General-purpose robots and autonomous machines are moving from research labs to real-world mass-market deployment, creating demand for compact, power-efficient AI supercomputers capable of running foundation models at the edge. 

To meet that need, NVIDIA today introduced the T3000 and T2000, new modules based on the NVIDIA Thor architecture that enable mass-market robotics and edge AI applications at scale.

Jetson AGX Thor is powering this next generation of humanoid and robotic systems, with growing adoption across industries. Leading companies — including 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi and Techman Robot — are building on the platform.

Unlocking Humanoid and Robotics Deployment With T3000

The hardware underpinning those capabilities starts with the Jetson and IGX T3000 modules, which delivers 865 FP4 teraflops of AI compute in a compact form factor roughly half the size and power of the T5000. Jetson T3000 combines an NVIDIA Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory and 273GB/s of memory bandwidth, along with 25 GbE connectivity. IGX T3000 delivers the same performance with integrated functional safety while seamlessly running the NVIDIA Halos for Robotics full-stack safety system for robots operating alongside humans.

Despite its smaller footprint, the T3000 achieves similar inference performance of the T5000 for multimodal workloads, including large language models, vision language models, vision language action models and world foundation models. Migrating to T3000 helps reduce costs amid high memory prices. 

Going Wide on Edge AI With T2000

The Jetson T2000 brings Thor architecture to a broader range of edge AI systems. With 400 FP4 teraflops of compute and 16GB of memory, it provides an entry point for developers building visual AI agents, autonomous mobile robots, industrial manipulators and other intelligent machines.

With the introduction of the new NVIDIA Jetson modules, NVIDIA now offers a scalable edge AI platform spanning performance from 70 TOPS to 2,000 teraflops, enabling developers to address virtually any edge AI workload.

New Agent Skills Automate Memory Optimization Across All Jetson Devices

AI agents are transforming developer productivity by automating memory optimization, system configuration and deployment tasks that previously required manual effort and deep domain expertise.

With the newly released Jetson agent skills, developers can optimize the entire software stack and achieve significant memory savings in days instead of weeks. These skills support the entire Jetson portfolio, including Jetson Thor and Jetson Orin, enabling developers to run more capable workloads on lower-memory configurations. 

The result is lower system cost, faster deployment and the flexibility to move down one memory SKU within the same product tier without compromising performance.

Companies across industries and regions have accelerated development while achieving substantial memory savings through software optimization.

Humanoid robotics leaders including UBTech and Agile Robots, along with industrial solutions provider Connect Tech, have reduced memory usage by up to 15GB, enabling them to move from NVIDIA Jetson AGX Orin 64GB to the 32GB module.

In smart retail, SandStar reduced memory usage by up to 4GB, enabling deployment on the NVIDIA Jetson Orin NX 8GB module instead of the 16GB configuration. In companion robotics, GROOVE X, creator of the LOVOT robot, uses Jetson’s heterogeneous AI accelerators to optimize workload distribution, reducing memory usage and enabling deployment on lower-memory configurations. 

In intelligent transportation, NoTraffic reduced memory usage by 30% on Jetson TX2 NX, creating headroom to add more AI capabilities into its smart traffic platform without increasing hardware requirements.

With agent skills simplifying development and NVIDIA NemoClaw blueprints orchestrating intelligent agents, Jetson is an agentic-ready platform for physical AI, enabling advanced reasoning, autonomous decision-making and task automation at scale.

Delivering Cosmos 3 Edge to NVIDIA Thor Lineup

NVIDIA today expanded its NVIDIA Cosmos 3 frontier open world foundation model family — built as a robot foundation model for embodied systems — with a lightweight model compatible with NVIDIA Thor platforms. Cosmos 3 Edge is a 4-billion-parameter model helping embodied systems see the world, reason over it in real time, and predict and generate actions through on-device inference. Using the open Cosmos framework, developers can post-train Cosmos 3 Edge for specific embodiments and sensors in about a day — closing the sim-to-real gap — then deploy on Jetson Thor for real-time vision analysis and on-device robot policy.

Start Development Today With Emulation Mode

Sharing the same chip architecture and software stack in the NVIDIA Thor family, the new modules provide a seamless development path. Developers can begin building today using the Jetson AGX Thor developer kit available through channel partners and emulate the performance of T3000 and T2000 modules.

Using NVIDIA’s full physical AI software stack — including NVIDIA Isaac for robotics simulation and perception — alongside open models such as NVIDIA Nemotron, Cosmos 3 and Isaac GR00T, developers can accelerate the development of next-generation robots, autonomous machines and visual AI agents.

Developers can begin using T3000 emulation mode later this month with JetPack 7.2.1. Support for T2000 emulation mode will follow in a future release. The Jetson T3000 and T2000 modules are scheduled to become available in Q1 2027.

ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK and YUAN are among other partners in the Jetson ecosystem already providing Thor-based solutions. Software partners such as Antmicro, Neurealm, REBOTNIX and RidgeRun will provide emulation and migration solutions for customers transitioning to the new modules.

As physical AI and embodied AI move toward mainstream deployment, the new NVIDIA Thor computers give developers a scalable foundation for bringing intelligent humanoids and autonomous machines into the real world.

Find a Jetson AGX Thor Developer Kit on the NVIDIA marketplace and start developing today.

Home to leading manufacturers, robotics pioneers, infrastructure builders and iconic gaming companies, of course, Japan is one of the world’s centers of AI — building across the full stack with NVIDIA technologies. This week NVIDIA and its partners in Japan are showcasing the AI ecosystem’s latest advancements. Check back here for updates.

Wednesday, July 15, 4 p.m. PT

Japan’s Leaders Advance Healthcare and Life Sciences With NVIDIA Agentic and Physical AI

Japan built the world’s most trusted names in medical technology and biopharma. Now the country’s healthcare leaders are engineering the next generational leap with AI, powered by NVIDIA.

From autonomous surgical robots to AI-accelerated CT systems, and from agentic drug discovery platforms to virtual cell models, Japanese innovators are deploying NVIDIA technology to reshape medicine at every level. 

Agentic AI Accelerates Japanese Drug Discovery

Japan’s pharmaceutical leaders are uniting around AI-powered drug discovery. Tokyo-1, the AI drug discovery consortium and platform operated by Xeureka, continues to expand, with Eisai joining this past April, bringing together leading pharma companies —  Astellas, Daiichi Sankyo and Ono Pharmaceuticals — all advancing drug discovery using NVIDIA BioNeMo.

Astellas has deployed nearly all BioNeMo NIM microservices within NVIDIA’s digital biology portfolio and is running BioNeMo Agent Toolkit, NVIDIA’s open platform that turns any AI agent into an autonomous life sciences scientist. It gives AI agents, software platforms and biopharma systems immediate access to NVIDIA’s full life sciences stack. 

Ono Pharmaceuticals is using the Boltz-2 NIM microservice to streamline and accelerate internal drug discovery. Daiichi Sankyo is conducting ultralarge-scale virtual screening on Tokyo-1 and leveraging NVIDIA RAPIDS to accelerate large-scale data processing. Xeureka is using NVIDIA BioNeMo to power its AI-driven drug discovery efforts, enabling researchers the flexibility to use the most appropriate models and tools across diverse discovery programs.

SyntheticGestalt announced two products: the molecular AI foundation model ZAO and the molecular generative model KOYA. ZAO is a foundation model that converts small molecules into data AI can use, through a “4D” representation that captures the multiple 3D conformations a molecule actually adopts; as a single general-purpose model, it ranked No. 1 on nine public drug-discovery benchmark tasks, achieving the world’s best performance. 

KOYA is a molecular generative model that designs novel, high-affinity ligands for a target protein while closely reflecting the user’s intent. Both products can be called from the NVIDIA BioNeMo Agent Toolkit, enabling AI agents to carry out everything from evaluating molecules to designing them, and to accelerate drug discovery in collaboration with researchers.

Biomy is pioneering a virtual cell foundation model with a massive clinical dataset from the Japanese Foundation for Cancer Research. Using NVIDIA single-cell RAPIDS, Biomy achieved 90% faster spatial transcriptomics analysis. Biomy will use NVIDIA Nemotron-powered agents to autonomously propose and orchestrate complex virtual experiments for drug development. 

Takeda recently announced a collaboration with Boltz to deploy the BoltzMol-1 and BoltzProt-1 biomolecular models across its research organization, giving scientists tools for structure prediction, affinity estimation and generative design that integrate into existing discovery workflows. NVIDIA accelerates these models through NVIDIA BioNeMo with libraries such as cuEquivariance. 

Physical AI Enters the Operating Room

Kawasaki Heavy Industries provides technology designed to improve the overall efficiency of hospital operations, including with its FORRO, Nyokkey and NURABOT robots.

The company plans to use NVIDIA Holoscan IGX, Isaac for Healthcare, Isaac GR00T and Cosmos to develop surgical support functions, nursing assistant and hospital transport robots.

Direava is developing a surgical vision language model for real-time surgical video understanding and natural language interaction with surgical scenes. Direava aims to evolve this technology into an intelligence layer for future surgical AI and physical AI in the operating room. 

NVIDIA Accelerated Computing Powers Japan’s Next-Generation CT

Two of Japan’s leading medical imaging companies are now shipping next-generation CT systems built on NVIDIA GPUs. 

Canon launched Japan’s first NVIDIA-accelerated photon-counting CT system, marking a step forward for the country’s next generation of medical imaging.

Fujifilm has commercialized Japan’s first whole-body CT system powered by NVIDIA Blackwell, using diffusion-based deep learning reconstruction to improve image quality.

The integration of AI and accelerated computing into medical imaging equipment contributes to improved image quality, enhanced accuracy, early detection and higher standards of medical care.

Together, these advances signal a new era: AI, and not just accelerated computing, is no longer an experiment in Japanese healthcare. It’s infrastructure. 

Wednesday, July 15, 4 p.m. PT

NVIDIA Metropolis Provides Developers Agent-Ready Libraries to Build NVIDIA Cosmos-Powered Vision AI Agents Faster 

As enterprises capture more video data across the physical world, vision AI is transforming beyond passive perception and dashboards into agentic systems that can understand, reason and act in real time. Powered by reasoning vision language models (VLMs) such as the NVIDIA Cosmos of open models, these agentic systems extract rich insights from video, whether on operations, environmental context or root causes for issues.

Building production-ready, high-accuracy vision AI agents can require thousands of developer hours across data collection, model training, validation and deployment. NVIDIA Metropolis now packages more than 80 new skills, including NVIDIA VSS Blueprint 3.2, NVIDIA DeepStream 9.1, NVIDIA TAO 7 and Physical AI Data Factory, that help developers use coding agents to speed that process by at least 6x.

Japan’s industrial and smart-space leaders including Asilla, AWL, Fujitsu, Hitachi, OMRON, Shimizu Corporation and Yazaki North America are using Metropolis to bring vision AI agents into factories, construction sites, stories, buildings and public spaces.

Metropolis Open Libraries and Skills Span the Vision AI Lifecycle

Metropolis provides a comprehensive set of open libraries and skills that span the entire vision AI development lifecycle, from creating data pipelines to generating synthetic data, fine-tuning models and deploying agents at scale. 

New libraries include:

• NVIDIA VSS Blueprint 3.2 helps developers build and operate vision AI agents that can see, reason and act over live or recorded video using natural language. New skills for coding agents make it faster to build and operate custom, always-on video agents that alert, summarize and search across large camera networks.
• NVIDIA DeepStream 9.1 helps developers create and deploy real-time, multi-sensor video analytics pipelines from edge to cloud for large-scale ingestion, multi-camera tracking and operations analytics.
• NVIDIA TAO 7 helps developers customize and optimize NVIDIA Cosmos and other vision AI models with agent skills for labeling, performance diagnostics, fine-tuning, data generation and automated machine learning. 
• NVIDIA Physical AI Data Factory skills help developers use NVIDIA Cosmos to automatically generate and augment synthetic image and video data to fill training gaps for rare or new product defects, environmental changes and other edge cases, pushing vision model accuracy to new levels.

Companies Advance Agentic Vision AI With NVIDIA Metropolis

Japan-based companies are using the new NVIDIA Metropolis technologies to bring real-time intelligence to physical operations.

For industrial inspection and operations, OMRON is enhancing automated inspections with VSS-powered video analytics agents.

DeepHow is helping Yazaki North America automate time and motion studies, reducing the current process from weeks to days and unlocking millions of dollars in annual savings.

For smart spaces and public safety, several Hitachi HMAX solutions use VSS-powered agents to generate actionable insights and identify issues in building and rail infrastructure, helping to reduce maintenance costs and energy consumption by 15% in rail applications alone. Fujitsu Kozuchi AI platform combines VSS with its Agentic Memory technology to transform long-duration video into operational knowledge, accelerating decision-making across manufacturing, logistics, retail and smart spaces. Meanwhile, Shimizu Corporation is piloting VSS for construction worker safety.

With DeepStream and VLMs, Asilla is monitoring public spaces and commercial facilities to detect incidents and improve response time, while AWL is building retail and manufacturing solutions with DeepStream.  

Developers can access NVIDIA VSS Blueprint 3.2 skills, NVIDIA DeepStream 9.1 skills and NVIDIA TAO 7 skills on GitHub. NVIDIA Physical AI Data Factory and synthetic data generation skills are available through GitHub and can be explored using Physical AI Launchables on NVIDIA Brev.

Wednesday, July 15, 4:00 p.m. PT

Japanese Megabanks Build Financial Intelligence With NVIDIA Nemotron and NVIDIA Agent Toolkit

Across Japan, leading banks and financial technology companies are building AI factories and models to deliver financial intelligence. NVIDIA Nemotron open models and NVIDIA Agent Toolkit are helping them turn regulated financial data into valuable intelligence.

In banking, the most powerful AI applications may not look like chatbots. They look like safer payments, smarter fraud detection, faster software development and more personalized financial services, all built on trusted data. 

Mizuho plans to build what is expected to be the largest on-premises AI factory in Japan’s financial industry, starting with NVIDIA DGX B200 systems and scaling toward a larger cluster. For a bank handling sensitive financial workloads, being on premises matters: it gives teams a foundation to develop agents with NVIDIA Agent Toolkit and NVIDIA NemoClaw blueprints, while keeping critical data close and secure.

With this secure foundation, Mizuho aims to safely expand the operational scope of these autonomous agents into core workflows, including information gathering, document creation, analysis and system development support, while ensuring rigorous governance and auditability. 

As the core IT company of SMBC Group, the Japan Research Institute (JRI) deployed an AI factory to transform financial data into intelligence using NVIDIA Nemotron open models. As one of Japan’s largest financial groups, SMBC Group’s adoption shows how open models and accelerated infrastructure can help established institutions move AI from experimentation into production-ready enterprise workflows. The initiative serves as a foundation for scaling AI adoption across the SMBC Group, improving productivity, accelerating innovation and delivering better financial services to customers.

Rakuten Bank brings digital-native scale to the same transformation. Using the Rakuten Group’s ecosystem, which spans more than 70 services and includes over 18 million banking accounts, 33 million credit cards and 14 million brokerage accounts, Rakuten Bank will develop transaction foundation models built with NVIDIA Agent Toolkit, helping turn high-volume consumer financial data into specialized intelligence for banking services.

Ippu Senkin is collaborating with a financial institution to build sovereign financial intelligence with NVIDIA Blackwell GPUs and Local AI Agent, a local coding agent developed by Ippu Senkin using NVIDIA Agent Toolkit, Nemotron and NemoClaw for secure payment operations within the institution’s group. The effort points to a broader ecosystem motion with AI-native partners helping financial services companies build local agents and applications that can run on local AI factories.

Japan’s financial services industry is moving from model pilots to AI infrastructure that can support regulated, domain-specific intelligence. Banks need performance, governance and proximity to data; digital banks need model-building capacity at transaction scale; and AI-native partners need a platform for local financial agents. 

NVIDIA provides a full stack across those paths, from accelerated computing and AI factory architecture to Nemotron open models and Agent Toolkit for building agents and specialized financial intelligence.

Learn more about how financial institutions are transforming financial data into intelligence with transaction foundation models.

Wednesday, July 15, 4:00 p.m. PT

NVIDIA Advances Japan’s World-Class Quantum and AI for Science Capabilities

NVIDIA is advancing a historic partnership between the U.S. and Japan, its first international partner in the Genesis Mission. 

Genesis Mission’s large-scale initiative to harness AI for scientific discovery calls on U.S. labs and industry, as well as international collaboration. 

NVIDIA and Japan are answering the call — from AI to quantum computing.  

NVIDIA and RIKEN Driving AI for Science 

At RIKEN, Japan’s leading national comprehensive research institute, two supercomputers driven by NVIDIA GB200 and NVIDIA Quantum-X800 are beginning operations. 

RIKYU, a new supercomputer for “AI for Science” development, deploying 1,600 NVIDIA Blackwell GPUs using the GB200 NVL4 platform, will support RIKEN’s development of open foundation models and contribute to accelerating AI adoption across broad fields, including life sciences, materials science and laboratory automation.  

JHPC-quantum GPU supercomputer “ROQUO” is a quantum-HPC system tightly integrating quantum processors with accelerated computing from 540 Blackwell GPUs accessed through the GB200 NVL4 platform. ROQUO is connected to on-premises quantum computers at RIKEN’s facilities in Wako and Kobe, Japan — including Quantinuum’s trapped-ion Reimei system, enabling hybrid quantum-HPC workloads. In ROQUO’s first months of operation, researchers are beginning to explore an evolutionary AI framework, developed with NVIDIA and integrated with the NVIDIA CUDA-Q platform for quantum-classical computing, to generate quantum circuits for the Reimei system. 

Building an Ecosystem That Brings AI to Quantum 

AI is the unlocking technology for scaling quantum processors into useful quantum-GPU supercomputers, but the adoption of AI in quantum computing workflows remains a key challenge.  

At the National Institute of Advanced Industrial Science and Technology’s (AIST) Global Research and Development Center for Business by Quantum-AI Technology (AIST G-QuAT), NVIDIA is working to bring state-of-the-art AI to the center’s current and future quantum processor systems. NVIDIA NVQLink provides the low-latency connection between GPUs and quantum processors, while NVIDIA Ising open models support automated QPU calibration and AI-based decoding for quantum error correction. 

Advancing Quantum Chemistry 

High-accuracy simulations of chemical systems are fundamental for next-generation research in areas such as materials science and drug discovery. AI approaches can expand what quantum algorithms are capable of, improving how these simulations scale. 

Mitsubishi Chemical, Mizuho Bank, Keio University, AIST, the University of Toronto and NVIDIA have demonstrated an AI- and GPU-driven workflow for harnessing quantum processors in molecular spectral analysis — a key tool for understanding the electronic structure and properties of molecules and materials. NVIDIA GPUs achieved a 13.4x speedup for this workflow over CPU-only nodes. Accelerating this analysis lets researchers apply it more quickly to early targets, like extreme ultraviolet photoresist for semiconductor manufacturing.

Developing useful quantum chemistry applications also means building workflows suitable for tomorrow’s large-scale hybrid quantum-GPU supercomputing systems. Fujitsu and NVIDIA are now investigating efficient ways to use NVIDIA CUDA-Q for large-scale quantum-chemistry simulation. Through the collaboration, Fujitsu has started the trial of NVQLink to determine if it can be utilized to realize efficient control of their quantum-classical hybrid computing environment.

Together, the U.S. and Japan are building on the NVIDIA platform to develop a shared foundation for useful, large-scale quantum computing and AI-driven science, and uniting industry, academia and government.

Wednesday, July 15, 4:00 p.m. PT

NVIDIA Expands Partnership With Toyota to Advance Physical AI Across Automotive, Robotics and Cities 

From self-driving cars to cities, the next era of mobility will be defined by AI-enabled systems that can perceive, reason and safely act in the physical world. Toyota and NVIDIA are working together to build that future — connecting AI across vehicles, infrastructure and industrial operations.

This builds on last year’s announcement that Toyota will develop next-generation vehicles with advanced driver-assistance capabilities (L2++) built on NVIDIA DRIVE AGX and running the safety-certified NVIDIA DriveOS operating system. 

NVIDIA has enabled Toyota to tap into NVIDIA accelerated computing, AI software and simulation technologies to develop safer, more intelligent vehicles, optimize automotive engineering workflows, fine-tune factory operations and power urban intelligence systems, in support of the company’s vision for safer mobility. 

“Physical AI will bring intelligence to every moving machine from cars, robots and trucks to the cities and factories they operate in,” said Rishi Dhall, vice president of automotive at NVIDIA. “Together, Toyota and NVIDIA are building the AI infrastructure for a new era of mobility, where vehicles can become more autonomous, manufacturing more AI-defined and urban environments more intelligent, responsive and safe.”

NVIDIA and Toyota’s latest work spans:

• Accelerating safe, intelligent vehicles: Toyota is building next-generation vehicles with advanced driver assistance capabilities using NVIDIA DRIVE AGX running the safety-certified NVIDIA DriveOS operating system. These vehicles will deliver L2++ functionality, enabling more intelligent, context-aware driving while maintaining Toyota’s rigorous safety standards.
• Software engineering: As vehicles become increasingly software-defined, Toyota is accelerating vehicle software engineering with a MISRA-compliant Code Assistant AI model, trained and fine-tuned using NVIDIA Megatron-LM, and referencing various datasets including NVIDIA Nemotron. By applying a custom automotive AI model to improve automotive-specific code generation and review, Toyota engineers can generate, review and validate safety-critical code more efficiently, accelerating development while adhering to stringent automotive compliance.
• Factory simulation: Toyota is bringing simulation to the manufacturing floor using NVIDIA Omniverse libraries and the NVIDIA Isaac Sim open framework for factory and robotics workflows, robot movement simulation and broader digital twin environments to optimize manufacturing operations. This simulation-first approach reduces downtime, improves efficiency, lowers costs and enables continuous optimization across production environments.
• Multimodal Vision Language Model: Woven by Toyota (a Toyota subsidiary), has developed Woven City AI Vision Engine, a multimodal vision language model for urban traffic intelligence, using NVIDIA H100 Tensor Core GPUs and Megatron-Core. The model is designed to help interpret real-world conditions, anticipate what happens next and support responses across mobility and infrastructure systems. 

Wednesday, July 15, 3 a.m. PT

NVIDIA and SEGA Celebrate 30 Years of Innovation, Bringing ‘VIRTUA FIGHTER CROSSROADS’ and Other Legendary SEGA Games to NVIDIA RTX Spark

NVIDIA and SEGA are celebrating more than three decades of collaboration by bringing VIRTUA FIGHTER CROSSROADS and future SEGA titles to NVIDIA RTX Spark — a new superchip for slim Windows laptops and compact desktop PCs. 

This builds on the companies’ long-standing relationship, which began 30 years ago when NVIDIA worked with SEGA on burgeoning graphics technology for arcade systems and gaming consoles — with the NVIDIA NV1 chip powering the first Virtua Fighter title on PC, among the world’s first 3D fighting games.

SEGA will support RTX Spark, giving gamers new ways to experience SEGA’s iconic franchises, including the upcoming VIRTUA FIGHTER CROSSROADS. 

Announced from the heart of Akihabara, a global gaming technology hub, at the original SEGA Akihabara Arcade (now GiGO Akihabara 3), VIRTUA FIGHTER CROSSROADS coming to RTX Spark reinforces the companies’ commitment to innovation and shows a glimpse of the future of gaming on a new era of Windows PCs designed for personal agents, AI, creating and gaming.

NVIDIA founder and CEO Jensen Huang joined SEGA CEO Haruki Satomi; SEGA chief operating officer Shuji Utsumi; Yu Suzuki, creator of Virtua Fighter; and former SEGA President Shoichiro Irimajiri, at the birthplace of countless arcade memories to celebrate the milestone. 

They showcased how technology partnerships can evolve across generations of hardware and software, connecting the gaming industry’s heritage with its future. 

The expanding NVIDIA RTX Spark ecosystem — including SEGA and other industry leaders — will offer gamers new experiences harnessing NVIDIA ray tracing, DLSS and AI technologies, while preserving and celebrating the iconic franchises they know and love.

Learn more about NVIDIA RTX Spark.

Editor’s note: This post is part of the Nemotron Labs blog series, which explores how the latest open models, datasets and training techniques help businesses build specialized AI systems and applications on NVIDIA platforms. Each post highlights practical ways to use an open stack to deliver real value in production — from transparent research copilots to scalable AI agents.

Enterprises have plenty of powerful models to choose from. The real test is whether the AI an enterprise builds uniquely addresses the needs of the business: improving workflows, tapping into domain knowledge and exceeding standards for accuracy and trust. 

Increasingly, competitive AI advantage comes from how organizations build with available models, more than which one they choose. 

Open models like NVIDIA Nemotron are built for customization — helping enterprises and nations build AI that’s controllable, trustworthy and tailored to their needs. 

From Using AI to Owning Intelligence

Specialized AI, such as autonomous agents and applications, are built with customized open models. These agents are built to do a defined task well, as the models used are tuned on proprietary knowledge and evaluated against real business outcomes.

That requires access to the model itself. Closed models advance what’s possible and continue to push forward the frontier of general intelligence, but also set a ceiling on what enterprises can inspect, tune and improve. Open models remove that barrier — providing complete ownership and control.

The most effective agentic AI applications are systems of models where open models work alongside leading frontier models, each fulfilling the job it does best. High-performance reasoning models can handle complex planning while smaller models execute on specialized tasks. This lets enterprises right-size inference costs, improve accuracy on specific tasks and maintain flexibility as workflows evolve. 

Customization Enterprises Can Trust

Open models give enterprises something closed models cannot: full control to customize, inspect and improve AI against business needs. Public benchmarks measure general capability — but business-specific evaluation lets teams test against their own data, workflows and definition of accuracy — then improve from there.

For example, the cost of a wrong answer is high for industries like healthcare and legal, where teams handle sensitive data and face strict accuracy requirements. Organizations in these sectors must have visibility into how a model was trained, how it performs and the ability to improve it when necessary. 

With open models, teams can inspect their applications, run private evaluations against their own criteria and stand up reinforcement learning environments tuned to their own workflows. No routing of their proprietary data through a third party is required.

Companies across industries are already specializing Nemotron for their domains:

• Abridge is customizing Nemotron to build the first foundation model purpose-built for clinical conversations.
• Glean built Waldo, an agentic search model that pairs Nemotron with larger closed models to deliver enterprise search at significantly lower latency and with fewer tokens.
• H Company built Holotron 3 Nano by post-training Nemotron 3 Nano Omni on proprietary computer-use data, achieving higher than 76% accuracy on OSWorld-Verified — a benchmark on computer tasks — and matching other leading frontier models at a fraction of the cost.
• Harvey post-trained Nemotron 3 Ultra on its legal benchmark and reached frontier-class accuracy — matching leading closed models on complex legal tasks at at least 10x lower cost per run.
• Heidi Health is delivering frontier-quality outcomes in clinical documentation without needing frontier-scale compute.
• YTL AI Labs post-trained a Nemotron model for the Malaysian language, putting locally customized AI in the hands of Malaysia’s developer community to further its AI capabilities.

Fine-Tuning Environments and Optimal Run Costs

Customization improves accuracy. When models are tuned for a specific harness or domain, they run more efficiently too. 

The NVIDIA NeMo suite of open libraries accelerates model customization and evaluation, in addition to agent optimization and governance. 

Partners like Prime Intellect and Unsloth are already enabling AI customization for enterprises building post-training pipelines on Nemotron, making it practical to run specialized AI at scale. 

LangChain tuned its Deep Agents harness for Nemotron 3 Ultra — adjusting prompts, tools and middleware, with no model retraining — and achieved top agent accuracy among open models at approximately 10x lower cost per run than leading closed alternatives.

Those cost advantages extend to infrastructure for optimal scalability. By post-training Nemotron on the NVIDIA Blackwell platform, Arcee AI achieved inference costs of roughly 90 cents per million output tokens — approximately 20x cheaper than comparable closed frontier models — while ranking second on PinchBench and remaining fully open weight.

Cost savings enable broader experimentation, more deployments and faster iteration.

Ecosystem Building on an Open Foundation

The shift from AI adoption to AI ownership is underway. The NVIDIA Nemotron Coalition is helping turn open model development into an ecosystem effort, bringing model builders and developers together to improve Nemotron through shared data, evaluations and domain expertise. In addition, hackathon submissions and community contributions generate reusable proof assets across industries.

Builders are adding Nemotron to their AI systems, proving value and sharing what works. The foundation is entirely open.

Learn more about NVIDIA Nemotron open models and try them at build.nvidia.com.

Power is AI infrastructure’s inescapable constraint. How many tokens an AI factory can generate within a fixed power budget determines its revenue and profitability. Because of this, performance per watt — a metric that can’t be gamed, only earned through real-world results — is the foundation for AI factories. 

As agentic AI drives token demand higher, the infrastructure decisions organizations make today will determine who scales and who doesn’t in a power-constrained world.

Virtually every frontier AI model today runs on a mixture-of-experts (MoE) architecture. Serving these large-scale models efficiently means GPU domain size — the number of GPUs connected over an ultrafast, scale-up interconnect — matters, and bigger is better. 

While the NVIDIA Hopper generation set the standard with an eight-GPU domain, the scale of frontier AI today has outgrown it. Serving MoE with a 72-GPU domain demands full-stack codesign and the operational depth earned from running these models under real production load.

With the NVIDIA Blackwell NVL72 platform, that foundation is already built and proven, delivering the highest performance per watt to maximize revenues and the lowest token cost to maximize profit margins. It’s this foundation that the NVIDIA Vera Rubin platform builds upon next to further elevate rack-scale energy efficiency.

Maximizing Performance per Watt for Frontier AI 

Each new generation of frontier models brings architectural changes that unlock greater intelligence while demanding new optimizations to run efficiently at scale. 

Across the newest generation of leading open models, NVIDIA GB300 NVL72 delivers up to 25x performance per watt compared with the NVIDIA Hopper generation — showcasing that MoE performance improves when moving from an 8-GPU to 72-GPU domain size. These numbers reflect where Blackwell stands today, a starting point that continues to improve. 

Any single number only tells part of the story. Different workloads demand different operating points: some optimize for latency, others for throughput and cost — and most need to move between the two. 

To best represent these operating points, NVIDIA showcases Pareto curves for each model rather than a single point and provides tools such as DynoSim to help teams find their optimal point on the Pareto frontier before spending a single GPU-hour on validation.

The performance per watt NVIDIA Blackwell delivers is a result of extreme codesign: every component of the rack-scale system, from silicon to software, designed together to maximize token throughput for AI inference workloads. That codesign touches every layer of the stack.  

For example, NVIDIA NVLink Switch, critical for rack-scale performance, is purpose-built to unlock massive scale-up GPU domains, not adapted from general-purpose networking. Now in its sixth generation with the Vera Rubin platform, its capabilities are designed specifically for AI workloads such as SHARP, which performs in-network computing directly in the switch, offloading work from the GPUs themselves.

NVIDIA’s inference software stack, including NVIDIA Dynamo and TensorRT LLM, as well as SGLang and vLLM, is built to run the full range of optimizations: NVFP4 quantization, disaggregated serving, large-scale expert parallelism, KV-aware routing, KV cache offloading and more. These stack together to multiply the performance each GPU delivers. Moreover, software keeps improving performance over time: On DeepSeek V4, performance per watt improved by up to 5x in a single month.

In AI factories, power lost to cooling and rack-level inefficiencies can mean only about 60% of the electricity pulled from the grid turns into useful AI work. NVIDIA DSX MaxLPS, the power-and-efficiency software in the NVIDIA DSX platform, closes that gap by shifting power between GPUs and racks in real time, supporting warm-water liquid cooling and using techniques like power steering to wring more performance. This enables operators to run up to 40% more GPUs within the same power budget.

Production Is Where It Counts

Rack-scale reliability at AI factory scale is hard-won. Rack-scale systems introduce failure modes that single-node deployments never encounter, and handling them requires engineering rigor and time in production.

NVIDIA Blackwell NVL72 systems continues to set the standard across a diverse range of models and production use cases delivering sustained performance, rack-level reliability and economics that hold under real traffic day after day. 

That’s why leading AI labs such as Anthropic, OpenAI and SpaceXAI use NVIDIA Blackwell NVL72 systems to run inference.

In addition, a variety of inference service providers and AI natives use the Blackwell platform to deploy open models in production.

CoreWeave has deployed Kimi K2.6 on NVIDIA GB300 NVL72, combining NVFP4 quantization and EAGLE3 speculative decoding to maximize inference performance. 

Perplexity runs Qwen3 235B and post-trained Qwen3.5-397B-A17B on NVIDIA GB200 NVL72 for its AI agent platform, serving millions of queries daily with the latency and reliability that consumers need.

Fireworks AI deploys GLM 5.2 on the NVIDIA Blackwell platform, enabling production deployments for customers including Cursor and Factory AI.

This accumulated production experience, built across generations of frontier models and real-world deployments, is what gives NVIDIA Vera Rubin its head start.

Learn more about the NVIDIA Vera Rubin platform in this technical blog and find details on the NVIDIA DSX AI factory-scale platform and DSX MaxLPS.

This GFN Thursday brings more games, more power and more ways to play on GeForce NOW. 

The cloud gaming service is expanding with a new GeForce RTX 5080-powered server in Toronto, bringing dedicated high performance in the cloud closer to members across the region.

NTE: Neverness to Everness also gets an update in the cloud, making it even easier to jump into the latest content from the supernatural adventure without a single download or more storage space needed. It leads the way for GeForce NOW bringing native touch control to the game, coming soon.

There’s even more to explore with three new games joining the GeForce NOW library this week. 

Take It to the Next Level in Toronto

The forecast is looking cloudy in Canada.

A new GeForce RTX 5080-powered GeForce NOW server is coming to Toronto, expanding service in the region and bringing dedicated cloud gaming performance closer to local members. The new server will roll out within days, giving more players access to top-tier cloud gaming across Canada.

Ultimate members can stream across PCs, Macs, handhelds, mobile devices, TVs and more with GeForce RTX 5080-class power in the cloud. Enjoy up to 4K resolution and beyond on supported ultrawide displays, up to 120 frames per second, plus NVIDIA DLSS, ray tracing and NVIDIA Reflex technologies. 

Bring on the ‘999 Nights’

Step into the surreal world of NTE: Neverness to Everness with the NTE Version 1.2 “999 Nights” update. 

This version introduces a massive gameplay evolution, plunging players into an immersive, tabletop-inspired fantasy role-playing game on the Warren Continent — a new permanent game mode featuring its own dedicated progression system.

This narrative and mechanical expansion is elevated by the debut of two powerful characters, Shinku and Iroi, alongside an unprecedented fashion upgrade featuring a sweeping collection of 19 new character outfits.

To top it off, exploration gets a high-octane upgrade with Draco, a revolutionary new motorcycle vehicle, making this version an absolute playground for combat strategy, stylish customization and high-speed urban traversal. 

Plus, GeForce NOW will soon be rolling out native touch controls to NTE: Neverness to Everness, which will make it even easier to explore the city’s mysteries from supported mobile devices.

Look for the game in the GeForce NOW app to seamlessly jump between devices and continue the adventure — no downloads, storage space or expensive additional hardware required.

Embark on Expanded Adventures

The journey continues in Granblue Fantasy: Relink with the Endless Ragnarok expansion. Known for its dynamic combat, diverse roster of Skyfarers and thrilling online co-op, Relink returns with fresh story content as mysterious beings known as the Ragnalia threaten the Zegagrande Skydom. Strange gateways, powerful new foes — including the mighty Beelzebub — and fresh challenges await across the skies.

The expansion also introduces additional ways to play, including summon abilities that add another layer of combat strategy, co-op quest tiers and a solo mode filled with unpredictable encounters. Master traits offer more opportunities to customize favorite characters, giving both longtime Skyfarers and newcomers plenty of reasons to take flight.

Jump into the latest action with the following games to play this week:

• Esports Manager 2026 (New release on Steam, available July 6)
• Assassin’s Creed Black Flag Resynced (New release on Steam and Ubisoft Connect, available July 9)
• Granblue Fantasy: Relink – Endless Ragnarok Demo (Steam)

As GeForce NOW continues to expand, so does the community discovering cloud gaming.

One new Ultimate member summed up their first impressions:

“Finally caved and got GFN Ultimate… and wow.”

Another longtime PC gamer shared how GeForce NOW completely changed their perspective after years of skepticism, calling it “a good alternative” for casual players and encouraging others to “try the one month and upgrade to the yearly before the sale ends.”

What are you planning to play this weekend? Let us know on X or in the comments below.

NVIDIA Nemotron 3 Ultra is offering leading performance at lower cost than top closed models with the largest and most widely adopted AI agent orchestration platform. 

LangChain tuned its Deep Agents harness for NVIDIA Nemotron 3 Ultra, achieving the highest accuracy among open models, while completing more tasks at higher throughput and running at 10x lower inference cost per run than leading closed models. 

Measured against LangChain’s Deep Agents benchmark, Nemotron 3 Ultra also achieved business task parity with the highest-scoring closed models. No model retraining was required. Every gain came from engineering the environment around the model, not the model itself. 

At a tenth of the cost, teams harnessing NVIDIA Nemotron 3 Ultra can run evaluations continuously, experiment faster and build specialized agents across more of their business. 

LangChain’s agent engineering platform has more than 200 million monthly downloads. By tuning its Deep Agents harness specifically for NVIDIA Nemotron 3 Ultra, it allows for high-performing agents that complete more tasks, run faster and give enterprises a fully open stack they can customize, own and run anywhere.

“The way to build better agents is to keep improving the system around the model,” said Harrison Chase, cofounder and CEO of LangChain. “Memory, tool use, evaluation and model behavior compound when teams can tune them together. Our work with NVIDIA shows that enterprises can get strong performance from an open stack while keeping control over the agent systems they are building.”

Abridge, Amdocs and Box are embedding specialized agents directly into their platforms and global systems integrator EY is expanding its NVIDIA implementation capabilities around NVIDIA NemoClaw blueprints for LangChain Deep Agents, helping clients customize, evaluate and govern specialized agents across high-value workflows. 

NVIDIA founder and CEO Jensen Huang recently sat down with Chase to discuss why the last six months have seen a leap in useful AI for enterprises.

Harness Engineering, Not Fine-Tuning

LangChain’s team ran Nemotron 3 Ultra against its public Deep Agents benchmark suite, then analyzed the deep agent’s execution traces to find exactly where it lost points. Instead of retraining the model, the team tuned the harness around it — adjusting system prompts, tool descriptions and middleware.

Every developer using LangChain Deep Agents with Nemotron 3 Ultra can put this to work today — the tuned profile is available directly through LangChain.

An Open Stack Built to Own

NVIDIA NemoClaw for LangChain Deep Agents is the open reference blueprint that packages this work for enterprises building their own specialized AI — systems of models, tools and runtime — tuned for their own workflows. It combines LangChain Deep Agents Code, tuned for Nemotron 3 Ultra, with the NVIDIA OpenShell secure runtime for executing agent actions safely.

An open model, an open harness and an open secure runtime means enterprises own the full stack, end to end. They can customize it around the expertise that sets their business apart, keep improving it and run it anywhere — their own infrastructure, their own cloud, their own governance. 

That distinction matters more as agents take on higher-stakes work. The shift from AI assistants that answer questions to agents that take action inside core systems changes what businesses get from their AI. 

NemoClaw for LangChain Deep Agents and the tuned Nemotron 3 Ultra model profile are available now. Developers can pull the tuned Deep Agents harness directly from LangChain, or use the NemoClaw for LangChain Deep Agents blueprint as a starting point for building specialized agents from scratch. 

How to Get Started

LangChain developers can access Nemotron 3 Ultra on Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius and Together AI  platforms, giving them a direct, hosted path to the tuned harness in production.

EY can help enterprises start building their own specialized agents today, using this open software stack.  

Learn more about NVIDIA NemoClaw for LangChain Deep Agents and NVIDIA Nemotron. 

Stay up to date on agentic AI, NVIDIA Nemotron and more by subscribing to NVIDIA news, joining the community, and following NVIDIA AI on LinkedIn, Instagram, X and Facebook.  

Explore self-paced video tutorials and livestreams.

Max single-threaded CPUs at scale are a new category of CPUs built for the agentic AI era. 

Across the creation and deployment of an agentic system, the CPU is on the critical path for reasoning, response time and learning. CPUs are the processor which executes the work the AI model commands: the tool calling, code execution, data processing, KV-cache and result analysis. 

For agents in AI factories, speed matters. 

The faster the CPU can run the tool, the faster the agent can perform the task at hand. 

For the AI factory, the utilization of GPU is the most valuable resource in the data center so any time waiting for a task to complete constrains the revenue of an AI factory — or worse, impacts the GPU utilization waiting for the CPU to finish its task. AI factories need a CPU with max single-threaded performance to maximize AI factory revenue and agent performance.

Today’s data center CPUs are not designed for speed at scale. 

While the world has fast CPUs for PCs and workstations, data center CPUs have been evolving in directions away from single-threaded performance. The advent of the cloud has pushed CPU makers to build higher core-count CPUs while minimizing cost at the expense of performance.  

Building CPUs that optimize costs per rentable core increased the number of cores per chip while taking away silicon area from what makes those cores run fast — like high-performance memory fabrics and faster instruction processing per core. The move to chiplet architectures further reduced cost but created a “chiplet tax” where each CPU’s cores can no longer can get access to the full memory performance of the chip.

AI agents need a CPU designed for max single-threaded performance at scale.

A max single-threaded CPU at scale keeps each agent step fast while the system is fully loaded. Every core completes the agent task at full performance without other cores slowing it down. Max single-threaded CPUs at scale are designed differently to deliver:

• Strong performance per core under load
• Enough memory bandwidth per core to keep active cores supplied with data 
• Predictable latency  

Every core can finish its task without any other core slowing it down, delivering excellent throughput and, more importantly, the fastest possible single-core task performance possible.

NVIDIA Vera exemplifies this new class of CPU design. 

How Max Single-Threaded CPUs at Scale Are Built to Run the Agentic Loop

An AI agent doesn’t stop running after a single request. It acts in a loop. The model reasons about the next step. The CPU executes the work around the model. The result comes back. The model decides what to do next. Then the loop runs again. 

That pattern creates a demand profile for which conventional CPUs were not optimized. Traditional CPU work is intermittent and user-driven, made up of short interactions triggered by people. Agentic work is persistent and parallel: swarms of agents running continuously, each advancing through a chain of steps where each step depends on the result of the one before it.

More cores in a CPU means more agent tasks per CPU, and data center CPUs need lots of cores to maximize throughput of tasks.

However, adding more cores to a CPU cannot shorten the time for each step inside a single agent loop. More cores can’t make any one task run faster. In fact, CPUs designed to maximize core count can even slow down the performance of each core as they contend for resources.   

Individual per-core performance matters to drive the speed of each step’s completion. The throughput of additional cores is useful but insufficient. And since each action is dependent on the previous result, per-core speed determines how fast the loop advances.

In the end, the best agentic CPU needs the best single-threaded performance per core, and every core needs to deliver that performance without compromise. The world counts in seconds. Agents count in nanoseconds. NVIDIA Vera is built for this new category — and speed — of work.

NVIDIA Vera Is the Max Single-Threaded CPU at Scale for Agents

NVIDIA Vera is a max single-threaded CPU at scale, designed from the ground up for the agent loop: the work that happens between model calls as agents use tools, process data, run code and check results.

At the core of Vera is Olympus, NVIDIA’s custom CPU core, which delivers 50% higher instructions per cycle than NVIDIA Grace. That matters because many agent steps are sequential. A tool call, code execution, test run or data-processing step must finish before the next model call can use the result. Faster cores move each loop forward faster.

Vera pairs those faster cores with up to 1.2TB/s of LPDDR5X memory bandwidth at less than 40 watts of memory power, plus a monolithic compute die that helps active cores stay fed and keeps data movement predictable with 3.4TB/s of core-to-core bandwidth, 3x greater than any other data center CPU. This enables all 88 cores with the full memory performance of the CPU without creating bottlenecks that slows down every core.

The result is faster agent loops. In loaded CPU workloads that represent agentic execution, Vera delivers 1.8x the sustained per-core performance of x86.

Those gains compound across tool calls, code executions, data-processing steps and verification passes, helping AI factories complete more agent work with the GPUs they already operate.

Perplexity tested Vera on the agentic work it runs every day. Running a real coding workflow — cloning a repository and running its test suite in sandboxes — Vera completed the job about 1.5x faster than x86, and started concurrent sandboxes up to 1.9x faster. Perplexity is now looking to deploy Vera in its upcoming production system. 

Agents also depend on data. They query, retrieve, filter and move information constantly, and Vera runs those CPU-side data workloads faster. Partners have measured 3x faster large-scale SQL analytics with Starburst and up to 6x lower latency on real-time streaming with Redpanda, both against leading x86 server CPUs.

Agent work isn’t one workload. An agent runs tools and sandboxes, processes data, serves requests and trains the next model with reinforcement learning — and all of it leans on the same strengths.

One Vera handles the whole range, rather than requiring a different CPU for each kind of work. And because Vera is the same CPU that hosts the GPUs in NVIDIA Vera Rubin and powers the NVIDIA BlueField-4 STX storage processor, the whole AI factory runs on one architecture and one toolchain.

And NVIDIA’s not done. NVIDIA’s next-generation Rosa CPU with the Rigel core will continue the company’s CPU roadmap for the agentic AI era. Rigel is NVIDIA’s next-generation Arm v9.2 CPU core, delivering higher per-core performance than Olympus while keeping the same silicon footprint. Key improvements include better instruction delivery, a larger L2 cache and more efficient memory handling.

Built for the Speed of Agents

In the agentic AI era, there will be billions of agents, and every one of them will turn to a CPU to act, check, retrieve, execute and verify. In this new market, completed agent work is the product. Faster agent loops help every GPU spend more time generating revenue producing work and less time waiting.

NVIDIA Vera is the CPU built for that future.

Learn more about the NVIDIA Vera CPU.

Open source AI has shown how quickly developers can innovate when models, data and tools are shared. Robotics has the same opportunity, but advancements in physical AI development can still be gated by costly and fragmented resources, from large datasets and robot foundation models to simulation, compute and validation tools.

NVIDIA and Hugging Face are collaborating to bring the NVIDIA Isaac GR00T 1.7 open, reasoning vision language action (VLA) model for humanoid robots and the NVIDIA Isaac Teleop framework to LeRobot — Hugging Face’s open source library for robotics — with NVIDIA Cosmos 3, a frontier model for physical AI, planned soon. Together, these integrations give developers a more accessible and standardized path for end-to-end robot development while driving innovation and collaboration across the open robotics community.

“Open source is how a field turns advanced research into something people can study, adapt and build on,” said Thomas Wolf, cofounder and chief science officer at Hugging Face. “With NVIDIA Isaac GR00T 1.7 and Isaac TeleOp in LeRobot today, robotics developers can use shared models, data and workflows to train and evaluate robots in the open. And with NVIDIA Cosmos 3 planned next, the community will have a path to bring frontier world models into that same collaborative loop.” 

An Open Pipeline for Robot Foundation Models

Hugging Face LeRobot is an open source robotics library for training, running and sharing robot datasets, models, policies and workflows. NVIDIA’s continued partnership with Hugging Face connects NVIDIA’s 3 million robotics developers with Hugging Face’s 16 million AI builders, expanding access to frontier physical AI tools through open workflows.

Bringing NVIDIA physical AI capabilities into LeRobot gives developers a common way to collect and standardize data, train and fine-tune robot foundation models, evaluate performance and deploy models through open workflows. The integrations include:

• NVIDIA Isaac Teleop, an open source framework for robot data collection, helps developers capture high-quality human demonstrations from external devices using standardized, interoperable formats, then expand and share datasets with the community, all directly in LeRobot. 
• NVIDIA Isaac GR00T 1.7, the first open and commercially viable robot foundation model, makes it easier to post-train and deploy models through LeRobot workflows, helping developers adapt GR00T to new robot embodiments and tasks with benchmarked performance.
• NVIDIA Cosmos 3, a frontier world foundation model for physical AI coming soon to LeRobot, will help developers generate and augment robotics data, simulate scenarios and support policy development when real-world data is limited or too expensive to collect.

These integrations build on a broader set of NVIDIA resources already connected to LeRobot to support the full robotics development loop, including: 

• The largest open source physical AI dataset, downloaded more than 15 million times, which includes more than 350,000 real and simulated trajectories and 57 million grasps to help developers kickstart their robotics workflows.
• NVIDIA Isaac Sim– and Isaac Lab-based simulation frameworks to help developers set up environments, generate robot data, test policies and validate behaviors before moving to physical robots.
• NVIDIA Isaac Lab-Arena in LeRobot Environment Hub to enable developers to quickly prototype complex simulation environments, register them in LeRobot EnvHub and seamlessly use them within the LeRobot ecosystem to train and evaluate generalist robot policies such as GR00T, Pi and SmolVLA.
• NVIDIA Jetson Thor integration with LeRobot’s Reachy 2 to support deployment of VLA models on open source humanoid robots.

Learn more about how to use Isaac Teleop, Isaac GR00T 1.7 and Isaac Lab-Arena with LeRobot for end-to-end humanoid development and explore detailed LeRobot integration workflows.

Every year, the International Conference on Machine Learning (ICML) reveals where thousands of AI researchers have decided to put their work. 

This year’s accepted papers reveal a clear direction: open frontier models and open AI infrastructure have become foundational to how modern AI science gets done.

NVIDIA had 74 papers accepted at ICML 2026. Approximately 2,000 accepted papers cite NVIDIA GPUs, and 145 cite NVIDIA Nemotron — a family of open models, including open datasets — as the foundation for new research. Hundreds more draw on NVIDIA Cosmos, NVIDIA Isaac GR00T, BioNeMo and other NVIDIA open model families, spanning physical AI, robotics, autonomous vehicles and biomedical research.

The Themes Defining This Year’s Research

Areas including vision and video generation, reinforcement learning for large language models (LLMs) and agent training as well as AI inference remained prominent themes across this year’s papers, reflecting sustained investment these fields command — while several new areas also broke through.

Robot world models drew significant attention, with papers like DreamDojo pushing the boundary of how AI systems learn to reason about and act in physical environments. DreamDojo, for example, learns how the physical world behaves from human video and builds on NVIDIA Cosmos open frontier models to predict how a robot would handle objects and operate in environments it was never trained on. It lets researchers evaluate policies, plan actions and teleoperate a virtual robot, accelerating development without the costs and risks of physical deployment.

AI for life sciences was fueled by NVIDIA BioNeMo open models and research contributions that help researchers understand protein function, molecular behavior and genetic code. Papers like FLIP2 introduce public benchmarks for testing how well AI predicts the effects of protein mutations. KERMT is a new BioNeMo open model for predicting molecular properties important to drug discovery. 

Synthetic data generation (SDG) drew particular interest at ICML this year with several Nemotron and physical AI open datasets, reflecting a broader shift in how researchers are thinking about training at scale without relying solely on human-labeled data.

The Open Research Stack

Open infrastructure gives researchers the tools to accelerate breakthroughs. 

The papers show Nemotron being used less like a single model release and more like a research stack: open weights to evaluate against, open datasets to train and adapt with, and open recipes for reasoning, tool use, safety, data curation and efficient inference.

Alongside the models, NeMo Curator and the open datasets it supports gives researchers a reproducible foundation for training data curation. SDG tools enable creating high-quality training sets at a scale and speed that would’ve been impractical just a few years ago.

The Cosmos 3 family of open, frontier omnimodels gives researchers and developers a generational leap in the ability to build robots, autonomous vehicles and vision AI that perceive, reason, plan and act in the physical world.

In addition, the NVIDIA Alpamayo open model family for autonomous vehicle development, NVIDIA Isaac GR00T for robotics and NVIDIA BioNeMo for biomedical help accelerate research and development across industries.

The Ecosystem Building on Top

The momentum extends beyond NVIDIA’s own research labs.

Basecamp Research developed a new DNA foundation model, EDEN, that helps researchers interpret and design genetic sequences.

Merck & Co., uses KERMT to predict how potential drug molecules may behave in the body, including whether they are likely to be effective, safe and developable.

Sakana AI — attending ICML this year — built its Fugu and Fugu-Ultra models directly on Nemotron 3 Ultra, using the open foundation to push forward its work on AI research automation.

KiloCode integrated Nemotron into its code-routing architecture, reporting token cost reductions of up to 90% — a result with real implications for the economics of deploying AI in production.

NAVER developed its own model using the Nemotron architecture, extending the foundation for Korean-language AI research.

Together AI is hosting Nemotron models on its platform, making them more accessible to researchers who need reliable, seamless access to open inference.

Humanoid, LG Electronics, NEURA Robotics and Noble Machines are adopting NVIDIA Isaac GR00T  models to accelerate industrial deployments of their humanoids, while 1X, Agility, Agile Robots, Boston Dynamics, Hexagon Robotics, and Mentee are building the next generation of humanoids using Cosmos world models, Isaac Sim and Isaac Lab to accelerate the development and validation of their robots.

Explore NVIDIA’s open models on Hugging Face.

Explore genomics and biology research at ICML’s GenBio Workshop on Friday, July 10.

Nations have long invested in domestic infrastructure to advance their economies, protect and use their data, and take advantage of technology opportunities in areas such as transportation, communications, commerce, entertainment and healthcare.

AI, the most important technology of our time, is turbocharging innovation across every facet of society. Countries are investing in AI capabilities so they can design, train and deploy models and applications, using domestic infrastructure, local datasets and homegrown expertise. This approach ensures AI solutions are tailored to local citizens, services and regulations.

Why AI Capabilities Matter

The urgency for countries to build and deploy AI capabilities has grown with the rise of generative and agentic AI, which is reshaping markets, inspiring new industries and transforming existing ones — from gaming to healthcare. It’s changing how people work, as many professions now use AI-powered copilots.

These efforts span physical infrastructure and data infrastructure. On the data side, countries are developing foundation models, such as large language models, built by local teams and trained on local datasets. This helps reflect regional dialects, cultural context and specific domains in the models’ outputs.

For example, speech AI models can help preserve, promote and revitalize indigenous languages. 

Large language models are not only used to understand and generate human language; they can also write software code, aid in drug discovery, help protect consumers from financial fraud, teach robots physical skills and much more.

As AI and accelerated computing become increasingly important for tackling climate change, boosting energy efficiency and defending against cybersecurity threats, national AI capabilities play a critical role in enabling every country to strengthen its resilience and sustainability.

Factoring In AI Factories

A new class of essential infrastructure for AI production has emerged: AI factories, where data comes in and intelligence comes out. These are next-generation data centers that host advanced, full-stack accelerated computing platforms for the most computationally intensive tasks.

Countries are building domestic computing capacity through various models. Some are procuring and operating AI clouds in collaboration with state-owned telecommunications providers or utilities. Others are sponsoring local cloud partners to provide shared AI computing platforms for public-private use.

“The AI factory will become the bedrock of modern economies across the world,” NVIDIA founder and CEO Jensen Huang said in a media Q&A.

Ingredients of a National AI Strategy

There are five ingredients of a national AI strategy:

• AI Imperative: Domestic AI capabilities are critical to economic growth, national security, cultural preservation and innovation — with responsible, trustworthy AI aligned to local policies as well as national goals.
• AI-Ready Workforce: A wide spectrum of local AI skills and talent, plus basic AI literacy across the population. Education is important at all levels, from early STEM programs through applied AI across industries.
• AI Models and Data: Foundation models and large language models trained and fine-tuned with local data, hosted and run on local infrastructure, subject only to local laws. The localization of models will ensure that AI factory outputs are fine-tuned to the language, culture and context for the intended workloads. 
• AI Ecosystem: A local ecosystem of AI investors, developers, scientists, entrepreneurs, enterprise customers and government organizations. 
• AI Factories: Highlighted above, AI factories are locally owned, operated and governed AI clouds for training and inference. The greatest utility of AI factories comes from the public-private partnerships that can scale infrastructure to meet the needs of growing innovation within countries and industries.

National AI Strategies Underway

Countries around the world are investing in AI capabilities tailored to their national needs. AI investments can help grow economies while delivering tangible social and environmental benefits for citizens.

Since 2019, NVIDIA’s AI Nations initiative has helped countries in every region build out their AI ecosystems and workforce, creating the conditions for engineers, developers, scientists, entrepreneurs, creators and public sector officials to pursue their AI ambitions at home.

In Europe, AI agents from ThinkDeep, built on the NVIDIA AI platform, are helping France’s Ministry of Economy and Finance automate complex public‑service workflows by processing millions of documents and data sources, cutting document search times from two days to two minutes, saving 2 million euros for 10,000 employees and reducing energy use through more efficient, in‑country infrastructure control. 

In Asia, India’s Sarvam platform, powered by NVIDIA GPUs and built entirely on domestic infrastructure, is delivering multilingual AI models and voice agents optimized for the country’s 22 official languages, enabling government and enterprise services to reach hundreds of millions of people in their own languages while keeping data, compute and governance under national control. 

In Latin America, AI solutions from Widelabs, running on NVIDIA‑accelerated infrastructure, are helping modernize and expand access to legal services for the Public Ministry of Rio Grande do Sul in Brazil, streamlining internal investigations and making justice records easier to find and use for more than 8 million citizens across nearly 500 municipalities — aligning advanced computing with more efficient, transparent and inclusive public administration. 

These efforts demonstrate how AI turns domestic infrastructure, local data and homegrown talent into solutions that can be used for social good.

Learn more by joining NVIDIA at the AI for Good Summit, running July 7-10 in Geneva, Switzerland, and read more about NVIDIA’s commitment to trustworthy AI.

Summer is heating up — and GeForce NOW is taking players along for the ride.

Start the month with Monopoly: Star Wars Heroes vs. Villains, bringing a galaxy far, far away to the iconic board-game franchise, alongside 12 new games joining the cloud this month. 

Plus, don’t let the sun set on the biggest GeForce NOW savings of the year. Level up for less before the deals disappear.

Light Side, Dark Side, Cloud Side 

Rule the board, you must. Choose a side in Monopoly: Star Wars Heroes vs. Villains, the classic property-trading board game reimagined with legendary characters, locations and rivalries from across the Star Wars universe.

Play as iconic heroes or infamous villains — each with unique abilities — to assemble a team and experience cinematic moments while competing with family and friends across locations from every era of the saga. Every roll of the dice brings new opportunities to build an empire and claim victory.

GeForce NOW makes it easy to take the battle between the light and dark sides across nearly any device. Jump into a match on a low-powered PC, Mac, phone, TV, tablet or handheld device and keep the fun going across the galaxy.

Check out what’s available this week:

• Monopoly: Star Wars Heroes vs. Villains (New release on Steam and Ubisoft, available June 30)
• Meccha Chameleon (Steam)

And look forward to the games coming throughout the month:

• Assassin’s Creed Black Flag Resynced (New release on Steam and Ubisoft Connect, available July 9)
• Denshattack! (New release on Steam and Xbox, available on Game Pass July 15)
• The Mound: Omen of Cthulhu (New release on Steam, available July 15)
• Heave Ho 2 (New release on Steam, available July 16)
• Fogpiercer (New release on Steam and Xbox, available on Game Pass July 17)
• ZeroSpace (New release on Steam, available July 20)
• The Planet Crafter (New release on Xbox, Available on Game Pass July 21)
• Carnival Hunt (New release on Steam, available July 23)
• The Ranchers (New release on Steam, available July 30)
• Corsair Cove (New release on Steam and Xbox, available on Game Pass July 31)

Juicy Extras From June

In addition to the 18 games announced last month, 10 more came to the cloud. 

• Deer & Boy (Steam)
• DOOM Eternal (Epic Games Store)
• Embers of the Uncrowned Demo (Steam)
• EMPULSE (Steam)
• The Elder Scrolls Online (Xbox, available on Game Pass)
• NBA THE RUN (Steam)
• SAND: Raiders of Sophie (Steam)
• Voidling Bound (Steam)
• Witchspire (Steam)
• World of Tanks: HEAT (Wargaming)

Last Call for Cloud Summer Savings

The final days of the GeForce NOW Summer Sale are here. Before the savings disappear, gamers can save $35 on a 12-month Performance membership or $70 on a 12-month Ultimate membership — unlocking GeForce RTX-powered gaming in the cloud across devices they already own.

The Performance membership delivers smooth, high-quality gaming with RTX-powered servers, making it easy to jump into favorite titles across PCs, Macs, phones, handhelds and TVs.

The Ultimate membership takes cloud gaming to the max with RTX 4080‑ or 5080‑class performance. Experience cinematic visuals, ultralow latency and responsive gameplay powered by technologies like NVIDIA DLSS, ray tracing and NVIDIA Reflex — all without the cost of a new gaming rig.

Hear directly from the GeForce NOW Community.

One GeForce NOW member recently called the Summer Sale “quite significant” after realizing the savings were even larger than expected in their local currency. By locking in a year of Ultimate, they calculated their monthly cost dropped from roughly 29 CAD to 17 CAD – showcasing how GeForce NOW continues to help gamers around the world enjoy the games they love, wherever they choose to play.

Plus, check out this spreadsheet, made by a community member, featuring discounted games streaming on GeForce NOW  and build out a bigger library at the best bargains during the Steam Summer Sale.

What are you planning to play this weekend? Let us know on X or in the comments below.

As AI moves from model development to production inference, compute demand is accelerating and shifting toward continuously operating AI factories that generate tokens at scale. This shift requires access to large‑scale, multi‑tenant accelerated computing that can come online quickly, stay highly utilized and support the economics of token‑scale AI services. 

Emerging AI companies historically have had limited access to capital-intensive infrastructure, with even long-term commitments insufficient to unlock financing for compute.

To address this, NVIDIA is introducing a new business model that opens up compute access to the fast‑growing AI ecosystem of startups, model builders, enterprises, research organizations and regional AI players. 

This new model enables AI clouds to procure NVIDIA infrastructure for AI-native, enterprise and ISV customers through economic alignment with a revenue-sharing and credit-support model. Through the partnership, AI clouds will sell NVIDIA-powered cloud services, with NVIDIA earning both standard product revenue and a share of the cloud revenue on the supported capacity. This structure accelerates adoption of NVIDIA platforms among the high-growth, high-conviction AI native sector, and provides NVIDIA with a recurring, usage-linked earnings stream.

For model builders, inference providers, agent platforms and enterprises scaling AI, it can mean faster access to full-stack accelerated computing without waiting through site selection, power procurement, construction and hardware bring-up.

NVIDIA AI Factory Capacity Built Around Demand

The initiative is already taking shape, with AI cloud companies building DSX AI factories designed to serve customers and workloads across regions. 

Sharon AI and Firmus are among the first companies to work with NVIDIA on this new business model. 

Sharon AI is deploying up to 40,000 NVIDIA Grace Blackwell GB300 GPUs.

“This strategic collaboration with NVIDIA marks a pivotal moment in Sharon AI’s mission to deliver sovereign, large-scale AI compute infrastructure,” said James Manning, cofounder and CEO of Sharon AI. 

Firmus is building a DSX AI factory campus in Batam, Indonesia. The campus is expected to scale to 360 megawatts and up to 170,000 NVIDIA GPUs.

“AI-native companies need access to scalable, energy- and cost-efficient compute infrastructure to compete globally,” said Tim Rosenfield, co-CEO of Firmus Technologies. “Firmus AI cloud is building a NVIDIA DSX-aligned AI factory, which will enable our cloud to help more customers access the compute they need to build and scale AI.”

AI natives such as Baseten, Fireworks AI and Together AI show where compute demand is headed: they need immediate access to AI cloud capacity to run model training, post-training, fine-tuning and high-volume agentic inference for developers, digital natives and enterprises building with AI.

Their customers need reliable access to large-scale NVIDIA accelerated computing as usage grows, but they also need commercial flexibility as products move from pilot to production. 

To secure compute capacity and build and deploy AI models, contact Sharon AI and Firmus. 

 Learn more about NVIDIA Cloud Partners and AI factories. 

Life sciences has entered an era of computational scale, and for more than a decade, NVIDIA has built the full GPU-accelerated computing stack — spanning hardware, frameworks, libraries, models, microservices and domain-specific tools — to help researchers run more sophisticated workflows and iterate faster.

This week, Anthropic announced Claude Science, an AI workbench for science research that lets scientists converse with agents in natural language to run their work end to end.

Claude Science integrates with NVIDIA BioNeMo Agent Toolkit as a resource that scientists can access within their workflow. The toolkit packages NVIDIA-accelerated capabilities as callable skills, enabling Claude Science to select the appropriate tool, prepare valid inputs and execute the workflow — all while connecting to NVIDIA compute resources deployed anywhere. This brings NVIDIA’s accelerated models, libraries and NVIDIA NIM microservices directly into the same environment where the rest of the research happens.

The world’s largest pharmaceutical companies use NVIDIA technologies to advance AI-enabled research across drug discovery, genomics, medical imaging, molecular design and protein engineering. Today, 18 of the top 20 pharmaceutical companies use NVIDIA BioNeMo, underscoring the breadth of its role across the ecosystem.

Advancing the Agentic Era of Scientific Discovery

Claude Science lets scientists use natural language to move their research from intent into action, without manually configuring models, endpoints, or software environments. NVIDIA BioNeMo Agent Toolkit extends that with access to accelerated workflows and models like Evo 2, Boltz-2 and OpenFold3, so the analyses that benefit from acceleration run faster. 

A scientist begins by describing a research task, such as analyzing a genomic sequence, predicting a protein structure or designing a potential binder, in natural language. Claude Science interprets the request and orchestrates the work through preconfigured domain-specialized agents that know established workflows across genomics, proteomics, single-cell analysis, cheminformatics and clinical research. 

BioNeMo Agent Toolkit gives these agents the context needed to connect each step with an appropriate NVIDIA scientific capability. Each skill includes information about its purpose and required inputs, helping agents prepare and execute the workflow and return outputs for review.

The result is an iterative loop between scientific reasoning and accelerated computational work. Scientists can inspect outputs, refine their questions and determine the next step while staying focused on the science.

One powerful example is generating better inhibitors of common cancer targets. In this workflow, a scientist starts with a known cancer-causing antigen mutation and asks Claude to design numerous potential inhibitors. Claude Science integrated with BioNeMo Agent Toolkit and NVIDIA NIM microservices accelerates high-throughput inhibitor prediction, optimization and validation.

A Scientific Foundation Built for Agents

AI agents reason, plan and use tools to complete tasks. In life sciences, those tools are often specialized computational workflows. 

An autonomous AI scientist agent doesn’t reason in isolation. It may need to fingerprint a library of compounds, cluster promising hits, generate conformers for top candidates, analyze genomic context and compare perturbation responses before recommending the next experiment. 

Each step relies on a scientific tool, and the agent can only work as fast as those tools run.

NVIDIA BioNeMo Agent Toolkit gives scientific agents the accelerated tools they need to operate at the speed of science. It includes:

• NVIDIA Parabricks accelerates genomic analysis from hours to minutes, so an agent can integrate genomic context into a decision in near real time.
• RAPIDS-singlecell, developed by scverse, compresses a 1.3-million-cell preprocessing and clustering workflow from 52 minutes to 25 seconds, so single cell analysis becomes part of the reasoning loop rather than an offline batch of jobs.
• nvMolKit accelerates cheminformatics operations like similarity search and conformer generation by up to 3,000x, so an agent iterating across a massive chemical space gets results at the speed of thought.
• NVIDIA BioNeMo open models deliver core biomolecular capabilities accelerated by NVIDIA libraries, so an agent has a purpose-built scientific model for each step of a workflow.
• BioNeMo NIM microservices package those models as enterprise-ready inference endpoints — containerized microservices with the full accelerated software stack pre-integrated and tuned for high-performance inference — so an agent can call a single stable application programming interface for production deployment.

NVIDIA BioNeMo Agent Toolkit is open and harness-agnostic, allowing the same scientific skills to work across agent frameworks and research platforms. The toolkit and its skills are available now through NVIDIA developer resources and GitHub.

Scientists can access BioNeMo-powered workflows through Anthropic’s Claude Science, which is entering public beta today. As part of the public beta, Anthropic is inviting researchers to provide feedback on additional domain specialists and integrations they need.

As organizations move from AI pilots to production AI factories, infrastructure decisions have shifted from peak chip specifications to cost per token: how many useful tokens they can deliver per dollar, per watt and within required latency targets.

Codesigned with NVIDIA GPUs, CPUs, networking and systems, and strengthened by a broad open source ecosystem, NVIDIA’s full-stack inference software continuously improves hardware performance. On the NVIDIA Blackwell platform, the software stack has already reduced token costs by up to 5x on the DeepSeek V4 model in just one month.

Leading companies and inference providers are already seeing the compounding value of NVIDIA’s inference software stack on Blackwell: 

• Baseten used the NVIDIA TensorRT-LLM open source library to serve DeepSeek V4 Pro on Blackwell GPUs for reasoning, coding and long-context workloads, applying proprietary runtime optimizations to deliver up to 50% more tokens per second.
• Cognition is using the NVIDIA Dynamo inference framework to manage inference GPUs, giving its team a ready-made path to scale reinforcement learning workloads without needing to build that infrastructure from scratch. 
• Deep Infra uses the NVIDIA inference software stack to serve frontier open source models performantly on Blackwell from day zero, including DeepSeek V4. 
• DigitalOcean helped Hippocratic AI use NVIDIA inference software on Blackwell GPUs to serve healthcare AI faster and more efficiently, increasing inference throughput by 30% while maintaining a sub-half-second time to first response across 10 million patient calls.
• Together AI used NVIDIA TensorRT-LLM on Blackwell to help Cursor accelerate the path from model optimizations to production endpoints for its real-time coding experience. 

Why Software Matters for Inference Economics

Traditional web, search and software-as-a-service workloads were relatively predictable: A user might load a page, refresh a feed or update a business record. These requests typically followed similar software paths, reading from or writing to a database, and scaled by adding more of the same servers. 

Agentic AI is different.

Agents can reason, plan, call tools, spin up specialist subagents and manage massive context across multi-turn workflows. They turn a single request into a distributed computing problem that can span hundreds of subagents, thousands of tasks and multiple large language models, running across GPUs, CPUs, DPUs and storage systems. 

The software stack determines whether that complexity turns into wasted capacity or lower cost per token.

Lower cost per token comes from turning individual optimizations into system-level performance. NVIDIA’s inference software stack does this by connecting three layers: 

• Production Operation: Coordinates distributed serving, orchestration, autoscaling and memory management so inference can run across the right compute and storage resources.
• Application Acceleration: Runs models with high performance while giving developers room to tune and customize, using runtime optimizations such as overlapping compute and communication and kernel fusion.
• Infrastructure Access: Exposes NVIDIA GPU, networking, memory and system capabilities without requiring developers to manage every device instruction set or data-transfer protocol directly.

When these layers work as one system, individual optimizations compound.

Disaggregated serving, large expert parallelism over NVIDIA NVLink interconnect technology, NVFP4 precision and multi-token prediction each deliver meaningful gains on their own. Combined, they increase throughput by up to 20x.

The chart below shows the result. Capturing that gain in production is complex, requiring coordination across the full inference stack — from production operations and model runtimes to kernels, communication libraries and hardware access. NVIDIA’s inference software stack is designed to make those layers work together so each optimization can build on the others. 

Open Source Amplifies the Full-Stack Advantage

That same full-stack foundation is amplified by the open source ecosystem. Many of today’s most widely used open source AI frameworks and inference projects are built natively on NVIDIA CUDA, which means new research and software optimizations run with leading performance on NVIDIA GPUs from day zero.

PyTorch is a leading example. Launched in 2016 with native CUDA support, PyTorch has coevolved with NVIDIA’s architecture, giving developers access to innovations such as Tensor Cores, Transformer Engine and NVFP4 directly through a familiar framework. 

When breakthroughs such as DFlash speculative decode, which delivers up to 15x more throughput on existing hardware, or FastVideo, which generates 1080p videos in less than five seconds, land in PyTorch, they can run instantly on NVIDIA, helping AI factories convert research progress into lower token costs.

The same open source momentum is why when a new frontier open model like DeepSeek V4 is released, leading inference frameworks like vLLM and SGLang have day-zero deployment recipes for the NVIDIA Blackwell architecture — making the model accessible across millions of Blackwell GPUs. It’s also why DeepSeek V4 performance on Blackwell improved by up to 5x within about a month across vLLM and SGLang frameworks, cutting token costs to roughly one-fifth of previous levels.

That’s the open source flywheel: more developers optimize CUDA-native inference paths, more production deployments feed back into the ecosystem and each software improvement increases delivered token output while lowering cost per token over time.

Explore how software multiplies hardware performance in this NVIDIA AI Podcast on tokenomics and this inference solutions page. 

When Jaiveer Singh talks about robots, he doesn’t begin with spectacle. He begins with infrastructure: the boards inside machines, the software that lets developers see through a robot’s cameras and the engineering required before a robot can leave a demo floor to do something useful.

As a robotics software engineer who leads the team behind NVIDIA Isaac ROS (Robot Operating System), Singh works on the connective tissue of the physical AI era. Built on the open source ROS 2 framework, Isaac ROS brings CUDA-accelerated libraries and AI models to developers building autonomous mobile robots, manipulation systems and humanoids. 

“My goal is to make sure everyone feels like they are a part of the robotics future,” Singh said.

For Singh, that future began in middle school, building with LEGO Mindstorms, a popular line of programmable robotics kits. After excelling in robotics competitions throughout high school, he studied electrical engineering, computer science and business at the University of California, Berkeley, before joining NVIDIA full time after an internship with the robotics team.

In a satisfying turn, the work he now leads began as his intern project.

“We wanted to see what would happen if we just released some software as open source that uses the NVIDIA Jetson platform and NVIDIA CUDA libraries for robotics. Would there be any value there?” Singh recalled. “And the answer was, of course, yes, because developers always want to be able to unlock the full power of their GPUs.”

The result was Isaac ROS.

The Building Blocks of a Robotics Revolution

Physical AI has long been a field of extraordinary imagination and stubborn, physics-bound realities. A clip of a robot dancing or executing complex balletics can travel the internet in hours. Building a system that works repeatedly, across sensors, platforms, factories and labs, is slower business. 

For Singh and the Isaac ROS team, the next era of robotics relies on a full stack: simulation, training, accelerated computing, AI models, middleware and edge deployment.

Isaac ROS supports manipulation, mobility and humanoids. It gives developers packages for perception, object detection, mapping, collision detection and motion planning, and it can run on workstations, NVIDIA DGX Spark personal AI supercomputers as well as NVIDIA Jetson edge systems. 

“Compared with the original Isaac SDK, Isaac ROS is completely modular,” Singh said. “We ship the software like a bunch of LEGO bricks — you get to assemble them however you want, and you can easily combine our packages with existing ROS code written by you or others in the global robotics community.”

NVIDIA is making it easier for many robot builders to move faster, Singh said, and to do so on a foundation they can inspect, adapt and trust.

“The main reason open source is valuable is because it gives people confidence that they can build upon this stack at this very initial stage,” Singh said. “Because the entire landscape can shift so rapidly, developers need the confidence that this platform is still going to be there to modify and improve two or three years into the future.”

That confidence matters because robotics is changing quickly. Humanoid robots, in particular, have moved from science fiction to an active engineering frontier.

Singh’s team has been making Isaac ROS better suited to this moment, including for developers using AI agents and for humanoid systems that need an end-to-end software stack.

NVIDIA’s long history of work in robotics and farsighted vision for the field is what initially attracted Singh to the company — and made him all the more confident in his work upon joining.

“NVIDIA was here and working on this problem before anybody else thought it was important,” he said. “We already had a stake in the ground.”

Open source, in Singh’s view, is a way of sharing both confidence and responsibility. If a robotics startup builds on a closed system, it must trust that the system will still match its needs years later. With open software, developers can inspect the code, change it, contribute fixes and carry it forward. One company’s bug fix becomes another company’s acceleration.

“When more people can build robots,” Singh said, “the future gets here faster.”

Follow @nvidialife on Instagram and learn more about NVIDIA life, culture and careers. 

Editor’s note: This post is part of Into the Omniverse, a series focused on how developers, 3D practitioners, and enterprises can transform their workflows using the latest advances in OpenUSD and NVIDIA Omniverse.

Vision AI agents are becoming a practical way to automatically turn video data from the physical world into operational intelligence in factories, cities, warehouses and transportation systems. 

That shift is accelerating as more AI workloads move closer to where data is generated. Gartner projects that more than two-thirds of enterprise-managed data will be created and processed outside the data center or cloud by 2028, and that over two-thirds of all enterprises globally will deploy edge AI by 2029, up from 10% in 2025 (1).

But more edge data doesn’t automatically create more intelligence. As much as 90% of existing edge data goes unprocessed, according to the same Gartner report.

Turning that data into useful action requires vision AI agents that can understand video, adapt to real-world conditions and connect insights to operational workflows. These agents often run near cameras, machines and sensors, where models must meet latency, power, cost and connectivity requirements while adapting to site-specific conditions.

To build those agents, developers need repeatable ways to generate training data, fine-tune models and deploy agentic video applications across edge and cloud environments.

NVIDIA Metropolis agent skills and blueprints give developers reusable workflows to build, operate and optimize vision AI agents across that lifecycle. 

For the simulation and synthetic data side of that work, Universal Scene Description, or OpenUSD, provides a common framework for describing, composing and reusing 3D worlds. Built on OpenUSD, NVIDIA Omniverse libraries help teams build simulation, synthetic data generation and digital twin workflows that model real-world environments and expand scenario coverage across conditions such as lighting, weather, traffic patterns, camera angles, occlusion and rare events.

Where Vision AI Agent Projects Can Get Stuck

As organizations move toward autonomous vision agents, three challenges often come up:

• Accuracy Plateaus With Data Gaps: Vision AI agents need to spot rare defects, abnormal events and changing environments. In manufacturing, for example, an inspection model may perform well on common scratches or dents but struggle with a new hairline crack not represented in the training data.
• Lack of Fine-Tuning Expertise: Once teams identify a performance gap, improving the model is rarely a simple handoff. Fine-tuning requires labeled datasets, training configuration, experiment tracking, evaluation and decisions about whether there’s improvement for the target use case. Many organizations building vision AI agents don’t have large in-house machine learning teams to manage that process quickly, especially across many sites, products or camera views.
• Complex, Time-Consuming Agent Assembly Workflows: Deploying a vision AI agent requires more than running inference. Developers have to stitch together video pipelines, AI models, metadata, embeddings, indexing, search, alerts, reporting and system integrations. Customizing that workflow for a specific environment adds significant time and requires specialized expertise. Without OpenUSD’s shared scene description layer, teams often rebuild 3D environments from scratch each time conditions or deployment sites change.

A Full-Lifecycle Approach to Vision AI Agents

NVIDIA agent skills and blueprints — used alongside NVIDIA Omniverse for OpenUSD-based simulation and synthetic data generation, NVIDIA Metropolis for model development and video AI deployment — give developers reusable starting points for key parts of those workflows: 

• The Defect Image Generation skill helps create synthetic defect data.
• The Video Data Augmentation skill helps expand scenario coverage.
• NVIDIA TAO skills enable model fine-tuning.
• NVIDIA video search and summarization (VSS) skills help turn video understanding into deployable workflows for alerts, reporting, stream management and more.

Instead of rebuilding every step from scratch, developers can use these reusable workflows to generate data, improve models and deploy vision AI agents faster.

Visual Inspection: Generating the Data That Production Lines Don’t Have

In manufacturing, the more successful a factory is at preventing defects, the harder it becomes to collect enough defect examples to train the next inspection model.

Roboflow is integrating the NVIDIA Defect Image Generation skill and NVIDIA Cosmos world foundation models into its vision AI platform to generate synthetic defect images for customers like Corning when real training data is scarce, enabling near-perfect detection performance while significantly reducing the need for daily manual image review. 

In a benchmark conducted with Corning’s optical fiber manufacturing engineering team, a model trained on just eight real defect images — augmented with synthetic data generated by the NVIDIA Defect Image Generation skill — reached an average precision of 95% and perfect recall on the most challenging defect class. This performance surpassed a baseline model trained solely on real data, effectively compressing a multi-quarter inspection project into just a few days.

Watch how synthetic data generation workflows help developers create the data needed to train and improve physical AI models:

Smart Cities: From Video Analytics to Autonomous Operations

Large-scale city operations show why vision AI agents need connected workflows, not just inference.

Linker Vision is building smart city AI systems with the NVIDIA Metropolis Blueprint for VSS to accelerate the deployment of video reasoning agents across city infrastructure. In this workflow, VSS skills can help package common video AI tasks such as search, summarization, alerts, reporting and stream management into reusable agent-executable workflows. 

OpenUSD-based NVIDIA Omniverse digital twins help model city environments and test how vision AI systems respond to varied traffic patterns, weather conditions, emergency events and infrastructure changes. Linker Vision uses NVIDIA Cosmos for video data augmentation and NVIDIA TAO for Cosmos model fine-tuning.

In Kaohsiung, Linker Vision reduced development effort by 85% using the VSS blueprint and reduced incident response times by up to 80%. Its newer AI-GRID expansion builds on this approach with NVIDIA NemoClaw blueprints for secure agentic AI, supporting autonomous video reasoning across city and transportation environments.

Industrial Operations: Reasoning Over Work as It Happens

In industrial environments, the challenge isn’t just detecting what appears in a video frame. Teams need agents that can: 

• Understand whether work is being performed correctly
• Compare execution against standard operating procedures
• Produce insights before defects move downstream.

At Foxconn, DeepHow’s Live Standard Operating Procedure (SOP) Verification agent uses the NVIDIA Metropolis VSS blueprint as the agentic video workflow layer for search, summarization and analysis across operational environments. NVIDIA Cosmos provides the reasoning capability that helps the agent interpret complex human activity and work sequences in context, such as whether assembly steps are performed correctly and in the expected order.

The solution has been used on the NVIDIA GB300 server production lines to improve first-pass yield by 3%, achieve 99% task-level accuracy in micro-action understanding of critical SOP steps and reduce redundant work by helping teams catch problems earlier.

To see how developers can build and deploy video analytics AI agents, watch this technical walkthrough on using NVIDIA VSS skills with coding agents.

Explore NVIDIA agent skills and blueprints to build, operate and optimize video analytics AI agents. 

Source: Gartner, Predicts 2026: Physical AI Pushes I&O to the Edge, 3 March 2026. Gartner is a trademark of Gartner, Inc. and/or its affiliates.

Anthropic’s Claude models in Microsoft Foundry — hosted on Microsoft Azure and running on NVIDIA GB300 Blackwell Ultra GPUs — are now generally available, giving Azure-native enterprises a powerful new way to build autonomous and domain-specific AI agents.

As agentic AI continues to drive enterprise innovation and becomes more autonomous, organizations need access to computing power to build and deploy specialized agents to accelerate essential business tasks. And having great inference performance and efficiency reduces total cost of ownership and drives positive company results.

With Claude in Foundry running on NVIDIA GB300 NVL72 systems with NVIDIA Quantum-X800 InfiniBand networking, enterprises can now build and run more powerful agentic systems, including autonomous and specialized sub-agents that can work across business domains to perform advanced tasks. 

A Growing Partnership

NVIDIA is working with Anthropic to extend developer capabilities by integrating NVIDIA tools into the Anthropic stack. That integration enables enterprises to give Claude agents domain-specific abilities. Through NVIDIA verified agent skills, enabled by access to NVIDIA accelerated computing, enterprises can embed AI agents deeply into their business and use them as the operating system for the organization.   

Enterprises can run Claude agents on Azure by using the NVIDIA Secure Agent Workspace Reference Design. It provides a blueprint for running autonomous agents in a governed environment where identity, network access, credentials and runtime policy are controlled at the infrastructure level.

Claude in Microsoft Foundry accelerated by NVIDIA GB300 GPUs on Azure builds on the strategic partnership Microsoft, NVIDIA and Anthropic announced in November to expand enterprise access to Claude and offer Anthropic models on NVIDIA accelerated computing. 

Get started by visiting Claude in Microsoft Foundry and learn more in Foundry documentation.