Real-time Update | NVIDIA GTC 2026 Conference Highlights Roundup

The NVIDIA GTC 2026 conference opened this week in San Jose, California, USA, and will run from March 16 to 19. Over 30,000 developers, researchers, and business representatives from 190 countries will attend the conference, which features over 1000 sessions.

Previous signals released ahead of the conference: NVIDIA is integrating the Groq technology acquired earlier into its product line; Samsung will for the first time manufacture AI chips for NVIDIA on behalf of NVIDIA; at the same time, OpenAI is expected to be one of the first customers for NVIDIA's next-generation inference chip. This series of actions shows that NVIDIA is further expanding into the inference chip market from being the AI training chip leader, and diversifying its supply chain to reduce reliance on TSMC.

This year, the conference also features a special OpenClaw area called "Build-a-Claw." Attendees can customize and deploy a sustainable AI agent on-site under the guidance of NVIDIA engineers.

As one of the most important annual AI industry technology release platforms, NVIDIA CEO Jensen Huang delivered a keynote speech at 2 a.m. Beijing time on March 17, the full text of which can be found at: "Jensen Huang GTC Keynote: Market Demand in 2027 Will Exceed a Trillion Dollars; Everyone Should Develop an OpenClaw Strategy". The BlockBeats AI Monitoring Team 1M AI News will provide real-time updates on hot topics and key points of the conference, with the latest developments below:

March 18

Mistral Launches Forge at GTC: Allowing Enterprises to Train Custom AI Models from Zero with Proprietary Data

According to 1M AI News monitoring, French AI company Mistral unveiled the enterprise platform Forge at the GTC conference, allowing enterprises and governments to use Mistral's open weight model library to train custom AI models from zero on proprietary data. Unlike fine-tuning and RAG, Forge supports full retraining and is suitable for non-English language processing, highly vertical domains, and scenarios requiring reinforcement learning training of Agent systems.

Forge is equipped with a synthetic data pipeline generation tool and is embedded on-site with customers by Mistral's Frontline Deployment Engineer (FDE) team to assist with data wrangling and model assessment. Early partners include Ericsson, the European Space Agency, Dutch chip equipment manufacturer ASML, and others. Mistral CEO Arthur Mensch revealed that the company's annual recurring revenue is expected to exceed $1 billion this year.

NVIDIA Resumes H200 Chip Production for China: Jensen Huang Says Approval Secured from Both US and China

According to 1M AI News monitoring, NVIDIA CEO Jensen Huang announced at GTC 2026 that the company is restarting production of the H200 chip for the Chinese market. He stated that approvals have been obtained from both the Chinese and American governments, and orders from several Chinese customers are already in hand, "our supply chain is coming up." The restart of manufacturing began several weeks ago.

The H200, based on NVIDIA's previous-generation Hopper architecture, is a China-market version designed to meet U.S. export restrictions. Previously, the Trump administration required export licenses for shipments to China starting in April last year, leading NVIDIA to record a $5.5 billion loss and halt sales to China. The policy shifted in December last year, allowing NVIDIA to resume sales, but with strict conditions: capped shipment volumes, mandatory third-party testing, and a 25% sales cut to the U.S. government.

Previously, Jensen Huang predicted that NVIDIA's Blackwell and Rubin series would generate over $1 trillion in revenue by 2027, but this target clearly does not include sales from the Chinese H200.

March 17

NVIDIA Builds Physical AI's "Data Factory": Cosmos Model Mass Produces Training Data, to Be Open-Sourced in April

According to 1M AI News monitoring, NVIDIA unveiled the Physical AI Data Factory Blueprint at the GTC conference, a set of open-source reference architectures for robots, visual AI agents, and autonomous driving that unify the generation, augmentation, and evaluation of training data into an automated pipeline. NVIDIA's VP of Omniverse and Simulation Technology, Rev Lebaredian, stated: "In this new era, compute is data."

The blueprint consists of three modular components:

1. Cosmos Curator: processes, refines, and annotates large-scale real and synthetic datasets

2. Cosmos Transfer: exponentially scales data volume through generation environments and lighting variants, covering long-tail scenarios hard to collect in reality

3. Cosmos Evaluator: automatically scores and filters generated data for physical accuracy, already open-sourced on GitHub

The orchestration framework OSMO has been integrated with AI programming agents such as Claude Code, OpenAI Codex, and Cursor, enabling automated resource management and bottleneck alleviation. Microsoft Azure and Nebius have integrated this blueprint into their respective cloud infrastructures, with FieldAI, Hexagon Robotics, Skild AI, Uber, Teradyne Robotics, among others, already using it.

NVIDIA itself has used this blueprint to train Alpamayo 1.5, touted as the first open-source Inference Vision-Language Action (VLA) model tailored for autonomous driving long-tail scenarios, supporting navigation guidance, conditional prompting, and multi-camera flexible configurations. Autonomous truck technology company PlusAI has announced the adaptation of Alpamayo to its truck platform. The full blueprint is expected to be open-sourced on GitHub in April.

NVIDIA Partners with Google DeepMind to Expand AlphaFold Database, nvQSP Accelerates Drug Simulation by 77x

According to 1M AI News monitoring, NVIDIA unveiled multiple medical AI advancements for the BioNeMo platform at the GTC Conference. The Proteina-Complexa protein drug design model accelerates structure-based drug discovery and therapeutic protein development, with Novo Nordisk, Viva Biotech, and Manifold Bio using the model to design target proteins and complete experimental validation.

NVIDIA collaborated with Google DeepMind, EMBL European Bioinformatics Institute, and Seoul National University to add about 30 million protein complex predictions to the AlphaFold protein structure database, where 1.7 million high-confidence predictions have been included to speed up new drug target and disease biology discoveries.

NVIDIA has also introduced nvQSP, a GPU-accelerated quantitative systems pharmacology simulation engine that, in benchmark tests, ran 77 times faster than traditional single-threaded CPU simulation. This allows researchers to explore hundreds of dosing levels and patient subgroups in computational models before clinical trials.

Samsung Unveils 7th-Gen HBM4E at GTC: Single-Pin 16Gbps, 4TB/s Bandwidth, 6th-Gen HBM4 in Mass Production for Vera Rubin

According to 1M AI News monitoring, Samsung Electronics unveiled its seventh-generation high-bandwidth memory, HBM4E, at the NVIDIA GTC conference, featuring a single-pin speed of 16Gbps and a total bandwidth of 4.0 TB/s. This marks the global debut of Samsung's HBM4E physical product.

The concurrently showcased sixth-generation HBM4 has entered mass production, designed specifically for the NVIDIA Vera Rubin platform, utilizing the sixth-generation 10nm-class DRAM technology (1c process). It boasts a stable speed of 11.7Gbps, exceeding the industry standard 8Gbps and peaking at 13Gbps. Samsung also presented its Hybrid-bonded Copper (HCB) technology, enabling stacking of 16 layers or more and reducing thermal resistance by over 20% compared to traditional Thermal Compression Bonding (TCB).

Samsung also exhibited a range of complementary products for NVIDIA's AI infrastructure: the mass-produced SOCAMM2 low-power server memory module (the industry's first mass-produced module), the PCIe 6.0-based PM1763 SSD, and the PM1753 SSD serving as a component of the NVIDIA BlueField-4 STX storage architecture. Yong Ho Song, President of Samsung Semiconductor's AI Center and EVP, will provide detailed insights into the application of Agent AI and Digital Twin in semiconductor manufacturing during a keynote speech on March 17.

NVIDIA Introduces BlueField-4 STX Storage Architecture and CMX Context Memory Platform: Agent AI's Long Context Inference Requires Next-Gen Storage

As reported by 1M AI News, NVIDIA unveiled BlueField-4 STX at the GTC conference, a modular storage reference architecture tailored for Agent AI. Traditional data center storage falls short in meeting the real-time contextual memory access needs of Agents across multiple steps, tools, and sessions. STX aims to keep data close to the GPU and readily available.

The first rack-scale product, the CMX Context Memory Storage Platform, is the core of STX, expanding GPU memory into a scalable high-performance context layer. Compared to traditional storage, it achieves up to 5x higher transactions per second throughput, 4x higher efficiency, and doubles data ingestion speed. At its core is the storage-dedicated BlueField-4 processor, integrating the Vera CPU and ConnectX-9 SuperNIC, along with Spectrum-X Ethernet.

CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure (OCI), and Vultr have planned adoption of STX for deploying context memory storage. On the storage vendor side, DDN, Dell Technologies, HPE, IBM, NetApp, Nutanix, VAST Data, WEKA, and 15 other companies are designing next-generation AI storage products based on STX. The platform is set to launch in the second half of this year.

NVIDIA GTC Announces Nemotron 3 with Three New Models: Ultra Focuses on Edge Inference, VoiceChat Integrates Speech Recognition, Large Models, and Speech Synthesis

According to 1M AI News monitoring, NVIDIA announced the expansion of the Nemotron 3 open model family at GTC, adding three multimodal models for AI agents:

1. Nemotron 3 Ultra: Positioned as cutting-edge intelligence, achieves 5x throughput efficiency in NVFP4 format on the Blackwell platform, targeting programming assistants, search, and complex workflow automation scenarios

2. Nemotron 3 Omni: Integrates audio, visual, and language understanding capabilities, efficiently extracting insights from videos and documents

3. Nemotron 3 VoiceChat: Supports real-time conversations, where AI can listen and respond simultaneously, integrating automatic speech recognition (ASR), large language model processing, and text-to-speech synthesis (TTS) into a single system

Simultaneously, NVIDIA also released Nemotron Security Models and Agent Retrieval Pipelines, with the former detecting unsafe content in text and images, and the latter enhancing the relevance and accuracy of Agent outputs. In addition, NVIDIA previously released Nemotron 3 Super on the 11th of this month, a 1200-billion-parameter (120 billion active parameters) hybrid Mamba-Transformer MoE model, natively supporting a 1 million-token context window, achieving over 5x throughput improvement compared to the previous generation. It scored 85.6% on the OpenClaw Agent benchmark test PinchBench, becoming the top-performing open model in its class.

Companies such as CodeRabbit, CrowdStrike, AI programming tool Cursor, Factory, ServiceNow, and AI search engine Perplexity have deployed the Nemotron model for Agent applications. The AI research platform Edison Scientific has integrated Nemotron into its autonomous AI scientist Kosmos, serving over 50,000 researchers, able to concurrently execute hundreds of research tasks, with claims to compress months of research into a single day.

NVIDIA GTC Robot Panorama: Cosmos 3 Unified World Model Released, GR00T N2 Tops Robot Policy Rankings, Disney Snowbot Takes the Stage

According to 1M AI News monitoring, NVIDIA unveiled a series of physical AI new products at the GTC conference and partnered with industrial giants in the global robot ecosystem, humanoid robot pioneers, and surgical robot manufacturers. Jensen Huang stated, "Physical AI has arrived, and every industrial enterprise will become a robot company."

Key product releases:

1. Cosmos 3: The first unified synthetic world generation, visual reasoning, and motion simulation foundational model for accelerating general-purpose robot intelligence development in complex environments

2. Isaac Lab 3.0: Early access version supporting large-scale robot learning on DGX-class infrastructure, built on the new Newton Physics Engine 1.0 and PhysX SDK, adding multi-physics simulation and complex dexterous manipulation support

3. GR00T N1.7: Early access version with commercial licensing, offering advanced dexterous control and other general skills for mass-produced robot deployments

4. GR00T N2 (Preview): Next-generation robot foundational model based on DreamZero research, adopting a new world action model architecture, achieving over twice the success rate of mainstream visual language action models in new tasks and environments, currently ranking first on the MolmoSpaces and RoboArena leaderboards, scheduled for release by the end of the year

In the industrial robot realm, with a global installed base of over 2 million units, FANUC, ABB Robotics, YASKAWA, and KUKA are integrating Omniverse Kit and Isaac simulation framework into virtual commissioning solutions, while integrating Jetson modules into controllers for edge AI inference. In humanoid robots, companies such as 1X, AGIBOT, Agility, Boston Dynamics, Figure, among others, are using Cosmos, Isaac Sim, and Isaac Lab to accelerate development. In medical robots, CMR Surgical is using Cosmos-H to train its Versius surgical system, Johnson & Johnson Medical is using Isaac Sim and Cosmos for training workflows for the Monarch urology platform, and Medtronic is exploring IGX Thor to provide functional safety for surgical robot systems.

One of the highlights of the conference came from Disney: Disney showcased Kamino, a GPU-accelerated physics simulator integrated into the Newton physics engine, using the NVIDIA Warp framework. Kamino was used to train the motion strategies of the Olaf snowman and BDX robot characters, allowing Olaf to learn self-heat management and reduce collision noise. Hwang In-hyun appeared on stage with the Olaf robot during the keynote speech, and Olaf is set to make its official debut at Disneyland Paris on March 29.

NVIDIA GTC Announces Nemotron 3 with Three New Models: Ultra focusing on edge inference, VoiceChat combining speech recognition, large models, and speech synthesis

According to 1M AI News monitoring, NVIDIA announced the expansion of the Nemotron 3 open model family at the GTC conference, adding three new multimodal models for AI Agents:

1. Nemotron 3 Ultra: Positioned as edge-level intelligence, achieving 5x throughput efficiency in NVFP4 format on the Blackwell platform, targeting programming assistants, search, and complex workflow automation scenarios.

2. Nemotron 3 Omni: Integrating audio, vision, and language understanding capabilities for efficient insight extraction from videos and documents.

3. Nemotron 3 VoiceChat: Supporting real-time conversations, where AI can listen and respond simultaneously, integrating automatic speech recognition (ASR), large language model processing, and text-to-speech synthesis (TTS) into a single system.

Simultaneously, NVIDIA also released Nemotron security models and an Agent retrieval pipeline, where the former detects unsafe content in text and images, and the latter enhances the relevance and accuracy of Agent outputs. In addition, NVIDIA had previously released Nemotron 3 Super on the 11th of this month, a hybrid Mamba-Transformer MoE model with 1.2 trillion parameters (120 billion active parameters), natively supporting a 1 million token context window. The throughput has been increased by over 5 times compared to the previous generation, achieving an 85.6% score on the PinchBench benchmark in the OpenClaw Agent, making it the best-performing open model in its class.

Companies such as CodeRabbit, CrowdStrike, AI programming tools Cursor, Factory, ServiceNow, and AI search engine Perplexity have deployed Nemotron models for Agent applications. The AI research platform Edison Scientific has integrated Nemotron into its proprietary AI scientist Kosmos, serving over 50,000 researchers, enabling the parallel execution of hundreds of research tasks, compressing months of research into a single day, according to their official statement.

NVIDIA Ventures into Space Computing: Introduces Vera Rubin Space-1 Module, AI Horsepower 25x H100

According to 1M AI News monitoring, NVIDIA announced its entry into the space computing field at the GTC conference, unveiling the Space-1 Vera Rubin module, designed for in-orbit data centers, integrating 2 Rubin GPUs and 1 Vera CPU, with AI inference horsepower peaking at 25 times that of the H100, enabling large language models and foundational models to run directly in orbit.

Jensen Huang stated, "Space computing, the final frontier, has arrived. With satellite constellations being deployed and deep space exploration advancing, intelligence must exist where data is generated." He also acknowledged that space cooling remains an unresolved engineering challenge: "In space, there is no conduction, no convection, only radiation, and 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 in-orbit autonomous analysis, real-time data processing, and scientific discovery. Initial partners include space solar power company Aetherflux, private space station developer Axiom Space, satellite communications company Kepler Communications, Earth observation company Planet Labs, Sophia Space, and satellite cloud computing company Starcloud. The specific listing date has not yet been announced.

Jensen Huang States "Claude Code and OpenClaw Triggered Agent Inflection Point": NVIDIA Releases OpenShell Secure Runtime, 17 Enterprise Titans Onboard

According to 1M AI News monitoring, NVIDIA unveiled the Agent Toolkit open platform at the GTC conference, with the core component being the open-source secure runtime OpenShell, providing policy-based security, networking, and privacy guardrails for self-running AI agents. Jensen Huang stated at the event, "Claude Code and OpenClaw triggered the agent inflection point, extending AI from generation and inference to action. Employees will be augmented by teams composed of cutting-edge, professional, and bespoke agents, and the enterprise software industry will evolve into a specialized agent platform, with the IT industry at the threshold of the next major expansion."

The Agent Toolkit also includes the open-source AI-Q Blueprint co-built with LangChain, adopting a hybrid architecture with cutting-edge model orchestration by Nemotron Open Model Research and query cost reduction of over 50%. The Agent developed by NVIDIA using the AI-Q Blueprint currently ranks first on both the DeepResearch Bench and DeepResearch Bench II leaderboards.

On the security front, NVIDIA is collaborating with Cisco, CrowdStrike, Google, Microsoft Security, and TrendAI to make OpenShell compatible with their network security and AI security tools. CrowdStrike simultaneously released the "Secure-by-Design AI Blueprint," embedding the Falcon platform's protective capabilities directly into NVIDIA's AI Agent architecture.

17 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. Among them, Salesforce will operate Agentforce Agent with Slack as the main interface and orchestration layer, while Siemens is launching the Nemotron-based Fuse EDA AI Agent for end-to-end automation of chip and PCB design.

NVIDIA Launches First Groq Chip LPX: Achieving Up to 35x Inference Efficiency Improvement per Megawatt Combined with Vera Rubin and Unveiling Next-Generation Kyber Prototype

According to 1M AI News monitoring, the Groq 3 LPU (Language Processing Unit) is NVIDIA's first chip launched after acquiring Groq, an AI inference chip startup, for approximately $20 billion last December, with shipments expected to start 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 per second of extended interconnect bandwidth. The company claims that when LPX is deployed with Vera Rubin NVL72, the maximum per-megawatt inference throughput can be increased by up to 35 times, unlocking the revenue potential for scenarios with trillion-parameter, million-token context inference. Jensen Huang described the two processors as "extremely different yet unified: one pursuing high throughput, one pursuing low latency, with LPX's on-chip memory significantly expanding the total memory capacity available to the model. The LPX rack is scheduled to be released in the second half of this year alongside the Vera Rubin platform.

During the conference, Jensen Huang also showcased a prototype of the next-generation rack architecture, codenamed Kyber. Kyber reimagines the 144-GPU compute tray into a vertical alignment to increase physical density, reduce latency, and will be deployed on the successor platform to Vera Rubin, Vera Rubin Ultra, set to launch in 2027.

NVIDIA Unveils DLSS 5: Traditional 3D Graphics Fused with Generative AI, Jensen Huang States This Path Will Sweep Across Industries

According to 1M AI News tracking, NVIDIA introduced DLSS 5 at the GTC conference, combining the structured data of traditional 3D graphics with generative AI models to enable GeForce GPUs to achieve real-time 4K photorealistic rendering natively, without rasterizing each scene element per pixel. Huang described this approach in his keynote as the "integration of controllable 3D graphics and probabilistic generative AI," labeling the former as "fully predictable" and the latter as "highly realistic," combining to empower developers to create content that is "both exquisite and controllable."

Huang positioned DLSS 5's technical trajectory as the starting point for a broader paradigm shift, stating that "the approach of fusing structured information with generative AI will replicate across industry after industry." He cited enterprise data platforms like Snowflake, Databricks, BigQuery, foreseeing future AI agents concurrently invoking structured and generative databases for processing tasks.

NVIDIA CEO Jensen Huang's Latest Keynote: Vera Rubin Series of Seven Chips in Full Production, Envisions $1 Trillion Compute Power Order

As monitored by 1M AI News, NVIDIA's founder and CEO Jensen Huang officially announced at GTC 2026 the full-scale production of the Vera Rubin platform, integrating seven new chips across five chassis systems, collectively designed as a supercomputer tailored for AI.

The flagship Vera Rubin NVL72 rack integrates 72 Rubin GPUs and 36 Vera CPUs interconnected via NVLink 6. Compared to the previous Blackwell platform, the GPU count required for training large hybrid expert models has been quartered, achieving up to 10 times the inference throughput per watt of Blackwell and reducing the cost per token to one-tenth.

The five types of rack systems constitute a complete AI Factory infrastructure:

- Vera Rubin NVL72 GPU Rack

- Vera CPU Rack (256 Vera CPUs, 2x more efficient than traditional CPUs, 50% speedup)

- Groq 3 LPX Inference Acceleration Rack

- BlueField-4 STX Storage Rack (designed for AI Agent key-value cache, up to 5x highest inference throughput)

- Spectrum-6 SPX Ethernet Rack

In terms of energy management, NVIDIA simultaneously released the DSX platform: DSX Max-Q can deploy 30% more AI infrastructure within a fixed power limit, and DSX Flex can unlock 100 gigawatts of previously unused idle grid capacity.

Cloud service providers such as AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, as well as system manufacturers like Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo, AMD, and others have all announced plans to launch Vera Rubin products in the second half of this year. Anthropic, Meta, Mistral AI, and OpenAI have explicitly stated that they will use this platform to train larger-scale models.

Jensen Huang said he predicts that the combined orders of Blackwell and Vera Rubin systems from 2025 to 2027 will reach at least $1 trillion, doubling the $500 billion forecast given at last year's GTC.

NVIDIA Introduces NemoClaw to Simplify "Shrimping"

Recently, the open-source AI agent known as "OpenClaw," popularly referred to as "Lobster," has gained significant traction, and NVIDIA (NVDA.O) has announced a simplified mode to assist users in "shrimping." NVIDIA CEO Jensen Huang announced at the GTC event on Monday the launch of NemoClaw for the OpenClaw agent platform, allowing users to install a deployment toolchain optimized for OpenClaw with just one command. NemoClaw uses the NVIDIA Agent Toolkit software to optimize OpenClaw with a single command. It installs OpenShell, providing open models and an isolated sandbox to add data privacy and security for autonomous agents.