I attended Build a Claw Tokyo, where Jensen Huang, CEO, also made a surprise appearance.
This page has been translated by machine translation. View original
Introduction
Hello, I'm Morishige from Classmethod's Manufacturing Business Technology Division.
On July 15, 2026, I attended "Build a Claw Tokyo," a community event hosted by NVIDIA. The name "Claw" comes from OpenClaw, a trending open-source autonomous AI agent, and NemoClaw, NVIDIA's enterprise edition that packages it together with a secure runtime. The event's theme was "meeting people who are actually trying, learning, and building" these kinds of agents, and it consisted of three tech talks, a live demo, and networking.
The venue was Happo-en in Shirokanedai. The first thing that surprised me was the sheer number of people. There were reports of 45-minute waits just to get in. While each demo initially had a limit of about 5 people, the lines kept growing and the organizers eventually dropped the restrictions midway through to keep things moving. And later on, CEO Jensen Huang made a surprise appearance as a special guest — something you wouldn't expect at a weekday evening community event.
In this article, I'll summarize the key points from the three tech talks, the moment CEO Jensen Huang took the stage, and the SOP Monitoring Blueprint exhibit that caught my attention the most — all as a participation report from the event.
Event Overview
Here are the basic details of the event:
- Date: Afternoon of July 15, 2026
- Venue: Happo-en (Shirokanedai, Tokyo)
- Organizer: NVIDIA
- Format: 3 tech talks (15:00–16:00) + Live demo (14:00–18:00) + Networking
The tech talks were a three-session lineup, each 20 minutes long.
| Session | Speaker | Main Topics |
|---|---|---|
| The Age of AI Agents Begins | Kenji Tanaka, NVIDIA | Nemotron 3, NVIDIA Agent Toolkit, OpenShell, NeMo Switchyard |
| At the Frontier of Physical AI Development | Ken Arai, NVIDIA | Data Factory, Cosmos 3, OSMO, Isaac, FOX |
| From Demo to Business System | Yo Sato, Fixstars | Enterprise deployment of agents, designing autonomy and reliability, leveraging NemoClaw |
After the sessions, there was time for demo exhibits and networking, followed by a surprise guest appearance and a raffle around 5 PM. Let's go through each in order.
"The Age of AI Agents Begins" — An Overview of Nemotron and Agent Toolkit
The first session was by Kenji Tanaka, Senior Developer Relations Manager for Generative AI at NVIDIA Japan. He began by looking back at the arc from ChatGPT in 2022, to the open-sourcing of reasoning with DeepSeek in 2025, to the emergence of long-running, self-evolving agents represented by OpenClaw in 2026 — framing the current moment as "the point where LLMs and harnesses have started turning together as two wheels." The harness layer repeatedly handles LLM inputs and outputs, inserts tools in between, observes results, and lets the model trial-and-error its way toward a goal over extended periods. The positioning was that we've entered an era where you can build open agents on your own.
From there, he introduced the NVIDIA Agent Toolkit for Enterprise — a reference architecture for building business-specific agents. It brings together NemoClaw as the agent runtime, OpenShell as the secure runtime beneath it, a suite of skills including AI-Q and cuOpt, and the open model Nemotron as the brain, all available openly. The demo showed the flow from a one-command installation to the agent booting up in a CLI-interactive style like Claude Code, with Nemotron 3 Ultra running in the background. What stood out was using the /goal command to set an objective and then letting it run autonomously for an extended period — with examples like implementing Andrej Karpathy's nanochat and automatically researching model optimization.
Nemotron 3, the open model serving as the agent's brain, comes in three sizes — Ultra (550B-A55B), Super (120B-A12B), and Nano (30B-A3B) — plus vision-based Omni, speech, and safety variants, with the flagship being Nemotron 3 Ultra announced in June. It was described as strong in agent productivity, instruction-following, and tasks requiring deep reasoning, while being honest that it falls short in some coding scenarios. Drawing on deep knowledge of GPU efficiency, the claim was 5x faster inference and up to 30% lower cost per task completion compared to open frontier-class models like Kimi K2.6 (1T), GLM-5.1 (744B), and Qwen3.5 (397B).
On the security side, OpenShell was introduced as a runtime that sandboxes the agent and controls what enters and exits via a gateway whitelist. The context was that even the smartest models can be subjected to prompt injection or jailbreaking, so enterprises using agents need a governance mechanism. Worth noting: it's not exclusive to NVIDIA agents — it can be used alongside others like Codex.
NeMo Switchyard was also touched on in the context of cost optimization — the idea of having an expensive frontier model handle only the planning, while delegating execution to a local Nemotron. It was introduced with the aside, "Claude Code has gotten pretty expensive lately, hasn't it?" I previously wrote a first-touch article on Switchyard, so if you're interested, check it out (as of July 3, 2026):
Overview of NVIDIA Agent Toolkit. Within the NemoClaw frame, harnesses like LangChain and OpenClaw are shown as interchangeable; OpenShell forms the foundation, with a suite of skills and models including Nemotron lined up on the right.
"At the Frontier of Physical AI Development" — Cosmos and Isaac Take Center Stage
The second session was on physical AI by Ken Arai, Senior Manager of Robotics Developer Relations at NVIDIA Japan. He began by framing VLA (Vision Language Action) — which takes vision and language as inputs to generate actions — as the core concept, expanding into autonomous driving, infrastructure, robotics, and healthcare. He outlined the direction of "aiming for the best of both worlds: the specialist-type automation Japan has excelled at, and the generalist-type flexibility of VLA."
The first challenge he raised was data. Real-world data is scarce and costly, and even supplementing it with simulation doesn't come close to the web-scale volume of data that LLMs have benefited from — physical AI simply doesn't have that. NVIDIA's answer to this is an approach centered on the motto "Compute is Data" — generating and augmenting data through computation — and the physical AI data factory that serves as that engine. OSMO, a workflow orchestration tool, and Cosmos, a world foundation model, were introduced as the core of this effort.
OSMO is an open-source framework for building, testing, and validating physical AI systems, allowing workflows that span different compute environments — simulation, training, and edge — to be described and executed in a single YAML file. There was also mention of handing off building and operations to coding agents like Claude Code or Codex, underscoring how agentic AI and physical AI are part of the same continuum. On Cosmos, the focus was on Cosmos 3, announced at GTC Taipei in June — described as an omni model with a Mixture-of-Transformers architecture combining autoregressive and diffusion models, capable of freely mixing text, images, video, audio, and actions as inputs and outputs. The vision is for a single model family to cover video reasoning, background replacement, future frame prediction, and robot policy learning. The example of generating rare dangerous scenarios like a child running into the road as training data for autonomous driving was a clear illustration of where world models shine.
For manufacturing, the NVIDIA Factory Operations Blueprint (FOX) was introduced — a reference blueprint described as giving factories an AI brain. Whereas factory floors have traditionally built up precise rule-based design and local optimization cell by cell, this approach has agents discover and integrate systems across the factory and take on cell management. Skills available include video analytics, defective product image generation, and digital twin construction from CAD data (CAD-to-SimReady). Compute resources were laid out in three tiers — DGX Spark for small factories, DGX Station as the core, and GB300 NVL72 for large-scale multi-plant deployments — but the addition of "it's fine to start with whatever environment you have on hand" struck me as genuinely honest.
The robotics development side was covered at a brisk pace. Isaac Sim 6.0 is now GA, with MCP server integration and a revamped physics engine. What personally caught my attention was the Newton physics engine available in Isaac Lab 3.0 (still in beta), co-developed with Google DeepMind and Disney Research, which can simulate particle and soft-body behavior. Also mentioned were Isaac SIL for testing in software before real-world deployment, and Isaac GR00T, a humanoid foundation that uses Cosmos Reason for rational judgment — Cosmos appearing here again. There was also a note that physical AI agent skills are growing on build.nvidia.com's Agent Skills section, which I plan to explore later.
The data pyramid slide. Real-world data at the apex is scarce and costly (24 Hrs / Robot / Day); simulation in the middle still has the Sim2Real problem; and the bottom tier of web data raises the question of whether it can be used as-is.
Fixstars on "From Demo to Business System"
The third session was by Yo Sato of Fixstars Corporation, an NVIDIA partner, titled "From Demo to Business System — Enterprise AI Agent Design in the NVIDIA NemoClaw Era." The key message was delivered upfront: the key to evolving AI agents from demos into business systems lies not in the agent itself, but in the system design that supports its autonomy. The core challenge of agents has shifted from intelligence to reliability, and the smartness of inferring 10 things from 1 translates directly into risk in a business system. A demo only needs to work once, but a business system needs to keep working — and that requires permissions, policies, fault handling, and auditability.
On the practical side, he used a co-developed ERM (Enterprise Risk Management) platform as the subject, presenting a combination of three elements: agents for interpreting ambiguous requests, workflows for routine processing, and humans for approving critical operations. Building on that, he noted that approvals alone can't prevent accidents like API key leaks, which is why they adopted NemoClaw and are actually using OpenShell's sandbox isolation and credential shielding. The framing of NemoClaw not as the agent itself but as the foundation for letting agents exercise autonomy safely resonated clearly with me. There were also plenty of design insights to take home: guardrails can only protect against anticipated scenarios and reducing the unanticipated is the designer's job; the risk of leaking information that entered a prompt can never be zero, so the decision is whether that's a risk you can accept.
Slide on enterprise design principles. Four principles — least privilege, critical operation control, execution tracking, and recovery from anomalies — are organized alongside the mechanisms that implement them and their purposes.
CEO Jensen Huang's Surprise Appearance
Midway through the demo time after the sessions, an announcement came: "Our special guest is currently on their way from Akihabara." Who it was was never revealed — until around 5 PM, when the real Jensen Huang CEO walked in.
Finally useful. After working on AI for 15 years, AI is now useful.
The speech that began with those words — "After 15 years of working on AI, AI is finally useful" — centered on "personal AI." Forty years ago, the personal computer revolution began; forty years later, everyone will be able to have their own personal AI in place of a personal computer. That's why he was happy to meet everyone who had come to build their own agents.
Then he picked up a DGX Spark. As a personal AI supercomputer carrying on the lineage of the original DGX-1 from 2016, he described it as "baby-sized but nearly equivalent performance — 128GB, 1 petaflops, NVFP4." He also shared the origin story of the first DGX-1: when he built it, there was no demand for AI and zero customers. The only "customer" was his friend Elon Musk, who despite being wealthy asked to have the first unit donated — and that donation went to OpenAI, which was a nonprofit at the time. As someone who verifies things daily on a DGX Spark of my own, hearing that origin story directly from him was genuinely moving.
CEO Jensen Huang talking about "personal AI" while holding an actual DGX Spark unit. In his trademark leather jacket, being able to see him at this distance was something only possible at a community event.
The speech closed with a raffle for a signed DGX Spark. Two winners were called out and received their boxes directly from Jensen. As for me — I didn't win, unfortunately. I at least got a video of it happening right in front of me, so I'll call that a win.
The SOP Monitoring Blueprint That Caught My Eye Most at the Exhibits
The demo exhibits featured nine setups, ranging from an introductory "Getting Started with Claw Agents," to browser operation by OpenClaw on a DGX Spark, live Isaac Sim control by a Claw agent, a Unitree G1 humanoid combined with Cosmos 3, and voice multimodal model personalization by SB Intuitions — a wide-ranging lineup. Among them, the exhibit I spent the most time at — as someone who handles manufacturing industry projects — was the SOP Monitoring Blueprint.
SOP (Standard Operating Procedure) refers to standardized work procedures used in manufacturing and other settings. In the demo, feeding a recorded video of a work task caused a VLM to analyze it, and the work steps appeared with timestamps along a video timeline in the UI. It then compared those against the correct procedure defined in the blueprint (9 steps in the demo) and gave a structured OK/NG judgment for things like missing parts or incorrect step order. When I asked at the booth, they said that even if you feed it a video of a different worker performing the task, as long as they follow the correct procedure, steps 1, 2, and 3 are properly recognized. The demo used pre-recorded video analysis; real-time streaming analysis depends on available computing power. What surprised me was that both the UI and the workflow definition are published as open source.
Looking at the repository after returning home, it's not a one-off demo app — it's structured to cover everything from training to inference end-to-end. There are two pillars: a training microservice that prepares a SOP-specific VLM and a temporal segmentation model, and a DeepStream-based inference microservice for low-latency inference — with a suite of agentic skills handling data augmentation, fine-tuning of Cosmos Reason-based VLMs, and evaluation. The overall workflow looks like this:
The flow involves annotating your own work videos with action start/end timestamps, generating QA pairs from those, and fine-tuning the VLM — so rather than relying on prompts from a general-purpose VLM, you can build a model specialized for your specific on-site procedures. NVIDIA's model and container distribution catalog NGC has a sample dataset based on server fan and power supply installation work, so you can try it out even without your own video footage. The code is licensed under Apache 2.0. Inquiries about verifying "whether work is being performed according to prescribed procedures" using factory camera footage are genuinely common, and having this workflow available openly feels quite practical.
The SOP UI in the demo. A video of fan installation on a server is segmented by step in color-coded segments along the timeline, with an SOP OK judgment displayed in the upper right.
A DGX Station Unit Was Also on Display
One other thing I personally wanted to see was an actual DGX Station unit. I have a DGX Spark on hand, but this was my first time seeing a Station in person. There's something convincing about having the very class of machine mentioned in Session 2 as the execution environment for FOX right there at the venue. (I forgot to take a photo of it...)
Summary
The three tech talks offered a sweeping overview spanning "the state of agents today," "physical AI," and "design for making it a business system," and the event as a whole — including the exhibits and networking — was remarkably dense. On a personal note, it was a pleasant moment when Tanaka-san mentioned during the tech talk closing that a talk session I appeared in for NVIDIA Japan's YouTube channel about OpenShell had been published. Above all, witnessing CEO Jensen Huang himself show up at a community event to talk about personal AI was well worth attending.
