NVIDIA's GTC 2026 conference opened today in San Jose, California, and will run from March 16 to 19. More than 30,000 developers, researchers, and business representatives from 190 countries will attend the conference, which features over 1,000 sessions.
The conference had previously released several important signals: Nvidia is integrating its previously acquired Groq technology into its product line; Samsung will manufacture AI chips for Nvidia for the first time; and OpenAI is expected to become one of the first customers for Nvidia's next-generation inference chips. These moves indicate that Nvidia is expanding further from its position as the leader in AI training chips into the inference chip market, and is reducing its reliance on TSMC through supply chain diversification.
This year's conference also featured a dedicated OpenClaw experience zone called "Build-a-Claw." Attendees could customize and deploy a sustainably running AI agent on-site under the guidance of NVIDIA engineers.
As one of the most important annual technology release platforms in the AI industry, NVIDIA CEO Jensen Huang delivered a keynote speech at 2:00 AM Beijing time on March 17th during the GTC conference. The BlockBeats AI monitoring team, 1M AI News, will provide real-time updates on the key highlights of the conference. The following is the latest update:
Nvidia's Jensen Huang's latest speech: Vera Rubin's seven chips are in full production, anticipating $1 trillion in computing power orders.
According to 1M AI News , NVIDIA founder and CEO Jensen Huang officially announced at GTC 2026 that the Vera Rubin platform is in full production, integrating seven new chips, covering five types of rack systems, and designed as a supercomputer specifically for intelligent AI agents.
The core rack, Vera Rubin NVL72, integrates 72 Rubin GPUs and 36 Vera CPUs interconnected via NVLink 6. Compared to the previous generation Blackwell platform, it reduces the number of GPUs required to train large hybrid expert models to one-quarter, achieves up to 10 times the inference throughput per watt of Blackwell, and reduces the cost per token to one-tenth.
Five types of rack systems constitute the complete AI factory infrastructure:
- Vera Rubin NVL72 GPU Rack
- Vera CPU rack (256 Vera CPUs, twice as efficient and 50% faster than traditional CPUs)
- Groq 3 LPX Inference Accelerator Rack
- BlueField-4 STX storage rack (designed specifically for AI Agent key-value caching, delivering up to 5x faster inference throughput)
- Spectrum-6 SPX Ethernet Rack
In terms of energy management, NVIDIA also launched the DSX platform: DSX Max-Q can deploy up to 30% more AI infrastructure within a fixed power limit, and DSX Flex can activate 100 gigawatts of previously unused grid capacity.
AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and other cloud service providers, as well as system vendors such as Cisco, Dell Technologies, HP Enterprise, Lenovo, and Supermicro, have all announced that they will launch Vera Rubin products in the second half of this year. Anthropic, Meta, Mistral AI, and OpenAI have explicitly stated that they will use the platform to train larger-scale models.
Jensen Huang stated that he predicts that Blackwell and Vera Rubin systems will receive at least $1 trillion in combined orders between 2025 and 2027, more than double the $500 billion forecast given by GTC last year.
NVIDIA GTC Robotics Panorama: Cosmos 3 Unified World Model Released, GR00T N2 Tops Robotics Strategy Rankings, Disney's Olaf Makes an Appearance
According to 1M AI News , NVIDIA launched a series of new physical AI products at the GTC conference and partnered with industrial giants, humanoid robot pioneers, and surgical robot manufacturers in the global robotics ecosystem. Jensen Huang stated, "Physical AI has arrived, and every industrial company will become a robotics company."
Core Product Launch:
1. Cosmos 3: The first unified world foundation model for synthetic world generation, visual reasoning, and motion simulation, designed to accelerate the development of general-purpose robotic intelligence in complex environments.
2. Isaac Lab 3.0: An early access version that supports large-scale robot learning on DGX-level infrastructure. Built on the new Newton Physics Engine 1.0 and PhysX SDK, it adds support for multi-physics simulation and complex dexterity manipulation.
3. GR00T N1.7: Early Access version with commercial license, providing general skills such as advanced dexterity control for mass-production robot deployment.
4. GR00T N2 (Preview): A next-generation robot foundation model based on DreamZero research, employing a novel world motion model architecture. Its success rate in new tasks and environments is more than twice that of mainstream visual language motion models. Currently ranked number one on the MolmoSpaces and RoboArena leaderboards, it is planned for release before the end of the year.
In the field of industrial robots, FANUC, ABB Robotics, YASKAWA, and KUKA, with over 2 million units installed globally, are integrating the Omniverse library and the Isaac simulation framework into their virtual debugging solutions, while also integrating Jetson modules into their controllers for edge AI inference. In humanoid robots, companies such as 1X, AGIBOT, Agility, Boston Dynamics, and Figure are using Cosmos, Isaac Sim, and Isaac Lab to accelerate their development. In medical robots, CMR Surgical uses Cosmos-H simulation to train its Versius surgical system, Johnson & Johnson Medical uses Isaac Sim and Cosmos post-training workflows to train its Monarch urology platform, and Medtronic is exploring IGX Thor to provide functional safety for its surgical robot systems.
One of the highlights of the conference came from Disney: Disney used Kamino, a GPU-accelerated physics simulator based on the NVIDIA Warp framework and integrated into the Newton physics engine, to train the movement strategies of Olaf and the BDX robot characters, enabling Olaf to manage its own heat and reduce collision noise. Jensen Huang appeared on stage with the Olaf robot during his keynote speech. Olaf will make its official debut at Disneyland Paris on March 29th.
NVIDIA GTC unveils three new Nemotron 3 models: Ultra focuses on cutting-edge inference, while VoiceChat integrates speech recognition, large models, and speech synthesis.
According to 1M AI News , NVIDIA announced at the GTC conference the expansion of the Nemotron 3 open model family, adding three new multimodal models for AI agents:
1. Nemotron 3 Ultra: Positioned as a cutting-edge intelligent processor, it achieves 5x throughput efficiency on the Blackwell platform using the NVFP4 format, targeting programming assistants, search, and complex workflow automation scenarios.
2. Nemotron 3 Omni: Integrates audio, visual, and language understanding capabilities to efficiently extract insights from videos and documents.
3. Nemotron 3 VoiceChat: Supports real-time conversations, with AI capable of simultaneously listening and responding, integrating Automatic Speech Recognition (ASR), large language model processing, and Text-to-Speech (TTS) into a single system.
NVIDIA also released the Nemotron security model and Agent retrieval pipeline at the same time. The former detects unsafe content in text and images, while the latter improves the relevance and accuracy of Agent output. In addition, NVIDIA had already released Nemotron 3 Super on the 11th of this month, a hybrid Mamba-Transformer MoE model with 120 billion parameters (12 billion active parameters), natively supporting a 1 million token context window, and achieving more than 5 times the throughput of its predecessor. It achieved a score of 85.6% on the OpenClaw Agent benchmark PinchBench, becoming the best open model in its class.
Companies such as CodeRabbit, CrowdStrike, AI programming tools Cursor and Factory, ServiceNow, and the AI search engine Perplexity have deployed the Nemotron model for agent applications. Edison Scientific, an AI research platform, has integrated Nemotron into its in-house AI scientist Kosmos, serving over 50,000 researchers and enabling the parallel execution of hundreds of research tasks. According to its official statement, it can compress months of research into a single day.
Nvidia ventures into space computing: Releases Vera Rubin Space-1 module, boasting 25 times the AI computing power of the H100.
According to 1M AI News , NVIDIA announced its entry into the space computing field at the GTC conference, releasing the Space-1 Vera Rubin module, designed specifically for on-orbit data centers. It integrates two Rubin GPUs and one Vera CPU, with AI inference computing power up to 25 times that of the H100, enabling large language models and basic models to run directly in orbit.
Jensen Huang stated, "Space computing, the final frontier, has arrived. With the deployment of satellite constellations and the advancement of deep space exploration, intelligence must reside where data is generated." He also frankly admitted that space cooling is an unsolved engineering challenge: "There is no conduction, no convection in space, only radiation. We must figure out how to cool these systems in space."
The Space-1 module is designed for size, weight, and power-constrained environments, supporting autonomous on-orbit analysis, real-time data processing, and scientific discovery. Initial partners include space-based solar energy company Aetherflux, private space station developer Axiom Space, satellite communications company Kepler Communications, Earth observation companies Planet Labs and Sophia Space, and cloud computing satellite company Starcloud. A specific release date has not yet been announced.
Jensen Huang stated that "Claude Code and OpenClaw triggered the inflection point for agents": Nvidia released the OpenShell security runtime, and 17 major enterprises have integrated it.
According to 1M AI News , NVIDIA released the Agent Toolkit open platform at the GTC conference. Its core component is the open-source security runtime OpenShell, providing policy-based security, network, and privacy safeguards for autonomous AI agents. Jensen Huang stated at the launch event, "Claude Code and OpenClaw have triggered the inflection point for agents, extending AI from generation and inference to action. Employees will be empowered by teams composed of cutting-edge, professional, and customized agents, the enterprise software industry will evolve into a specialized agent platform, and the IT industry is at the tipping point of the next major expansion."
The Agent Toolkit also includes the open-source AI-Q Blueprint, co-developed with LangChain. This blueprint employs a hybrid architecture where state-of-the-art models handle orchestration and Nemotron's open models handle research, reducing query costs by over 50%. NVIDIA's agents developed using the AI-Q Blueprint currently rank first on both the DeepResearch Bench and DeepResearch Bench II leaderboards.
In terms of security, NVIDIA is collaborating with Cisco, CrowdStrike, Google, Microsoft Security, and TrendAI to make OpenShell compatible with their cybersecurity and AI security tools. CrowdStrike also released the "Secure-by-Design AI Blueprint," which directly embeds the protection capabilities of the Falcon platform into NVIDIA's AI Agent architecture.
Seventeen software platform vendors have integrated the Agent Toolkit: Adobe, Amdocs, Atlassian, Box, Cadence, Cisco, Cohesity, CrowdStrike, Dassault Systèmes, IQVIA, Palantir, Red Hat, SAP, Salesforce, Siemens, ServiceNow, and Synopsys. Salesforce will use Slack as the main interface and orchestration layer to run the Agentforce Agent, while Siemens has launched the Nemotron-based Fuse EDA AI Agent for end-to-end automation of chip and PCB design.
Nvidia unveils its first Groq chip, the LPX: when paired with Vera Rubin, it delivers up to 35x higher inference efficiency per megawatt, and showcases a next-generation Kyber prototype.
According to 1M AI News , the Groq 3 LPU (Language Processing Unit) is the first chip launched by NVIDIA after acquiring AI inference chip startup Groq for approximately $20 billion last December, and is expected to begin shipping in the third quarter of this year. The Groq 3 LPX rack can accommodate 256 LPUs, equipped with 128GB of on-chip SRAM and 640TB of extended interconnect bandwidth per second. NVIDIA claims that when deployed with the Vera Rubin NVL72, LPX can increase inference throughput by up to 35 times per megawatt, unlocking revenue potential for scenarios with trillions of parameters and millions of tokens for contextual inference. Jensen Huang describes the two processors as "extremely different yet unified: one pursues high throughput, the other low latency, while LPX's on-chip memory significantly expands the total memory capacity available to the model." The LPX rack is planned to launch in the second half of this year along with the Vera Rubin platform.
At the conference, Jensen Huang also showcased a prototype of the next-generation rack architecture codenamed Kyber. Kyber changes the compute trays of 144 GPUs to a vertical arrangement to increase physical density and reduce latency. It will be used in Vera Rubin Ultra, the successor platform to Vera Rubin, and is expected to launch in 2027.
NVIDIA releases DLSS 5: a fusion of traditional 3D graphics and generative AI; Jensen Huang claims this approach will sweep across various industries.
According to 1M AI News , NVIDIA released DLSS 5 at the GTC conference. By fusing structured data from traditional 3D graphics with generative AI models, DLSS enables GeForce GPUs to achieve real-time 4K photorealistic rendering locally, without the need for pixel-by-pixel rasterization of each scene element. In his presentation, Jensen Huang described this approach as a "fusion of controllable 3D graphics and probabilistic generative AI," stating that the former is "completely predictable" and the latter is "highly realistic," allowing developers to create content that is both "beautiful and controllable."
Jensen Huang positions DLSS 5's technological path as the starting point for a broader paradigm shift, stating that "the practice of integrating structured information with generative AI will be replicated in industry after industry." He cites enterprise data platforms such as Snowflake, Databricks, and BigQuery as examples, predicting that future AI agents will simultaneously invoke both structured and generative databases to process tasks.
NVIDIA launches NemoClaw to simplify shrimp farming.
Recently, OpenClaw, the open-source AI agent colloquially known as "Lobster," has become incredibly popular, and NVIDIA (NVDA.O) has announced a simplified version to help users "raise lobsters." NVIDIA CEO Jensen Huang announced NemoClaw for the OpenClaw agent platform at the GTC event on Monday. Users can install a deployment toolchain deeply optimized for OpenClaw with just one command. NemoClaw uses the NVIDIA Agent Toolkit software, optimizing OpenClaw with a single command. It installs OpenShell, providing an open model and an isolated sandbox, adding data privacy and security for autonomous agents.
Click to learn about BlockBeats' job openings.
Welcome to the official BlockBeats community:
Telegram subscription group: https://t.me/theblockbeats
Telegram group: https://t.me/BlockBeats_App
Official Twitter account: https://twitter.com/BlockBeatsAsia
