUs, racks, and storage shape overall AI performance and scalability. Bring your own laptop and experience AI factory networking in action. In AI factories, the network is critical to maximizing performance. You’ll see how networking choices shape AI workload performance and scalability, then configure key pieces yourself in guided labs. We start with a level set session that walks through essential communication paths in an AI factory deployments: east-west GPU-to-GPU communication across
Fundamentals of GPU-Accelerated Workflows with CUDA Python
This course delivers a hands-on introduction to GPU-accelerated computing in Python, empowering developers to build fast, scalable applications using NVIDIA’s CUDA ecosystem. Through guided notebooks, participants master CuPy for array acceleration, cuDF for GPU DataFrames, and the cuda-python API to write custom kernels—all without leaving Python. Real-world exercises in data science, machine learning, and scientific computing emphasize performance, interoperability, and end-to-end efficiency. Learners progress from drop-in speedups to fully integrated GPU pipelines, mastering data movement,
Adding New Knowledge to LLMs
In today's AI landscape, even powerful Large Language Models (LLMs) face limitations when confronted with specialized business knowledge, technical domains, or cultural contexts absent from their training data. While retrieval-augmented generation can mitigate some gaps, true domain mastery requires deeper model adaptation. This comprehensive workshop equips developers with hands-on skills to transform open-source LLMs into domain-specialized AI assets. Through five interconnected modules, you'll master the complete lifecycle of model customization. By workshop completion, you'll possess the c
Building Observable and Scalable Multi-Agent Workflows for Asset Lifecycle Management
Learn to develop multi-agent workflows tailored for Asset Lifecycle Management use cases within industrial process applications. You'll be introduced to the NVIDIA open-source NeMo Agent Toolkit (NAT) and its key features, enabling the creation of a comprehensive and customizable reference solution for this use case. You'll also be guided in effectively utilizing other components of NVIDIA AI stack, including NVIDIA Inference Microservices (NIM), cloud endpoints, and NVIDIA Tesseract time series foundational models. These components are used to design agents capable of executing frequently per
Building AI Agents with Multimodal Models
Just like how humans have multiple senses to perceive the world around them, computers have a variety of sensors to help perceive the human world. In the health industry, computed tomography (CT) scans provide a 3D representation used to detect potentially dangerous abnormalities. In the robotics industry, lidars are used to help robots see depth and navigate the complex topology around them. In this course, learners will develop neural network based multimodal models that can understand many different data types by exploring different fusion techniques. Certificate: Upon successful completion
Building LLM Applications with Prompt Engineering
With the incredible capabilities of large language models (LLMs), enterprises are eager to integrate them into their products and internal applications for a wide variety of use cases, including (but not limited to) text generation, large-scale document analysis, and chatbot assistants. The fastest way to begin leveraging LLMs for diverse tasks is by using modern prompt engineering techniques. These techniques are also foundational for more advanced LLM-based methods such as Retrieval-Augmented Generation (RAG) and Parameter-Efficient Fine-Tuning (PEFT). In this workshop, learners will work wi
OpenUSD Crash Course: Build 3D Data Pipelines for Physical AI
Discover how OpenUSD principles and Python scripting can build scalable 3D data pipelines for manufacturing, robotics, and physical AI applications. Throughout this full-day workshop, you'll learn composition arc techniques, asset hierarchy design, and performance optimization strategies essential for production pipelines. Develop practical skills in automating 3D asset workflows, managing complex scene data, and preparing simulation environments for physical AI training. This workshop directly prepares you for the NVIDIA-Certified Professional: OpenUSD Development exam. Certificate: Upon succ
How to Simulate, Train, Validate, and Deploy an End-to-End Robotics Workflow with NVIDIA Isaac
In this full-day workshop, we'll guide you through an end-to-end robotic workflow—from simulation-based training to real-world robot deployment. You'll simulate the robot in NVIDIA Isaac Sim™, train the policies in Isaac Lab, validate the trained skills through software-in-the-loop testing, and deploy them to physical robots. Through practical exercises, you'll gain end-to-end experience in the techniques that bridge the sim-to-real gap: synthetic data generation, policy training and refinement, hardware testing on edge devices, and real-robot deployment. This workshop demonstrates how simulat
Deploying and Optimizing AI Inference at Scale
As foundation models move toward deeper test-time computation, inference becomes the dominant scaling constraint. Latency, throughput, and cost are governed by a small set of forces: autoregressive decoding, KV-cache growth, memory bandwidth, and scheduling under contention. This workshop frames large-scale inference through these emerging laws of inference, starting from first principles and building toward real systems. Learners begin with monolithic and gateway-based vLLM deployments on Kubernetes to establish baseline behavior, then transition to NVIDIA Dynamo to operate aggregated and dis
利用 NVIDIA Cosmos VSS 构建智慧交通（ITS）违章检测系统
与传统的计算机视觉模型相比，基于 VLM 的 Cosmos VSS 服务在视频分析领域的意图识别相关场景中表现出色。在本次演示中，我们将深入探讨一个实际的商业项目，以介绍 VSS 在实际应用中的使用方式，以及如何利用它来构建能够检测交通违规行为的服务。本次演示将带您了解我们在部署 Cosmos VSS 时所面临的挑战、资源需求，以及我们如何利用真实数据集对其进行微调以满足客户特定的业务目标。
面向灵巧操作的高效强化学习框架
当前的具身基础模型，仍然难以在真实世界任务中同时满足接近 100% 的成功率与严格的执行周期的要求。强化学习被认为是弥合这一差距的关键技术路径，然而在灵巧操作场景中，强化学习仍面临奖励稀疏、样本效率低以及真实世界试错成本高等核心挑战。我们提出一个高效强化学习框架，它建立在一个大一统的具身基础模型之上，不仅实现了视觉和语言的理解，在同一架构中同时也建模图像生成、动作生成以及价值生成。这样的一个统一模型范式带来了多重好处，它结合了VLA和世界模型的能力，通过语言、2D视觉、3D结构以及本体状态的时间空间思维链大幅提升了模型跨时间、跨模态的联合推理，从而整体提升了策略的泛化能力；同时也为强化学习提供了过程级的稠密信息，这些信息不仅仅包含动作的，同时也包含了视觉和语言的隐空间，显著提升了灵巧操作在真实世界中的强化学习效率。这个学习框架已经在NVIDIA Thor上实现了端侧的部署，展示了其在具身基础模型规模化落地与量产应用中的潜力。
十载相伴，NVIDIA 赋能创业公司在 AI 时代加速前行
NVIDIA 初创加速计划深耕中国十年，陪伴上千家创业企业从技术探索走向商业落地。2025 年，我们与生态伙伴持续携手，为中国创业公司提供了一系列产品、技术、市场、资本与业务对接的全链路支持，同时迎来了会员规模突破 3000 家的重要里程碑。本视频将回顾 2025 年该项目如何赋能中国创业者加速成长，并展望 2026 年在生态协同与创业扶持上的新机遇。同时，3 家 NVIDIA 初创加速计划会员企业也将分享他们如何借助 NVIDIA 的创新技术快速实现业务跃迁，为行业与社会创造更大价值。
ROLL：集成 NVIDIA Megatron、高效易用的大规模 LLM 强化学习框架 An Efficient and User-Friendly Large-Scale RL Framework Integrating Megatron
ROLL 是专为 LLM 设计的高性能分布式强化学习库，旨在兼顾大规模集群处理千亿参数 LLM 的线性扩展、系统容错和高利用率。它深度集成 Megatron-LM 实现强化学习工作流与分布式训练原语无缝互操作，结合 Megatron-Core、SGLang 及 vLLM 构建端到端训推加速链路。ROLL 在偏好对齐、复杂推理及多轮智能体交互等场景显著优化了大规模集群性能。此外，ROLL 实现了千亿规模 MoE 训练在数千卡 GPU 节点的拓展，支持长达两周的生产级无中断运行，展现了卓越的可扩展性和容错能力。ROLL is a high-performance RL library designed for LLMs to balance scalability, fault tolerance, and high utilization for trillion-parameter models on large-scale clusters. It integrates Megatron-LM to achieve seamless interoperability between RL and distributed training primitives, it leverages M-Core, SGLang, vLLM for an E2E training and in
基于 NVIDIA AI Enterprise 全栈赋能：“AI Factory” 重塑智造新标杆
金盘科技携手 NVIDIA，针对传统制造业面临的效率瓶颈、数据孤岛及安全管理被动等痛点，构建了软硬一体的“AI 工厂”。依托 NVIDIA AI Enterprise 算力底座，融合 NIM 推理微服务、NeMo Agent Toolkit 智能体编排、VSS 视频分析及 Omniverse/Isaac 数字孪生平台，企业成功落地了多项关键场景。在工厂智能体分析方面，利用 VSS 视觉大模型技术，将工程现场非结构化的监控视频实时转化为结构化数据，主动识别 PPE 穿戴合规性与设备异常，实现从“事后回溯”到“事前预警”的跨越，不仅使项目安全水平提升 70%，运营效率也同步提升 50%。此外，通过标书生成智能体、职能数字员工、供应链智能控价等应用，标书制作周期从周级压缩至小时级并辅助中标 12 亿元，合同审查效率提升 10 倍，数字化解决方案对外输出斩获超 9 亿元 订单，成功打造了制造业从“人力驱动”向“数据与硅基劳动力驱动”转型的行业标杆。
基于 NVIDIA 全栈技术打造代理式 AI 与物理 AI 的未来基石
从帮助团队快速搭建和优化多智能体工作流的 NVIDIA NeMo Agent Toolkit，到围绕 Data Flywheel 与 NVIDIA Agentic Blueprint 构建可自我提升的代理式 AI 系统，再到面向物理 AI 的 NVIDIA Isaac GR00T 基础模型、辅助驾驶以及 NVIDIA Cosmos 世界模型等平台能力。NVIDIA 技术专家将系统讲解如何借助 NVIDIA 完整技术栈，更高效地设计、训练与部署新一代代理式 AI 与 物理 AI 应用。同时，3 家 NVIDIA 初创加速计划会员企业将通过生动的案例展示如何借助这些技术突破业务瓶颈，推动行业变革与发展。
工业级具身智能，正在从概念走向现实
依托合成数据驱动的具身 VLA 大模型，Galbot G1 在通用性与泛化能力上实现关键突破，真正打通了通用机器人从感知、决策到执行的闭环，并在复杂、非结构化的真实世界环境中验证了其可靠能力。打造高质量仿真数据是我们的核心优势。通过与 NVIDIA Isaac Sim 的深度合作，我们构建了一套可规模化、具备确定性的训练与验证路径，使机器人能够系统性地掌握全身协同控制、灵巧操作等高阶技能。与此同时，随着 NVIDIA Jetson Thor 的性能释放， Galbot G1 Premium 的自主工作能力得到进一步提升，在严苛的工业环境中显著提升了实时规划能力与整体执行性能，推动具身智能迈向更高水平的工业应用。
面向大规模 Agent RL 的高并发沙箱系统构建与实践
随着 Agent 从对话式演进为能够与真实世界交互的自主智能体，Agent 强化学习（AgentRL）逐渐成为核心训练范式。与传统 RL 或监督学习不同，AgentRL 对基础设施提出了前更高要求和挑战，包括：超大规模的环境并发、环境的快速启动与销毁、强隔离安全要求，以及 GPU&CPU 的高效利用。 本次分享将聚焦于支撑大规模 AgentRL 的基础设施设计与工程实践，我们将从系统工程视角，拆解 AgentRL 背后的关键 Infra 挑战，并分享可复用的架构模式与设计原则。 内容将覆盖 AgentRL 任务编排、基于NVIDIA Dynamo的模型部署、环境生命周期管理、执行隔离机制以及多层次加速技术，重点介绍如何将环境启动延迟从分钟级压缩到毫秒级，如何支撑数十万级并发 Agent 沙箱环境运行，以及如何提高资源利用。
通过触觉 AI 解锁高自由度灵巧操作
随着集成触觉感知的高自由度（High-DoF）灵巧手产品出现——例如 SharpaWave——机器人灵巧操作以及高质量数据采集迎来了新的发展机遇。 本次分享将介绍如何利用触觉信息来：1)训练模型，使机械手能够根据接触状态自适应调整行为，从而实现精细的操作动作，2)提升操作数据采集的效率与质量，3)缩小在接触丰富的任务中从仿真到现实（Sim-to-Real）的差距。演讲还将涵盖一系列实用方法与实现方案，包括：1) 在 NVIDIA Isaac Sim 中对触觉灵巧机械手进行建模与仿真，2)通过仿真中的强化学习生成训练数据，3)在真实硬件实验中进行模型训练与验证
AI 重塑影视产业：打造基于 NVIDIA 全栈的 AI 算力池，生成式 AI工作流与 OpenUSD 资产平台
上海电影集团成立于76年前，是中国电影产业的重要先驱者。2023年，我们迈出了关键一步——投入建设AI算力体系。 我们围绕上海电影集团昊浦智慧产业社区，构建并规模化落地一套面向影视内容生产、产业协同与生态共创的 NVIDIA 全栈智能基础设施体系。构建面向 CG、AIGC、虚拟拍摄与大模型服务的统一GPU算力池，实现跨企业、跨团队的资源池化、服务化与精细化调度；基于 NVIDIA Omniverse 与 OpenUSD 打造影视行业级数字资产平台，将场景、角色与内容资产从“项目文件”升级为“可复用、可流转、可进化的产业级生产资料”；依托 NVIDIA AI Enterprise 与 NVIDIA NIM，将AI能力深度嵌入创作、制作、交付与运营全流程，形成可复制、可审计、可规模化推广的影视AI生产体系。
基于具身智能加速即时物流机器人研发：从大语言动作模型训练到数字孪生验证
基于 PI0.5 架构与英伟达 GPU 集群的大语言动作模型训练：在 Isaac Lab 平台上，依托 PI0.5 架构训练灵巧手操作及换电夹爪模型，在仿真环境中实现 90% 的抓取成功率；融合 100 余个真实场景运动捕捉数据集与 Isaac GR00T 拟态技术，完成 10 倍数据增广，有效缩小仿真到现实的落地差距。 面向快速原型开发的物理数字孪生技术：在物理原型制作前，于 Isaac Sim 平台完成无人机机器人对接及换电站的设计验证；以 0.1 毫米的精度仿真 500 余种对接场景（如对位偏差、振动干扰等）；硬件落地前，基于数字孪生实现 75% 的首次对接成功率，将现场测试故障率降低 30%。 实际场景性能提升：通过数字孪生验证替代 80% 的物理测试，研发周期缩短 50%；经 200 次数字孪生迭代优化后，货物转运成功率达 92%。
多模态大模型训练的性能优化 Performance optimization for training Multimodal Large Language Models
与大语言模型相比，多模态大模型（MLLM）能够处理文本、图像和音频等多种类型的数据，更接近人类对世界的感知方式。然而，不同模态之间的数据和模型结构存在差异，这使得传统的模型训练方式难以高效地训练多模态大模型。DistTrain 通过解耦编码器、主干网络和生成器的训练并行策略，实现了更高的训练效率。NVIDIA 与 Stepfun 合作，在开源框架 Megatron-LM 上实现了 DistTrain 方案，使社区能够更轻松且高效地训练多模态模型。 Multimodal large language models (LLMs) can process and generate different types of data, such as text, images, and audio. However, combining these modalities creates challenges for model training due to differences in their data and model structures. DistTrain introduced optimizations that separate the training parallelism of different model components (encoders, backbon
LLM 如何赋能量化投资
LLM 作为 AI 技术的核心分支，正在重塑量化投资范式。LLM 具有超乎寻常的语义理解与推理能力，可以作为强大的工具模型嵌入量化投资流程。 从 LLM 的输出随机性出发，我们可以依赖 LLM 进行因子挖掘，并基于 NVIDIA NIM 和 TensorRT-LLM 加速挖掘； 从 LLM 的语义理解能力出发，我们基于 GPU 本地部署LLM解读海量文本、提取情感信息，构建或优化量化选股策略； 从 LLM 的多模态思维链出发，我们可利用 LLM 对K线图像进行交互式技术分析。LLM 正在引发一场量化投资技术更迭的革命......
Liquid-Cooled AI Server Platforms for Next-Generation Performance (Presented by ASRock Rack, Inc.)
Explore ASRock Rack AI server solutions from edge to cloud and discover how ASRock Rack adopts liquid-cooling approaches in next-generation AI platforms to unlock higher performance and scalability. Learn how these liquid-cooled server designs are validated in ASRock Rack’s Liquid Cooling Lab to ensure reliability, efficiency, and readiness for modern AI workloads.
Scale AI Data Pipelines With High-Capacity JBODs and NVIDIA BlueField-4 DPUs (Presented by Seagate Technology)
Learn how to move data-intensive applications from proof of concept (PoC) to production scale using a reference architecture that pairs high-capacity JBOD storage with NVIDIA BlueField 4 DPUs to accelerate data access, reduce CPU overhead, and improve end-to-end efficiency. This session presents a practical blueprint for solution architects designing cost-effective, high-throughput storage systems to support modern, data-driven and AI-enabled applications at scale.
How to Run AI-Powered Computer-Aided Engineering Simulations
In this lab, we explore how to build AI surrogate models for crash simulations, one of the most computationally intensive tasks in computer aided-engineering (CAE). Learn how modern AI architectures can reproduce high-fidelity physics simulations at a fraction of the cost, enabling faster design exploration and decision-making. We'll walk through the full workflow for developing an AI surrogate for CAE applications, including data preparation and analysis, training and optimization, loading and running inference with pre-trained models, and evaluating model accuracy and uncertainty. We'll also
Assembling Industrial Digital Twins with OpenUSD and NVIDIA Omniverse Libraries
In this hands-on lab, participants will explore how to build industrial digital twins from manufacturing facilities to AI factories, using OpenUSD and NVIDIA Omniverse libraries, drawing on real-world collaboration with industry leaders. The session walks through an end-to-end workflow, from CAD models to full digital twin assembly, demonstrating how manufacturers are transforming design and production through simulation and interoperability. Prerequisite(s): Intermediate experience with 3D content concepts: meshes, materials, transforms, scene hierarchies, and USD fundamentals Familiarity wit
Create Vision AI Applications With Generative AI Coding Agents
Computer vision is a cornerstone of the generative AI revolution, providing the essential visual metadata required by vision language models (VLMs) to extract actionable insights. In this two-hour instructor-led workshop, you will learn how to GPU-accelerate your video analytics workflow by combining the NVIDIA DeepStream SDK with the power of prompt engineering. Discover how to generate complete, GPU-accelerated DeepStream pipelines using simple natural language prompts. Whether you are a developer new to DeepStream or an advanced user looking to accelerate your prototyping, this course demon
Fine-Tune a Telco Reasoning Model: A Guide to Synthetic Data, Tool Calling, and Evaluation
Network operation centers (NoC) are the central nervous system of telecommunications, but they are often overwhelmed by "alarm storms." In a traditional setup, engineers manually validate alarms, swivel between multiple dashboards, check topologies, and perform root-cause analysis. This manual process is time-consuming and prone to fatigue. To solve this, we are moving toward zero-touch, self-healing networks. In this tutorial, we'll walk through creating a fine-tuning playbook that integrates synthetic data generation, training, and evaluation pipelines. We'll demonstrate how to build an AI-d
Synergy at Scale: Unify Compute, Power and Cooling (Presented by COMPAL)
High-density AI workloads demand a paradigm shift, from siloed infrastructure to integrated ecosystems. This presentation reveals how a combined entity is leveraging its expanded scale and expertise to deliver the first truly total solution for next-gen data centers. We will explore how unifying compute, power, and liquid cooling into a single, validated solution stack eliminates integration risk, slashes deployment time, and ensures optimal performance for NVIDIA's GPU architecture. Learn how to transition from fragmented purchasing to a seamless, factory-integrated, plug-and-play AI infrastr
The Agentic AI Data Factory: Why Agents Need a GPU-Native Data Platform to Create Real Value (Presented by Capgemini)
Agentic AI is evolving fast—but most organizations can’t fully capitalize on it, because their data foundations weren’t built for real‑time, AI‑driven decision-making. In this session, we introduce InsightGrid—a GPU‑accelerated data factory built on NVIDIA technologies that unifies all enterprise data into a single, AI‑ready foundation. InsightGrid brings together batch, streaming, structured, and unstructured data, and embeds unstructured content alongside core business entities to give agents a complete, contextual view of your business. The result: agentic systems that don’t just retrieve i
Get NVIDIA-Certified at GTC for Free: Associate Certification
Register separately to book an exam slot and take your chosen exam onsite. (Exam registration link coming soon) Maximize your GTC experience by earning an industry-recognized Associate NVIDIA certification at no additional cost. Conference attendees can register to take an Associate-level certification exam onsite -choosing from Generative AI LLMs, Gen AI Multimodal, Data Science, and AI Infrastructure and Operations. Advance registration is required to book an exam slot. We recommend preparing in advance using official NVIDIA study guides and the on-demand certification overview session which
Advance AI Data Center Infrastructure Design and Construction to Accommodate GW Scale (Presented by EdgeConnex)
Today’s data center industry is evolving faster than ever, with record demand for gigawatt (GW)-scale designs now paired with the accelerated delivery intervals needed to meet today’s ever-changing technology demands. This race to build large-scale data center infrastructure has fundamentally shifted the industry’s approach, from traditional design and construction methods to ones that are highly optimized for accelerated delivery at GW scale. In this session, explore the drivers and overall impacts of designing and building large-scale GW data center infrastructure that serves as the building
Practical Applications of NVIDIA 3 Computers in Vision, Robotics, and Edge (Presented by EDOM Technology)
This session explores how NVIDIA 3 Computers—Train, Simulate, and Deploy—enable enterprises to turn AI models into real-world applications across computer vision, robotics, and edge AI. Learn how this end-to-end architecture accelerates AI deployment, reduces risk, and helps organizations scale AI from experimentation to production.
Supercharging Postgres for Agentic Analytics with NVIDIA RAPIDS and Apache Iceberg (Presented by EDB)
As data volumes increase, the primary bottleneck for high-performance AI agents will shift from the model to the data. This increases the importance of the underlying data engine’s ability to process massive enterprise datasets in real time. This scaling problem is further amplified by the desire to make the latest transactional business data seamlessly available to agentic processing. Join the experts from EDB for a technical deep dive into how to overcome these scaling and transactional integration hurdles using the world's most popular open-source database. We will showcase the architecture
Leverage Acceleration for AI and HPC in Algorithmic Trading
Join this interactive session with NVIDIA experts to discover how the convergence of artificial intelligence and high performance computing (HPC) is reshaping algorithmic trading. As trading strategies grow in complexity, the need for high-throughput processing and advanced modeling capabilities becomes critical. Speak with our experts about the latest methods to leverage acceleration for data-intensive tasks, from training sophisticated AI models to running massive-scale simulations. Important: Connect With the Experts sessions are interactive sessions that give you a unique opportunity to me
Scaling Out and Across: Networking Innovations for Giga-Scale AI Systems
As AI factories grow to hundreds of thousands, and soon millions, of interconnected GPUs, the network has become the defining architecture of large-scale AI. Achieving uncompromised performance requires fabrics that not only scale up and out within a single data center, but scale across multiple sites with predictable latency, lossless throughput, and end-to-end resilience. In this session, NVIDIA unveils the full breadth of networking innovations that enable the construction of giga-scale AI infrastructure. Learn how advancements in networking, including ultra-high-radix switches, next-genera
AI-First Design: Build Smarter From Day Zero Through Every Funding Stage
The decisions AI-first companies make on Day Zero will radically shift based on initial resources. Does a $5 million seed round change the fundamental AI strategy compared to a bootstrapped $0 start? Absolutely. This session explores four distinct Day Zero scenarios, each representing a different starting capital: $0 (Bootstrapped), $500,000 (Pre-Seed), $5 million (Seed/Series A), and $50 million (Series B+). For each scenario, we'll highlight high-leverage AI-first tactics, illustrating how capital immediately dictates where to spend resources and how fast to build proprietary technology. You
Scale Sports Intelligence and Insight With Foundational Models
Imagine that you could generate the world’s most in-depth dataset for sports using the same video you see on your TV screen. In this presentation we’ll show you how we combine computer vision and our own proprietary multimodal foundation model that takes away the need for in-venue hardware and processing and scales infinitely across the game of soccer and beyond; uncovering and explaining the language of sport. We’ll show how GPU acceleration allows us to model entire matches as unified multi-agent sequences, and how this foundation model powers the next generation of soccer intelligence - fro
From Concept to Production: Humanoid Robotics at Scale
Humanoid robots are evolving from prototypes to production-ready systems, powered by breakthroughs in reasoning AI, large-scale simulation, and real-time edge computing. But to get there, the industry must bridge the gap between digital intelligence and physical reality. Hear from luminaries at the world’s leading robotics companies on the future of physical AI, and what it will take to unlock truly capable humanoids.
Turning Models Into Engines: The AI Factory Era
Step inside Mistral’s journey to industrial-scale AI as we share lessons learned and the principles guiding our approach in creating Mistral Compute. Built on four key pillars—open-source foundations, security by design, sustainability at scale, and digital sovereignty—we’ll share how our approach unlocks bare-metal GPU performance, why that matters in building trusted, European-hosted future-ready AI systems, and how we’re bringing collaboration to the next level by giving innovators access to the same infrastructure powering our models.
Big Compute Powering the Energy Value Chain
Herlinde Mannaerts-Drew, SVP oil and gas technology, will discuss how bp is working with NVIDIA to help enhance seismic imaging and reservoir modelling to enable faster exploration and reduce uncertainty.
The Era of GPU Data Processing: From SQL to Search and Back Again
This session delivers a technical state of the union on GPU-accelerated data processing across SQL/DataFrames, vector search, ML, and decision optimization. Learn how GPU-native engines enable interactive analytics on massive lakehouse-scale datasets, real-time semantic and vector search over billions of embeddings, and makes the hardest ML and decision science workloads tractable, cost-efficient, and energy-efficient. The talk highlights the implications for high-impact scientific and enterprise computing, then looks ahead to what’s in flight for 2026 and beyond, outlining concrete architectu
From Research to Production: How Alpamayo Accelerates Autonomous Vehicle Development
Discover the latest advancements in end-to-end autonomous driving as the industry accelerates toward scalable, Level 4 AI-driven systems. This session explores how NVIDIA’s Alpamayo—particularly reasoning vision-language-action models (VLA)—enable vehicles to interpret, reason about, and navigate complex real-world scenarios with greater humanlike understanding. We’ll highlight how this approach is strengthened by a continuous self-improving loop powered by NVIDIA Cosmos and safeguarded by NVIDIA Halos, NVIDIA’s comprehensive full-stack safety system for physical AI. Finally, we’ll discuss how
Open Source AI Shaping the Next Era of Intelligent Digital Workers
Open source AI is redefining how we build and scale intelligence. In this session, we explore how AI factories unlock a new generation of digital workers who learn continuously, collaborate with humans, and drive breakthroughs across every industry.
Integrating AI and Quantum Computing to Accelerate the Future of Supercomputing
The path to fault-tolerant quantum computing is an AI and supercomputing challenge. This talk outlines a vision to accelerate scientific discovery for everyone, everywhere, through the complete convergence of AI with quantum hardware and systems. We will present the blueprint for a collaborative, open quantum ecosystem that transcends boundaries. By leveraging open standards, CUDA-Q, and NVQLink, we are creating a unified, hybrid infrastructure where academia, labs, startups, and industry can collaborate to drive the future of supercomputing
How Open Models, Agents, and Physical AI Are Fueling a Flywheel of Healthcare Innovation
Healthcare is entering a once-in-a-generation leap as AI, accelerated computing, and robotics converge to turn data into decisions and accelerate breakthrough discovery. At GTC 2026, NVIDIA will showcase how open foundation models, agentic AI, and physical AI are creating entirely new categories of applications across care delivery, drug discovery, and laboratory operations. AI for science is closing the loop between digital dry labs and autonomous wet labs, catalyzing biology’s transformer moment. Healthcare robotics that combine digital and physical agents are closing care gaps around the wo
The Developer’s Roadmap to Physical AI: Bridging Digital Models and Real-World Systems
The era of generative AI has mastered digital content, but the next frontier lies in the physical world. For developers comfortable with LLMs and retrieval-augmented generation (RAG), the leap into physical AI—robotics, autonomous systems, and edge computing—can feel like starting from zero. This session will bridge that gap. We will deconstruct physical AI workflows, showing you how to translate digital reasoning into physical action. We'll explore how to move beyond text-based inputs to incorporate multi-modal sensing and spatial intelligence, turning static models into interactive, real-wor
A Playbook – Operating Cloud AI Factories at Scale
Discover how to build and operate AI factories that scale with your ambitions. In this session, NVIDIA shares hard-earned lessons from working alongside Cloud Partners and ecosystem ISVs to power massive AI infrastructure around the world. You’ll gain a practical framework—built on open, composable infrastructure software—for managing the full lifecycle of large AI systems, from reliability and performance to multi-tenant orchestration. Walk away ready to accelerate your time-to-production and deliver sustainable scale for your own AI operations.
CUDA: New Features and Beyond
The CUDA platform is the foundation of the GPU computing ecosystem. Every application and framework that uses the GPU does so through CUDA's libraries, compilers, runtimes and language—which means CUDA is growing as fast as its ecosystem is evolving. Presented by one of the architects of CUDA, at this engineering-focused talk you will learn about all that's new and what's coming next for both CUDA and GPU computing as a whole.
Accelerating Industrial Engineering: From Product Design to Manufacturing in the AI Supercomputing Era
Accelerated computing and AI supercomputing are redefining industrial engineering, spanning computational engineering (CAE) to semiconductor design and manufacturing. Together, these fields drive how the world conceives, simulates, and builds everything from advanced machinery to next-generation chips increasingly in the digital world before physical production begins. In this address, NVIDIA’s Tim Costa, general manager for Industrial and Computational Engineering, will show how CUDA-accelerated platforms, AI physics, DGX and cloud-scale AI infrastructure, and NVIDIA Omniverse are reshaping w
Leverage the Power of Thor in Industrial Applications (Presented by Advantech)
As physical AI becomes more prevalent in industrial applications, performance, longevity, and reliability are critical. Enter NVIDIA Jetson Thor solutions from Advantech. In this session, Advantech will provide an overview of both NVIDIA IGX and Jetson Thor, discuss use cases in a variety of industries, and showcase how NVIDIA Thor platform and Advantech together solve some of the biggest challenges facing you today.
Industrial Takeoff: Reinventing Operations with Physical AI at Scale (Presented by Accenture)
Industrial operations are entering a decisive new phase. As manufacturers face rising pressures on resilience, sustainability, productivity, and geopolitical exposure, emerging technologies—Physical AI, Robotics, and next generation digital twins—are converging to redefine how factories and supply chains operate. Discover how Accenture and Airbus are reinventing operations—from foundational technologies to transformative business applications. With NVIDIA Omniverse and Isaac Sim, Accenture can enable high-fidelity simulation, synthetic data generation, and scalable AI training across interoper
Improving AIDC Operational Efficiency with AI Hardware Diagnostic Innovations (Presented by Aivres)
This session will dive into the current difficulties in addressing common data center conundrums, with a specific focus on the intricate world of AI hardware interaction and failure analysis. With a comprehensive overview of existing strategies for effective troubleshooting and a detailed case study of real-world data center diagnostic scenarios, we will analyze the limitations of current methodologies. Furthermore, this talk will explore forward-looking perspectives beyond present-day solutions and outline the desired features and functionalities that next-gen AI hardware could incorporate to
Real Cloud Infrastructure for Real AI Workloads: Training and Inference at Production Scale (Presented by CoreWeave)
As dense frontier and mixture-of-experts models move toward multi-trillion parameters, AI pioneers need real cloud infrastructure ready for real-world AI workloads spanning from training trillion-parameter models to long-context inference and reinforcement learning. In this session, we will provide a deep dive on how we optimized every layer of our full stack architecture from infrastructure and orchestration to observability to run large-scale training and inference workloads efficiently to power all AI workloads, especially agentic AI. We will examine the architectural breakthroughs, and new