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 many sessions while being busy with exhibitions and hackathons, but thankfully over 700 sessions are available on demand.

I caught up on a few before returning to Japan, 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 thoughts as someone who has hands-on experience with the DGX Spark.

Please note that the content of this article is based on my own listening of the on-demand videos. There may be mishearings or notation errors, so please refer to each session's video and 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 three tiers: Nano (30B/3B active), Super (120B/12B active), and Ultra (~500B), with Super having just been released. It scored 36 points on the Artificial Analysis Intelligence Index, surpassing GPT OSS 120B's 33 points.

What particularly struck me was the phrase "Acceleration is intelligence." If you make a model faster, you can process more tokens in pre-training, run more rounds in reinforcement learning, and use more thinking cycles in inference. Speed directly leads 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 a necessity in achieving a context length of 1 million tokens.

Pre-training with NVFP4 (4.75 bits) is also noteworthy. Within the publicly available information, NVIDIA is reportedly the first to perform 4-bit pre-training at this scale. It's a proactive investment looking ahead to the enhanced FP4 throughput of Blackwell Ultra, and I'm curious to see how this benefit will manifest on the Blackwell GPU of the DGX Spark.

There was also an announcement of the Nemotron Coalition. 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 initiative, just like CUDA's 20 years of investment, felt very NVIDIA-like.

The rollout of the Nemotron family has continued after the session. Nemotron-Cascade 2 was released around the time of GTC. While it shares the same parameter configuration as Nano—a 30B MoE / 3B active model—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 surpassing Qwen3.5-35B-A3B with only 3B active parameters. The direction of "intelligence density"—matching frontier models with 20x fewer parameters—is a concrete example of exactly what Bryan Catanzaro was saying about "faster models are smarter models."

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

Building Domain-Expert Agents (S81707)

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

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

An interesting figure shared is that only 12% of companies are currently 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 can be as low as "a few dollars."

In Aible's case, fine-tuning for Txt2SQL significantly improved accuracy. The Yum! Brands case introduced a "hybrid agent" approach 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 cost optimization in mind from the start.

As part of the NeMo toolkit, the following were introduced: Agent Toolkit for capturing execution traces, Data Designer for generating synthetic data, Customizer for fine-tuning with LoRA etc., Evaluator for model evaluation, and Guardrails for ensuring safety. I've had experience fine-tuning Nemotron Nano on a DGX Spark, so it's reassuring to see this full-stack pipeline coming together.

Edge AI and Inference Optimization

Optimizing 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 2B and 8B model sizes and supports Chain of Thought reasoning and an input token length of 256K.

What I found practically useful was inference optimization with NVFP4. It was noted that 4-bit quantization improves latency and throughput by approximately 20% while also reducing memory usage. Memory has always been a concern when running VLMs on the DGX Spark, so this optimization seems directly useful.

EVS (Efficient Video Sampling) is also an interesting technology. By dynamically skipping redundant frames from video—such as scenes with little movement—it improves latency while maintaining accuracy. It's a practical answer to the question "do we really need to process every frame?" in video analysis applications.

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 reduce work that used to take a human expert 45 hours to under 10 hours.

What 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 an encouraging message for DGX Spark and Jetson users.

AI Factory and Digital Twin

Building an Enterprise AI Factory (S81851)

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

What was interesting in this session was the "F1 car vs. SUV" analogy. Some companies need a cutting-edge, ultra-large-scale AI factory, 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.

Specific examples of business outcomes were also cited, such as a case where fine-tuning an open-source model on proprietary data reduced fraud rates from 1% to 0.18%, and a case where Agentic AI improved productivity by 10x.

There was also a discussion about how the key to deploying an AI factory is reducing Time to Value—the time to realize value—to around 45 days. The point that cloud, NCP (network computing providers), and on-premises options are "and" rather than "or" was also something I could practically relate to.

Hitachi Energy's AI Factory Power Solution (EX82366)

This is a session on a power supply solution for AI factories developed by Hitachi Energy and Hitachi Digital in collaboration with NVIDIA. As a Theater Talk, it is short at 15 minutes, but the content is dense.

The catchphrase "Velocity of Volts" sticks in the mind. The approach is to dramatically shorten the planning process for connecting a 1 GW-scale AI factory to the power grid—from months—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 on a scale a bit beyond me as a DGX Spark user, but I found it refreshing to see power challenges for AI infrastructure addressed with such concrete solutions.

Improving Data Quality with NeMo Curator and Cosmos (S81964)

This is a session by Mr. Takahin from APTO. Practical use cases using NeMo Curator's filtering capabilities and Nemotron-Personas-Japan were introduced.

The figure of a 83-hour reduction in man-hours for LLM synthetic data creation is concrete and compelling, but what I personally found most interesting was the safety improvement case. 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 approximately 20 minutes on 2 A100s. If this ease of use can produce these results, strengthening safety with proprietary data is quite a realistic option.

Data curation for robot imitation learning data 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 the supercomputer "Seiran" 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 HPLMXP score being 17 times higher than the standard score succinctly demonstrates the high AI computing performance of Blackwell.

By using the Ansys Fluent GPU solver, a 50–60x speedup compared to CPU was achieved, enabling urban-scale fluid analysis that used to take days to weeks to complete in just 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, it's a framework that allows students to learn the latest technologies through hands-on practice. With the maturation of CAE software, even students with limited expertise can now run high-quality simulations, and I found myself envying the 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, achieving up to several hundred times faster speeds compared to MuJoCo (MJX). It is integrated as the physics backend for Isaac Lab 3.0 and significantly accelerates 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 at this time. I'd like to check back once the video is released.

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

Summary

Having selected 7 sessions based on my own interests, the three keywords that emerged were "open model customization," "edge inference optimization," and "AI factory productization."

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

GTC on-demand sessions can be viewed with a free Virtual Pass. With over 700 sessions, it's impossible to watch them all, but if any sessions in this article caught your interest, please do check them out.

Reference Links

• NVIDIA GTC On-Demand

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