I picked up 7 interesting sessions from GTC 2026 on-demand sessions

Hello, I'm Morishige from Classmethod's Manufacturing Business Technology Department.

I attended NVIDIA GTC 2026 in person, held in San Jose in March 2026. I missed quite a few sessions while being caught up with exhibits and the hackathon, but thankfully over 700 sessions are available on-demand.

I caught up on a few before returning home, so I'll introduce the 7 sessions I watched, divided into 3 themes: "Nemotron and Open Models," "Edge AI and Inference Optimization," and "AI Factory and Digital Twin." I've summarized the key points of each session along with my impressions from the perspective of someone who has hands-on experience with DGX Spark.

Please note that the content of this article is based on what I personally heard while watching the on-demand videos. There may be mishearings or notational errors, so please refer to each session's video or official materials for accurate information.

Nemotron and Open Models

Nemotron Unpacked (S81719)

This is a session where Bryan Catanzaro, NVIDIA's VP of Applied Deep Learning Research, discusses the full picture of the Nemotron project.

Nemotron is not just a model, but an open ecosystem that includes models, datasets, and training methods. It is structured in 3 tiers: Nano (30B/3B active), Super (120B/12B active), and Ultra (~500B), with Super having just been released recently. It achieved a score of 36 on Artificial Analysis's Intelligence Index, surpassing GPT OSS 120B's score of 33.

What particularly struck me was the phrase "Acceleration is intelligence." Making a model faster means you can process more tokens in pre-training, run more rounds in reinforcement learning, and use more thinking cycles in inference. Speed directly translates to intelligence — a perspective unique to a hardware manufacturer.

On the technical side, the Mamba2 transformer hybrid architecture is interesting. It is largely composed of linear-time attention layers, with quadratic attention layers kept to a minimum. The expression "10x less quadratic" was used. I think this design was inevitable in achieving a context length of 1 million tokens.

Pre-training with NVFP4 (4.75 bits) is also noteworthy. As far as publicly available information goes, NVIDIA is reportedly the first to conduct 4-bit pre-training at this scale. It's a forward investment with an eye toward enhanced FP4 throughput on Blackwell Ultra, and I'm curious to see how this benefit will manifest on the Blackwell GPU in DGX Spark.

The Nemotron Coalition was also announced. It's a framework for advancing model development in collaboration with multiple companies including Mistral, Perplexity, Cursor, and Black Forest Labs. The first project is pre-training with Mistral, and for post-training, each company will participate according to their own needs. The statement that Nemotron is positioned as a long-term endeavor, just like CUDA's 20-year investment, felt very characteristic of NVIDIA.

Development of the Nemotron family has continued after the session. Around GTC, Nemotron-Cascade 2 was released. While it has the same parameter configuration as Nano at 30B MoE / 3B active, it significantly enhances reasoning performance through Cascade RL and multi-domain on-policy distillation. It has achieved gold medal-level scores at IMO 2025 and IOI 2025, delivering performance that surpasses Qwen3.5-35B-A3B with just 3B active parameters. The direction of achieving "intelligence density" rivaling frontier models with 20x fewer parameters is a concrete example of what Bryan Catanzaro was saying about "faster models are smarter models."

For the edge, an even smaller Nemotron 3 Nano 4B has also been released, configured for lightweight deployment.

Building Domain-Expert Agents (S81707)

This is a practical fine-tuning session on Nemotron 3, by Aible and NVIDIA.

The concept that repeatedly comes up in this session is the "data flywheel." For complex agentic tasks where prompt engineering alone hits a ceiling in accuracy — such as tool-calling and structured output — the session concretely explains how to cycle through fine-tuning using the NeMo toolkit.

An interesting statistic is that currently only 12% of companies are fine-tuning models. Even so, it was emphasized that using NeMo Data Designer, you can generate large amounts of synthetic data from a small number of high-quality examples, and the cost of fine-tuning comes down to "a few dollars."

In the Aible case study, fine-tuning Txt2SQL significantly improved accuracy. In the Yum! Brands case study, a "hybrid agent" approach was introduced where they initially operated with a large reasoning model, then used its output data to fine-tune a smaller model and gradually transitioned to it. I found this to be quite a realistic configuration with the final cost optimization in mind.

The NeMo toolkit covered included the Agent Toolkit for capturing execution traces, Data Designer for generating synthetic data, Customizer for fine-tuning with LoRA, Evaluator for model evaluation, and Guardrails for ensuring safety. Having personally fine-tuned Nemotron Nano on DGX Spark, it's reassuring to see this full-stack pipeline being built out.

Edge AI and Inference Optimization

Optimize Performance of Vision AI Models on the Edge (S81833)

This is a session by NVIDIA's Louise Huang on VLM optimization for edge devices.

Cosmos Reason 2 is an open VLM with over 2 million downloads on Hugging Face, specialized for physical AI use cases. It comes in model sizes of 2B and 8B, supporting Chain of Thought reasoning and an input token length of 256K.

What I found practically useful was inference optimization with NVFP4. With 4-bit quantization, latency and throughput improve by approximately 20%, and memory usage is also reduced. Since memory has always been a concern when running VLMs on DGX Spark, this optimization looks directly applicable.

EVS (Efficient Video Sampling) is also an interesting technology. By dynamically skipping redundant frames from video — for example, scenes with little movement — it improves latency while maintaining accuracy. It's a practical answer to the question that often comes up in video analysis: "Do we really need to process every frame?"

Regarding fine-tuning, it was demonstrated that PEFT using LoRA can achieve accuracy equal to or better than full fine-tuning. Furthermore, automating hyperparameter search with AutoML can reportedly reduce a task that used to take a human expert 45 hours to under 10 hours.

Something I quietly appreciated was that all benchmarks in the session were measured on Jetson Thor. The stance of highlighting practical utility in an edge environment rather than a data center GPU felt like a reassuring message for DGX Spark and Jetson users.

AI Factory and Digital Twin

The Builder's Toolkit: Scaling Enterprise AI Factories (S81851)

This is a session explaining the overall picture of AI factories, presented by two leaders from NVIDIA's Enterprise AI Factory division.

What was interesting about this session was the "F1 car vs. SUV" analogy. Some companies need cutting-edge, hyper-scale AI factories, but what most companies need is a more accessible configuration. NVIDIA is helping OEM partners bring products to market quickly through reference architectures and certified systems.

Concrete examples of business outcomes were also given, such as a case where fine-tuning an open-source model with proprietary data reduced the fraud rate from 1% to 0.18%, and a case where Agentic AI improved productivity tenfold.

There was also discussion about how the key to AI factory deployment is shortening Time to Value — the time to realize value — to around 45 days. The point that cloud, NCP (Network Computing Provider), and on-premises options are "and" rather than "or" also resonated with me from a practical standpoint.

The Velocity of Volts: Hitachi Energy's AI Factory Power Solutions (EX82366)

This is a session on power supply solutions for AI factories, developed through collaboration between Hitachi Energy, Hitachi Digital, and NVIDIA. It's a Theater Talk, so it's only 15 minutes, but the content is substantial.

The catchphrase "Velocity of Volts" sticks in my mind. The approach involves dramatically shortening the planning process — which takes months — for connecting a 1 GW-class AI factory to the power grid, using NVIDIA's solver technology. The configuration combining NVIDIA's three platforms — DGX (training), OVX/RTX (simulation), and IGX (real-time asset management) — to manage the entire lifecycle of an AI factory is a world a bit too large-scale for me as a DGX Spark user, but it was refreshing to see the power challenges of AI infrastructure translated into such concrete solutions.

Improving Data Quality with NeMo Curator and Cosmos (S81964)

This is a session by Mr. Takami of APTO. It introduced practical case studies using NeMo Curator's filtering capabilities and Nemotron-Personas-Japan.

The specific figure of a 83-hour reduction in man-hours for creating LLM synthetic data is also compelling, but what I personally found most interesting was the safety improvement case study. Using Nemotron-Personas-Japan, which reflects Japanese culture and demographics, questions were generated from the perspective of elderly people, and with just 504 data points, the Attack Success Rate (ASR) was brought down to 0%. Only 0.15% of all parameters were updated through LoRA tuning, and training time was about 20 minutes on 2 A100s. If this kind of effect can be achieved with such ease, strengthening safety using proprietary data is a very realistic option.

Curation of imitation learning data for robots using Cosmos Curator was also introduced. The approach of detecting anomalous motion frames using Mahalanobis distance is solid and practical.

Urban CFD Simulation and Digital Twin (S81972)

This is a session on large-scale urban simulation using Omniverse and Ansys Fluent, leveraging "Seiran," the supercomputer at Tokyo University of Technology.

Seiran is a cluster configured with 12 nodes of DGX B200, totaling 96 Blackwell GPUs, ranking 374th on the TOP500 and 22nd on the AI-specialized HPLMXP. The fact that it scores 17 times higher on HPLMXP than on the standard benchmark succinctly demonstrates the high AI computing performance of Blackwell.

By using the Ansys Fluent GPU solver, they achieved 50–60x acceleration compared to CPU, and urban-scale fluid analysis that used to take days to weeks can now be completed in a few hours. The footage of wind environment analysis covering tens of thousands of buildings in Chuo Ward using the Ministry of Land, Infrastructure, Transport and Tourism's PlaTO data was quite impressive.

There was also an introduction to NVIDIA's Student Ambassador Program. With NVIDIA engineers serving as mentors, students can learn the latest technologies in a hands-on setting. It was noted that with the maturation of CAE software, even students with limited expertise can now run high-quality simulations, and I found myself envying students who get to learn in such an environment.

Sessions Not Covered This Time

GTC 2026 also featured many announcements related to Physical AI, and in the robotics field, Isaac Lab 3.0 (Early Access) and the new GPU physics engine Newton 1.0 were announced. Newton is an open-source physics engine co-developed with Google DeepMind and Disney Research, reportedly achieving up to hundreds of times faster performance compared to MuJoCo (MJX). It is integrated as the physics backend for Isaac Lab 3.0, significantly accelerating reinforcement learning for deformable body simulation and contact-rich manipulation tasks. The related session (DLIT81700) was in a hands-on lab format, so I have not been able to confirm on-demand viewing availability at this time. I'd like to check back once the video is released.

Also, the case study by Hitachi Building Systems on improving workplace safety using VLM and VSS (S81896) is another session I'm curious about, but the video had not yet been published either.

Summary

Having selected 7 sessions based on my own interests, the keywords that emerged were "customization of open models," "optimization of edge inference," and "practical deployment of AI factories."

The fact that Nemotron is designed not merely as a model release but as a platform supporting an entire ecosystem. The fact that VLM optimization at the edge has entered the practical stage with concrete technologies like NVFP4 and EVS. The fact that AI factories are being offered as comprehensive solutions encompassing even power infrastructure. Steady progress was visible at each layer.

GTC on-demand sessions can be viewed with a free Virtual Pass. There are over 700 sessions, so there's no way to watch them all, but if any sessions in this article caught your interest, please check them out.

Reference Links

• NVIDIA GTC On-Demand

The sessions introduced in this article can be viewed by searching by session ID (e.g., S81719) on the GTC On-Demand site.