NVIDIA GTC Berlin 2026 Drops 4 New AI Hardware Gamechangers (H2-X ROBOT)

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: NVIDIA GTC Berlin 2026 Drops 4 New AI Hardware Gamechangers (H2-X ROBOT) · Channel: AI News · 07:41 · Watch on YouTube

Topics: AI agents, AI infrastructure and hardware, Local and open-source models, Robotics and autonomous systems, AI safety

Summary

The video covers Unitree’s H2X humanoid robot and four NVIDIA GTC 2026 announcements: the Agile One reference robot, RTX Spark laptops for local AI-agent workflows, Vera Rubin DSX AI-factory pods, and the Vera Rubin architecture. It describes the H2X as an apparent H2-based design with a claimed eight-hour runtime, though Unitree has not released its specifications, price, or release date. NVIDIA presents its hardware as supporting AI agents from development on laptops to large-scale deployment in AI factories. The video also highlights OpenShell and Century as security measures for isolating and monitoring agents.

Key points

Unitree has shown the H2X humanoid, but has not published a specification sheet, price, or release date; security and emergency response are cited as possible use cases.
The H2X appears to build on the H2, which the video describes as about six feet tall, weighing around 154 pounds, and having 31 degrees of freedom.
The H2X is said to run for about eight hours, compared with about three hours for the H2.
Agile One is a humanoid reference design from Germany’s Agile Robots, with availability announced for the first quarter of 2027.
Agile One combines Jetson Thor with NVIDIA Isaac Groot software; NVIDIA says the setup time can be reduced from weeks to days.
RTX Spark is a line of Windows laptops with Blackwell RTX graphics and up to 128 GB of unified memory, intended for building and running AI agents locally.
The Vera Rubin DSX pod packages seven liquid-cooled Vera Rubin NVL72 racks in transportable containers; NVIDIA says deployment could take as little as 14 days once site utilities are ready.
NVIDIA says Vera Rubin is in full production and describes an NVL72 rack with 72 Rubin GPUs and 36 Vera CPUs.
OpenShell provides an isolated sandbox for each agent, while the Century monitor is described as able to quarantine a rogue agent in milliseconds.

Chapters

Start	Chapter	Summary
00:00	Unitree’s H2X humanoid	The video introduces the H2X and outlines the H2 specifications it may build on, including its claimed endurance and potential applications.
02:15	Agile One reference robot	NVIDIA’s Agile One reference design pairs a humanoid body with Jetson Thor and Isaac Groot for robotics development.
03:50	RTX Spark for local AI agents	The video presents RTX Spark laptops and demonstrates a workflow for teaching a small robot to see and walk.
04:53	Vera Rubin DSX pod and architecture	NVIDIA’s AI-factory pod, its deployment timeline, and the Vera Rubin rack architecture are described.
07:07	Agent workloads and security	The video describes Vera Rubin’s workloads and NVIDIA’s OpenShell and Century tools for isolating and monitoring AI agents.

Quotes

Is this the beginning of the embodied AI explosion? — 00:00

Unitry says the H2X can run for about 8 hours. — 01:04

It’s called RTX Spark, and the pitch is to build, train, and run AI agents and their tools locally on a laptop instead of only in the cloud. — 03:50

on-site deployment goes from several months to days once the site utilities are ready — 05:26

each agent runs inside its own isolated sandbox. — 07:41

Concepts and entities

Unitree, H2X, Agile Robots, Agile One, Jetson Thor, RTX Spark, OpenShell, Vera Rubin DSX pod, Vera Rubin, Isaac GR00T

Full transcript (original language)

00:00 The next major industry, the next great industry, and this one is likely to be extremely large, is human or robotics. Is this the beginning of the embodied AI explosion? Today on AI news, Nvidia just unveiled four new breakthroughs in AI hardware, plus one autonomous AI gamecher. But first, Nvidia’s robotics partner, Unitree, just dropped a first look of its newest H2X humanoid, which it calls the kung fu legend, Guardian, because it’s built to guard, rescue, and

00:31 take an incredible beating along the way. So, the real question is, what’s inside? And Unitry hasn’t published an H2X spec sheet, but the H2X looks to be built on the H2. So, here’s what we do know about that body. The H2 stands at about 6 feet tall and weighs around 154 lbs with 31° of freedom, six in each leg, seven in each arm, and three in the waist and then two in the head. And its peak torque is 360 new m in the legs and 120 new m in the arms. And Unitree lists

01:04 an arm payload of about 15 lb, peaking at about 33 lb. And the robot’s frame is made of aircraft grade aluminum, titanium alloy, and high strength plastics. And reports say it works from minus20 to 60° C. But a body like that only matters if it lasts. Which brings us to the battery. Because while the H2 runs just 3 hours with 15 A of power, Unitry says the H2X can run for about 8 hours. But beyond raw physical power and endurance, the H2X will likely feature a

01:40 brain to match, as the H2 already features an Intel Core i5, binocular cameras, a microphone array, Wi-Fi 6, and Bluetooth 5.2. Plus, the EDU version adds a Core i7 for developers. And the top option is Nvidia’s Jetson AGX Thor with 2,70 terra operations per second, which for the top level spec costs $29,900. But the H2X has no price yet or release date with Unitree pointing to security and emergency response as use cases. So,

02:15 tell us in the comments how much you’d be willing to pay for the H2X, or would you rather this separate NVIDIA robotics partner, as Nvidia just revealed four brand new pieces of AI hardware at GTC 2026 this week in Berlin, starting with another humanoid, but with a twist. Them smallmedium businesses can now show the robot what needs to be done. The robot because it has general capabilities and general intelligence would be able to learn and adapt

02:45 quickly. That twist is that this humanoid is actually a reference design. It’s the Agile One built by Germany’s Agile Robots. And Jensen Hong says it will be available in the first quarter of 2027 pre-assembled with the computer inside and the software integrated. And the idea is to hand researchers a starting point instead of a pile of parts. So, first let’s start with the body. The version called Agile 1 SN stands at over 4.5 feet tall and weighs just under 95 lbs and has 31° of freedom with a

03:20 payload of over 6.5 lbs per arm and five fingered hands with tactile sensing. Next, there’s the brain which is Jetson Thor running Nvidia’s Isaac Groot software. And Nvidia says that cut setup time from weeks to days. But tell us down below how much you think this kind of robotics development kit should cost for consumers and would you use it to build? Or would you save your money instead for Nvidia’s second GTC 2026 reveal, which is a brand new genre of

03:50 personal computer that’s built for the agent era. It’s called RTX Spark, and the pitch is to build, train, and run AI agents and their tools locally on a laptop instead of only in the cloud. And to show what that means, Nvidia used a robot. First, the hardware. RTX Spark is a line of slim Windows laptops with Blackwell RTX graphics and up to 128 GB of unified memory. And they also say that many have shipped with unlocked bootloadaders so developers can

04:20 customize local agents. And it pairs them with OpenShell, which is a new runtime that we’ll get into in just a moment. But next is the demo because Jensen Huang introduced Micro Duck which is Hugging Face’s small bipeedal robot and the goal was to teach it to see and walk on RTX Spark. An agent calls Hermas and Nvidia’s Neatron model plans the course and lists every object needed. Then OpenAI’s GPT6 Astra takes over from there, modeling objects in Frecad, adding materials in Blender, composing the scene in Isaac Sim, and rigging the

04:53 duck’s joints simulated with Newton physics engine. Then training in the cloud where thousands of ducks learn in parallel in Isaac lab. The policy comes back to the RTX Spark gets evaluated in simulation and is deployed to the real duck which sets off its own in the clip. But if a laptop can run that whole pipeline, what happens when you need something bigger than a laptop? The answer is a pod. Nvidia’s third reveal is the Vera Rubin DSX pod. Basically an AI factory delivered in a container. So, first what’s inside are seven liquid

05:26 cooled Vera Rubin NVL72 racks connected with 17 km of Spectrum X cables integrated, fully cabled, and factory tested before it ships in transportable containers. Next, the speed. As Nvidia’s video says, on-site deployment goes from several months to days once the site utilities are ready, with Nvidia saying it can be as little as 14 days down from about 6 months. Then there’s the design, which Nvidia calls responsible with 800 volt DC power and dry cooling instead of

05:58 evaporative cooling. And pods are expected in early 2027 with Dell, HPE, and Lenovo among the first server partners with no price yet. But a pod is only as good as the chips in its racks. So what makes Vera Rubin so fast? It’s the fourth reveal, which is the architecture behind it all. Vera Rubin, which Nvidia says is in full production. And with Jensen Hang calling it another giant step on top of Blackwell. But to see how big of a step it is, you have to start with his claim. So first the speed. Hang says Blackwell was about 30

06:32 times faster than Hopper at large language model processing and that Vera Rubin is another 30 times greater than Blackwell in just 2 years. Next is the hardware which is a Vera Rubin NVL72 rack that links 72 Reuben GPUs and 36 Vera CPUs with 88 cores each. And according to Nvidia’s preliminary specs, each GPU has about 288 GB of HPM4 for about 3.6 exoflops of low precision NVFP4 inference per rack. Hang also says it takes under 5 minutes to assemble

07:07 versus 2 hours for Grace Blackwell nodes. Then there’s what it’s for. Because Nvidia says it extends past language models to processing the entire agent, supplying the GPU, networking, and storage acceleration to run NVIDIA’s CUDA X libraries for computational lithography, fluid dynamics, and data processing. And this last reveal from Nvidia just tied it all together because robots, laptops, and AI factories can all run autonomous agents. And Nvidia says that takes extremely secure tech, which is why they introduced OpenShell,

07:41 which is a secure runtime, and it’s open- source and tracks agents on Vera CPUs with Nvidia saying each agent runs inside its own isolated sandbox. On top of that is a hardwarebased monitor that Nvidia calls Century, which runs Blue Field 4 chips outside the host and can quarantine a rogue agent in milliseconds. Anyways, make sure to like and subscribe for more of the latest in AI and robotics news. And thanks for watching.