NVIDIA GTC Berlin 2026, Explained for Small Business Owners

Key Takeaways

• NVIDIA's GTC Berlin 2026 keynote ran over 90 minutes. Most of it was data center hardware. Three ideas in it apply to a ten person business.
• Jensen Huang broke an AI agent into five parts: the model, the harness, the sandbox, knowledge and tools. Every agent you buy or build has all five, whether anyone names them or not.
• Telling an agent "please don't delete anything" is not a permission system. NVIDIA's answer, OpenShell, enforces what each agent can touch from outside the agent.
• An agent gets useful when it can read your company's own knowledge and use real tools. The model alone is not the difference.
• Robots that learn by being shown, and a laptop with the compute NVIDIA handed OpenAI in 2016, are coming. Neither is something you need to buy this quarter.

Every week there is another AI keynote, and every one of them is an hour and a half of chips, racks and partner logos. You run a business. You do not have 90 minutes to find out whether any of it changes how you answer the phone on Monday.

So we cut NVIDIA's GTC Berlin 2026 keynote down to 20 minutes and put it on the On Point Tech YouTube channel. This post is the short version of the short version: what was announced, and the three parts worth your attention if you are a Go High Level Bay Area business already using AI in your inbox, on your phones or in your follow up.

What NVIDIA announced, in one pass

Here is the list, kept to what was said on stage:

• Vera Rubin, the new data center computer, is in full production. It is 100 percent liquid cooled, has no cables inside, and NVIDIA says a node now takes under five minutes to assemble instead of two hours.
• DSX Pod ships an AI factory built, cabled and tested at the factory, in containers. NVIDIA's estimate is on-site deployment going from several months to days once the site is ready.
• OpenShell and Sentry, security for AI agents. More on this below, because it is the part that matters most.
• Open models plus the Agent Knowledge Foundation, a set of libraries for turning a company's documents and databases into something an agent can use.
• Hyperion, a robotaxi platform, now joined by VW, JLR, Volvo and Mercedes.
• Agile ONE, a reference humanoid robot built with Agile Robotics, due in the first quarter of next year.
• RTX Spark, a new line of desktops, laptops and workstations built with Microsoft over four years.
• A live demo of Micro Duck, Hugging Face's little robot, trained in simulation and then asked to do a somersault on stage.

The five parts of every AI agent

The most useful minute of the keynote was a list. Huang said an agent has five elements: the model, the harness, the sandbox the harness sits in, knowledge and tools.

Put that in small business terms and it maps onto the AI you may already run. Say you have an AI answering your website chat or your missed calls.

• The model is the AI itself, the thing that reads and writes.
• The harness is the setup around it: its instructions, its goals, when it hands off to a person.
• The sandbox is what it is allowed to touch. Can it book on your calendar? Change a contact? Send a text?
• Knowledge is what it knows about you: services, service area, hours, prices, policies.
• Tools are the things it can actually do, like booking the appointment or tagging the lead.

When an AI agent disappoints a business owner, it is almost never the model. It is a thin knowledge base, a vague harness or a sandbox nobody thought about. That is the useful part of this framing. It tells you where to look.

Asking nicely is not a permission system

NVIDIA's OpenShell segment opened on a line every owner who has tried an AI agent will recognize. Right now, they said, the only thing stopping most agents from misusing their access is a polite request in the prompt not to delete files or touch the keys. Then the line that sums it up: "hope isn't security."

OpenShell flips that. You write a policy for what each agent can touch, and the enforcement happens outside the agent, not inside its instructions. Each agent runs in its own isolated sandbox. Every outbound request goes through a supervisor that checks it against your policy and denies it if it is not allowed. The keys stay with the supervisor, so a compromised agent cannot leak a key it never had. Anthropic and NVIDIA also announced they are partnering on Claude OpenShell agents.

You are not going to install OpenShell at a dental office. The principle still applies to every AI you turn on:

1. Give each agent only the actions it needs. A chat agent that books appointments does not need to issue refunds.
2. Keep a person on anything public or expensive: pricing exceptions, review replies on low ratings, anything that goes out under your name.
3. Do not hand an agent logins it has no reason to hold.
4. Write the limits down in your SOP, not only in the agent's prompt.

Knowledge is what makes an agent worth having

Huang was direct about what turns an agent into a useful one: access to your company's knowledge. NVIDIA's Agent Knowledge Foundation is built to read unstructured documents like PDFs and structured data in databases, and pull meaning out of both. After that, he said, it still comes down to the tools.

At small business scale this is good news, because it is the cheap part. Your edge is not a better model than the business down the street. It is the price sheet, the service area, the "we don't do commercial jobs" rule and the answers to the twenty questions your front desk hears every week. Load that in and the same AI gets noticeably better.

Robots you show, and a petaflop on a laptop

Two segments are further out, but worth knowing about.

On humanoid robots, Huang named small and medium businesses directly. A business without a robotics team, he said, could show the robot what needs to be done, and a robot with general capabilities could learn and adapt quickly. The Agile ONE reference robot is the first step, due early next year. Treat this as something to watch, not something to budget for.

RTX Spark is closer. It is one petaflop with 128 GB of memory. Huang pointed out that in 2016 he delivered that same one petaflop and 128 GB to a small company called OpenAI. Now it comes in a laptop. The Micro Duck demo ran on one, using Hermes as the agent harness, NVIDIA's Nemotron and OpenAI's GPT-6 Astra as models, and FreeCAD, Blender and Isaac Sim as tools, bursting to the cloud only for the big training runs. Running serious AI on your own machine is getting real.

Why a Go High Level Bay Area business should care

None of this asks you to buy NVIDIA hardware. What it does is confirm where the work is. The agents in your CRM, your Conversation AI, your Voice AI, your follow up, all live or die on the same five parts NVIDIA put on a slide. The model keeps getting better on its own. The harness, the limits, the knowledge and the tools are the part somebody has to build and keep up.

That is the part we build. On Point Tech sets up AI agents inside a Go High Level account you own, with written limits on what each one can do, a knowledge base built from your real business, and a person in the loop on anything that goes out under your name.

Watch the 20 minute cut

Original keynote, footage and audio: NVIDIA. Watch the full keynote. Highlights edit: On Point Tech Solutions.

Final Thoughts

Big keynotes are mostly about other people's data centers. The piece that reaches your business is small and practical: an agent is five parts, the limits have to be enforced rather than requested, and your own knowledge is what makes it worth turning on.

On Point Tech Solutions is a Go High Level consultant in Los Gatos, serving Go High Level Bay Area businesses from San Jose up the Peninsula. We build AI and marketing automation that you own: the account, the workflows, the logins. We can build it for you, or teach you to run it yourself with an SOP.

Want to build this yourself? Join the On Point Tech Academy. It's free, and we build live every Tuesday and Thursday.