This year, I was among the lucky ones who got to attend NVIDIA’s annual flagship conference, GTC 2025, in person. It was my first time at the event, so I really didn’t know what to expect.

The conference took place over five days in San Jose, California – right in the heart of Silicon Valley. While the main expo was hosted at the San Jose Convention Center, the sheer scale of the event meant that it took over an entire city block. Multiple hotels, a park, and even a museum served as official venues.

The kickoff keynote was delivered by NVIDIA CEO Jensen Huang at the SAP Arena. In his usual direct and personal style, he presented the key innovations across AI, graphics, quantum computing, and other domains.

Major announcements included:

• The Blackwell GPU architecture has now entered full production and offers up to a 40x speedup in training and inference for AI models compared to the previous generation. Especially striking was the new Blackwell Ultra system—a rack-sized AI supercomputer built specifically for the emerging class of “reasoning AI” models.
• Another highlight was a new open-source software component called Dynamo—described as the “operating system for AI factories.” It dramatically boosts inference efficiency by allowing multiple GPUs to collaborate in the background to process complex queries more quickly and cost-effectively.
• Robotics announcements also drew considerable attention: NVIDIA introduced its first open-source humanoid robot model, Isaac GR00T N1, designed to learn general motion patterns and behaviors. This was paired with the unveiling of Newton, a new physics simulation engine that allows robots to be trained digitally in simulated environments.
• The theme of simulation continued to grow: through new Omniverse blueprints, it’s now possible to digitally design, optimize, and test AI data centers, logistics systems, or even manufacturing lines.
• NVIDIA presented its new photonic data center switches (Spectrum-X Photonics), designed to efficiently interconnect massive GPU clusters—reaching into the millions—while reducing power consumption and latency. This is a key building block toward enabling full-scale AI factories.
• A special mention goes to the new GeForce RTX 5090, which can render fully ray-traced scenes in real-time with AI-assisted image generation. Its graphical power takes not only gaming, but also filmmaking, design, and simulation to a whole new level.
• On the professional side, the new RTX Pro Blackwell series was announced, offering up to 96 GB of memory and tens of thousands of CUDA cores. These are designed for workstations where high-performance AI and 3D workflows need to run locally.
• Jensen Huang also shared a five-year roadmap with the audience, outlining the planned rollout of NVIDIA’s next-generation AI architectures. As part of this vision, the company introduced a new platform named Vera Rubin, which signals the beginning of the post-Blackwell era. According to NVIDIA, Vera Rubin is expected to deliver unprecedented levels of performance and bandwidth—paving the way for the next phase of AI scaling.
• Lastly, new solutions were announced for enterprise environments: NVIDIA’s AI Data Platform and NIM microservices now allow AI models to run directly on corporate data sources, within secure infrastructure, including support for Oracle Cloud.

After the keynote, the crowd moved on to the exhibition area, where the technological deep dive continued. One of the central elements of the event was the show floor. The first large expo hall—home to NVIDIA’s main booth—brought together a wide variety of major hardware manufacturers, data center providers, and companies across multiple industries: automotive, healthcare, architecture, design, industrial technology, and large-scale services. The common theme was showcasing the infrastructure and technologies required to power modern AI workloads. There was also strong interest in robotics and automated manufacturing solutions.

The expo floor itself was buzzing and densely packed. NVIDIA’s booth had two main sections: a hardware showcase where visitors could get up close with everything from the smallest to the largest NVIDIA devices—resulting in long queues to enter—and a software-focused demo area, where technology experts explained the capabilities of each solution and answered questions. There were also multiple robotics demos featuring both industrial and home-use robots.

For me, the most exciting highlight was the compact DGX Spark. The idea that something which used to require an expensive cloud or server environment can now fit in the palm of your hand is a game-changing innovation. It fundamentally reshapes what’s possible. This device makes it accessible for a broader audience to work with and develop AI in a completely private and secure way. These machines are capable of running language models with up to 200 billion parameters—and by pairing two devices, even 405B models can run smoothly. What truly makes it revolutionary is that every DGX Spark runs the same DGX OS and full software stack as NVIDIA’s largest systems. That means anything developed on Spark can be transferred directly to full-scale AI factory infrastructure without any modification. I can’t wait to see what innovations developers come up with once these become widely available.

Another interesting detail was how NVIDIA hardware appeared at many other booths as well—encouraging visitors to engage more deeply with partner displays. Hardware was everywhere. Major manufacturers brought plenty of equipment, which attendees were eager to photograph and explore. I was especially intrigued by the growing presence of liquid cooling technologies, with thick fluid-carrying tubes now clearly part of the standard setup for cooling energy-hungry AI servers.

I also really enjoyed the many robotics demos. From robotic arms and robot dogs to full humanoid robots, it was thrilling to see the future up close. It was also exciting to see how AI is being integrated into the daily workflows of industrial companies and design studios alike.

A smaller, but densely packed exhibition hall hosted a large number of AI startups, many of them part of the NVIDIA Inception program. Key themes here included AI security, RAG, agentic AI, and AI avatars. On the other side of the hall, software and OS vendors showcased their own AI-augmented platforms.

The expo also included a wide range of talks and presentations—from deep technical dives to lighter educational sessions. Partners, vendors, satisfied customers, and NVIDIA engineers all shared fascinating insights. Because there was so much happening at once, I had to accept that I’d be catching up on many of the sessions via recording later. Of course, official NVIDIA trainings and even onsite certification exams were also available.

Talks started as early as 8 AM and often ran past 6 PM—though it wasn’t unusual to find something happening afterward too. For those needing a break, the nearby park offered a relaxing spot with live concerts and food trucks.

One smaller, dedicated area focused on VR and 3D technologies. Though it was surprisingly small compared to other sections, it was still great to see ongoing innovation in that space.

Networking was another key part of the event. I was fortunate to connect with interesting professionals from both Europe and the U.S. It was fascinating to realize that, regardless of geography, everyone is wrestling with similar challenges when it comes to AI. Most companies are still experimenting, trying to get their first real AI use-case into production. I was especially happy to hear a Hungarian accent—that’s how I met Csabi, a Principal Machine Learning Engineer who’s been living in the U.S. for over a decade. Check out his blog (https://devquasar.com/) and Hugging Face (https://huggingface.co/DevQuasar) profile, where he regularly publishes deep-dive posts on AI agents, edge-based inference, fine-tuning, and open-source model development.

One of the most memorable moments was when Jensen Huang himself walked through the expo floor, signing servers, posing for selfies, and even autographing attendee badges and freshly purchased Jetson Nano dev kits. It was heartwarming to see how approachable and patient he was with everyone—and although it wasn’t easy, I even managed to get my badge signed.

All in all, it was an inspiring, content-rich experience. I came home filled with knowledge and stories. If you ever get the chance, I wholeheartedly recommend attending any GTC event—and letting yourself be swept up in the magic of this world.

It’s worth watching our video from the event as well, where we showcase the atmosphere of GTC up close.