A new creature-catching adventure is ready to stream from the cloud this week. Pawprint Studio’s Aniimo arrives on GeForce NOW at launch, inviting gamers to explore the vibrant continent of Idyll across supported devices.

Also this week, 007 First Light receives a path-tracing update on GeForce NOW, alongside a smashing limited-time Deluxe Edition sale on Steam and Epic Games Store.

Gaijin Network’s fractured-multiverse action game Active Matter and Annapurna Interactive’s acclaimed space mystery Outer Wilds are also among 11 new titles joining the cloud this week.

Explore Idyll With the Cutest Companions

The path to Idyll begins now. Aniimo is a free-to-play, open-world creature-catching role-playing game where every creature encountered can become a companion. Catch Aniimo with Aniipods, then Twine with them to take on their form and use unique skills to solve puzzles, win battles and overcome challenges.

Glide, dive and burrow across Idyll, then return to a personal RV to build a warm, interactive Homeland alongside Aniimo companions. Stream Aniimo on a Steam Deck, in the newly supported Firefox browser and across any other supported devices — without waiting through its 45GB download or making room in local storage. There’s always room in the Homeland for one more Aniimo.

Choose Your Path

In the critically acclaimed, multimillion-selling 007 First Light, follow James Bond as a young, resourceful and sometimes reckless recruit in MI6’s training program, and discover a reimagined origin story of the world’s most famous spy. Incorporating IO Interactive’s signature stealth gameplay with world-class Bond action, players can embark on missions in breathtaking locations around the globe, drive iconic vehicles and dive into a cinematic adventure in pursuit of a rogue agent who’s always one step ahead.

Path tracing has now been added to 007 First Light, introducing extra-detailed lighting, shadows and reflections, making its levels and set pieces even more cinematic, immersive and realistic. And NVIDIA DLSS 4.5 Ray Reconstruction ensures the fidelity, clarity and accuracy of these additions are at their absolute best.

Ultimate members can stream with GeForce RTX 5080-class performance, with up to 5K high dynamic range and cinematic-quality streaming. DLSS 4.5 Super Resolution and Dynamic Frame Generation accelerate frame rates and enhance image quality.

A limited-time opportunity awaits: the 007 First Light Deluxe Edition is on sale on Steam from Sept. 15-29 and on Epic Games Store from Sept. 3-18. The mission is on sale. The getaway car is in the cloud.

Matter of Time

Gaijin’s Active Matter, a realistic military shooter set in a fractured multiverse, has launched on GeForce NOW. As an operative stuck in a time loop, join dangerous raids for loot or intense player vs. player battles.

Fight against rivals from other timelines, survive physics-breaking anomalies and try to stay alive. Harvest active matter, gather loot and extract to a safe place before the whole zone ceases to exist.

Ultimate members can take on each raid with GeForce RTX 5080-class performance in the cloud, delivering high frame rates, advanced graphics features and low latency across supported devices. The zone won’t wait — and neither does the next loop.

Let’s Get Loopy

Outer Wilds joins the GeForce NOW library this week. As the newest recruit of Outer Wilds Ventures, search for answers across a strange, ever-changing solar system trapped in an endless time loop. Gamers can trace mysterious signals, decipher alien writing and discover hidden locations before an underground city is swallowed by sand or a planet crumbles beneath their feet.

There’s even more to stream this week:

• Active Matter (New release on Steam and Gaijin, Sept. 15)
• Aniimo (New release on Steam, Sept. 15)
• RuneScape: Dragonwilds (New release on Xbox, Sept. 15, available on Game Pass)
• Train Sim World 7 (New release on Steam, Sept. 15)
• The Guild 1 Remake: Europa 1410 (New release on Steam, Sept. 17)
• The Blood of Dawnwalker (GOG.com)
• Mixtape (Steam)
• Neon White (Steam)
• Outer Wilds (Steam)
• Stray (Steam)
• What Remains of Edith Finch (Steam)

Dates listed above reflect when games are released on their respective stores. GeForce NOW availability may vary, as games are onboarded after they’re released and added throughout the week. Keep an eye on GeForce NOW channels and GFN Thursdays for availability updates to announced titles.

Ready to take this week’s new games for a spin? Start with a day pass to try premium cloud gaming before committing to a membership. Even better, the cost of the day pass can be applied toward a first monthly membership — making it easy to level up with GeForce RTX-powered cloud gaming.

What’s on the playlist this weekend? Let us know on X or in the comments below.

System performance, efficient infrastructure scaling and continuous software optimization are key levers that determine AI inference economics. Higher system performance means more tokens generated, resulting in higher revenue. Efficient scaling means throughput grows proportionally as hardware gets added, requiring fewer resources to serve users at scale. Continuous optimization means generating more value from infrastructure investments. 

Underlying all three is platform fungibility: the same infrastructure runs any model, any workload, from training to inference, recommender to reasoning, language to video, keeping utilization high.

The NVIDIA platform is purpose-built to optimize across all these, as highlighted by MLPerf Inference v6.1 results released today:

• NVIDIA Vera Rubin NVL72 system debuts with leading performance: In its first MLPerf Inference preview submission, NVIDIA Vera Rubin NVL72 delivers up to 3.7x better throughput than GB300 NVL72.
• NVIDIA GB300 NVL72 scales with leading efficiency: A 288-GPU submission across four GB300 NVL72 racks achieved 99% scaling efficiency, with throughput growing nearly linearly from a single-rack baseline.
• Continuous software optimizations drive performance gains: Software optimizations in NVIDIA’s MLPerf Inference v6.1 submissions delivered up to 1.6x higher performance over v6.0. Optimizations continued post-v6.1 submission, delivering further performance gains.

For organizations making AI infrastructure decisions, performance, scaling efficiency and software velocity are important considerations that determine long-term inference economics. 

Vera Rubin NVL72 Makes MLPerf Inference Debut With Leading Performance

NVIDIA submitted Vera Rubin NVL72 preview results on two of the most demanding benchmarks in the MLPerf Inference v6.1 suite: DeepSeek-R1 and Qwen3-VL. 

Vera Rubin NVL72 delivers up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL across offline, server and interactive scenarios, using vLLM with the NVIDIA Dynamo open source inference framework. On DeepSeek-R1, using the NVIDIA TensorRT-LLM library, throughput is up to 2.5x higher than GB300 NVL72. These early results showcase NVIDIA’s accelerated pace of innovation and how performance will improve with continuous software optimizations. 

This performance means each Vera Rubin NVL72 rack delivers significantly more tokens, serves more users and generates more revenue than a GB300 NVL72 rack, while lowering cost per token.

The results reflect full-stack codesign across hardware and software. Vera Rubin’s enhanced Tensor Cores and Transformer Engine accelerate both the prefill and decode stages of inference, while NVFP4 precision reduces memory footprint across model weights, attention and KV cache — increasing throughput with minimal loss of output quality. 

Vera Rubin submissions heavily used disaggregated serving, separating prefill and decode along with large-scale expert parallelism for maximum efficiency across the mixture-of-experts layers that power models like DeepSeek-R1 and Qwen3-VL. 

The NVL72 scale-up domain — powered by sixth-generation NVIDIA NVLink and NVLink Switch to deliver 10x higher packet rates and 3x lower latency than off-the-shelf Ethernet — provides the interconnect foundation that makes these techniques effective at rack scale. 

This codesign extends to NVIDIA’s partner ecosystem: Nebius also submitted Vera Rubin NVL72 preview results and demonstrated excellent performance.

AI agents, which reason, plan and act across multiple steps, are reshaping how inference performance is measured. In benchmarks designed to capture this shift, such as SemiAnalysis AgentX, Vera Rubin NVL72 delivered 30x better performance than GB300 NVL72 in preview testing. In addition, the upcoming MLPerf Endpoints benchmark will bring standardized measurement to agentic inference workloads, beyond what traditional throughput benchmarks capture.

NVIDIA GB300 NVL72 Scales With Leading Efficiency

Scaling efficiency — how effectively additional GPUs translate to throughput gains — is a key measure of AI infrastructure productivity. NVIDIA delivers this with high-bandwidth, low-latency scale-up interconnects within each rack, high-bandwidth networking between racks and efficient request orchestration across nodes.

NVIDIA’s DeepSeek-R1 (DSR1) submission scaled from a single GB300 NVL72 rack (72 GPUs) to four racks (288 GPUs), achieving 99% scaling efficiency in the offline scenario. Throughput grew nearly in proportion to the hardware added.

Scaling efficiency is key because more GPUs don’t automatically mean proportionally more throughput. If adding nearly double the GPU count delivered only a single-digit percentage improvement in throughput, the infrastructure cost would far outpace the performance return. The architecture, interconnect and software must all scale together.

GB300 NVL72 also demonstrated rack-scale efficiency on the WAN 2.2 text-to-video benchmark, reaching 0.65 720p videos per second at 5.7 seconds per video — 9x higher throughput and 7.5x lower latency than a single node.

Software Optimizations Drive Continuous Gains

NVIDIA platform undergoes continuous software development, delivering performance and feature improvements. 

In v6.1, GB300 NVL72 performance on Qwen3-VL improved up to 1.6x over v6.0 results. The gains came through lower KV cache precision, additional kernel fusion, better kernels and disaggregated serving with vLLM and NVIDIA Dynamo. 

Software optimization continued past the v6.1 submission deadline as well. Post-submission results, not yet verified by MLCommons, on GPT-OSS-120B and DLRMv3 show further performance gains. 

AI Inference at Every Scale

Beyond the NVIDIA Grace Blackwell and Vera Rubin NVL72 platform results, NVIDIA submitted Jetson AGX Thor results using NVIDIA TensorRT Edge-LLM on the newly introduced Edge-Agentic benchmark with Qwen3.6-27B.

The NVIDIA partner ecosystem participated broadly, with 19 partners — eight of them on multi-node Blackwell NVL72 systems — demonstrating excellent performance. This includes ASUS, Azure, Cisco, CoreWeave, Crusoe, Dell Technologies, Fujitsu, Giga Computing, HPE, Inventec, Lambda, MiTAC Computing, Nebius, Oracle Cloud Infrastructure, Quanta Cloud Technology, Red Hat, ScitiX, Supermicro and Wiwynn.

From compact edge devices to the largest AI factories, NVIDIA continues to advance performance across the full technology stack with an annual cadence of platform architectures, continuously improving software and an ecosystem built to deliver it at scale.

Learn more about the NVIDIA Vera Rubin platform.

AI factories are the infrastructure of the intelligence era. Scaling them responsibly will depend as much on innovation across the grid as inside the data center. 

Today, Emerald AI, Google and NVIDIA announced the launch of the AI Energy Management Alliance (AEMA), a first-of-its-kind coalition advancing data centers that can dynamically manage their electricity use in response to grid conditions.

This power flexibility can help unlock faster, larger connections for AI infrastructure while supporting the energy systems and communities that make its growth possible. Getting more watts out of existing infrastructure reduces environmental impacts per watt and supports energy affordability.

The objective is clear: build AI infrastructure that doesn’t just connect to the grid but works with it.

Flexibility Is an Energy Imperative

Power has become a defining constraint on the expansion of U.S. AI infrastructure.

Traditional interconnection processes were designed around facilities with flat, static electricity demand. They weren’t built for computing infrastructure capable of responding intelligently when the power system is constrained.

A flexible data center can adjust its electricity drawn from the grid in several ways — shifting computing workloads, discharging storage, using paired generation or responding to system contingencies. These capabilities allow a large electricity customer to serve as a controllable resource rather than an inflexible load.

Used effectively, flexibility can make more efficient use of existing grid capacity, reduce demand during periods of system stress, and avoid or defer costly infrastructure upgrades. It can also give utilities and grid operators greater confidence to connect AI facilities on shorter timelines.

Technology-Neutral, Performance-Based Requirements

AEMA is technology-neutral and performance-based. Its focus is on the measurable service a facility can deliver — including response speed, duration, predictability and behavior during an emergency — rather than the specific hardware or software used.

Reliability remains paramount. The alliance’s principles call for:

• Defining ride-through, curtailment and contingency-response obligations before a facility connects — meaning the alliance is setting clear rules for facilities regarding staying connected during brief grid disturbances, reducing power use when needed and responding to emergencies.
• Standardizing technical requirements, performance metrics and operational data sharing.
• Creating faster, risk-adjusted pathways for customers that make credible and verifiable flexibility commitments.
• Allocating interconnection costs in a way that reflects actual system impacts and benefits, such as avoided upgrades and improved ramping capability.

These measures can reduce uncertainty for developers while giving system operators the information and control needed to preserve reliability.

Convening the Full AI and Power Value Chain

AEMA convenes the full value chain across computing and power — including AI platforms, infrastructure providers, data center operators, technology companies, power producers, utilities and regional grid operators.

The founding members will be joined by launch partners from across the ecosystem. Together, AEMA will develop technical and operational approaches, collaborate with utilities on interconnection solutions and advocate for policies that recognize grid-responsive demand.

Accelerating US AI Infrastructure 

AI factories transform energy and data into intelligence. Power-flexible design gives them the potential to support the grid as they do it.

NVIDIA and Emerald AI are already working with energy and infrastructure leaders on AI factories that can respond to grid conditions in real time. AEMA will broaden that work by bringing the technology, energy and policy communities together around models that can be deployed across the U.S.

The rules governing power for AI are being written now. By creating a common framework for performance, reliability and collaboration, AEMA aims to help the U.S. build the infrastructure of intelligence at the speed and sustainability the moment demands.

Learn more about AEMA and membership opportunities.

Air pollution is a serious public health risk, contributing to an estimated 30,000 deaths in the U.K. alone last year. Data-driven insights can help — but computing air quality with traditional chemistry-based models is expensive, which limits how detailed they can be and how regularly they can be run. 

David Topping, a professor in the University of Manchester’s department of Earth and environmental science, saw that the NVIDIA Earth-2 family of open AI models and tools had cracked a related problem for weather forecasting — and asked whether the same generative frameworks could work for pollution fields.

“The biggest challenge is the compute required to forecast air quality,” said Topping. “Once you put chemistry into weather models, they get really, really slow. So I said, why don’t we try using the generative frameworks that NVIDIA develops for climate and weather for pollution fields?”

Working with the NVIDIA Earth-2 team, Topping and colleagues generated training data from existing chemistry-climate simulations, then trained Earth-2 CorrDiff — a generative downscaling model — on Isambard-AI, the U.K.’s national AI supercomputer in Bristol. 

The model worked on the first attempt. 

The team has since added Earth-2 StormCast, a model that enables time-dependent forecasts that directly use air quality observations, and showed the test-training and inference workflows running on the NVIDIA DGX Spark personal AI supercomputer.

“To improve human health, it’s essential that we understand the impact of environmental stressors in the air we breathe,” said Topping. “Our U.K.-wide pollution model allows us to model potential future scenarios, such as predicting what would happen if different pollution-related government policy changes went into effect.

Video credit: Bristol Centre for Supercomputing (BriCS) @ University of Bristol

Another potential application is proactive air quality insights for healthcare organizations. Topping envisions a scenario where regional and national healthcare services could reach out to patients with conditions like asthma to let them know that air pollution is going to be high in their area tomorrow, or next week. 

The team is also exploring how the air pollution model could pair with data from edge AI devices to ingest real-time air quality data and drive real-time decision making, such as in the event of a wildfire. 

“The fact that this model trained in two days on Isambard-AI — and can now run on a DGX Spark sitting on a desk — changes who can do this science and how quickly,” said Niall Robinson, developer relations manager for weather and climate at NVIDIA. “We’re just at the beginning of what these open workflows can do globally.” 

“The ability to switch from one NVIDIA framework to another was really impressive,” said Hao Zhang, a doctoral student at the University of Manchester who trained StormCast on Isambard-AI. “We’re only just starting to explore how to use these frameworks in different ways to model complex pollution fields.”

From National Supercomputer to the Desktop

To retrain the Earth-2 model for air pollution, Topping and the team used a year’s worth of U.K. pollution data simulated at hourly intervals to generate a detailed, U.K.-wide pollution model at a resolution of 2-3 square kilometers.

Running on a single, eight-GPU node on Isambard-AI — the U.K.’s most powerful AI supercomputer, packed with 5,448 NVIDIA GH200 Grace Hopper Superchips delivering 21 exaflops of AI performance — the process took just two days.

“Earth-2 CorrDiff has shown an incredibly efficient use of the world-class NVIDIA hardware inside Isambard-AI,” said Simon McIntosh-Smith, director of the Bristol Centre for Supercomputing at University of Bristol and cofounder of Isambard-AI. “It’s fitting that, for a climate-based project, the GPU hours used were relatively low, requiring less power from the supercomputer to run the workloads.”

In addition to providing a look back at air pollution over the past year, the model can help predict future air pollution scenarios for the U.K. The team plans to further increase the resolution of its model by incorporating additional open data, enabling researchers to understand air pollution at street scale.  

The same generative pollution workflow also runs on the NVIDIA GB10 Grace Blackwell superchip-powered DGX Spark desktop AI system for inference and smaller training runs. Topping now has an DGX Spark system in his office retraining models.

“You can now invest a few thousand dollars to get started developing powerful AI models,” Topping said.

Open Science, Agentic Future

The team plans to release open source training data and workflows for the pollution models will be released so that similar models can be trained for other countries and regions. 

“Our aim is to offer this workflow to the entire world,” he said. “We hope that every global country and every major city with a small burst of supercomputer AI time will be able to produce their own detailed pollution models with their own local data.” 

Looking five years out, Topping sees the endpoint as something simpler still: an agentic interface where a clinician or government agency asks the question and the chain of models handles everything else. 

“With better open access to air quality observations, someone could ask our pollution model running on DGX Spark: what’s the pollution going to be like in this neighborhood tomorrow?” he said. “And a whole chain of interactions will deliver an answer, grounded on the science these frameworks represent.”

Hear more from Topping in the webinar, AI for Science: From the Lab to the Frontier, taking place September 30.

Learn more about NVIDIA Earth-2 climate and weather AI.

Know everything. Do anything.

That was the message NVIDIA founder and CEO Jensen Huang brought to Salesforce Dreamforce Tuesday, joining CEO Marc Benioff onstage in an appearance that coincided with the announcement of Koa — Salesforce’s first CRM reasoning model, built on NVIDIA Nemotron 3 Super.

Huang didn’t just take the stage. He walked into the crowd — threading through the audience at Moscone Center, mic in hand, with Benioff alongside. “We’re all following you, bro,” Benioff told Huang at one point, drawing applause. 

“With electricity, we could power everything,” Huang said, tracing the arc of industrial revolutions. “With the internet, you can find anything. And now, with artificial intelligence as an infrastructure layer across the planet, we can know everything and do anything.”

That arc led directly to the announcement. Salesforce’s first reasoning model was built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset drawn from nearly three decades of enterprise CRM deployments.

AI Safety

When Benioff asked Huang about AI safety, Huang was direct. 

“Safety is paramount in a lot of ways. It’s job one,” Huang said. “However, safety is an engineering problem. We’re developing computing systems after all … If you build a product or a service and you’re not confident in its functionality, capability or safety, then don’t release it.”

He reiterated that innovation and leadership should coincide with safety and security. “Innovation, speed and safe products — it’s a false choice,” Huang said. “You could definitely have both at the same time. So run as fast as you can. But if you feel at any given point in time the company’s out of control, or the product’s not going to be safe, take a pause and make sure you get it right.”

Throughout his talk, Huang offered a positive vision for the future of AI. He noted he was among the first to push back on the notion that AI would end the software industry. He also pushed back on the idea that AI is a threat to jobs. 

“As a result of our ambition, and with the productivity boost we get from AI, the sky’s the limit,” he added. “Engage AI. Don’t get left behind.”

“Every company, every enterprise, every country would become an AI company,” Huang said. “It’s going to be an agentic enterprise.”

“The sky is especially the limit for our Trailblazers,” Benioff said, referring to the company’s customer and developer community. “I think this technology, the way it empowers people … the technology can partner with them to do this in incredible new ways.”

Because NVIDIA Nemotron is open, Salesforce was able to fine-tune and run the model entirely within its own infrastructure. Salesforce controls the weights, the model runs on Salesforce’s systems and no customer data is used during training or inference.

“Not a single byte of customer data was used,” said Rohan Kumar, Salesforce’s president of platform and engineering, who introduced Koa before Huang took the stage.

Open models went from 30% at the beginning of last year to now some 70%, Huang explained.

“People are both adopting closed models at exponential rates, but also building their own custom AIs — because every single software company is an AI company, and every single enterprise is going to be an agentic and AI enterprise.”

NVIDIA is one of them — running on Salesforce across sales, service, marketing and operations, and currently piloting Agentforce for customer-support workflows.

Koa was built using supervised fine-tuning and reinforcement learning with NVIDIA NeMo RL, NeMo Gym and NeMo AutoModel. 

The training corpus covers synthetic enterprise scenarios across 14+ industries — manufacturing, financial services, healthcare and travel. 

In Salesforce’s CRM Bench, a model benchmark that includes a suite of real-world tasks like updating an opportunity, routing a case, or scheduling a follow-up, Koa already matches or exceeds leading model performance on CRM actions with 3x fewer errors.

In Use Now

Koa is already running inside Salesforce, powering an employee agent in Slack, and is moving into customer pilots in October as a customer-selectable model in Agentforce, starting with Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine and Xero. General availability is expected winter 2026 in U.S. regions.

On a sweltering August evening in Silicon Valley, as the sun dropped and air conditioning loads spiked, Silicon Valley Power sent a signal to an AI factory to adjust its power consumption.

Varun Sivaram was watching on Zoom with about forty others — his team at Emerald AI in their San Francisco conference room, engineers at the data center and people from the utility itself. Nobody touched anything.

Emerald AI’s Conductor platform — a grid-orchestration platform from NVIDIA partner Emerald AI, and an early example of the kind of flexibility NVIDIA DSX Flex is built to deliver  — receives signals about grid conditions and adjusts the data center’s flexible computing workloads. Work that can wait is slowed or rescheduled, while higher-priority services continue operating. 

The goal is to reduce electricity demand when the grid is constrained without interrupting critical AI workloads— exactly what Silicon Valley Power needed,

When the reduction showed on screen, everyone cheered.

“We were watching with bated breath,” Sivaram said. “It was our first time deploying across thousands of NVIDIA GPUs.” His head of product, Mansi Shah, was emotional. “This feels kind of like a SpaceX rocket launch,” she said.

Silicon Valley Power has since sent more than 200 demand signals to that AI factory. It worked every single time. 

This is grid flexibility in production. And it points at something much bigger than one facility in Santa Clara: a path to unlocking the power America’s AI factories need, without waiting a decade to build new transmission lines.

At the AI Infra Summit on Tuesday, Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, made AI factory efficiency the centerpiece of his infrastructure keynote. 

Results from cloud provider Lambda’s first validation in a deployment environment, released the same day, put numbers to it: a fixed power budget can support 24% more token throughput when managed intelligently.

“With our proof of concept, we believe we’ve moved beyond the limitation of fixed power budgets,” said Dave Ward, president of cloud services at Lambda. “NVIDIA DSX MaxLPS paves the way to reclaiming stranded capacity and converting it into real-world usage, with significantly more compute density in the same footprint.” 

That August evening, when SVP called, Conductor executed against a predefined workload hierarchy: lowest-priority jobs yielded, high-priority inference kept running, and power fell from four megawatts to three. Automated. No operator required.

In the AI factory economy, power is the constraint. Work per gigawatt is the metric. Data center operators are meticulous about efficiency — every watt put to work is a watt delivering productive compute, and the industry has driven remarkable gains at every layer of the stack, from facility design to rack-level power conversion.

DSX extends that discipline into the AI workload itself. Smarter rack provisioning puts power where workloads actually need it. Operational intelligence — tighter scheduling, faster restarts, leaner checkpointing — keeps GPUs running rather than waiting. The goal is the same one operators have always pursued: more work from the power you have.

“A one-gigawatt factory will never become a two-gigawatt factory,” NVIDIA founder and CEO Jensen Huang has said.

The answer engineers reach when systems hit physical limits is always the same: stop optimizing the parts and start designing the whole. 

Introduced at GTC Taipei in May, NVIDIA DSX is that answer for the AI factory — and the early deployments are already proving it out. 

The full platform spans networking, cooling, water efficiency and facility design; the sections below focus on some of the results so far in power management and grid participation.

More Compute, Same Budget: DSX MaxLPS

Lambda’s results, released at the AI Infra Summit, are the first validation of DSX MaxLPS on NVIDIA HGX B200 GPU Servers. 

DSX MaxLPS monitors GPU and rack-level power consumption and reallocates headroom across nodes based on workload type, recovering capacity that static provisioning would leave stranded. Training and inference draw power differently; MaxLPS optimizes allocation in AI factories running both.

Lambda, a GPU cloud provider serving more than 10,000 customers from AI-native startups to hyperscalers, ran the software on a five-rack, 19-node cluster. 

What they found: by running 19 nodes within the same power budget as 16 nodes at full power, Lambda achieved 24% more cluster-wide token throughput — from roughly 4 million tokens per second to 5 million. Performance per watt improved by 23%.

Based on NVIDIA’s projections, DSX MaxLPS can enable up to 40% more GPU capacity for next-generation Vera Rubin NVL72 AI factories within the same megawatt power budget in suitable deployment environments.

Automated Demand Response, Proven in Production

The Santa Clara story isn’t a DSX Flex installation — it’s something earlier and more important: proof that the concept works at commercial scale.<
NVIDIA’s Eos AI factory is running Emerald AI Conductor as a participant in Silicon Valley Power‘s Flexible Load Interconnect Program, the first commercial grid utility program designed to treat AI factories as dispatchable resources. 

When Silicon Valley Power sends a signal, Conductor responds in under a minute. The factory that’s willing to flex gets to run bigger.

That’s the pattern DSX Flex is built to generalize — with Emerald AI Conductor integrating into DSX Flex as the platform matures. The first dedicated DSX Flex commercial deployment will be the Manassas, Virginia, facility: a 96-megawatt Vera Rubin AI factory at NVIDIA’s AI Factory Research Center, building on five prior demonstrations across two continents.

The Next Power Architecture Layer: 800V DC Power Architecture

The gains inside today’s AI factory are real and deployable now. The next layer is how power is delivered to denser accelerated computing racks.

As AI factories scale, traditional lower-voltage power paths add conversion complexity and distribution constraints.

NVIDIA’s 800 VDC architecture is designed to reduce conversion complexity, improve power delivery efficiency and support denser accelerated computing racks.

NVIDIA DSX is incorporating 800V DC into its reference designs. 

The Whole Factory, Not the Parts

No single component can optimize an AI factory on its own. A faster GPU still waits on the network. Power can be stranded by bad provisioning. Cooling overhead still diverts electricity from GPUs; GB200 NVL72 racks running direct liquid cooling carry ~120 kW of heat that has to go somewhere before that power reaches compute.

The only reliable path to more tokens per megawatt is to optimize the whole factory — DSX Sim before the first rack goes in, DSX OS and DSX Exchange once it’s running, DSX Reference Designs so builders start from a validated architecture rather than from scratch. (See sidebar for the full DSX suite at a glance.)

The Gigawatt Infrastructure Standard

It all comes down to one question: how much useful work does the factory produce per megawatt consumed? 

NVIDIA DSX gives infrastructure builders the reference designs, simulation tools, operational software, and power-management technology to compete on that metric, on current hardware and into the next generation.

When the grid needed relief, the factory gave it without dropping a job, without asking for more power. 

With NVIDIA DSX, that’s the new baseline for what an AI factory is supposed to do.

Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, Tuesday spoke on AI factory efficiency at the AI Infra Summit, the Santa Clara Convention Center event that has morphed into a Coachella of infrastructure tech.

Before a packed audience — with more than 8,000 attendees this year, up from 3,500 last year — Buck discussed new collaborations across NVIDIA platforms and more.

• Amazon’s Annapurna Labs is working with NVIDIA on the NVHBM custom high-bandwidth memory technology.
• d-Matrix is integrating with NVLink Fusion to combine NVIDIA Vera CPUs with d-Matrix Raptor XPUs to deliver ultra low-latency inference at scale.

NVIDIA and partners unveiled new results as well: 

• Emerald AI and NVIDIA demonstrated a commercial AI factory flexible-load program, working with Silicon Valley Power.
• Lambda improved performance per watt by 23% with NVIDIA DSX MaxLPS.
• Pinterest is using the NVIDIA Blackwell platform and NVIDIA Dynamo inference software to bring conversational AI to visual discovery. 

The news comes as agentic AI is driving a new class of workloads that demand more performance, efficiency and scale from AI infrastructure. 

NVIDIA addresses that challenge with a full-stack AI factory platform spanning Vera Rubin systems, Dynamo inference software, NeMo libraries and NVIDIA networking — including NVIDIA NVLink for scale-up computing, Spectrum-X Ethernet and ConnectX SuperNICs for connecting thousands of nodes, BlueField-powered context-memory storage and BlueField DPUs for infrastructure security. 

“Infrastructure that’s fungible, that’s reliable, that’s going to last 10 years and really becomes an asset for computing the world’s computing problems in the world’s industries — they can build that with Vera Rubin, with DSX and all the innovations that we have here,” said Buck.

The metric for AI infrastructure is fast shifting from peak performance to validated agentic tokens per megawatt. AI factories must now be codesigned from silicon to grid. NVIDIA DSX MaxLPS can deliver up to 1.4x more tokens per megawatt through factory-wide power optimization, while NVLink helps unite large-scale accelerated computing into a single high-performance system. 

The result is AI infrastructure designed to generate more tokens, improve efficiency and help customers get more value from every megawatt of power. 

Wednesday, Sept. 16, 8:00 a.m. PT

MLPerf Inference v6.1: NVIDIA Vera Rubin NVL72 Sets New Benchmark for AI Throughput

AI inference economics increasingly come down to three factors: system performance, infrastructure scaling and software optimization. Together, they determine how many tokens can be generated, how efficiently AI services scale and how much value organizations extract from their infrastructure investments.

The NVIDIA platform is designed to optimize across all three, while giving enterprises the flexibility to run any model and workload on the same infrastructure, from training and inference to recommender systems, reasoning models and generative AI.

New results from MLPerf Inference v6.1 — the MLCommons consortium’s long-standing industry benchmark, spanning a broad range of models with every result peer-reviewed before publication — highlight that advantage. In its first MLPerf Inference preview submission, the NVIDIA Vera Rubin NVL72 system delivered up to 3.7x higher throughput than NVIDIA GB300 NVL72, demonstrating the performance gains possible with NVIDIA’s next-generation AI infrastructure.

NVIDIA GB300 NVL72 also showcased industry-leading scalability. A 288-GPU submission spanning four GB300 NVL72 racks achieved 99% scaling efficiency, with throughput growing nearly linearly from a single-rack baseline.

The results also underscore the impact of software innovation. NVIDIA’s MLPerf Inference v6.1 submissions delivered up to 1.6x higher performance than v6.0 through software optimizations alone, with additional gains achieved after the benchmark submission period.

For organizations evaluating AI infrastructure, performance, scaling efficiency and software velocity remain critical drivers of long-term inference economics.

Learn more about the NVIDIA Vera Rubin NVL72 benchmark results.

Tuesday, Sept. 15, 9:55 a.m. PT

AI Factory Flexible-Load Program Unlocks More Grid Power

Silicon Valley Power operates a flexible-load interconnection program that enables AI factories to support grid flexibility. 

Through this program, Emerald AI worked with NVIDIA to demonstrate automated load reduction at Silicon Valley Power. The system successfully responded to hundreds of demand signals from Silicon Valley Power while protecting AI workload performance. 

NVIDIA partner Emerald AI is planning to use NVIDIA DSX Flex for its Conductor grid-responsive power management software that dynamically adjusts an AI factory’s energy consumption based on real-time electricity grid signals and hybrid energy sources. 

Using DSX Flex to autonomously load balance, Emerald AI can demonstrate its ability to automatically throttle back power on low-priority AI jobs and then return to normal. 

DSX Flex can receive grid signals — load-shedding requests, demand-response events and pricing signals — and automatically act within a predefined workload hierarchy. The most critical jobs keep going, while everything else pauses temporarily and then resumes. In this way, facilities can participate in energy efficiency for the grid. 

This capability enables AI factories to operate as flexible grid resources, so they can reduce demand when the grid needs relief, protect priority AI workloads and prove that controllable AI load can help unlock more grid capacity for growth. 

Learn more about the Emerald AI partnership with Silicon Valley Power.

Tuesday, Sept. 15, 9:55 a.m. PT

Lambda Maximizes Performance per Watt With NVIDIA DSX MaxLPS

AI cloud provider Lambda released results at the AI Infra Summit, providing the first validation of NVIDIA DSX MaxLPS on NVIDIA Blackwell servers. 

DSX MaxLPS continuously monitors power consumption across GPUs and racks, dynamically shifting available power where it’s needed most and reclaiming capacity that static provisioning leaves unused. Because training and inference workloads have different power profiles, MaxLPS optimizes power allocation across mixed-workload AI factories. 

The results were significant: Lambda ran 19 nodes within the same power budget typically allocated to 16 full-power nodes, increasing cluster-wide token throughput by 24% — from about 4 million to 5 million tokens per second — while improving performance per watt by 23%. For next-generation NVIDIA Vera Rubin NVL72 AI factories, DSX MaxLPS can enable up to 40% more GPU capacity within the same megawatt budget in the right deployment environments.

This means AI factory operators can increase AI capacity and token throughput within the same power envelope, helping maximize the productivity and economic value of every megawatt.

Learn more about Lambda’s results. 

Tuesday, Sept. 15, 9:55 a.m. PT

Vera Rubin and Groq 3 LPX Turn More Power Into Tokens

In AI factories, power is the constraint. Every watt counts, so squeezing more tokens from every megawatt is what matters most for AI infrastructure.

NVIDIA does this with a full-stack AI factory built on the Vera Rubin NVL72 — systems, networking, software and power management working as one. At the factory level, NVIDIA DSX MaxLPS dynamically shifts power across racks as demand rises and falls. 

The payoff is significant:

• Up to 40% more GPUs within the same site-power envelope
• Up to 35% higher token throughput — no new power lines required

Inside each rack, Intelligent Power Smoothing software and expanded energy buffering absorb short spikes, letting systems run closer to sustained demand. Stranded headroom becomes productive compute.

For agentic AI — where agents chain reasoning steps and tool calls — latency and context length compounds quickly. That’s where NVIDIA Groq 3 LPX comes in, adding deterministic ultralow-latency inference to Vera Rubin and complements DSX MaxLPS. The combined platform delivers up to 35X higher token throughput per megawatt than GB200 NVL72 for 2-trillion-plus-parameter models at long context.

The numbers tell the story. On a 100K-context Qwen 3.8 27B workload, Groq 3 LPX hit 2,529 output tokens per second per user. That headroom lets agents run more reasoning steps and tool calls within the same response budget, even as workloads scale.

Learn more about how NVIDIA Vera Rubin NVL72 and Groq 3 LPX can produce more tokens with less power. 

Tuesday, Sept. 15, 9:55 a.m. PT

Vera Rubin NVL72 Delivers 30x Better AI Factory Throughput on SemiAnalysis AgentX

SemiAnalysis AgentX measures inference on recorded real-world agentic coding sessions, with actual context growth, tool calls delay and sub-agent spawning preserved, rather than on single-request benchmarks. NVIDIA Vera Rubin NVL72 agentic performance results are now live on the SemiAnalysis AgentX dashboard. On the DeepSeek V4 Pro model, Vera Rubin NVL72 delivers up to 30x higher throughput per megawatt than NVIDIA GB300 NVL72.

Agentic workloads have a different shape from the request-response tasks that inference benchmarks have historically measured. A single session can accumulate hundreds of thousands of input tokens as agents reason, call tools and spawn sub-agents — roughly 15x the token volume of a simple chat request — with wide variability in input and output length across requests. Capturing that end to end takes a benchmark built around full agent trajectories. AgentX does that, complementing standardized suites like MLPerf Inference that measure per-request performance across a broad range of models and scenarios. 

The efficiency gain translates directly into AI factory economics. Up to 30x higher throughput per megawatt means up to 30x more agentic work from the same energy footprint, and the AgentX results show up to 45x lower cost per million tokens. In power-constrained deployments, throughput per megawatt determines how much revenue an AI factory can generate, and cost per million tokens determines the margin on it.

This performance is the product of extreme codesign including the NVL72 scale-up domain and sixth-generation NVLink interconnect, NVFP4 precision on fifth-generation Tensor Cores, and an inference stack spanning NVIDIA TensorRT LLM and NVIDIA Dynamo. Vera Rubin is in full production and scaling across the ecosystem, and performance will continue to improve with ongoing software optimization.

Learn more about the platform architecture behind these results.

Tuesday, Sept. 15, 9:55 a.m. PT

Startups Celebrate Performance Results With NVIDIA Vera CPU

Startups are putting the NVIDIA Vera CPU to the test for running agent workloads and other demanding applications, celebrating the results. 

Perplexity benchmarked the Vera CPU for its new SPACE secure sandbox platform for agentic AI, showing 1.9x faster sandbox starts. 

Daytona put the Vera CPU through agentic workloads, citing “serious gains for agentic execution.” 

ClickHouse shared results for Vera CPU on ClickBench, its open benchmark for analytical databases. Vera was the fastest machine the team had measured so far. It’s a “strong signal of what’s ahead for CPU performance for data-intensive workloads,” ClickHouse said.  

DeepInfra ran a number of benchmarks on the Vera CPU, showing it winning on all metrics including 2.2x faster orchestration step latency.

Prime Intellect found that the Vera CPU maintained high bandwidth and low, consistent memory latency as more workloads ran in parallel — the kind of predictable performance needed for agentic AI.

Redpanda said its benchmark of the Vera CPU showed 5.5x lower latencies and 73% higher throughput than other CPUs.

Starburst released benchmark results that showed the Vera CPU delivered 3x faster query throughput than other CPUs.

Kinetica saw 2.7x faster analytical query performance with Vera CPUs compared with traditional CPUs. 22

NVIDIA NVLink 6 Keeps AI Factories Running at Massive Scale

As AI factories scale to hundreds of thousands of GPUs, reliability becomes a performance feature. 

Training and inference workloads depend on continuous operation, making transient errors, signal degradation and hardware failures inevitable challenges. NVIDIA NVLink 6 addresses these challenges with a multilayer resiliency architecture designed to detect, contain and recover from faults before they affect applications.

At the physical layer, custom forward error correction, physical layer retry and universal physical layer recovery help maintain a lossless fabric with minimal latency impact. At the network layer, credit-based flow control, dynamic routing and link rebalancing contain faults locally and prevent cascading stalls that can reduce AI factory throughput.

Learn more about NVIDIA NVLink 6.

As local models become more capable, AI agents can handle more work directly on a PC while keeping sensitive information on the device.

Portable Computer is a local version of the agent Perplexity Computer that plans and carries out multistep tasks. Accelerated by NVIDIA GPUs, it uses local models to analyze data, bring together information across files and handle recurring work. Sensitive information stays on device, and locally completed work doesn’t consume Perplexity Computer credits. Users can also orchestrate work up to cloud models for more advanced research and reasoning.

Today, Perplexity is adding Portable Computer in the Perplexity app for Windows on compatible NVIDIA GeForce RTX PCs and NVIDIA RTX PRO Workstations, bringing powerful agentic AI to more Windows PC users. The release builds on existing support for NVIDIA DGX Spark systems and RTX PCs running Linux.

Local Agents Powered by NVIDIA RTX

Perplexity brings local and cloud AI together in one app, letting users work with sensitive files on their PCs and take advantage of Computer’s built-in tools such as the built-in browser and proprietary SPACE sandbox.

For tasks that call for more advanced reasoning, Portable Computer can also identify when a task needs cloud support, asking the user for permission before sending information off-device.

The app simplifies setup with a local model, such as Qwen 3.8 27B, that is post-trained to work with Perplexity Computer and optimized for NVIDIA RTX GPUs. Users can put the agent to work without having to research models or configure the complex software stack typically required to run local AI.

Connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub extend that experience across the files and apps already part of users’ daily workflows.

For example, the agent can help with:

• Engineering: Review open pull requests in a connected GitHub project, organize them by status and identify next steps. Computer can also flag outdated documentation and submit proposed updates for review.
• Finance: Point Computer at two years of brokerage summaries, consolidated 1099s and tax returns, and have it trace the recurring holdings creating the most avoidable fees and tax drag, with every figure cited to the exact file and page — all without a document ever reaching a chatbot.
• Startups: Ask Computer why activation went flat, and the agent analyzes the funnel export locally to find where new signups drop off between install and first completed task, then posts the top insights straight to the team’s Slack channel.

Try Portable Computer on Windows PCs Today

Portable Computer is available for NVIDIA GeForce RTX and RTX PRO GPUs with 24GB or more of VRAM. NVIDIA DGX Station support is expected to come soon. 

Try Perplexity Portable Computer today.

#ICYMI: The Latest Updates From NVIDIA Local AI

Z.ai’s GLM 5.3 Flash provides impressive performance and visual intelligence at low cost, optimized for DGX Station and dual DGX Spark systems.

DeepSeek-v4.1 Flash significantly reduces key-value cache memory demands and operating costs for complex AI agent workloads, delivering remarkable intelligence per dollar.

Qwen has released Qwen3.8-Flash-Next, an open-weight multimodal mixture-of-experts model, and an early preview of Qwen4, which can run locally on a single DGX Spark with NVFP4 and punches well above its weight.

GLM 5.3 is Z.ai’s 744-billion-parameter flagship model, tuned for agent sessions that run for hours on DGX Station and a cluster of four DGX Spark systems. 

Follow NVIDIA RTX Spark on X, Instagram, TikTok and Facebook — and stay informed by subscribing to the NVIDIA Local AI newsletter. Follow NVIDIA Workstation on LinkedIn and X. 

See notice regarding software product information.

Manufacturing floors, warehouses and production lines rarely stay fixed — tasks change, layouts shift and new products arrive, and most robots can’t keep up without significant reprogramming.

Skild AI’s new S1 robot foundation model helps address this, designed to learn previously unseen, long-horizon tasks from a single video demonstration. The model, launched last week, uses video as input to understand and execute the task without updating its weights or undergoing task-specific post-training — a technique called in-context learning.  

Skild built S1 and conducted the research on NVIDIA AI infrastructure, part of a broader collaboration spanning synthetic data generation, model training, simulation and real-world physical AI deployment. The companies are working together to move adaptable robot intelligence from the lab into factories and other dynamic operating environments.

“Learning by experience, and not preprogramming, is the step change that has happened in robotics,” said Deepak Pathak, cofounder and CEO of Skild AI. “NVIDIA Isaac Lab and NVIDIA Cosmos technologies help Skild create the scalable, diverse experience its robots need to learn across many scenarios and embodiments.”

The launch comes as the company reached a $100 million annual revenue run rate 10 months after its first commercial deployment. In that time, Skild has built more than 60 deployment partnerships with work spanning manufacturing, logistics, inspection, security, food preparation and other applications. 

Learning New Work From One Video

Most industrial robots are built for fixed jobs, so each new product, process or layout requires more data, retraining and validation.

S1 takes a different approach: An operator records a video of the desired task and provides it to the model as a prompt. It interprets the demonstrated intent, objects and sequence, then maps them into actions for the robot in front of it — with no retraining — and often for a task not covered by its pretraining dataset. 

S1 can perform unfamiliar tasks lasting up to 10 minutes, including plant potting, pancake making, pour-over coffee brewing and kit assembly. These tasks can span dozens of manipulation steps and require the robot to compose skills in sequences it hasn’t previously performed. 

In one plant-potting test, the Skild AI team moved from recording the demonstration to autonomous execution on hardware in just 11 minutes. The model can also adjust when objects move, recover from errors and combine skills in sequences that weren’t explicitly programmed.

In Skild’s tests on new, multistep tasks, its S1 robot succeeded about 66% of the time at each step, compared with 9% for a similar AI system — a more than sevenfold improvement. Skild also estimates that showing the robot one short video example can be as useful as giving it roughly 380 hands-on training examples. A person collecting those examples manually could take 50-100 hours.

From Research to Factory Work

S1 breaks the cycle of needing to constantly retrain robots for new factors by letting operators demonstrate new tasks directly without requiring a new dataset or training run for every change. Where customer agreements permit, experience from Skild’s commercial deployments can inform the broader model and help accelerate future deployments.

That work is already in action on the factory floor. Skild, NVIDIA and Foxconn are deploying the Skild Brain on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems. In one demonstrated workflow, a robot installs a busbar and limit block, fastens 16 screws and adapts to disturbances across a multistep task. The work requires precise motion, contact-aware control, sequence tracking and recovery when the scene differs from the plan.

NVIDIA Technology Across the Development Cycle

NVIDIA accelerated computing gives Skild the scale to train its shared robot brain using simulation, human video, teleoperation and, where permitted, deployment data. NVIDIA Cosmos open world foundation models help diversify training data and turn video into structured descriptions, while Cosmos Curator helps annotate, filter and organize data at scale.

Skild is extensively using NVIDIA’s open simulation frameworks to train and validate its robot brain before real-world deployment. NVIDIA Omniverse libraries and the NVIDIA Isaac Sim framework provide physically based virtual environments for generating data, testing edge cases and validating behaviors.

Skild further strengthens the skills of its brain through reinforcement learning in Isaac Lab, an open modular robot learning framework. Powered by the Newton physics engine, Isaac Lab helps Skild’s engineers accurately model various physical parameters, such as forces, contact, collision and pressure, and reduce the simulation-to-reality gap.

Skild and NVIDIA are also jointly developing new GPU-accelerated simulation solvers that quickly and accurately model how robots physically touch, grip and manipulate solid objects. They’ll soon be made available to all developers as part of Newton. 

As models move toward production, NVIDIA Nsight tools help engineers find performance bottlenecks during training, and the NVIDIA TensorRT software development kit optimizes inference so robots can respond quickly in the physical world. Together, these technologies connect the data, simulation, training and deployment stages instead of treating them as separate systems.

Read Skild AI’s S1 research and explore the NVIDIA Isaac robotics platform.

The global robotaxi market — physical AI’s first commercial breakthrough — is projected to reach $400 billion by 2035, with over 6 million commercial vehicles in operation as driverless fleets are already moving people through some of the world’s busiest and most complex streets.

Deploying a driverless vehicle is one challenge. Scaling a fleet is a next-level computing challenge; it means delivering the same safe, reliable performance across thousands of vehicles. 

Meeting those demands requires enormous amounts of compute across the robotaxi development lifecycle, from preparing and training AI models to simulating and validating driving behavior, as well as real-time processing in the vehicle. 

NVIDIA provides an open platform for AI training, simulation and safety validation, with libraries, software development kits, workflows and models that developers can use alongside their own technology stacks.

Every major robotaxi program operating at commercial scale today is running on NVIDIA’s modular stack, spanning AI training, simulation, in-vehicle computing — or a combination of the three — to develop and deploy fleets at scale. 

What Is a Robotaxi Technology Stack?

A robotaxi technology stack is the end-to-end set of technologies used to develop, validate and deploy autonomous vehicles (AVs) — from data and AI model training to simulation, safety validation and real-time in-vehicle computing. 

NVIDIA’s robotaxi and AV platform brings these capabilities together in a three-computer solution: the model training computer, simulation and validation computer, and in-vehicle computer.

1. Training Computer: NVIDIA DGX

Robotaxi intelligence advances as programs turn growing volumes of fleet data into increasingly capable models. Driving models can be trained on NVIDIA DGX systems. 

The NVIDIA Alpamayo portfolio of open reasoning vision language action (VLA) models, simulation frameworks and physical AI datasets gives developers building blocks they can adapt to their own data, requirements and technology stacks. Its reasoning models help address long-tail AV challenges by breaking complex driving situations into smaller steps, reasoning through each one and selecting the safest trajectory. 

NVIDIA also provides physical AI datasets, reinforcement learning blueprints and recipes for post-training and distillation, helping developers optimize models for their target vehicles.

2. Simulation and Validation Computer: NVIDIA Omniverse and Cosmos on NVIDIA RTX PRO 

Robotaxi programs can’t rely on physical miles alone to capture rare, long-tail driving scenarios. NVIDIA Omniverse NuRec models reconstruct real-world driving scenarios from sensor data, while NVIDIA Cosmos world foundation models generate physically based variations of them, enabling developers to turn thousands of real-world corner cases into millions of combinations of driving behavior, traffic, weather, lighting and sensor conditions.

Running on NVIDIA RTX PRO Servers, NVIDIA Omniverse and Cosmos support closed-loop simulation and validation. The NVIDIA AlpaSim simulation framework extends the workflow for training and evaluating reasoning-based autonomous-driving models, helping developers identify weaknesses before deployment.

3. In-Vehicle Computer and Sensor Architecture: NVIDIA Hyperion With DRIVE AGX

NVIDIA Hyperion is NVIDIA’s modular in-vehicle compute and sensor reference architecture for level-4-ready robotaxis. Hyperion 10 pairs dual NVIDIA DRIVE AGX Thor systems-on-a-chip, built on the NVIDIA Blackwell platform, with 14 high-definition cameras, nine radars, three lidars and 12 ultrasonics for real-time, 360-degree sensor fusion. Its redundant compute and sensing design supports fail-operational driving if a sensor or compute component fails.

The dual DRIVE AGX Thor is designed to run modern AI workloads — including VLA models — for perception, reasoning, path planning and driving actions.

NVIDIA Halos provides a production-ready safety foundation through Halos OS, and a broader validation and certification framework spanning independent inspection, system validation, large-scale simulation and continuous testing from cloud to car.

Robotaxi Leaders Adopting NVIDIA’s Robotaxi Technology Stack

NVIDIA’s robotaxi ecosystem spans every region where commercial robotaxi services are emerging today: Asia, Europe, the Middle East and North America. Across these markets, mobility providers, AV developers and automakers are adopting NVIDIA’s three-computer architecture to train AI models, simulate and validate driving behavior, and deploy autonomous vehicles at scale.

Scaling Robotaxi Services Globally

• Uber is scaling its fleet of NVIDIA Hyperion, with plans to reach 28 cities by 2028. Uber and NVIDIA are also building a robotaxi AI data factory on NVIDIA Cosmos to curate fleet driving data for rare scenarios. Together, Uber and NVIDIA are collaborating with Autobrains, Avride, Lucid, May Mobility, Mercedes-Benz, Momenta, Nissan, Nuro, Pony.ai, Stellantis, Waabi, Wayve, WeRide and Zoox to bring NVIDIA-powered robotaxi services to the Uber platform.
• May Mobility is planning to operate autonomous ride-hailing services through Uber’s network, while developing its software stack on the NVIDIA DRIVE platform. May Mobility has also launched an autonomous ride-hailing pilot with Lyft in Atlanta, built on the NVIDIA DRIVE platform.
• Bolt uses NVIDIA technologies to develop and scale AVs across Europe.
• Lyft plans to use NVIDIA Hyperion as a reference architecture for future autonomous fleets.
• Through its partnership with Grab, WeRide plans to bring its Hyperion- and DRIVE AGX Thor-based GXR to key markets across Southeast Asia.
• Waymo partners with NVIDIA to help build its autonomous computing system.

Building Robotaxi Intelligence 

Behind these services, AV developers are using NVIDIA accelerated computing, simulation and in-vehicle platforms to build the intelligence that operators and automakers deploy.

• Wayve, Nissan and Uber are developing a global robotaxi program using a prototype vehicle that combines Nissan’s vehicle engineering, Wayve’s embodied AI and the NVIDIA Hyperion platform.
• Autobrains is developing robotaxi programs with Uber in Munich and VinFast in Southeast Asia, built on NVIDIA Hyperion and enabled by Autobrains’ Agentic AI technology.
• Zoox uses NVIDIA DRIVE for in-vehicle computing and cloud-based training and simulation. 
• Momenta is developing its software stack based on NVIDIA DRIVE AGX running on DriveOS. 
• Pony.ai developed its new-generation autonomous-driving domain controller with NVIDIA Hyperion and DRIVE AGX Thor. 
• Tensor is developing its level 4 Robocar with eight NVIDIA DRIVE AGX Thor systems-on-a-chip in its in-vehicle supercomputer. 
• Waabi expands into the robotaxi market through a deployment collaboration with Uber; its Waabi Driver platform is built on NVIDIA DRIVE AGX Thor.
• TIER IV and Isuzu are deploying level 4 autonomous buses built on NVIDIA Hyperion and DRIVE AGX Thor.
• Lenovo is supplying its NVIDIA DRIVE AGX Thor-based AD1 level 4 domain controller for a next-generation robotaxi program with SWM.
• DeepRoute.ai is developing a new generation of robotaxis built on the NVIDIA Hyperion platform with DRIVE AGX Thor.

Bringing Robotaxis Into Production

As these systems move from development into production, automakers are integrating NVIDIA technology into autonomous and robotaxi-ready vehicle programs.

• Tesla trains its autonomous-driving neural networks on NVIDIA supercomputers. 
• Mercedes-Benz and NVIDIA are collaborating with Uber to develop a robotaxi ecosystem based on the new S-Class, built on the NVIDIA Hyperion architecture, full-stack NVIDIA DRIVE AV L4 software, and NVIDIA Alpamayo open AI models, simulation tools and datasets to support reasoning-based, safety-first autonomy.
• Stellantis, Wayve and Uber are collaborating to develop and deploy L4 driverless mobility services, leveraging NVIDIA Hyperion and AI computing technologies.
• Lucid, Nuro and Uber are developing a global robotaxi service using NVIDIA DRIVE AGX Thor, part of the Hyperion platform.
• Hyundai Motor and Kia are expanding their collaboration with NVIDIA to develop data-driven autonomous-driving systems built on NVIDIA Hyperion. NVIDIA will also explore expanded collaboration with Hyundai Motor Group’s joint venture, Motional, to advance level 4 robotaxi services.
• Geely, alongside its ecosystem partners, plans to develop and commercialize robotaxis using Hyperion.
• Zeekr, a Geely Auto Group brand, has adopted DRIVE AGX Thor for a centralized domain controller.

From cloud to car, nearly every layer of the robotaxi platform is being developed on NVIDIA accelerated computing. 

Explore NVIDIA’s complete platform for robotaxi development.

AI inference chipmaker d-Matrix today announced it will use NVLink Fusion to connect its next-generation Raptor XPUs to NVIDIA’s AI infrastructure platform — joining a growing roster of ecosystem partners.

By connecting Raptor to NVIDIA NVLink scale-up and Spectrum-X scale-out networking, the NVIDIA MGX rack architecture and the broader NVIDIA AI platform, NVLink Fusion gives d-Matrix an accelerated, lower-risk path from custom silicon to large-scale deployment. 

“Demand for inference is soaring, but capital, time and energy remain finite,” said Sid Sheth, cofounder and CEO of d-Matrix during a press briefing yesterday. “With NVLink Fusion and MGX, we can integrate our Raptor XPUs into a broadly deployed, liquid-cooled architecture, giving customers a faster, lower-risk path to deploy and scale ultralow-latency inference.”

The NVIDIA AI platform is vertically integrated and horizontally open. NVLink Fusion extends this openness to XPUs and CPUs, allowing silicon companies to focus on their processor innovations while using NVIDIA infrastructure to deploy them at AI factory scale.

From Custom Silicon to Rack-Scale Deployment

Building an XPU is only the first step. Deploying it at AI factory scale requires a complete platform spanning networking, rack architecture, power, cooling, software and a proven supply chain. Each of the steps — sourcing chips, integrating high-speed interfaces, validating a scale-up networking solution, designing and certifying rack architecture — adds time, costs and risks.

NVLink Fusion lets silicon innovators connect directly into NVIDIA’s proven platform. It’s the high-bandwidth, low-latency technology that connects custom XPUs and CPUs to the NVIDIA stack.

By adopting NVLink Fusion, d-Matrix can tap into the NVIDIA MGX ecosystem’s mature, validated rack designs, supply chain, power and cooling infrastructure. By standardizing on a common rack, data centers can be built once and support GPUs, CPUs and XPUs — without requiring a separate rack architecture for each processor type.

NVLink Fusion Integrates d-Matrix XPUs Into AI Factories

Using NVIDIA NVLink, d-Matrix plans to connect its XPUs in a single high-bandwidth, low-latency scale-up domain. Its racks can also work alongside NVIDIA GPU-based systems like NVIDIA Vera Rubin NVL72 for disaggregated inference. 

d-Matrix also plans to integrate NVIDIA Vera CPUs, NVIDIA ConnectX-9 SuperNICs, NVIDIA BlueField-4 DPUs and NVIDIA Spectrum-X Ethernet networking. Together with NVLink and MGX, these technologies give d-Matrix a proven foundation to deploy specialized inference alongside NVIDIA systems within flexible, unified AI factories.

NVLink Fusion Opens NVIDIA AI Factories to Specialized XPU Architectures

The NVIDIA full-stack AI factory platform includes NVIDIA Vera Rubin NVL72, Groq 3 LPX, the Vera CPU rack, Vera BlueField-4 STX storage and Spectrum-6 SPX Ethernet networking. It’s designed to be completely fungible — running every AI workload, model and model architecture — with the best performance per watt and lowest cost per token. 

NVLink Fusion gives customers the flexibility to match the right compute to each workload within a common AI factory platform, opening access to NVIDIA networking, systems, software and global supply chain. Silicon innovators like d-Matrix can use NVLink Fusion to increase performance, accelerate time to market and reduce the risk of deploying semi-custom AI factories.

Gear up: The latest PC games and major updates are ready to play on GeForce NOW this week. WARDOGS drops onto the cloud at early-access launch, alongside the Valheim 1.0 Deep North update and Bus Simulator 27 — part of nine new titles joining the cloud.

The newest PC releases can demand serious hardware, storage space and upgrade budgets. GeForce NOW puts GeForce RTX-powered performance in the cloud, so members can jump into the newest titles across supported devices without needing expensive PC upgrades to keep up.

From a 100-player tactical firefight to a frozen Viking frontier and a growing bus empire, every adventure is ready when members are — no downloads or storage management required.

Who Let the ‘WARDOGS’ Out?

Take on the Control Zone: WARDOGS launches on GeForce NOW today, bringing BULKHEAD’s large-scale tactical first-person shooter to the cloud at early-access launch.

Up to 100 players across three teams battle to control randomized zones in a large-scale military sandbox, where tactical gunplay, combined-arms combat, building and destruction can change the course of a match.

There’s no single route to victory. Outfit a loadout, coordinate a squad, fortify a strategic position or take the fight to the enemy in vehicles. Every team play action earns cash, giving players more ways to invest in gear and vehicles that can turn the tide.

Ultimate members can stream the fight with GeForce RTX 5080-class performance in the cloud for a tactical advantage. Pick up the battle across supported devices without waiting on a huge game install or clearing local storage first.

Ice, Ice in the Deep North, Baby

Sharpen the axes and raise the shields – Valheim has left early access with its 1.0 update, letting members stream the final biome in Iron Gate’s Viking survival saga through Steam and Xbox Store. 

Set sail for the Deep North, a frozen frontier filled with new enemies to face, weapons to craft and strongholds to build before the cold closes in. The 1.0 launch also arrives alongside new platform versions with full crossplay support, making it easier to rally the clan.

Whether returning to a long-running world on a PC or Mac or beginning a fresh voyage on a Steam Deck or Amazon Fire TV Stick, members can head to the Deep North across supported devices without waiting for downloads.

The Wheels on the Bus Go ‘Round in the Cloud

Hop into the driver’s seat. Bus Simulator 27 is now streaming from GeForce NOW. 

Drive a fleet of more than 45 officially licensed buses from 13 manufacturers through Felicia Bay, a sunny region with two cities, 20 districts and a full day-and-night cycle. Build routes, manage timetables and grow a successful bus company in Story, Career or Sandbox mode.

And there’s even more to stream from the GeForce NOW library this week:

• Bus Simulator 27 (New release on Steam, Sept. 8)
• Halloween: The Game (New release on Steam, Sept. 8)
• Honeycomb: The World Beyond (New release on Steam, Sept. 8)
• Wanderburg (New release on Steam, Sept. 8)
• Shroom and Gloom (New release on Steam, Sept. 10)
• Train Sim World 7 – Advanced Access (New release on Steam, Sept. 10)
• WARDOGS (New release on Steam, Sept. 10)
• Welcome to Elderfield (New release on Steam, Sept. 10)
• Death Stranding Director’s Cut (Epic Games Store)

Dates listed above reflect when games are released on their respective stores. GeForce NOW availability may vary, as games are onboarded after release and added throughout the week. Keep an eye on GeForce NOW channels and GFN Thursdays for availability updates on announced titles.

One GFN community member recently returned after six years, saying NVIDIA DLSS 5 technology renewed their interest in GeForce NOW cloud gaming. Ready to try it out, too? Start with a day pass to test premium cloud gaming before committing to a membership. For those who decide to keep playing, the cost can be applied toward their first monthly membership.

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

At the IBC conference, running Sept. 11-14 in Amsterdam, the creative, technology and business communities are coming together to turn ideas into action and discuss innovations across the media and entertainment industries. More than 44,000 attendees from 170+ countries are gathering to explore 1,300+ exhibitions in 14+ halls and outdoor spaces, with over 600 speakers delivering insights.

Read on to learn more about what NVIDIA’s highlighting at the show.

NVIDIA AI for Media Brings Real-Time Intelligence to Broadcast, Sports and Production Workflows

Media companies are increasingly integrating AI into live production, sports, news and streaming workflows to unlock richer performance insights, verify video authenticity, and enhance and localize content — all without disrupting trusted broadcast environments.

At IBC 2026 in Amsterdam, NVIDIA announced a major expansion to NVIDIA AI for Media — a collection of GPU-accelerated software development kits (SDKs), NVIDIA NIM microservices, playbooks, and blueprints that enhance audio, video and augmented-reality effects for media and entertainment workflows — to unlock new ways to understand motion, verify and enhance video, localize programming and build AI-powered media applications. 

The NVIDIA Synthetic Video Detector (SVD) NIM microservice, announced earlier this year at SIGGRAPH, helps organizations assess the probability of whether footage is authentic or AI-generated, giving editorial, content-authentication, digital-forensics and media-integrity teams another point of analysis in their review process.

Since its initial release, SVD’s accuracy has reached 99.3% for text-to-video content and 97.7% for image-to-video content, with especially large gains on difficult image-to-video cases. 

Dalet is integrating SVD into a secure, cloud-hosted verification workflow for news organizations. This allows editorial teams to submit footage through SVD, inspect and review the resulting scores and metadata within a Dalet interface. 

TwelveLabs announced the general availability of Compliance by TwelveLabs, its first application built on the company’s video intelligence platform, helping media and broadcast teams rapidly screen content against regional and custom compliance standards. The solution integrates SVD to add frame-level authenticity signals and confidence scores, enabling media teams to identify potentially synthetic media within the same compliance workflow.

Wowza, whose Wowza Streaming Engine media server technology powers more than 35,000 video deployments across over 170 countries, will distribute SVD through the Wowza Video Intelligence Framework. The solution, powered by NVIDIA-accelerated infrastructure, will enable broadcasters, streaming providers and other organizations to analyze live video feeds and extract data around detected objects, scenes and signs of AI generation in real time. It can be deployed and run on premises, at the edge, in the cloud, across hybrid deployments or fully air-gapped, giving organizations greater control over critical media workflows.

NVIDIA 3D Body Pose estimates 2D and 3D human joint locations and angles from video captured by a single camera, helping turn motion into structured data without marker-based capture systems.

For sports organizations, that data can support player and athlete movement tracking, biomechanics and performance analysis, replay enhancement, officiating and adjudication workflows, player-safety applications, and virtual interaction and immersive experiences. 

The technology can also provide structured human-motion data for content-creation workflows. When mapped to a compatible character rig, joint and motion data can serve as input for animation blocking, digital doubles, character retargeting and virtual-production experiences.

Vizrt is using Body Pose technology in live virtual-studio environments, with tracked body movement driving real-time 3D lighting effects such as reflections, shadows and environmental rendering.

Video Frame Generation (VFG) makes video motion appear smoother by using generative AI to create new frames between the original frames of a video. It can increase frame rates by 2x or 4x while preserving visual quality and temporal consistency, enabling more fluid sports, slow-motion replays, live media and other high-motion video experiences. VFG also supports frame-rate conversion and frame boosting for generative AI video workflows.

Ross Video is integrating VFG into its Rio Replay platform to create AI-assisted slow-motion video for sports production.

The work supports 6x slow-motion generation for sports replay. Development is underway toward 8x interpolation, meaning generated intermediate frames can give replay teams smoother motion without requiring every frame to be captured by an ultrahigh-frame-rate source camera. 

NVIDIA Video Super Resolution (VSR) uses AI to upscale video while reducing noise, blur and compression artifacts. New streaming modes let developers choose between real-time performance and higher image quality, while adjustable controls help achieve the desired level of enhancement. VSR also adds 10-bit video support and improves overall performance and quality. VSR is available through the NVIDIA Video Effects SDK and a NIM microservice for use in streaming, broadcast, conferencing, video playback and content-creation applications. 

The technology can support video players, conferencing applications, creator tools, streaming services, transcoders and broadcast systems through a common interface.

NVIDIA TrueHDR converts standard-dynamic-range video into high-dynamic-range output in real time, reaching up to approximately 2,000 nits while preserving local contrast and adapting brightness to the content.

VSR, VFG and TrueHDR can be combined within a single video-effects pipeline — helping media companies enhance existing content libraries for streaming, transcoding, gaming and creator workflows.  

The NVIDIA LipSync and Active Speaker Detection NIM microservices help developers build localization systems for interviews, news, sports, entertainment and other programming where multiple people may appear on screen.

LipSync transforms mouth movement in an input video to match a target audio track while preserving natural head pose, blinking and body movement. The new release improves facial occlusion handling and better preserves teeth, lip and facial textures.

The new Active Speaker Detection NIM microservice no longer requires speaker diarization for multiple audio tracks, adds voice activity detection and expands NIM microservice deployment support through a gRPC interface and broader GPU compatibility.

NDI is using NVIDIA AI for Media, including the NVIDIA LipSync NIM microservice, to enable real-time translation, lip-synced dubbing and regional language adaptation within existing broadcast workflows. By generating multiple language experiences from a common media stream, the approach can help broadcasters reach global audiences while reducing the bandwidth, infrastructure and production complexity traditionally required for multilingual distribution.

Studio Voice includes new Microphone Profiles built on NVIDIA Studio Voice NIM microservices, giving users more control over the tonal character of enhanced speech.

The capability is designed to suppress background noise, reduce room reverberation and improve speech clarity, then shape the enhanced output into a selected microphone profile for more polished live communications, streaming, podcasting and content creation.

Try NVIDIA AI for Media NIM microservices. See the latest NVIDIA and partner workflows at IBC 2026.

NVIDIA Holoscan for Media Provides Open Media Exchange Layer to Build and Connect Live Media Applications

As broadcasters, streaming services and sports organizations adopt software and AI, the infrastructure behind live content is becoming more flexible, more connected and increasingly built on shared accelerated computing.

The integration of Media Exchange Layer (MXL) with NVIDIA Holoscan for Media accelerates that transition — providing the common exchange layer that helps media applications connect and operate together. 

Holoscan for Media is an open reference architecture and developer toolkit for building AI-powered media functions and applications for software-defined live production. MXL adds an open way for those software-based media functions to exchange live video, audio and data across a distributed environment. 

As production functions move into software, developers can build applications that share accelerated infrastructure, connect dynamically and evolve independently. That can help media companies use infrastructure more efficiently, introduce new capabilities faster and reduce the amount of custom integration required between applications.

The integration also creates a stronger foundation for AI in live media. AI processing, video applications and traditional media functions can increasingly operate on the same accelerated infrastructure and in the same software-defined environment.

For technology vendors, this expands the opportunity to build applications that can work across broader, multi-vendor ecosystems. For media companies, it creates a path toward infrastructure that can adapt as formats, applications and AI capabilities change.

See the demo at IBC in EBU Stand 10.D21. ​Learn more about Holoscan for Media.

NVIDIA Sports Intelligence Playbooks Chart a Path to Multimodal AI for Sports

Sports is becoming a proving ground for a broader shift in AI: from general-purpose models toward fine-tuned open models built on proprietary data.

NVIDIA Sports Intelligence Playbooks are designed to accelerate that transition. They give leagues, media companies and technology providers structured frameworks to fine-tune NVIDIA open models on their own sports footage and annotations, creating multimodal AI that can understand the rules, players, scoring, strategy and context unique to a sport.

Sports organizations hold large volumes of proprietary video, metadata and performance information that are difficult for competitors to replicate. The playbooks provide a practical blueprint for converting those assets into AI capabilities that can underpin new analytics products, media experiences, automation tools and revenue streams.

The playbooks span the AI lifecycle, including data preparation, fine-tuning, inference, evaluation, optimization and deployment, and bring together NVIDIA technologies including Nemotron, NeMo AutoModel, Megatron Bridge, NIM microservices and NVIDIA accelerated computing. 

By providing an integrated path from model customization to production, Sports Intelligence Playbooks can reduce the cost and complexity of building specialized sports AI while increasing demand across its compute, software and inference stack.

Early testing demonstrates the potential of domain specialization. When evaluated on previously unseen footage using question formats similar to those used in training,  multiple-choice accuracy increased from approximately 53% to 94% and open-ended evaluation from approximately 5.7% to 66%.

Machina Sports is integrating Sports Intelligence Playbooks with its sports-native data, evaluation and agent infrastructure, enabling rights holders to turn proprietary media and expertise into private, deployable intelligence for live production, content and fan experiences.

The opportunity also expands as agentic AI becomes increasingly adopted. With the NVIDIA AI-Q Blueprint, organizations can use their domain-specific sports models as expert intelligence within agents that reason across video, enterprise data and software systems, extending the playbook from sports understanding into decision-making and automation.

Wowza is integrating vision language models, including NVIDIA Cosmos 3 and Nemotron, into the Wowza Video Intelligence Framework, fine-tuned through NVIDIA Sports Intelligence Playbooks to detect sports-specific moments in live streams and reduce time to action. 

Explore NVIDIA Sports Intelligence Playbooks.

NVIDIA Brings Multilingual Content Localization to Live Broadcast

Reaching global audiences with live programming requires more than translating words. Language nuances, voice, timing, facial movement, captions and onscreen graphics must work together in real time, while preserving the editorial intent and production quality of the original program.

To help broadcasters, sports leagues, rights holders and streaming services bring these elements into a unified, software-defined, real-time localization workflow, NVIDIA is bringing its Content Localization technologies to the NVIDIA Holoscan for Media developer toolkit. Designed for broadcast and streaming developers, the reference workflow enables captions, translated audio, dubbing, synchronized video and localized graphics.

Content Localization with Holoscan for Media provides a reference for how localization technologies can work together in software-defined broadcast applications. Developers can select the capabilities needed for each program, market or distribution channel rather than deploying separate infrastructure for every localized version.

Content Localization with Holoscan for Media incorporates the latest advancements from NVIDIA AI for Media, including improved LipSync when faces are partially obscured and enhanced Active Speaker Detection to help applications identify who’s speaking in multi-person scenes.

Expanding the Reach of Live Programming

Localization can transform the reach and economics of live programming. A shared, composable workflow can help media companies introduce regional coverage faster, serve more audiences and tailor experiences for individual markets — while preserving the timing, visual context and editorial control required for live production.

Technologies from AI-Media, CAMB.AI, Chyron and Panjaya each address a specific part of content localization with Holoscan for Media, from adapting voice and onscreen delivery to creating multilingual captions and translated audio, localizing graphics, and preserving expression and identity across live and on-demand content.

The Content Localization technologies also support file-based, streaming and post-production applications. Developers can use application programming interfaces for on-demand workflows and the Holoscan for Media reference workflow when localization must run as part of a live media environment. Together, they provide a consistent foundation for building multilingual media services across production and distribution.

Learn more about NVIDIA Holoscan for Media and AI for Media.

Frontier intelligence is going local. At IFA 2026, NVIDIA, Microsoft and its partners are teaming up to provide faster inference and new tools that make agents easier to set up and run locally on NVIDIA hardware. New compact NVIDIA RTX Spark Windows PCs are also coming in October to give AI enthusiasts, developers and creators more ways to run capable agents locally and securely. 

Today’s announcements include:

• Simplified local AI support for NVIDIA GPUs is coming in Hermes Agent, OpenClaw and Perplexity Portable Computer.
• Up to 1.9x faster local inference — new llama.cpp and vLLM optimizations are available now directly and through LM Studio and Ollama. 
• NVIDIA PAIR — a Personal AI Router tool that intelligently distributes AI inference across the PCs on a user’s local network.
• NVIDIA RTX Spark arrives in October  — with new Windows PCs from Lenovo and Acer. Electronic Arts, Embark and Ubisoft are among the latest game publishers and developers bringing their blockbuster titles to NVIDIA RTX Spark.

Also, August was a busy month for local AI:

• Nemotron 3.5 Lightning — which can run on NVIDIA RTX PCs, RTX PRO Workstations, DGX Spark and Jetson — is a 30-billion parameter model that has been launched. Get started with Nemotron 3.5 Lightning today. 
• Z.ai’s GLM-5.3-Flash is a multimodal mixture-of-experts (MoE) model that’s bringing agentic AI to DGX Station.
• Qwen has released Qwen3.8-Flash-Next, an open weight multimodal MoE model, which can run locally on DGX Spark and DGX Station, along with Qwen3.8-27B, a 27-billion-parameter open model optimized for local agentic and coding workloads on NVIDIA GPUs. 
• LTX’s LTX 2.5 is an open-world video generation model optimized for NVIDIA RTX GPUs, DGX Spark and DGX Station, with new NVFP4, FastVideo and ComfyUI enhancements for faster, more memory-efficient local generation.
• MiniMax-H3 is an open-weight video generation model with synchronized audio that can run locally on NVIDIA GPUs through ComfyUI. FastVideo teamed up with NVIDIA researchers to improve this further by releasing FastH3 — an open-weight, four-step distilled version that improves performance by 7x. Optimized FastVideo recipes for NVIDIA RTX GPUs and DGX Spark are coming soon.
• Meta’s Muse Glimmer is a 30-billion-parameter open-weight model for coding and agentic workloads that can run locally on GeForce RTX PCs, DGX Spark, DGX Station and Jetson. NVIDIA has also released NVFP4 quantization with DGX Spark support for more memory-efficient local deployment.
• DeepSeek v4 Flash is a 284-billion-parameter MoE model with 13 billion active parameters that can run locally on 2x DGX Spark cluster and DGX Station.

A Simpler Start for Local Agents

Getting a local agent up and running with local models required some effort — choosing a model, finding a compatible inference server, dialing in quantization settings and keeping everything updated. That friction is disappearing on RTX and DGX systems.

Three of the most widely used agent apps will offer simplified local model setup on Windows, each built on llama.cpp and incorporating NVIDIA’s latest inference optimizations. The new setup experiences are designed to reduce manual configuration and make it easier to get local agents up and running. 

Last month, Perplexity introduced its Portable Computer agent, giving users a simple way to run Perplexity locally on Linux systems like NVIDIA DGX Spark with the models, orchestration and tools packaged into a single app experience.

Perplexity Portable Computer is available on NVIDIA RTX GPUs with at least 24GB VRAM running on Linux, with support on Windows coming soon, bringing that same streamlined setup to a broader group of PC users. Users can run complete workflows locally without consuming credits, while selectively escalating parts of a task to one of 15+ frontier models in the cloud when additional research or reasoning is needed. Portable Computer asks for permission before sending content to the cloud, helping users keep sensitive information on their device. Here’s some example use-cases:

• Engineering: Review open PRs in a connected GitHub repo and sort them into ready, blocked, stale, and needs review, each tagged with the next step. Docs that fell out of sync with the latest merge get caught and fixed, with a PR opened for the changes.
• Finance: Point the agent at two years of brokerage summaries, consolidated 1099s and tax returns and have it trace the recurring holdings creating the most avoidable fees and tax drag, with every figure cited to the exact file and page — all without a document ever reaching a chatbot.
• Startups: Ask why activation went flat, and the agent analyzes the funnel export locally to find where new signups drop off between install and first completed task, then posts the top insights straight to the team’s Slack channel.

Try Portable Computer today.

Hermes Agent — developed by Nous Research — is a general-purpose agent used by millions that excels at reliability and self-improvement. Model- and provider-agnostic, Hermes is built to run all day on local systems, making RTX PCs, RTX PRO workstations and DGX Spark a natural fit.

Configuring a local model in Hermes will provide users with one-click setup across RTX and DGX systems on Windows. The agent will automatically detect the NVIDIA GPU, select an appropriate model and configuration, and run it through integrated llama.cpp with NVIDIA inference optimizations already in place, eliminating manual model downloads and tuning. Support for Linux is coming soon.

Once it is running, Hermes works the way it does anywhere else. It uses tools, maintains context across tasks, remembers information between sessions and creates reusable skills over time, allowing the agent to become more capable with continued use. Running the model locally on a GPU keeps performance fast while keeping data on the system.

One-click local model setup is available now on Windows, with support coming soon to Linux. Learn more about Hermes Agent.

OpenClaw has become one of the defining projects of the open-agent movement — the largest AI project on GitHub, with more than 380K stars and a fast-growing community that’s building tools and skills across research, engineering, project management and everyday productivity.

NVIDIA, Microsoft and OpenClaw have been working together to make that experience easier to set up on Windows PCs. To reduce onboarding friction, the OpenClaw Windows App simplifies the process of setting up an optimized local model on any RTX GPU with at least 24GB of VRAM.

Learn more in the OpenClaw blog.

Faster Inference Gives Local Agents a Boost

Inference performance is critical to keeping local agents responsive. NVIDIA is continuing to collaborate with the open-source llama.cpp and vLLM communities to accelerate agentic workloads across local NVIDIA platforms.

llama.cpp delivers up to 1.9x higher throughput through kernel optimizations on a GeForce RTX 5090, enhanced speculative decoding techniques and faster prefill. ​

vLLM delivers 1.2x on RTX PRO 6000 Blackwell Workstation Edition and up to 1.4x on two DGX Spark clusters. New XQA attention kernels in FlashInfer and backend optimizations help to accelerate inference across both platforms.

These gains are available on the llama.cpp and vLLM inferencing backends. 

Users can also experience these via the LM Studio and Ollama applications.

Tap Idle PCs for More Local AI Compute With NVIDIA PAIR

More than half of U.S. households have two or more PCs, and much of that computing power sits idle throughout the day. NVIDIA Personal AI Router (PAIR) is a free, open source software tool that puts those systems to work together for local AI.

Agentic workflows often break complex tasks into smaller jobs that can run in parallel, but performance can slow when every request is competing for the same GPU. PAIR automatically discovers compatible PCs on a local network and routes independent inference requests to whichever system has capacity. It works with Ollama and LM Studio and can adapt as devices join or leave the network.

For example, a user could ask Hermes to create a “Sunday Reset” plan by sorting through a cluttered inbox and prioritizing what needs attention now, what can wait and what can be skipped. Hermes can split that work across multiple subagents, while PAIR distributes those jobs across available PCs instead of having them all wait on a single GPU.

The result is more compute for local agents, with more tasks running in parallel and the flexibility to move AI workloads to another PC while the main system is being used for gaming, creating or other work.

The NVIDIA PAIR beta is available for Windows, macOS and Linux through both graphical and terminal interfaces, supporting NVIDIA GeForce RTX 20 Series GPUs and newer, NVIDIA RTX PRO workstation GPUs (Turing architecture and newer), NVIDIA DGX Spark and Apple M4 or newer silicon.

Check out the NVIDIA tech blog to get started with NVIDIA PAIR. 

Powerful On Device Photo Editing With Cyberlink PhotoDirector AI PC Mode on RTX Spark

Open image and video models enable artists to experiment with Creative AI models on PCs.  This enables artists to iterate and explore concepts and ideas, without the dreaded token anxiety and keep more of their creative work private and on-device.

CyberLink’s new PhotoDirector AI PC Mode is one of the first applications to integrate these diffusion models directly into a creative software, and turn them into a creative tool at the finger tips of the artists. Coming to PhotoDirector 365 and optimized for NVIDIA RTX Spark when it launches, AI PC Mode users are getting AI-powered editing tools for generative editing, image enhancement, object and distraction removal, background removal and replacement, portrait refinement and the creation of entirely new visuals — with the flexibility to choose between local or cloud processing, depending on the task.

On NVIDIA GPUs, PhotoDirector uses TensorRT-RTX and FP8 to accelerate local AI.

Start using Cyberlink’s PhotoDirector 365 photo editing software and learn more about PhotoDirector AI PC Mode, launching with RTX Spark in October.

NVIDIA RTX Spark Windows PCs Arrive October 2026

NVIDIA RTX Spark is coming this October— and at IFA 2026, partners are showing off their hardware. At IFA, newly announced designs join the existing six OEMs shipping in October. Acer showed its compact desktop RTX Spark concept, and Lenovo announced its Yoga Pro  9n and Yoga 9n 2-in-1.  

RTX Spark is a new beginning for Windows PCs. One PC built for creators, gamers and AI agents. With a powerful 1 Petaflop RTX Blackwell GPU, up to 128GB of unified memory and a highly efficient 20-core Grace CPU, RTX Spark delivers incredible performance and efficiency. This superchip enables high performance thin laptops with all day battery life and compact desktops to power always-on agents. Paired with the new Windows Agent framework, it enables agents that run safely in the background under OS level control.

Last week at Gamescom, Electronic Arts, Embark and Ubisoft were among the latest game publishers and developers bringing their blockbuster titles to NVIDIA RTX Spark Windows PCs. They join the publishers that announced RTX Spark support at COMPUTEX in May, including KRAFTON, NetEase, Riot Games and XBOX. Read more.

Sign up to be notified when RTX Spark laptops and desktops are available.

#ICYMI: More Updates From NVIDIA Local AI 

NVIDIA Brings New RTX Tech and Games to Gamescom — NVIDIA released DLSS 4.5 Ray Reconstruction, featuring a new second-generation transformer model for improved image quality in ray-traced and path-traced games. Gamescom also brought new RTX announcements for titles including 007 First Light, CONTROL Resonant and Gears of War: E-Day, plus expanded game support for the upcoming NVIDIA RTX Spark.

Introducing DeepSeek Harness — DeepSeek’s new open source harness pairs with DeepSeek-V4-Flash to power local agentic coding workflows on NVIDIA DGX Station and multi-DGX Spark setups.

MLPerf Client v2.0 Expands AI PC Benchmarking — MLCommons released MLPerf Client v2.0, developed in collaboration with NVIDIA and other industry leaders. The update adds new benchmarks for agentic AI and image generation, alongside expanded LLM testing for real-world local AI workloads.

Follow NVIDIA RTX Spark on X, Instagram, TikTok and Facebook — and stay informed by subscribing to the NVIDIA Local AI newsletter. Follow NVIDIA Workstation on LinkedIn and X. 

See notice regarding software product information.

September is here with 28 more games streaming on GeForce NOW this month, led by a slam dunk: NBA 2K27 with the NVIDIA DLSS 5 3D-Guided Neural Rendering feature.

Through NVIDIA’s close collaboration with Visual Concepts and 2K, DLSS 5 brings a new level of lifelike lighting and material detail to the court — tuned by developers to support the game’s authentic broadcast-style presentation.

Members can also walk the line between human and vampire in Rebel Wolves’ The Blood of Dawnwalker and step into genma-haunted Kyoto in Capcom’s Onimusha: Way of the Sword — part of the 10 games joining the GeForce NOW library this week.

Take the Court With DLSS 5

NVIDIA, Visual Concepts and 2K have brought DLSS 5 3D-Guided Neural Rendering to NBA 2K27, delivering a new level of sports realism for players streaming from the cloud at launch. 

DLSS 5 enhances and tunes lighting and materials to make the action feel even more lifelike — from how light catches hair and skin to the detail preserved in the facial geometry of real-world athletes — achieving new levels of authenticity. 

GeForce NOW Ultimate members in NVIDIA-operated regions can stream NBA 2K27 across PCs, Macs, handhelds, mobile devices, TVs and more. DLSS 5 is available when streaming from a GeForce RTX 5080-powered rig in the cloud.

Tip off without waiting for downloads or managing storage, and experience NBA 2K27 from the cloud with the visual fidelity of DLSS 5.

Walk the Line Between Day and Night

Rebel Wolves’ newly launched The Blood of Dawnwalker is streaming from GeForce NOW. The open-world, dark-fantasy action role-playing game transports players to 14th-century Europe, where war, plague and rising vampire powers have reshaped history.

Play as Coen, a young Dawnwalker fighting to save his family while caught between his humanity and cursed vampiric strength. His abilities — and the choices available to him — change between day and night, shaping how players explore, fight and uncover the secrets of the world around them.

Sink teeth into the fresh adventure from nearly any supported device, with Ultimate members tapping into GeForce RTX 5080-class performance in the cloud. Skip the installs and let GeForce NOW handle the heavy lifting, leaving more time to take a bite out of everything Coen’s new world has to offer.

Skip the Downloads and Sharpen the Sword

Capcom’s Onimusha: Way of the Sword arrives on GeForce NOW at launch, giving members a way to jump straight into the series’ return instantly from the cloud.

Fight across a twisted vision of Edo-period Kyoto as a lone samurai wielding the mystical Oni Gauntlet. Master precise swordplay, unleash powerful abilities and absorb the souls of fallen enemies while battling relentless Genma beneath mysterious clouds of Malice.

Ultimate members can stream every clash with GeForce RTX 5080-class performance in the cloud across PCs, Macs, handhelds, mobile devices, TVs, the newly supported Firefox browser and more. Cut to the chase by skipping any downloads and storage management and step into battle.

In addition, members can look for the following titles streaming this week:

• Breathedge 2 (New release on Steam, Aug. 31)
• Serious Sam: Shatterverse (New release on Steam, Aug. 31)
• Crimson Moon (New release on Steam, Sept. 1)
• The Blood of Dawnwalker (New release on Steam, Sept. 2)
• NBA 2K27 (New release on Steam, Sept. 3)
• SpeedRunners 2: King of Speed (New release on Steam and Xbox, available on Game Pass, Sept. 3)
• Onimusha: Way of the Sword (New release on Steam, Sept. 4)
• How to Fish (Steam)
• Shelves and Sorcery: Tidy Up the Enchanted Shop (Steam)
• VHOLUME (Steam)

And look forward to the games coming throughout the month:

• Bus Simulator 27 (New release on Steam, Sept. 8)
• Halloween: The Game (New release on Steam, Sept. 8)
• Honeycomb: The World Beyond (New release on Steam, Sept. 8)
• Wanderburg (New release on Steam, Sept. 8)
• Shroom and Gloom (New release on Steam, Sept. 10)
• WARDOGS (New release on Steam, Sept. 10)
• Welcome to Elderfield (New release on Steam, Sept. 10)
• Active Matter (New release on Steam, Sept. 15)
• Aniimo (New release on Steam, Sept. 15)
• Train Sim World 7 (New release on Steam, Sept. 15)
• The Guild 1 Remake: Europa 1410 (New release on Steam, Sept. 17)
• CONTROL Resonant (New release on Steam, Sept. 24)
• Nivalis Nights (New release on Steam, Sept. 29)
• Mixtape (Steam) 
• Neon White (Steam) 
• Outer Wilds (Steam) 
• Stray (Steam)
• What Remains of Edith Finch (Steam)

All the Extras From August

In addition to the 26 games announced last month, 18 more joined the GeForce NOW library. 

• Cat Mail Co. (Steam)
• Monster Hunter Wilds Prologue Demo (Steam)
• Stars Reach (Steam)
• DIVE or DIE – Children of Rain (Steam)
• Escape the Backroom (Xbox, available on Game Pass)
• Hell Is Us (Xbox, available on Game Pass)
• Parcel Simulator (Steam)
• The Thaumaturge (Xbox, available on Game Pass)
• Tomb Raider IV-VI Remastered (Epic Games Store)
• STAR WARS Zero Company (EA app and Steam)
• Baldur’s Gate 3 (GOG)
• Clair Obscur: Expedition 33 (GOG)
• Cold Fear (GOG)
• IRON NEST: Heavy Turret Simulator (Steam)
• Kingdom Come: Deliverance II (GOG)
• Mount & Blade: Bannerlord II (GOG)
• No Man’s Sky (GOG)
• The Witcher 3: Wild Hunt (GOG)

Dates listed above reflect when games are released on their respective stores. GeForce NOW availability may vary, as games are onboarded after release and added throughout the week. Keep an eye on GeForce NOW channels and GFN Thursdays for availability updates on announced titles.

New to GeForce NOW? Start with a day pass to try premium cloud gaming before committing to a membership. Better yet, the cost of the day pass can be applied toward a first monthly membership — making it easy to try GeForce RTX-powered cloud gaming with the latest releases, then level up for more.

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

I’m excited to announce that NVIDIA has agreed to acquire Hugging Face for $12,930,300,000. Together, we will scale Hugging Face’s platform, strengthen its infrastructure and expand access to AI for developers and institutions worldwide.

Over the past decade, Clem, Julien, Thomas and the team at Hugging Face have built something remarkable: a vibrant home for the open model developer community.

More than 18 million developers, researchers and creators use Hugging Face to share more than 3 million models, 500,000 datasets and 1 million applications. More than 200,000 companies use the platform to discover, evaluate, customize and deploy AI.

Hugging Face will remain an open platform for the entire AI ecosystem. Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want. NVIDIA compute will not be required to build on or deploy through Hugging Face.

Hugging Face will continue to support open source and open weight models from across the ecosystem, from every model builder. It will continue to support multi-cloud and multi-accelerator development and deployment, so builders can use the hardware and infrastructure that best fit their work.

Recently, I coauthored an open letter on the importance of open weights to the AI economy. Joined by leaders from across the industry, we made a simple point: open weights broaden access to AI and help ensure that AI leadership is distributed across companies, institutions and communities.

Open models let startups, businesses, universities and public institutions build on advanced capabilities without training every model from scratch. They enable organizations to match the right model to the right job. That is how AI can advance safely, strengthen cybersecurity and sovereignty, accelerate innovation, and reach factories, hospitals, farms, classrooms and Main Street businesses around the world.

AI advances faster when people can build together.

NVIDIA has been committed to open weight models for years, demonstrated by multiyear investments and major contributions to open source platforms, including Hugging Face. NVIDIA has said that open models, data and tools broaden access to AI, and it has contributed hundreds of open models and datasets to Hugging Face as part of that effort.

• NVIDIA is the largest contributor of open models and data to Hugging Face, and our contributions continue to grow.
• NVIDIA has released more than 500 models on Hugging Face and more than 250 open datasets.
• We build our own models, libraries and tools in the open so developers everywhere can use them, modify them and build on top of them.

As the opportunity for open models accelerates, Hugging Face can serve the global AI community at unprecedented scale. NVIDIA’s infrastructure, engineering and global reach can help improve platform reliability, safety, model evaluation, inference and deployment capabilities, while preserving the open ecosystem that made Hugging Face foundational.

I am honored that Clem came to me as he considered the next chapter of Hugging Face and believed NVIDIA would be a great home for the company, its community and the future of open models. We share this vision, and the Hugging Face team will now bring their passion and expertise to a much larger canvas, with their same iconic brand.

To the millions of builders on Hugging Face: thank you for pushing the boundaries of what is possible. We can’t wait to build the future together with you. Together, we will make AI more open, more capable and more accessible to people and institutions around the world.

“We’re at an inflection point in cybersecurity,” Jensen Huang told a sold-out crowd at CrowdStrike’s Fal.Con 2026 in Las Vegas Tuesday. Attacks are now automated. Defense has to be, too. 

The NVIDIA founder and CEO joined CrowdStrike CEO and founder George Kurtz to announce CrowdStrike SafeMind, its agentic cybersecurity system developed by the CrowdStrike Cyber Superintelligence Lab.

“This is the beginning of a new age of cybersecurity,” Huang told the crowd of 10,000 security professionals. “On the one hand, the adversaries are going to be more armed than ever. On the other hand, all of you are going to be more armed than ever.” 

SafeMind combines CrowdStrike’s purpose-built, beyond frontier-capable models and customized agentic harnesses, with defensive models built on NVIDIA Nemotron, in a continuous coevolution loop where offense and defense repeatedly challenge and improve each other. 

CrowdStrike also announced CrowdStrike Falcon IQ to operationalize Project QuiltWorks through agentic workload automation and expanded its CrowdStrike Guardian AI safety solution.

“We have asymmetric advantages because we have a large community of cybersecurity experts who want to work with each other and keep the world safe,” Huang told the crowd. 

CrowdStrike’s annual conference drew security leaders from financial services, healthcare, the public sector and critical infrastructure.

“The real gap that I saw was that the attackers had frontier AI, and the defenders didn’t,” Kurtz told them. “And that changes now.”

SafeMind

CrowdStrike built SafeMind’s defensive model using NVIDIA Nemotron open models, post-trained with CrowdStrike’s cyber experience and threat data. The SafeMind models are paired with proprietary cybersecurity harnesses optimized to work as an agentic stack.

The result ships natively in the CrowdStrike Falcon platform as SafeMind, CrowdStrike’s agentic cybersecurity system. SafeMind brings offensive and defensive AI together in a continuous coevolution loop, where each side adapts to and strengthens the other. This process continuously hardens the security of the customer environment until attacks are unsuccessful.

“Your decade and a half of security data that we can train on — we can take a frontier model and make it essentially a super AGI that is incredibly good at cybersecurity,” Huang told Kurtz.

“Together with NVIDIA, we built cybersecurity’s first complete agentic system for cybersecurity, including the first frontier models and harness purpose-built for defenders,” Kurtz said. “This isn’t a copilot baked into someone else’s intelligence. It’s not a chatbot with a security skin. It is a frontier-class model built and trained by CrowdStrike on our data in partnership with NVIDIA.”

NVIDIA Nemotron 3 Ultra orchestrates the defensive agent harness. A fine-tuned Nemotron 3 Super powers SafeMind’s rule-generation sub-agent. 

By post-training Nemotron with CrowdStrike data, CrowdStrike internal evaluations showed that the Blue Solano model — based on Nemotron 3 Super — delivered higher accuracy rates than leading frontier models at 99% lower cost. 

While SafeMind can operate as a complete system, the models can be used independently to empower defenders to stay ahead of the adversary. Security experts can also pair their own models with CrowdStrike’s custom harnesses, giving customers the flexibility to use the right models and capabilities for their environment.

“The harness is essentially the exoskeleton of the large language model,” Huang said. “The large language model is the brain. The exoskeleton turns it into an agent — and this exoskeleton doesn’t have to be the same shape and capability for every domain.”

When AI Is the Defense

AI-enabled attacks rose 89% in the past year, and the fastest eCrime breakout time has reached 27 seconds, according to CrowdStrike. Human-speed response isn’t defense. It’s documentation.

“There are many applications in the world where you must have the ability to fine-tune, to post-train — to create an AI that is super good at a particular domain,” Huang said. “Nemotron was created for precisely that. Completely free. Incredibly fast. You have the ability to have an asymmetric advantage against whatever comes your way.”

With open Nemotron as the base, CrowdStrike’s security teams post-trained on their own threat data without sending it to an outside provider, and customized the AI to their environment. 

That’s not possible with a closed frontier model, and in security, the ability to inspect what’s defending matters. 

Red vs. Blue

NVIDIA announced its work testing the CrowdStrike SafeMind models and harnesses in a high-fidelity cyber agent environment running as a simulation of the NVIDIA network. 

The testing runs SafeMind in an offensive-defensive loop for adversarial coevolution. An offensive red-team agent finds the exploit, a blue-team defensive agent closes it and the findings become actionable detections to block attacks. 

The red-agent harness runs Recon, Assault and Compromise sub-agents executing attack paths inside the cyber agent environment. The blue-agent harness monitors via Falcon sensors, generates detection candidates, validates them and promotes them. 

CrowdStrike built the test environment with NVIDIA: a digital twin of NVIDIA’s own accelerated computing infrastructure, validated against NVIDIA’s real threat landscape.

“The basic framework of SafeMind — an adversarial model acting on a digital twin of the environment, with a defender model in a continuous cat-and-mouse loop, eventually learning how to secure itself — this basic framework applies to robotics, edge computing, enterprise computing and just about everything,” Huang said.

CrowdStrike also announced Falcon IQ. NVIDIA Nemotron models help to power the agentic engine at the heart of Charlotte AI AgentWorks, CrowdStrike’s no-code agent development platform where Falcon IQ runs. 

Falcon IQ uses more than 50 agents working together as a unified agentic workforce to automate the most time-intensive workflows in assessment, prioritization and remediation. 

Partners use Falcon IQ to deliver customized findings, recommendations and executive outputs to customers. Charlotte AI AgentWorks enables every Falcon user to build their own agentic security workforce.

The Full Stack

CrowdStrike has thousands of customer organizations generating trillions of daily security events. 

With NVIDIA’s full-stack accelerated computing platform, the collaboration runs from the chips up through the models to the harnesses acting on what those models find. For Kurtz, that’s the point. 

“The crowd in CrowdStrike,” Kurtz added, “is the asymmetry that puts the defenders in a unique position to defeat the adversary.”