NVIDIA GTC Keynote 2026

NVIDIA GTC Keynote 2026 — NVIDIA

NVIDIA

AI-assisted summary of NVIDIA GTC Keynote 2026 — original source. Check the original for full context.

Source published:

Key ideas

  • Jensen Huang projected over $1 trillion in visible compute demand and purchase orders through 2027, driven by a one-million-fold surge in computing demand as workloads transition from training to agentic inference. (52:23)
  • NVIDIA unveiled the Vera Rubin platform featuring seven chips, five rack architectures, 3.6 exaflops of compute, and 45°C warm-water cooling, marking a 40-million-fold compute expansion over the 2016 DGX-1. (67:56)
  • Following the acquisition of Groq's engineering team and IP, NVIDIA introduced a disaggregated inference architecture that routes prefill to Vera Rubin and token decode to Groq LPUs, yielding a 35x performance boost at interactive latency tiers. (88:15)
  • To support autonomous agent workflows and facilitate the transition from legacy SaaS to 'Agentic-as-a-Service,' NVIDIA partnered with OpenClaw creator Peter Steinberger to introduce NemoClaw, an enterprise-grade sandboxed runtime. (112:48)

Detailed notes

## Background This keynote presentation was delivered by NVIDIA Founder and CEO Jensen Huang and published on YouTube by NVIDIA on March 16, 2026. According to the video description, the keynote unveils breakthroughs across accelerated computing, agentic AI, industrial AI factories, and physical robotics powering next-generation intelligent systems.

## TL;DR Jensen Huang declares that the global AI landscape has shifted decisively from training to inference, driving an estimated one-million-fold surge in computing demand and creating over $1 trillion in visible hardware visibility through 2027. To meet this inflection point, NVIDIA introduced the Vera Rubin platform alongside the integration of Groq's low-latency LPUs, creating a disaggregated architecture capable of scaling token generation up to 350x within a 1-gigawatt power envelope. The presentation expands NVIDIA's full-stack strategy with the DSX digital twin platform for AI factory simulation, the open-core NemoClaw enterprise framework for autonomous agent swarms, and open frontier model ecosystems like Nemotron and Cosmos. The keynote concludes by demonstrating physical AI in action, highlighted by real-time autonomous vehicle reasoning via Alpamayo, widespread industrial robotics deployments, and an on-stage bipedal robotic appearance by Disney's Olaf powered by Jetson and the Newton physics engine.

## Atmosphere & style The presentation carries the electrifying, highly polished momentum of a peak-era tech vision keynote, balancing massive industrial numbers with Huang's signature self-deprecating humor and candid audience banter. Moving briskly through a sprawling technical agenda, Huang transitions smoothly between high-level macroeconomics, dense architectural slides, and unscripted moments—such as fumbling a vertically sliding server tray and teasing the audience about rising GPU values. The atmosphere culminates in a sense of theatrical wonder during the final act, where abstract physics simulations and AI models materialize into physical reality as a small robotic snowman strolls onto the stage.

## Bulletin - [52:23](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3143s) Jensen Huang projected over $1 trillion in visible compute demand and purchase orders through 2027, driven by a one-million-fold surge in computing demand as workloads transition from training to agentic inference. - [01:07:56](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4076s) NVIDIA unveiled the Vera Rubin platform featuring seven chips, five rack architectures, 3.6 exaflops of compute, and 45°C warm-water cooling, marking a 40-million-fold compute expansion over the 2016 DGX-1. - [01:28:15](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5295s) Following the acquisition of Groq's engineering team and IP, NVIDIA introduced a disaggregated inference architecture that routes prefill to Vera Rubin and token decode to Groq LPUs, yielding a 35x performance boost at interactive latency tiers. - [01:52:48](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6768s) To support autonomous agent workflows and facilitate the transition from legacy SaaS to 'Agentic-as-a-Service,' NVIDIA partnered with OpenClaw creator Peter Steinberger to introduce NemoClaw, an enterprise-grade sandboxed runtime.

## Key takeaways - [00:08](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=8s) Computing has transformed into an industrial process centered on generating tokens as the fundamental commodity of artificial intelligence across science, enterprise software, and physical industries. - [06:02](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=362s) CUDA's 20-year architectural flywheel maintains backward compatibility across hundreds of millions of deployed GPUs, allowing older hardware such as Ampere to appreciate in cloud rental value via software optimizations. - [13:50](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=830s) Trustworthy generative and agentic AI depends on grounding probabilistic neural network models in structured enterprise dataframes and relational tables accelerated by cuDF and cuVS. - [29:00](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1740s) NVIDIA's core strategy combines deep vertical integration across algorithms, libraries, systems, and silicon with an open horizontal deployment model across cloud, on-premises, edge, and sovereign environments. - [46:13](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=2773s) AI has progressed across three waves—generative AI, reasoning AI, and agentic AI—with autonomous agents increasing compute demand per task by roughly 10,000x through long internal reasoning traces. - [52:23](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3143s) Visible demand through 2027 exceeds $1 trillion, split roughly 60% across top hyperscalers and 40% across sovereign clouds, regional cloud service providers, enterprise IT, and robotics. - [01:00:34](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3634s) Under fixed gigawatt utility power constraints, token cost economics dominate; Grace Blackwell NVLink 72 delivers 50x higher performance-per-watt and 35x lower cost per token compared to Hopper H200. - [01:07:56](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4076s) The Vera Rubin platform integrates seven chips and five rack systems to deliver 3.6 exaflops of compute and 260 TB/s of all-to-all fabric bandwidth, supported by 45°C warm-water cooling and cable-free rack backplanes. - [01:17:38](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4658s) The Kyber rack architecture uses a vertical midplane design to connect 144 Rubin Ultra GPUs in a single unified NVLink domain without relying on horizontal copper cabling bundles. - [01:28:15](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5295s) Dynamo orchestrates disaggregated inference by running compute-heavy prefill and attention on Vera Rubin GPUs while offloading memory-bandwidth-bound token decode streaming to Groq LPUs. - [01:40:08](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6008s) NVIDIA DSX creates physics-based digital twins in Omniverse to simulate mechanical, thermal, electrical, and optical networking behaviors of gigawatt-scale AI factories before physical construction. - [01:52:48](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6768s) Enterprise software is transitioning from SaaS into 'Agentic-as-a-Service' (AaaS), requiring enterprise guardrails like NemoClaw and OpenShell to manage agent execution rights, sensitive data access, and network egress. - [01:57:11](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7031s) NVIDIA committed to long-term iterative development across six open frontier model families—Nemotron, Cosmos, Alpamayo, GR00T, BioNeMo, and Earth-2—backed by the global Nemotron Coalition. - [02:05:00](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7500s) Autonomous vehicle technology has reached an inflection point, with automakers producing 18 million passenger vehicles annually adopting NVIDIA's Robotaxi-Ready platform alongside an Uber rideshare partnership. - [02:08:41](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7721s) Physical AI development relies on a three-computer loop connecting AI supercomputer training (DGX / Vera Rubin), Omniverse simulation using the Newton physics engine (OVX), and edge vehicle or robot control (DRIVE Thor / Jetson).

## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=8s) The Genesis of Tokens and the Intelligence Factory

The keynote opens with an atmospheric, cinematic visual prologue illustrating the transformation of modern computing into an industrial process of intelligence. A narrator describes a new breed of factory designed not for physical goods, but as a generator of tokens—the fundamental building blocks of artificial intelligence. These digital tokens synthesize human knowledge, clean energy systems, and cosmic exploration, bridging the gap between simulated and physical reality.

The sequence showcases tokens directing robotic learning in virtual environments before those skills are perfected in the physical world, from autonomous agricultural harvesters cutting through fields to life-critical applications in healthcare, working where human hands cannot so that "the smallest hearts beat stronger." Reaching outward toward deep-space operations such as "Star Cloud One," the film frames token generation as an unprecedented planetary-scale endeavor designed to empower global industry.

As the film concludes, the narration summarizes the scope of this transformation: “Tokens have opened a new frontier, turning data into knowledge and drawing on all we have learned.”

- **Token Applications Highlighted:** Autonomous robotics, space exploration, planetary clean energy management, precision agriculture, and medical diagnostics. - **Star Cloud One:** Featured as a symbolic milestone for autonomous systems operating well beyond manual human reach.

---

[03:18](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=198s) Welcome to GTC 2026 and NVIDIA's Platform Strategy

Jensen Huang steps onto the stage to an enthusiastic arena audience, humorously reminding the crowd that despite the rock-concert energy and people lining up before dawn, “I just want to remind you this is a tech conference.” He frames the event around NVIDIA’s foundational mission as a platform company, noting that while NVIDIA has historically centered its narrative on CUDA-X and its computing systems, the company is formally establishing a third pillar: the AI factory platform.

Before unpacking the architecture, Huang expresses gratitude to his pre-game show hosts—Sarah Guo of Conviction, Alfred Lin of Sequoia Capital (NVIDIA’s first venture capital backer), and Gavin Baker (NVIDIA’s first major institutional investor)—praising their technological depth and industry perspective. He turns to the sheer scale of the gathering, noting that 450 corporate sponsors, 1,000 technical breakout sessions, and 2,000 speakers are in attendance, effectively representing all sectors of the modern $100 trillion global economy.

Huang articulates the conference roadmap across what he terms the "five-layer cake" of artificial intelligence: 1. **Land, power, and shell** (the physical facilities and energy footprint) 2. **Infrastructure** (networking, cooling, and mechanical integration) 3. **Chips** (accelerators and processors) 4. **Platforms and models** (frameworks, runtime environments, and foundational AI) 5. **Applications** (the ultimate economic engine that drives industry adoption)

Underscoring the holistic scope of the conference, Huang says: “This conference is going to cover every single layer of the five layer cake of artificial intelligence from land, power, and shell to infrastructure to chips to the platforms, the models.”

---

[06:02](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=362s) 20 Years of CUDA and the Accelerated Computing Flywheel

Marking the 20th anniversary of CUDA, Huang reflects on two decades of unwavering commitment to a single computing architecture. He highlights CUDA's roots in Single Instruction, Multiple Threads (SIMT), a paradigm that allowed developers to write simple scalar code that automatically executed as massively parallel multithreaded tasks—making it vastly more programmable than traditional SIMD (Single Instruction, Multiple Data) approaches. Over time, NVIDIA expanded the programming model with mathematical tiles to make it intuitive to program Tensor Cores and handle the tensor mathematics foundational to modern AI. Today, the CUDA ecosystem encompasses thousands of tools, compilers, libraries, and hundreds of thousands of public repositories.

```mermaid flowchart TD accTitle: The NVIDIA CUDA Flywheel accDescr: A cyclical feedback loop showing how CUDA installed base drives software breakthroughs and market expansion.

subgraph HardwareBase["Installed Base Engine"] A["Massive Installed Base<br/>(Hundreds of Millions of GPUs)"] E["Extended Useful Life &<br/>Decreased Computing Costs"] end

subgraph SoftwareEcosystem["Developer & Algorithm Loop"] B["Global Developer Community<br/>& CUDA-X Libraries"] C["Algorithmic Breakthroughs<br/>(e.g., Deep Learning)"] D["Emergence of New Markets<br/>& Commercial Ecosystems"] end

A --> B B --> C C --> D D --> A A -.-> E E -.-> B ```

Huang displays a classic architectural flywheel diagram, calling it the foundation of “100% of NVIDIA's strategies.” The hardest component to build was the bottom layer: an installed base spanning hundreds of millions of CUDA GPUs across every hyperscale cloud, server manufacturer, and technical industry. That base attracts researchers and developers, whose algorithmic breakthroughs create entirely new commercial markets. These markets bring in new ecosystem partners, which further expands the installed base. Because every generation remains architecturally backward-compatible, software updates consistently improve the performance of previously deployed chips. Huang reveals that cloud rental pricing for the six-year-old Ampere architecture is actually increasing rather than depreciating, because software optimizations continuously expand its throughput.

Making the case for architectural continuity, Huang explains: “Meanwhile, as we continue to nurture and continue to update software over its life, not only do you get the first time pop, you get the continuous cost reduction of accelerated computing over time, and we're willing to nurture willing to support every single one of these GPUs in the world because they're all architecturally compatible.”

- **Key Tenets of CUDA's Longevity:** - SIMT execution model enabling straightforward multithreaded scaling. - Integration of tile-based abstractions for Tensor Core mathematical structures. - Architectural backward-compatibility that increases residual value and useful lifecycle across older hardware deployments like Ampere.

---

[10:23](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=623s) GeForce: The 25-Year Foundation and AlexNet Big Bang

Huang traces the roots of CUDA back 25 years to the launch of GeForce and the invention of the programmable pixel shader. He jokes that GeForce was NVIDIA’s greatest long-term marketing strategy: attracting enthusiastic young gamers long before they had purchasing power of their own. As Huang puts it to laughter from the audience: “Your parents paid for you to be NVIDIA customers. And every single year they paid up year after year after year until someday you became an amazing computer scientist and became a proper customer, a proper developer.”

That initial invention of the programmable accelerator paved the way for CUDA five years later. Huang recalls that the financial commitment to integrate CUDA into every consumer GeForce GPU was so immense that it consumed the vast majority of NVIDIA’s operating profits at a time when the company could barely afford it. Yet that stubborn persistence built the hardware footprint that sparked the modern AI boom. It was consumer GeForce GTX cards that allowed Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton, and Andrew Ng to train AlexNet and unlock deep learning. Around eight to ten years ago, NVIDIA introduced RTX, integrating hardware-accelerated real-time ray tracing and Tensor Cores, banking on the idea that AI would eventually reinvent computer graphics just as graphics hardware had invented modern AI.

Highlighting that historic turning point, Huang reflects: “GeForce therefore enabled Alex Krizhevsky and Ilya Sutskever and Geoff Hinton, Andrew Ng, and so many others to discover that the GPU could be their friend in accelerating deep learning.”

---

[13:50](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=830s) DLSS 5: Neural Rendering and the Fusion of 3D Graphics with Generative AI

Huang presents the next leap in visual computing: DLSS 5 and the dawn of neural rendering. DLSS 5 merges deterministic, controllable 3D graphics pipelines with generative, probabilistic AI models. Following a visual demonstration showing hyper-realistic, fully dynamic rendering in real time, Huang deconstructs the underlying philosophy. Traditional 3D graphics represent the structured ground truth of virtual worlds—geometry, physics, spatial coordinates, and deterministic lighting. Generative AI, by contrast, operates probabilistically to produce visual richness and detail that would be computationally prohibitive to simulate manually.

By anchoring generative models to structured geometric inputs, the rendering engine avoids hallucinations and maintains precise physical control while synthesizing photorealistic imagery on the fly. Huang emphasizes that this architectural pattern—grounding generative diffusion or transformer models in rigid, structured datasets—is not limited to video games. It represents the exact technical template that every enterprise sector must adopt to build reliable, auditable artificial intelligence.

Huang delivers the core takeaway for the enterprise audience: “Structured data is the foundation of trustworthy AI.”

---

[16:19](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=979s) Structured Data: The Ground Truth of Enterprise AI

Huang warns the audience to prepare themselves as he transitions to an intricate technical architectural diagram: “This is going to scare you a little bit. I'm going to flip the slide and don't gasp.” He jokes that his engineering and presentation teams repeatedly begged him to cut the complex schematic from the keynote deck, but he insisted on keeping it, teasing the crowd: “These seats are free, for some of you. So this is your price of admission.”

The schematic maps out the structured data ecosystem: SQL, Apache Spark, Pandas, Velox, and data-lakehouse platforms including Snowflake, Databricks, Amazon EMR, Microsoft Azure Fabric, and Google Cloud BigQuery. Huang points out that enterprises run entirely on dataframes—massive tables and structured spreadsheets that log supply chains, financial transactions, customer interactions, and inventory. This structured information constitutes the literal ground truth of enterprise computing. With autonomous AI agents entering business operations, structured data processing can no longer remain a batch-oriented, overnight task; it must be accelerated by GPUs so queries can run continuously and cost-effectively throughout the day.

Huang stresses the central role of these business tables in guiding enterprise AI: “These data frames are giant spreadsheets and they hold all of life's information. This is the structured data, the ground truth of business.”

- **Enterprise Data Platforms Highlighted:** - **Frameworks & Engines:** SQL, Apache Spark, Pandas, Velox. - **Cloud Data Ecosystems:** Snowflake, Databricks, Amazon EMR, Microsoft Azure Fabric, Google Cloud BigQuery. - **Strategic Imperative:** Accelerating dataframe manipulation on GPUs to enable AI agents to query corporate ground truth in real time at sustainable costs.

[17:51](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1071s) Unstructured Data, Vector Search, and Foundational Libraries: cuDF and cuVS

Looking toward the evolution of software, Jensen Huang explains that structured data tables will increasingly be queried not by human analysts, but directly by autonomous AI agents that operate at speeds vastly exceeding human capabilities. Beyond relational schemas lies unstructured data—the generative domain of PDFs, videos, spoken audio, and arbitrary documents that makes up approximately 90% of all information created annually. Historically, this data remained essentially inert: people saved files to disk, unable to search or query them meaningfully because unstructured information cannot be indexed without first grasping its semantic intent.

Just as modern AI conquered multi-modal perception and understanding, that same intelligence is now directed at reading complex files like PDFs, comprehending their context, and embedding that meaning into high-dimensional vector representations that can be indexed and queried. To anchor this transition, NVIDIA introduced two foundational GPU-accelerated libraries analogous to what RTX represents for 3D graphics:

- **cuDF:** Accelerates DataFrames and relational, structured data processing pipelines. - **cuVS:** Accelerates vector stores, semantic search, and unstructured data indexing for AI workloads.

Because enterprise data infrastructures evolved over decades across countless legacy platforms, deep ecosystem integration has required immense engineering effort. Jensen emphasizes how vital these software layers have become, noting: “These two platforms are going to be two of the most important platforms in the future.”

```mermaid flowchart TD accTitle: Data Processing Architecture for AI accDescr: Diagram showing how structured and unstructured data feed into cuDF and cuVS libraries, powering enterprise cloud and edge AI platforms.

subgraph InputData["Enterprise Data Types"] SD["Structured Data (Relational tables, DataFrames)"] UD["Unstructured Data (PDFs, videos, audio ~90% of data)"] end

subgraph Foundation["NVIDIA Accelerated Libraries"] cuDF["cuDF (Accelerated DataFrames & SQL)"] cuVS["cuVS (Accelerated Vector Search & Semantics)"] end

subgraph Ecosystem["Integration Layer"] Engines["Databases, Data Warehouses & Vector Stores"] CloudDeploy["Cloud Services, OEMs & Hybrid Systems"] end

SD --> cuDF UD -->|"Multimodal Semantic Embedding"| cuVS cuDF --> Engines cuVS --> Engines Engines --> CloudDeploy ```

---

[20:16](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1216s) IBM Watsonx.data and Accelerating Enterprise Data Processing

Marking a major milestone in enterprise data acceleration, NVIDIA partnered with IBM—the pioneer of System/360, relational data warehousing, and SQL—to modernize core database engines for artificial intelligence. A keynote video highlights how IBM Watsonx.data SQL engines are being accelerated directly by NVIDIA GPU computing libraries. As Jensen states in the narration: “Data is the ground truth that gives AI context and meaning.”

CPUs are no longer able to keep pace with the massive ingestion demands required to feed real-time enterprise AI. To illustrate the tangible impact, the keynote examines Nestlé's global supply chain operations across 185 countries. Nestlé's order-to-cash datamart aggregates every single order and logistics milestone to guide thousands of daily business decisions. On legacy CPU-based systems, refreshing this datamart was so computationally expensive that it could only run a few times per day. By deploying accelerated Watsonx.data powered by NVIDIA GPUs, Nestlé achieved:

- **5x faster execution** for data processing workloads. - **83% lower compute costs** compared to standard CPU infrastructure.

---

[21:53](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1313s) Beyond Moore's Law: Driving Cost Down and Speed Up

Jensen points out that accelerated computing applies equally to on-premises enterprise data centers and public cloud providers. In the hardware domain, Dell integrated cuDF and cuVS into the Dell AI Data Platform, yielding major performance leaps on customer workloads such as those run by NTT DATA. In public clouds, NVIDIA's long-standing collaboration with Google Cloud has expanded from Vertex AI into BigQuery. As a concrete benchmark, accelerating BigQuery data pipelines allowed Snapchat to slash computing costs by nearly 80%.

Jensen connects these cost reductions directly to the fundamental economics of computing. For decades, the industry relied on Moore's law to double performance at flat capital costs every couple of years, effectively cutting compute expenses in half on a predictable schedule. With physical scaling hitting physical limits, classic Moore's law has stalled. Jensen explains: “Accelerated computing allows us to take these giant leaps forward.” Because NVIDIA writes and optimizes algorithms continuously across a massive, uniform installed hardware base, performance increases compound over time without requiring hardware reinvestment at every step.

---

[23:55](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1435s) Landing Developers Across the Global Cloud Ecosystem

NVIDIA’s go-to-market architecture follows a repeatable flywheel: build accelerated computing platforms, layer optimized domain libraries on top (such as RTX, cuDF, and cuVS), integrate them into cloud platforms and OEM systems, and funnel major software developers directly onto those clouds. Demonstrating this strategy, Jensen walks through the major cloud partnerships:

- **Google Cloud:** Accelerating Vertex AI, BigQuery, and both PyTorch and JAX/XLA environments. NVIDIA acts as the bridge landing enterprise developers and applications—including Baseten, CrowdStrike, Puma, and Salesforce—onto Google Cloud infrastructure. - **Amazon Web Services (AWS):** Deep integrations across Amazon EMR, SageMaker, and Bedrock. Jensen announces that NVIDIA is bringing OpenAI workloads onto AWS to help alleviate OpenAI’s severe compute bottlenecks while driving massive consumption on AWS. - **Microsoft Azure:** Tracing back to the very first NVIDIA-built A100 supercomputer installed at Azure (which enabled the initial training breakthroughs for OpenAI), the partnership now extends across Azure Cloud, Azure AI Foundry, and Bing search. - **Oracle Cloud Infrastructure (OCI):** Jensen humorously recalls that when setting up their relationship, NVIDIA became Oracle's very first AI customer after he explained the concept of an AI cloud to them, paving the way for OCI to host Cohere, Fireworks AI, and OpenAI. - **CoreWeave:** Recognized as the world's first dedicated AI-native cloud company built from scratch for the singular purpose of hosting and provisioning GPU compute clusters.

A critical technical capability underpinning these deployments across all regions is hardware-enforced **confidential computing**. Implemented directly in NVIDIA GPUs, confidential computing ensures that cloud operators and hypervisors cannot inspect or tamper with customer data or proprietary model weights. This trust boundary allows high-stakes frontier models from creators like OpenAI and Anthropic, alongside mission-critical electronic design automation (EDA) and computer-aided engineering (CAE) workflows from partners like Synopsys, to be securely deployed into shared public clouds and sovereign regional data centers.

Reflecting on how cloud providers view this dynamic, Jensen jokes about their constant appetite for new GPU-accelerated tenants: “And so as you could see, most of our cloud service providers love working with us, and they're always asking us to land the next customer on their cloud. And I just want to let you know, there are a lot of customers. We're going to accelerate everybody.”

---

[29:00](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1740s) Air-Gapped Sovereign AI and Vertical Integration

Deploying AI extends beyond centralized hyperscalers into sovereign data centers and field operations. Jensen highlights a joint venture between Palantir, Dell, and NVIDIA that deploys the Palantir Ontology and AI Platform on-site into air-gapped, sovereign, or tactical military environments anywhere in the world. This capability relies on NVIDIA’s complete stack, spanning confidential computing hardware, system-level design, and end-to-end acceleration libraries for structured tables and vector spaces alike.

Summarizing NVIDIA’s broader identity, Jensen explains that the company operates under a unique architectural model: “NVIDIA is vertically integrated the world's first vertically integrated but horizontally open company.” He emphasizes that computing cannot be sped up by looking at components in isolation: “Accelerated computing is not a chip problem. Accelerated computing is not a systems problem. Accelerated computing has a missing word. We just never say it anymore. Application acceleration.” If a chip could magically run arbitrary legacy code faster without full-stack redesign, it would simply be a standard CPU—an approach that has reached its historical limits.

[30:47](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1847s) Vertically Integrated, Horizontally Open: NVIDIA’s Computing Philosophy

Jensen Huang explains the structural philosophy behind NVIDIA's long-term strategy. The path forward for sustained application speedups and dramatic cost reductions cannot rely on generic hardware scaling alone; it demands deep domain-specific and application-specific acceleration. Over time, NVIDIA shortened that concept simply to "accelerated computing." Because of this imperative, NVIDIA had to build library after library, conquer domain after domain, and penetrate vertical after vertical. This mandates operating as a vertically integrated computing company, mastering everything from foundational algorithms down to microchips, whether the target deployment environment is a centralized hyperscale data center, an on-premises enterprise cluster, an edge node, or a real-time robotic system.

```mermaid flowchart TD accTitle: NVIDIA Vertical Integration and Open Platform Architecture accDescr: Diagram showing how NVIDIA integrates down from domain algorithms to hardware while remaining horizontally open across deployment platforms.

subgraph SW["Vertical Integration: Software & Algorithms"] ALGO["Domain Algorithms & Applications"] CUDAX["CUDA & Domain Libraries (CUDA-X)"] end

subgraph HW["Vertical Integration: Systems & Chips"] CHIP["Processors & Silicon Architectures"] SYS["Full-Stack Computing Systems"] end

subgraph DEP["Horizontally Open Deployment Platforms"] CLOUD["Cloud & Data Center"] ONPREM["Enterprise On-Prem"] EDGE["Edge & Telecommunications"] ROBOT["Robotics & Embedded Systems"] end

ALGO --> CUDAX CUDAX --> CHIP CHIP --> SYS SYS -.->|"Integrates into"| CLOUD SYS -.->|"Integrates into"| ONPREM SYS -.->|"Integrates into"| EDGE SYS -.->|"Integrates into"| ROBOT ```

While NVIDIA insists on full vertical integration internally to optimize the computational pipeline, its market strategy remains horizontally open. NVIDIA does not lock customers into proprietary end-to-end silos; instead, it packages algorithms into accessible software libraries and modular systems that integrate seamlessly into any external platform or enterprise infrastructure.

As Jensen emphasizes: > “We are a vertically integrated computing company. There is no other way. We have to understand the applications. We have to understand the domain. We have to understand fundamentally the algorithms, and we have to figure out how to deploy the algorithm in whatever scenario it wants to be deployed, whether it's a data center, cloud on prem at the edge, or in a robotic system.”

---

[32:08](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=1928s) Transforming Global Verticals: From Quant Finance to $50T Manufacturing

Looking out at the GTC audience, Jensen highlights the sheer breadth of industries present. Both upstream suppliers and downstream partners span firms that are 50, 70, or even 150 years old, many having achieved record-breaking business years by plugging into NVIDIA's computing ecosystem. Notably, the largest percentage of attendees at GTC 2026 comes from the financial services sector—a development Jensen quips about, hoping the audience is made of technical developers rather than day traders. Quantitative finance is undergoing a profound paradigm shift: moving away from classical machine learning driven by manual human feature engineering toward autonomous supercomputers ingesting massive datasets to discover underlying patterns, effectively experiencing their own deep learning and transformer moment.

Across industry after industry, tailored domain libraries are unlocking massive economic transformations: - **Healthcare & Life Sciences:** Entering its "ChatGPT moment," addressing AI biology and physics for de novo drug discovery, intelligent clinical diagnostic copilots, and hospital-grade customer service agents. - **Industrial & Manufacturing:** Powering the largest industrial build-out in human history to fabricate AI factories, semiconductor fabrication facilities, and automated computer manufacturing plants across a $50 trillion sector. - **Media, Entertainment & Gaming:** Deploying real-time AI pipelines using NVIDIA Holoscan for low-latency live broadcast translations, gaming graphics augmentation, and interactive streaming. - **Quantum Computing:** Collaborating with 35 different quantum hardware and software companies to engineer next-generation hybrid quantum-GPU platforms. - **Retail & Consumer Packaged Goods (CPG):** Driving autonomous supply-chain logistics and multi-agent customer shopping systems across a $35 trillion market. - **Robotics & Physical AI:** Capitalizing on a decade of NVIDIA robotics research, building the foundational "three-computer" architecture (simulation, training, and real-time edge execution), with 110 physical robots featured across the conference floor. - **Telecommunications:** Re-architecting the $2 trillion global telecommunications footprint. Base stations, which previously performed only single-function radio signal processing, are converting into edge AI computing nodes running the NVIDIA Aerial AI-RAN platform in deep partnership with Nokia and T-Mobile.

Jensen reiterates that algorithm invention remains the soul of NVIDIA: > “The libraries is the crown jewels of our company. It is what makes it possible for that platform, the computing platform, to be activated in service of solving a problem making impact.”

At this GTC alone, NVIDIA is releasing or updating approximately 70 to 100 specialized libraries and 40 AI foundation models, building on the revolutionary legacy of libraries like cuDNN, which initially ignited the modern AI revolution.

---

[38:26](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=2306s) The CUDA-X Montage and the Power of Pure Simulation

NVIDIA screens a video showcase celebrating 20 years of CUDA evolution, illustrating how a unified architecture morphed into thousands of CUDA-X libraries serving every discipline of scientific discovery and engineering. Jensen narrates the sequence, listing flagship algorithmic frameworks: - **cuOpt:** For high-scale logistics and combinatorial decision optimization. - **cuLitho:** Accelerating computational lithography for advanced semiconductor manufacturing. - **cuDSS:** Direct sparse solvers handling complex mathematical systems. - **cuEquivariance:** Accelerating geometry-aware neural networks for molecular dynamics and structural chemistry. - **Aerial:** Driving software-defined AI-RAN for next-generation telecommunications. - **Warp:** Differentiable physics simulations for robotics and graphics. - **Parabricks:** High-throughput genomics sequencing and bioinformatics.

Returning to the stage, Jensen reflects on the striking photorealism and dynamic fidelity of the video reel, pointing out that not a single frame was hand-animated or mechanically rigged. Every visual was an absolute mathematical simulation driven by first-principles physics solvers, surrogate AI physical models, or physical AI robotic controllers.

Summarizing NVIDIA's mission through the lens of computational science, Jensen states: > “Everything was completely simulated. That is what fundamentally NVIDIA does. It is through the connection of understanding of the algorithms with our computing platforms that we're able to open up to unlock these opportunities.”

---

[43:15](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=2595s) The Rise of AI Natives and the Trillion-Token Economy

While historic industry titans like Walmart, L’Oréal, JPMorgan Chase, Roche, and Toyota anchor global industry, a vast and explosive wave of new companies—termed "AI natives"—has emerged. Jensen playfully displays a slide crowded with an unreadably dense cluster of startup logos, joking that he shrank them so small that no one's feelings would be hurt by omissions. Beyond pioneering frontier labs like OpenAI and Anthropic, hundreds of specialized AI-native startups are disrupting distinct software and industry verticals.

This sector witnessed unprecedented capital concentration over the past year, absorbing $150 billion in global venture capital—the largest venture investment volume in human history. Jensen explains that the structural economics of these startups are completely unlike those of software companies of the past:

```mermaid flowchart LR accTitle: Historical Computing Revolutions vs AI-Native Infrastructure accDescr: Progression of tech shifts showing how the AI shift requires massive upfront compute and token infrastructure.

PC["PC Revolution<br/>(Local Compute)"] --> NET["Internet Wave<br/>(Client-Server)"] NET --> MOB["Mobile & Cloud<br/>(Hyperscale APIs)"] MOB --> AI["AI Native Era<br/>(Massive Compute & Tokenomics)"] ```

In previous technology waves—the PC era, the internet, and mobile cloud computing—startups raised modest seed rounds in the millions or tens of millions of dollars to pay modest engineering payrolls. In the AI era, capital rounds immediately reach hundreds of millions or billions of dollars out of structural necessity: - Every AI-native company fundamentally requires massive scale compute to survive. - Their core business runs on raw tokens—either pre-training and generating their own proprietary token streams, or integrating and adding high-value business logic on top of foundation models provided by Anthropic, OpenAI, and open-source models. - The computing platform shift guarantees that just as previous transitions forged titans like Google, Amazon, and Meta, this era is currently forging the consequential frontier enterprises of tomorrow.

Framing the historical magnitude of the current moment, Jensen observes: > “This is also the first time that the scale of the investments went from millions of dollars, tens of millions of dollars to hundreds of millions of dollars and billions of dollars. And the reason for that is this is the first time in history that every single one of these companies needs compute, and lots and lots of it.”

[46:13](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=2773s) The Three Waves of AI: Generative, Reasoning, and Agentic

Jensen Huang reflects on why computing demand has escalated so sharply over the prior two years, identifying three sequential breakthroughs that transformed the nature of computing. The era began at the close of 2022 into 2023 with ChatGPT, ushering in generative AI. Huang emphasizes that computing previously centered on data retrieval, whereas modern systems generate unique, context-aware content on the fly. He shares that he uses ChatGPT every single morning and urges everyone in the audience to do the same, noting that this generative capability is fundamentally re-architecting how computers are designed, provisioned, and built out.

```mermaid flowchart LR accTitle: Evolution of AI Computing Workloads accDescr: The progression from perception to generative AI, reasoning models, and autonomous agentic execution. subgraph Wave1["1. Generative AI (2022–2023)"] A["ChatGPT"] --> B["Generate Content & Synthesize"] end subgraph Wave2["2. Reasoning AI (2024)"] C["o1 & o3"] --> D["Decompose, Plan & Self-Reflect"] end subgraph Wave3["3. Agentic AI (Present)"] E["Claude Code, Codex, Cursor"] --> F["Tool Use, Coding, Autonomous Execution"] end B --> C D --> E ```

The second major inflection came with reasoning models, pioneered by OpenAI's o1 and further expanded with o3. Rather than merely synthesizing answers, these models reflect, plan, break down intricate questions into manageable sub-problems, and ground their findings in research to become trustworthy. Although the model weight size itself was only slightly larger, reasoning demanded vastly expanded context inputs and long chains of output tokens to allow the model to "think to itself," which dramatically inflated computational requirements.

The third wave is agentic AI, typified by tools like Claude Code. Instead of querying an AI for facts, users instruct it to build, code, execute, test, and iterate. Huang reveals that 100% of NVIDIA's software engineering workforce now relies on AI coding assistants, using combinations of Claude Code, OpenAI Codex, and Cursor:

> “An AI that was able to perceive became an AI that could generate, an AI that could generate, became an AI that could reason, an AI that could reason now became an AI that can actually do work, very productive work.”

[50:34](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3034s) The Inference Inflection and the Million-Fold Compute Surge

Because artificial intelligence has transitioned from answering conversational prompts to conducting continuous, multi-step tasks, its underlying workload has flipped decisively from training to inference. Every step an AI agent takes—reading files, evaluating logic, formulating plans, running compiler tools, and self-correcting—requires generating tokens, meaning inference must execute at massive scale across every phase.

Huang points out that while the computational demand per task grew roughly 10,000 times due to extended reasoning traces and complex contextual prompts, user adoption and token usage concurrently rose about 100-fold. Compounding these two factors leads to a staggering conclusion: total computing demand has surged by an estimated one million times over the span of two years. This shift explains why GPU spot pricing remains elevated and cluster availability is tight worldwide despite NVIDIA continuously expanding production shipments. Startups, OpenAI, and Anthropic find themselves in a compounding flywheel where acquiring more compute capacity translates directly into higher token generation, increased enterprise revenue, and smarter baseline models.

[52:23](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3143s) $1 Trillion in Visibility and Platform Fungibility

Huang recalls that at GTC 2025, he projected seeing $500 billion in high-confidence demand and purchase orders for Blackwell and Rubin through 2026. Looking back at the audience's unbothered reaction, he quips that everyone had record years and was difficult to impress. He then delivers a major forecast update:

> “Well, I'm here to tell you that right now where I stand a few short months after GTC, one year after last GTC right here where I stand, I see through 2027 at least $1 trillion.”

He predicts NVIDIA will still be short of supply because demand will far exceed that figure. Addressing the economic reality of deploying $1 trillion in capital expenditures, Huang highlights that NVIDIA's infrastructure provides long useful life across training, post-training, and inference, driving token costs to the lowest levels in the industry. The platform is entirely fungible, powering diverse modalities across speech, computer vision, protein folding, generative biology, robotics, and classic large language models. He underscores that Anthropic and Meta have made massive commitments to NVIDIA architectures, collectively representing a third of the world's frontier AI compute alongside the rapid rise of open-source models.

```mermaid flowchart TD accTitle: NVIDIA AI Infrastructure Revenue Allocation accDescr: Breakdown of NVIDIA AI infrastructure demand across hyperscalers and broader enterprise markets. A["NVIDIA AI Infrastructure Demand"] --> B["60% Top-5 Hyperscalers"] A --> C["40% Broad AI Ecosystem"]

subgraph Hyper["Hyperscaler Workloads"] B --> D["Internal Upgrades: LLM Recommender Systems (RecSys) & Search"] B --> E["External Clouds for AI Natives & Frontier Labs"] end

subgraph Broad["Diverse Markets"] C --> F["Regional & Sovereign Clouds"] C --> G["Enterprise, Industrial, Edge & Robotics"] C --> H["Supercomputing & On-Prem Servers"] end ```

- **Hyperscaler Mix (60%):** Top-five hyperscalers deploy NVIDIA hardware both for external developer clouds and for mission-critical internal workloads, including migrating recommender systems (RecSys) from static tables and collaborative filtering over to deep learning and large language models. - **Diversified Enterprise Mix (40%):** Spans sovereign national clouds, regional cloud service providers, enterprise datacenters, edge deployments, and robotics supercomputers. - **Architectural Flexibility:** A single software and hardware stack runs on-premise, in all public clouds, and across hybrid sovereign borders, providing customers the confidence that multi-billion-dollar investments will not end up stranded by shifting algorithm styles.

[58:46](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3526s) Engineering the Inference Breakthrough: Grace Blackwell NVLink 72

Huang outlines the calculated infrastructure bets made when NVIDIA designated 2025 as its internal "year of inference." Even while the Hopper architecture was enjoying peak market dominance, NVIDIA chose to radically rethink server topology. Rather than remaining constrained by standard eight-GPU node boundaries interconnected via legacy layouts, the team broke apart conventional system limits to engineer the Grace Blackwell NVLink 72 rack-scale architecture.

This transition introduced NVLink 72, which fuses 72 Blackwell GPUs into a single massive unified compute domain capable of handling massive mixture-of-experts (MoE) models and reasoning engines without running into inter-node networking bottlenecks. Alongside the hardware overhaul, NVIDIA introduced NVFP4—a precision format supported by dedicated Tensor Core hardware capable of executing inference and model training without numerical accuracy loss while providing massive gains in throughput and energy efficiency. To support the software side, NVIDIA developed Dynamo and TensorRT-LLM, while dedicating billions of dollars to its own DGX Cloud supercomputing cluster to automatically generate, optimize, and tune high-performance inference kernels across the entire stack.

[01:00:34](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3634s) The Economics of Tokens: Throughput, Smartness, and the Gigawatt Constraint

Jensen addresses a persistent industry misconception about AI workloads, noting that people used to tell him inference was easy. In reality, as he emphasizes, inference is both the ultimate engineering challenge and the primary driver of enterprise revenue. Presenting findings from SemiAnalysis—described as the largest, most comprehensive AI inference sweep ever conducted—Jensen breaks down the core metrics of modern token factories. The vertical axis represents tokens per watt, a vital figure of merit because every single data center and factory is fundamentally power-constrained by the laws of physics. A one-gigawatt facility cannot magically become two gigawatts, meaning operators must extract the maximum volume of output tokens from fixed megawatt budgets. The horizontal axis measures interactivity or the speed of each inference. Faster inference not only produces quicker user responses, but directly unlocks larger models, broader context windows, and deeper multi-step reasoning.

```mermaid flowchart TD accTitle: AI Factory Token Economics Tradeoff accDescr: Diagram showing how fixed data center power feeds into compute throughput and determines token intelligence and cost.

subgraph Facility["Physical Factory Constraints"] Power["1 Gigawatt Power Budget<br/>$40B amortized 15-year facility shell"] end

subgraph Compute["AI Supercomputing System"] Arch["Extreme Hardware/Software Co-Design<br/>(Grace Blackwell / Vera Rubin)"] Power --> Arch end

subgraph Tradeoff["Token Production Dynamics"] Throughput["Tokens Per Watt<br/>(Factory Volume & Revenues)"] Smartness["Interactivity & Thinking Time<br/>(Model Context & Reasoning Depth)"] Arch --> Throughput Arch --> Smartness end

subgraph Outcome["Economic Realities"] Cost["Lowest Cost Per Token"] Throughput --> Cost Smartness -.->|"Smarter AI = longer compute"| Throughput end ```

Jensen points out the inherent tension between throughput and model capability: the smarter the AI, the longer it must think, which naturally lowers raw token throughput. However, maximizing intelligence per watt across a fixed footprint determines the financial viability of an entire business. Jensen predicts that going forward, every chief executive in the world will analyze their enterprise through this exact operational framework:

> “Every CEO in the world will study their business from now on in the way I'm about to describe, because this is your token factory.”

While traditional Moore's Law scaling would have yielded roughly a 1.5x performance gain from the Hopper H200 generation, NVIDIA achieved a radical generational leap. When Jensen initially claimed that the Grace Blackwell NVLink 72 system delivered 35 times higher performance per watt, skeptics doubted him. Yet independent benchmarking by Dylan Patel at SemiAnalysis confirmed that Jensen had actually understated the numbers: Dylan accused him of sandbagging because real-world performance reached 50 times higher perf-per-watt, resulting in an industry-leading cost per token that is 35 times lower. Jensen reiterates his longstanding economic rule:

> “I've said before, if you have the wrong architecture, even if it's free, it's not cheap enough.”

Because building and amortizing a one-gigawatt factory over 15 years costs approximately $40 billion before adding a single computing rack, deploying sub-optimal computing silicon wastes tens of billions of dollars in sunk facility expenses.

- **Data center capital amortization**: An empty one-gigawatt shell amortized over 15 years represents approximately $40 billion in baseline capital expenditure. - **SemiAnalysis validation**: Dylan Patel determined NVIDIA's real-world inference leap reached 50x performance per watt over Hopper, exceeding Jensen's conservative 35x guidance. - **Inference token metrics**: Trade-off balances tokens per watt (throughput/revenues) against token generation latency (model smartness and reasoning depth).

---

[01:05:31](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3931s) Extreme Co-Design and the Global Standard for Inference at Scale

NVIDIA pairs its vertical hardware and software integration with an open horizontal deployment model, packaging optimizations directly into leading inference service providers like Fireworks and Together AI. Jensen notes that these fast-moving AI-native startups have seen their inference workloads grow by 100x over the prior year. For these token factories, production efficiency and unit cost per token represent their entire margin structure.

By applying full-stack updates across libraries, kernel algorithms, and scheduling software without altering the underlying physical silicon, NVIDIA transformed the operational throughput of these production fleets. Systems previously delivering an average of 700 tokens per second surged to nearly 5,000 tokens per second—an immediate sevenfold performance improvement achieved solely through full-stack software co-design.

Jensen stresses that historical corporate data centers were architected primarily to store and retrieve static files, but the modern data center has transitioned into an industrial manufacturing plant whose sole commodity is intelligence rendered as tokens. Because land, grid interconnects, and building shells remain strictly finite, cloud service providers and enterprise operators must treat compute efficiency as their primary top-line revenue driver:

> “Your data center, it used to be a data center for files, it's now a factory to generate tokens.”

- **Inference provider growth**: Leading inference service providers expanded production token volumes over 100x in a single year. - **Software-driven speedup**: Full-stack algorithmic updates accelerated production token speeds from ~700 tokens/sec to nearly 5,000 tokens/sec (~7x increase) on identical installed hardware. - **Industrial transformation**: Data centers now function as power-constrained factories converting electrical megawatts into commercial token streams.

---

[01:07:56](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4076s) A Decade of AI Infrastructure Innovation: From DGX-1 to Vera Rubin

A retrospective keynote video narrates NVIDIA's ten-year journey transforming data center infrastructure into a unified supercomputing fabric, starting on April 6, 2016, with the debut of the DGX-1. Designed specifically for deep learning researchers, the DGX-1 integrated eight Pascal GPUs over the first-generation NVLink to deliver 170 teraflops of compute. With Volta, NVIDIA introduced the NVLink Switch, enabling 16 GPUs to communicate with full all-to-all bandwidth as if they were a single giant GPU. As models rapidly outgrew single chassis boundaries, Mellanox joined NVIDIA, culminating in the 2020 DGX A100 SuperPOD—the first GPU supercomputer to balance scale-up connectivity (NVLink 3) with scale-out networking (ConnectX-6 and Quantum InfiniBand).

```mermaid flowchart LR accTitle: Evolution of NVIDIA AI Infrastructure (2016 to 2026) accDescr: Timeline flowchart tracking architectural milestones from DGX-1 Pascal up through Vera Rubin NVLink 72.

subgraph Gen1["2016 - 2018: Single-Node Scaling"] DGX1["DGX-1 (Pascal)<br/>8 GPUs, NVLink 1<br/>170 Teraflops"] Volta["DGX-2 (Volta)<br/>16 GPUs, NVLink Switch<br/>All-to-All Fabric"] DGX1 --> Volta end

subgraph Gen2["2020 - 2022: Scale-Up Meets Scale-Out"] A100["DGX A100 SuperPOD<br/>NVLink 3 + Mellanox<br/>Quantum InfiniBand"] Hopper["Hopper H100 / H200<br/>FP8 Transformer Engine<br/>ConnectX-7, BlueField-3"] Volta --> A100 A100 --> Hopper end

subgraph Gen3["2024 - 2026: The Agentic Factory Era"] Blackwell["Blackwell NVLink 72<br/>72 GPUs, 130 TB/s Fabric<br/>Grace CPU, Spectrum-4"] Rubin["Vera Rubin NVLink 72<br/>3.6 Exaflops, 260 TB/s Fabric<br/>BlueField-4, Groq 3 LPX"] Hopper --> Blackwell Blackwell --> Rubin end ```

The progression accelerated through Hopper, which introduced the FP8 transformer engine, NVLink 4, ConnectX-7, BlueField-3 DPUs, and Quantum-2 InfiniBand, launching the generative AI wave. Blackwell subsequently redefined system design with NVLink 72, binding 72 GPUs across 130 terabytes per second of all-to-all bandwidth alongside Grace CPUs, ConnectX-8, and Spectrum-4 Ethernet. With compute demand continuing to compound across pre-training, post-training, test-time inference, and autonomous agentic workflows, the narration introduces the Vera Rubin platform:

> “The new Vera Rubin platform: seven chips, five rack-scale computers, one revolutionary AI supercomputer for agentic AI—40 million times more compute in just ten years.”

- **2016 (Pascal DGX-1)**: 8 GPUs, first-generation NVLink, 170 teraflops compute. - **2018 (Volta)**: First NVLink Switch architecture, unifying 16 GPUs into a single shared memory domain. - **2020 (Ampere DGX A100 SuperPOD)**: Co-designed scale-up (NVLink 3) and scale-out (ConnectX-6 and Quantum InfiniBand) cluster networking. - **2022 (Hopper)**: FP8 Transformer Engine, NVLink 4, ConnectX-7, and BlueField-3 DPUs. - **2024 (Blackwell NVLink 72)**: 72 GPUs linked with 130 TB/s all-to-all NVLink bandwidth, Grace CPU integration, and Spectrum-4 Ethernet. - **2026 (Vera Rubin Platform)**: 3.6 exaflops compute, 260 TB/s all-to-all NVLink bandwidth, Vera CPU, STAT storage rack with BlueField-4, Spectrum-X with co-packaged optics, and Groq 3 LPX integration delivering 35x higher throughput per megawatt.

---

[01:11:29](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4289s) NVIDIA Vera Rubin: Systems Architecture for the Agentic Era

Returning to the stage, Jensen contrasts earlier eras—where holding up a solitary chip like Hopper was sufficient—with modern supercomputing, explaining that Vera Rubin must be understood as an end-to-end, vertically integrated system architecture. Agentic AI places radically distinct demands on hardware: beyond raw next-token generation within expanding foundational models, autonomous agents continuously pound memory hierarchies, retrieve multi-modal KV caches, query structured and unstructured databases, and execute software tools. Unlike human users who tolerate perceptible tool latency, autonomous AI agents require instant response times when interacting with web browsers, programmatic APIs, and cloud-hosted virtual PCs.

To eliminate host-orchestration bottlenecks, NVIDIA engineered the Vera CPU specifically for extreme single-threaded execution, high-bandwidth data movement, and maximum energy efficiency. It stands as the world's first data center processor equipped with LPDDR5 memory, delivering unmatched single-thread responsiveness and performance-per-watt for agentic loop orchestration.

Jensen highlights the mechanical and physical evolution of the Vera Rubin NVLink 72 rack. The system is 100% liquid-cooled and completely eliminates external copper cable harnesses on the backplane, slashing assembly and deployment times from two days down to just two hours while dramatically compressing manufacturing turnaround. Furthermore, the supercomputer utilizes 45°C warm-water cooling, shifting thermal overhead away from facility chillers so that data center operators can repurpose that electrical power directly for compute:

> “This is also a supercomputer that is cooled by hot water, 45 degrees, which takes the pressure off of the data center, takes all of that cost and all of that energy that's used to cool the data center and makes it available for the system.”

Underpinning the entire platform is NVIDIA's sixth-generation NVLink switch fabric, an interconnect standard that Jensen distinguishes from conventional Ethernet or InfiniBand as NVIDIA's definitive proprietary scale-up advantage.

- **Vera CPU design**: Engineered with LPDDR5 memory to maximize single-thread performance, agent tool execution, and host data processing efficiency. - **Cooling specifications**: 100% liquid-cooled architecture utilizing 45°C warm-water cooling, drastically lowering data center PUE and cooling infrastructure power requirements. - **Rack assembly turnaround**: Fully cable-free rack backplane reduces physical installation and deployment time from 48 hours to 2 hours. - **Interconnect generation**: Houses NVIDIA's proprietary sixth-generation NVLink scale-up switching fabric, running separate from standard InfiniBand and Ethernet networks.

[01:14:59](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4499s) Hardware Innovations: Groq LP 30, Co-Packaged Optics, and the Vera CPU

Jensen Huang showcases a sequence of groundbreaking hardware systems, beginning with a fully liquid-cooled NVLink platform and the brand-new Groq system housing eight Groq chips. Highlighting the LP 30, he notes that while the world has only seen version one, NVIDIA is already in volume production with this third-generation hardware. Alongside it, Huang presents the world's first Co-Packaged Optics (CPO) Spectrum-X switch, also in full volume production. The optics interface directly with silicon on the chip, translating electrons to photons without intermediary board trace losses. Huang points out that NVIDIA co-invented this revolutionary process technology, called COOP, in partnership with TSMC, making NVIDIA the only company currently producing it at scale.

```mermaid flowchart TD accTitle: CPO Spectrum-X and Compute Hardware Stack accDescr: Diagram illustrating optical silicon integration with TSMC COOP alongside Vera CPU and BlueField-4 storage platforms. subgraph Optics ["Co-Packaged Optics (CPO)"] A["Photons"] <-->|"Direct Silicon Interface"| B["TSMC COOP Process"] B <--> C["Spectrum-X Switch Silicon"] end subgraph Compute ["Compute & Storage Engines"] D["Vera Standalone CPU\n(2x Perf/Watt)"] E["CX9 SuperNIC\n(Powered by Vera)"] F["BlueField-4 STX\n(Storage Platform)"] G["Groq LP 30\n(8-Chip Gen 3 System)"] end Optics --- Compute ```

Turning to general-purpose computing, Huang reveals the Vera system, boasting twice the performance-per-watt of any rival CPU in the world. Addressing NVIDIA's shifting business model, he reflects with genuine satisfaction: “Well, you know, we never we never thought we would be selling CPUs standalone.” Yet standalone Vera shipments have already turned into a confirmed multi-billion-dollar business. The lineup expands further with the Vera-powered ConnectX-9 (CX9) network adapter and the BlueField-4 STX storage platform. Huang also examines the dense cabling infrastructure powering these systems, demonstrating structured cabling racks for both NVLink and a 256-node liquid-cooled Ethernet rack designed for rapid, clean data center deployment.

- The Groq LP 30 system features eight third-generation chips in full volume production. - Spectrum-X CPO switch implements TSMC COOP technology, directly translating electrons to photons on-chip. - The Vera CPU achieves double the performance-per-watt of competing CPUs and represents a multi-billion-dollar standalone business. - BlueField-4 STX serves as the foundational architecture for the new high-throughput storage platform. - The Ethernet liquid-cooled rack aggregates 256 nodes into a single structured cabling envelope.

---

[01:17:38](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4658s) Rubin Ultra and the Kyber Rack Architecture

Huang shifts the keynote's physical hardware focus to Rubin Ultra, presenting a dramatic departure from standard rack configurations. While standard Rubin compute trays slide horizontally into cabinets, Rubin Ultra pairs with an entirely new rack design named Kyber. Kyber scales the NVLink fabric to connect 144 GPUs into a single unified NVLink domain. Rather than using conventional horizontal slots and expansive copper bundles that hit strict physical distance limits, Kyber relies on a massive vertical midplane.

Huang gestures toward the sheer physical scale of the tray, remarking on its mass: “This is one compute node and it slides into the Kyber rack vertically.” Compute nodes dock vertically into the front of the Kyber midplane using specialized top-mounted NVLink connectors, while the corresponding NVLink switch nodes dock vertically along the back of the same midplane. This front-to-back sandwich transforms the 144-GPU cluster into a single giant computer while eliminating the reach limitations of external copper cabling. As stagehands wheel the heavy hardware offstage to reveal the video screens, Huang jokes about the unscripted moment with assistant Jeanine: “This is what happens when you don't practice.”

- Kyber rack expands the single NVLink domain to 144 GPUs. - Vertical midplane docking replaces traditional long-run copper cabling assemblies. - Compute nodes insert vertically in the front, directly meeting vertical NVLink switch modules in the rear.

---

[01:19:42](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=4782s) The Economics of AI Factories: Throughput vs. Token Speed

Re-displaying a fundamental slide from the prior year's keynote, Huang underscores that AI data centers must be treated as industrial token manufacturing plants. He tells the audience that repetition is deliberate because executive teams often overlook how data center power translates into top-line revenue: “This is probably the single most important chart for the future of AI factories.” AI factories are constrained by fixed utility envelopes—such as an ISO-power ceiling of one gigawatt—meaning operators must rigorously balance vertical throughput against horizontal token delivery speed.

```mermaid flowchart LR accTitle: Token Market Segmentation and Pricing Strategy accDescr: Flowchart showing how token throughput and latency speed map to consumer tiers and pricing models. subgraph Fast ["Low Latency / Interactive Tier"] H["Critical Path & Research\n(High Token Rate)"] --> P3["$45 to $150 / M Tokens"] end subgraph Balanced ["Medium Latency Tier"] M["Large Models & Long Context\n(Standard Speed)"] --> P2["$3 to $6 / M Tokens"] end subgraph Bulk ["High Throughput Tier"] L["High-Throughput / Batch\n(Lower Token Rate)"] --> P1["Free to Low Cost"] end ```

Huang breaks down the emerging commoditization and tiered market segmentation of intelligence tokens. As models expand in parameters and context windows swell from 100,000 tokens to millions of tokens for both prompt inputs and reasoning outputs, token valuation shifts dramatically: - **Free Tier:** Prioritizes high volume and bulk throughput at lower interactive speeds, functioning primarily as a customer acquisition channel. - **Medium Tier:** Delivers larger context windows and higher speeds at market prices around $3 to $6 per million tokens. - **Smart / Interactive Tier:** Designed for advanced models that iterate and reason extensively, commanding pricing around $45 per million tokens. - **Premium Research Tier:** Provides ultra-high speeds for interactive, mission-critical research paths where customers willingly pay $150 per million tokens. At 50 million tokens a day, Huang notes that this expense remains negligible relative to the value created by elite research teams.

---

[01:23:30](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5010s) Generational Leap: Hopper, Grace Blackwell, and Vera Rubin

Plotting architectural generations onto the throughput-versus-token-speed frontier, Huang demonstrates how architectural shifts alter factory monetization. The Hopper generation established the baseline curve at standard token rates. When Grace Blackwell entered the market, it expanded bulk throughput in the entry tier while simultaneously lifting throughput at the primary monetization tier by 35x.

Vera Rubin takes this progression further by shifting the entire operating curve upward and outward, securing a 10x throughput leap in the highest average selling price (ASP) segment. Huang explains that achieving these high token speeds requires extreme co-design and the massive fabric bandwidth of NVLink 72. To demonstrate the real-world financial return, Huang outlines a model where a 1-gigawatt AI factory splits its power equally, allocating 25% to free, medium, high, and premium tiers: “Assuming this simplistic example allows Blackwell to generate five times more revenues. Vera Rubin to generate five times.” Moving quickly to Vera Rubin slashes the underlying cost per token while drastically boosting revenue output within identical electrical budgets.

| Architecture | Primary Monetization Tier Improvement | High-ASP Segment Improvement | Overall 1 GW Revenue Multiplier | |:--- |:--- |:--- |:--- | | **Hopper** | Baseline | Baseline | Baseline (1x) | | **Grace Blackwell** | **35x** vs. Hopper | Establishes High Tier | **5x** vs. Hopper | | **Vera Rubin** | Extended frontier | **10x** vs. Blackwell | **5x** vs. Blackwell |

---

[01:26:32](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5192s) The Groq Integration: Resolving the FLOPs vs. Bandwidth Conflict

Huang tackles the core physics challenge governing processor design: throughput demands massive mathematical computing density (FLOPs), whereas low-latency token streaming demands extreme memory bandwidth. Because physical chip surface area and silicon packaging are strictly finite, optimizing simultaneously for maximum FLOPs and lowest latency creates an architectural impasse. Huang summarizes the friction plainly: “Computers don't like extreme amounts of flops. Extreme amount of bandwidth. Because there's only so much surface area for chips that any systems has. And so optimizing for high throughput and optimizing for low latency are in fact enemies of each other.”

To break this bottleneck, NVIDIA acquired the Groq engineering team and licensed its technology to combine architectures. Integrating Groq LPX processors alongside NVIDIA's computing architecture yields an additional 35x performance increase at the most demanding interactive tiers. While NVLink 72 remains unbeatable for high-throughput batch and standard interactive workloads, it hits a bandwidth wall when customers demand extreme throughput rates such as 1,000 tokens per second for real-time human interaction or autonomous agents. Merging Groq’s specialized bandwidth architecture directly into the platform resolves that latency bottleneck and extends NVIDIA's reach across the entire token generation spectrum.

- High FLOP density and extreme memory bandwidth compete directly for limited die and package surface area. - NVIDIA integrated the acquired Groq design team and licensed architecture into the broader platform. - Groq integration delivers up to a 35x performance boost in the highest-speed interactive tier. - Addresses extreme operating points beyond 400 to 1,000 tokens per second where standard GPU interconnects face memory bandwidth limits.

[01:28:15](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5295s) Integrating Groq with Vera Rubin via Disaggregated Inference

Jensen Huang explores how adding Groq technology to NVIDIA Vera Rubin extends performance well beyond the boundaries of standard NVLink 72 domains, radically shifting data center economics. Comparing the setup to Blackwell, Huang explains that Vera Rubin alone yields a 5x boost for high-throughput computing. For data centers handling heavy software engineering and coding workloads—where super-smart, highly valuable tokens are at a premium—he recommends integrating Groq across approximately 25% of the data center while keeping the remaining 75% on pure Vera Rubin.

```mermaid flowchart LR accTitle: Disaggregated Inference Pipeline with Dynamo accDescr: Pipeline showing prefill on Vera Rubin, split decode between Rubin and Groq, and coordination by Dynamo. subgraph Rubin["NVIDIA Vera Rubin"] P["Prefill Phase"] A["Decode: Attention Layer<br/>(Heavy Math / KV Cache)"] end subgraph Groq["Groq LPU (LP 30)"] FFN["Decode: Feed-Forward Network<br/>(Token Generation / SRAM)"] end subgraph Orchestration["Dynamo AI Factory OS"] D["Software Pipeline Manager"] end

P --> A A <-->|"Ethernet (Low-Latency Mode)"| FFN D -.-> Rubin D -.-> Groq ```

Huang details why Groq’s architecture was so uniquely appealing: it is a deterministic, compiler-scheduled data flow processor. Instead of relying on runtime dynamic scheduling, its compiler statically plans out the exact arrival of compute and data simultaneously. The hardware trades standard DRAM for massive pools of on-chip SRAM explicitly tuned for low-latency inference. However, contrasting the two chips reveals an extreme physical disparity: a single Groq chip houses only about 500 MB of SRAM, whereas a single Vera Rubin GPU features 288 GB of high-bandwidth memory. Storing the trillions of weights and massive key-value (KV) caches of modern foundation models entirely in Groq chips proved impractical for mainstream adoption. NVIDIA solved this through disaggregated inference managed by a software runtime called Dynamo:

> “What if we re-architected the way that inference is done in the pipeline, so that we could put the work that makes perfect sense on Vera Rubin, and then offload the decode generation, the low latency, the bandwidth limited challenged part of the workload for Groq.”

By splitting the pipeline, prefill and the compute-heavy attention portion of decode—along with the enormous KV cache footprint required by long-context agentic AI—run on Vera Rubin. Meanwhile, the memory-bandwidth-bound feed-forward network (FFN) token generation executes across a pool of Groq chips. Coupled over optimized low-latency Ethernet that cuts network latency roughly in half, this hybrid architecture delivers a 35x performance increase and opens up entirely new low-latency tiers of token generation.

- **Vera Rubin vs. Groq Memory Profile:** 288 GB of high-bandwidth memory per Rubin GPU versus ~500 MB of ultra-fast SRAM per Groq LPU. - **Interconnect:** Coupled over optimized low-latency Ethernet mode that reduces latency by roughly 50%. - **Dynamo Software Layer:** Operates as an AI factory operating system to schedule disaggregated pipeline phases seamlessly across GPUs and LPUs. - **Throughput Gains:** Yields an overall 35x increase in inference performance when pairing Rubin and Groq.

---

[01:32:50](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5570s) Groq LP 30 Production, Supply Chain Scaling, and the Vera Storage Ecosystem

Announcing the hardware realization of this architecture, Huang unveils the Groq LP 30 chip, extending direct gratitude to Samsung for manufacturing the silicon as production ramps toward shipments targeted for the second half of 2026 (around Q3). Turning to Vera Rubin, Huang highlights how smooth early customer sampling has been compared to the complex bring-up of Grace Blackwell NVLink 72 systems. In fact, Satya Nadella had already confirmed via text message that Microsoft Azure brought up its first operational Vera Rubin rack. To meet demand, NVIDIA has prepared an industrial-scale manufacturing pipeline capable of delivering thousands of racks every week—amounting to multi-gigawatt AI factory deployments per month spanning both Vera Rubin and GB300 systems.

At the heart of these racks sits the Vera CPU, tailor-made for agentic tool use, autonomous reasoning loops, and modern data processing engines. Alongside ConnectX-9 (CX-9) and the BlueField-4 DPU stack, Vera establishes an accelerated storage architecture that has drawn universal backing from enterprise vendors. Huang points out that enterprise storage faces unprecedented pressure: historically, storage infrastructure served human-driven database queries, but AI agents running accelerated libraries like cuDF and cuVS will hammer data lakes non-stop while streaming massive KV cache states across tiers.

> “The storage system is going to get pounded. It's going to get pounded because we used to have humans using the storage systems. We used to have humans using SQL. Now we're going to have AI using these storage systems.”

Huang contrasts this extreme architectural co-design with historical chip scaling. Under conventional Moore’s Law, two years might yield incremental factor improvements across transistor counts, raw FLOPS, and memory bandwidth. In contrast, in a one-gigawatt data center footprint over two years, NVIDIA's end-to-end full-stack integration leaps token generation from 2 million to 700 million tokens per second—a 350x increase.

| Metric / Dimension | Traditional Component Scaling | NVIDIA Full-Stack AI Factory | |:--- |:--- |:--- | | **Token Generation Rate (1 GW Factory)** | ~2,000,000 tokens/sec | ~700,000,000 tokens/sec (350x increase) | | **Supply Chain Output** | Staggered multi-quarter deployments | Thousands of racks/week (Multi-GW / month) | | **Storage Demands** | Interactive human SQL queries | Agentic cuDF, cuVS, and heavy continuous KV caching | | **Compute Pairing** | Homogeneous general-purpose nodes | Vera CPU, Rubin GPU, Groq LPU, CX-9, and BlueField-4 |

---

[01:36:28](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=5788s) The Architectural Roadmap: From Rubin to Rubin Ultra and Feynman

Huang details NVIDIA's multi-year system architecture roadmap, reiterating NVIDIA’s strict commitment to backward compatibility. The Oberon rack platform, originally deployed for Blackwell, continues forward into the Rubin era. This allows customers to adopt new architectures without overhauling physical data center layouts. Addressing industry debates around scale-up interconnect technologies, Huang emphasizes that NVIDIA will not pick one medium over another—it will deploy both copper scale-up and optical scale-up aggressively across its product tiers:

> “And so there's a lot of conversation about is NVIDIA going to copper scale up or optical scale up? We're going to do both.”

In the Rubin generation, copper scale-up runs at NVLink 144 using Kyber racks, while Oberon combines NVLink 72 over copper backplanes with optical scale-up to stretch up to NVLink 576 domains. Moving to Rubin Ultra, silicon is already in tape-out alongside a newly developed companion LPU: the LP 35. The LP 35 incorporates NVIDIA’s NVFP4 precision formats directly into its tensor core processing, unlocking several additional factors of execution speed. Oberon configurations for Rubin Ultra employ Spectrum-6 switches featuring co-packaged optics (CPO).

Looking further out, Huang unveils the next-generation architecture: **Feynman**. Scheduled on NVIDIA’s annual cadence, Feynman introduces a completely new GPU, the next-generation LP 40 processor co-engineered by the unified NVIDIA and Groq teams, and the custom **Rosa** CPU (named in tribute to Rosalind Franklin). System networking steps up to BlueField-5 paired with SuperNIC-10, deployed across both copper-based Kyber and the first-ever Kyber CPO scale-up systems. Huang cautions the hardware ecosystem that the sheer volume of accelerated infrastructure will require massive global capacity increases across copper interconnects, discrete optical transceivers, and co-packaged optics manufacturing lines.

- **Oberon Rack System:** Maintains backward compatibility; supports NVLink 72 copper backplanes with optical scale-up reaching NVLink 576 domains. - **Kyber Rack System:** Delivers NVLink 144 over copper, evolving into Kyber CPO for hybrid scale-up topologies. - **Rubin Ultra Silicon:** Features the Rubin Ultra GPU and the LP 35 LPU, introducing native NVFP4 precision to LPU architectures. - **Spectrum-6:** World’s first switch silicon to integrate co-packaged optics (CPO) in high-volume production. - **Feynman Platform Architecture:** Introduces Feynman GPU, LP 40 LPU, Rosa CPU (Rosalind), BlueField-5, and SuperNIC-10.

---

[01:40:08](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6008s) NVIDIA DSX and Omniverse: Virtual Co-Design for Gigawatt AI Factories

Closing this phase of the architecture reveal, Huang charts NVIDIA’s evolution from a graphics chip vendor into a holistic AI factory builder. Modern gigawatt-scale data centers can squander vast amounts of electrical and thermal energy if components operate as disconnected silos. Historically, hardware, electrical, and facility vendors never interacted until their physical equipment landed on the data center floor. To eliminate systemic inefficiencies, NVIDIA introduced Omniverse DSX (Data Center Simulation at Scale), a collaborative physics-based digital twin platform where vendors model and co-design gigawatt facilities long before breaking ground.

Within Omniverse, engineering teams simulate mechanical layouts, liquid and air cooling, thermal dynamics, structural stresses, electrical buses, and multi-tier networking topologies in real time. These digital twins also tie into local utility grids, allowing bidirectional telemetries to dynamically rebalance data center load against fluctuating utility power availability. Inside the facility, NVIDIA implements Max-Q dynamic tuning across compute racks, chassis fans, chillers, and coolant pumps, fine-tuning operational thresholds to prevent power throttling and wasted thermal margins:

> “There's no question in my mind there's a factor of two in here and a factor of two at the scale we're talking about is gigantic.”

- **Omniverse DSX Capabilities:** Unified digital twin simulation covering mechanical, thermal, electrical, and optical networking behaviors. - **Grid Integration:** Bidirectional communication links facility power drawing directly with regional utility power grids to shave peaks and conserve energy. - **Max-Q Data Center Orchestration:** Dynamically coordinates chip clocks, power capping, and facility cooling loops to extract maximum token throughput per megawatt.

[01:42:02](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6122s) How AI Factories Maximize Tokens, Power, and Profit With NVIDIA DSX

Jensen introduces the NVIDIA DSX platform, explaining that like all NVIDIA platforms, it is built across three synchronized layers: hardware, acceleration libraries, and an expansive ecosystem. A video demonstration narrated by Jensen details how the modern digital industrial revolution is fundamentally constrained by electricity. Because AI factory revenue is directly determined by tokens produced per watt, leaving megawatts stranded or unoptimized directly bleeds capital. To solve this, NVIDIA developed DSX as an Omniverse digital twin blueprint engineered specifically to design, commission, and orchestrate gigawatt-scale AI factories for peak token throughput, operational resilience, and energy efficiency.

```mermaid flowchart TD accTitle: NVIDIA DSX AI Factory Architecture accDescr: End-to-end workflow of the NVIDIA DSX platform from design simulation to live AI factory operations.

subgraph Design["Design & Simulation"] A["PTC Windchill PLM<br/>SIM-Ready Assets"] --> B["Dassault Systèmes<br/>3DEXPERIENCE MBSE"] B --> C["Jacobs Omniverse App<br/>Finalized Layout"] C --> D["Multi-Physics Simulation<br/>STAR-CCM+, Cadence, ETAP, DSX Air"] end

subgraph Commission["Construction"] D --> E["Procore<br/>Virtual Commissioning"] end

subgraph Operations["Live AI Factory Operations"] E --> F["NVIDIA Omniverse<br/>Digital Twin Operator"] F --> G["Phaedra Agent<br/>Cooling & Power Telemetry"] F --> H["Emerald AI Agent<br/>Grid Stress & Demand"] G --> I["DSX Max-Q<br/>Dynamic Compute Optimization"] H --> I end ```

The DSX environment exposes key developer APIs: DSX-M for physical, electrical, thermal, and network simulation; DSX Exchange for real-time factory telemetry and operational data; DSX Flex for bidirectional, secure dynamic power negotiation with municipal electrical grids; and DSX Max-Q to dynamically modulate compute loads to maximize token throughput within fluctuating thermal and power envelopes. The design pipeline starts with simulation-ready assets managed through PTC Windchill PLM, undergoes model-based systems engineering in Dassault Systèmes 3DEXPERIENCE, and is compiled inside Jacobs’ custom Omniverse application. It is rigorously validated across external thermodynamics using Siemens STAR-CCM+, internal chassis airflow via Cadence Reality, electrical distributions with ETAP, and data fabric topology using NVIDIA’s DSX Air network simulator, before undergoing virtual commissioning in Procore to eliminate construction delays. Once operational, autonomous software agents take over: Phaedra’s agent monitors cooling and power distribution to inform DSX Max-Q, while Emerald AI agents interpret grid stress signals to dynamically balance industrial power consumption against real-time token generation.

Following the video, Jensen marvels at how Omniverse has expanded from modeling Earth systems to hosting planetary-scale digital twins of the most complex computing facilities ever conceived. Reflecting on the sprawling coalition of industrial partners on screen, he emphasizes that many of these construction, electrical, and simulation vendors were entirely outside NVIDIA's orbit only two years prior. As Jensen puts it: “AI factory revenues are equal to tokens per watt, so with power constraints every unused watt is revenue lost.”

- **Core DSX APIs**: DSX-M (physics/thermal/network modeling), DSX Exchange (operational telemetry), DSX Flex (grid power negotiation), and DSX Max-Q (real-time token throughput maximization). - **Design & Simulation Partners**: PTC Windchill (PLM asset management), Dassault Systèmes 3DEXPERIENCE (model-based systems engineering), Jacobs (Omniverse integration), Siemens STAR-CCM+ (facility thermals), Cadence Reality (rack thermals), ETAP (electrical infrastructure), and Procore (construction commissioning).

[01:45:40](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6340s) Vera Rubin Space-1 Module: Taking Data Centers to Orbit

Jensen pivots briefly to off-planet computing, highlighting that NVIDIA’s silicon has already graduated into space environments. The NVIDIA Drive Thor platform has earned radiation approval and currently powers advanced imaging payloads aboard orbital observation satellites. Looking forward, the next step is establishing off-world data processing centers. NVIDIA is collaborating with aerospace partners on a specialized computing platform dubbed Vera Rubin Space-1, designed specifically to operate in the harsh environment of low Earth orbit and deep space.

Building computing clusters in space fundamentally rewrites data center cooling architectures. Because space lacks an atmosphere, convective cooling with fans or circulating air is physically impossible, leaving thermal radiation as the sole mechanism for dissipating heat. Despite these extreme mechanical and thermodynamic challenges, Jensen expresses complete confidence in the engineering teams tackling orbital heat rejection. Highlighting this audacious frontier, Jensen states: “And so we have we're working with our partners on a new computer called Vera Rubin Space one. And it's going to go out to space and start data centers out, out in space.”

[01:46:23](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6383s) OpenClaw: The Open-Source Operating System for Personal Agents

Turning his attention back to software, Jensen welcomes Peter Steinberger, who is seated in the audience, acknowledging him as the creator of OpenClaw. Jensen remarks that Steinberger likely did not foresee the sheer magnitude of what he unleashed. Within weeks of its release, OpenClaw surged past every record to become the most popular open-source software project in human history, racking up adoption faster than Linux did over three decades. Jensen demonstrates its extreme accessibility: a user simply types a one-line command into their console, which automatically pulls OpenClaw, installs dependencies, synthesizes an autonomous agent, and stands ready to execute complex objectives.

A video montage showcases how the developer community embraced OpenClaw. Clips feature Andrej Karpathy highlighting autonomous research agents that receive an experiment prompt before bedtime, execute a hundred empirical trials overnight, and catalog what worked by morning. Steinberger is shown sharing an anecdote about a 60-year-old father who hooked his brewing equipment up to OpenClaw via Bluetooth, automated the fermentation, and had the agent code and launch an online storefront for ordering fresh lobsters, prompting hundreds of customers to queue up outside a lobster outlet in Shenzhen. The clip concludes by spotlighting ClawCon, a global community convention spawned almost instantly around the framework.

Jensen breaks down the architectural mechanics that make OpenClaw so transformative. Rather than functioning as a standard wrapper or static script, OpenClaw operates as a comprehensive agentic framework: it schedules background cron jobs, manages system resources, interacts with local file systems, makes dynamic tool calls, and routes prompts to large language models. It autonomously decomposes open-ended natural language prompts into sequential sub-tasks, delegates intermediate goals to spawned subagents, and handles rich multimodal I/O—communicating seamlessly over text, email, speech, or visual gestures. Jensen argues that this design mirrors the core primitives of an operating system. Just as Microsoft Windows created personal computing by providing an abstraction layer for hardware and applications, OpenClaw has provided the world with an open-source operating system for autonomous personal agents. Jensen issues a challenge to tech leaders: “For the CEOs, the question is, what's your open source strategy?”

- **Growth Milestone**: OpenClaw eclipsed the cumulative adoption benchmarks of 30 years of Linux development in mere weeks. - **Architectural Primitives**: Native job scheduling, cron workflows, task decomposition, subagent delegation, file system execution, and multimodal I/O. - **Strategic Imperative**: Jensen compares the urgency of adopting an open agentic strategy directly to the historical imperatives of adopting Linux, HTTP/HTML, and Kubernetes.

[01:52:48](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6768s) The Shift from SaaS to AaaS and Enterprise Security with NemoClaw

Jensen contrasts traditional enterprise IT with the emerging post-agentic landscape. Conventional enterprise data centers were built to warehouse structured business files, transaction ledgers, and systems of record. Software vendors layered tools and codified business logic over that data, which digital human workers operated manually, guided by Global System Integrators (GSIs) and enterprise consultants for governance, security, and regulatory compliance. With the rise of autonomous agents, that entire paradigm is shifting. Jensen declares that legacy software models are obsolete and predicts a complete structural migration: “Every single SaaS company will become Agaas company, an Agentic-as-a-service company.”

```mermaid flowchart LR accTitle: NemoClaw Enterprise Security Stack accDescr: Security and policy integration between OpenClaw agents and enterprise corporate infrastructure.

subgraph AgenticCore["Agentic Core"] A["OpenClaw Agent<br/>Autonomous Workflows"] --> B["OpenShell<br/>Sandboxed Execution"] end

subgraph EnterpriseGuard["Enterprise Protection Layer"] B --> C["Enterprise Policy Engine<br/>Corporate SaaS Logic"] C --> D["Privacy Router &<br/>Network Guardrails"] end

subgraph CorporateEnv["Corporate Infrastructure"] D --> E["Internal Systems<br/>HR, Supply Chain, Finance"] D --> F["NVIDIA Frontier Models<br/>Nemotron Open Models"] end ```

However, deploying open-source autonomous agents into corporate intranets introduces severe enterprise security liabilities. By definition, an autonomous agent needs broad permissions: it must access sensitive files, execute arbitrary code, and communicate across external networks. In an enterprise environment, an untamed agent could access proprietary financial models, HR personnel data, or supply chain trade secrets and inadvertently broadcast them outside the company firewall. Jensen warns the audience: “Agentic systems in the corporate network can have access to sensitive information. It can execute code and it can communicate externally. Just say that out loud.”

To solve this vulnerability, NVIDIA collaborated directly with Peter Steinberger and leading cybersecurity researchers to engineer an enterprise-grade reference architecture called NVIDIA NemoClaw. The cornerstone of NemoClaw is OpenShell, a sandboxed security and runtime layer integrated into OpenClaw. OpenShell enforces corporate network guardrails, privacy routers, and strict policy execution, allowing enterprise IT administrators to bind their existing SaaS policy engines directly to agent runtimes. This guarantees that corporate agents can orchestrate complex internal workflows without leaking confidential information. Furthermore, NemoClaw allows companies to connect their secured agentic workflows directly to domain-specific, custom models developed under NVIDIA's Nemotron open model initiative.

- **Enterprise Security Risks**: Unregulated agents pose tri-fold risks of unrestricted data access (finance, HR, supply chain), arbitrary local code execution, and unmonitored outbound external communications. - **NVIDIA NemoClaw Architecture**: Built on OpenClaw, it introduces OpenShell, network guardrails, privacy routing, and enterprise SaaS policy engine integration. - **Model Customization**: Native hookups to NVIDIA Nemotron and open frontier models for proprietary, fine-tuned corporate intelligence.

[01:57:11](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7031s) NVIDIA’s Open Frontier Models Across Specialized Domains

Jensen Huang highlights that the sheer diversity of the physical and digital world makes it impossible for any single foundational model to satisfy every industry. NVIDIA contributes extensively to the open-source artificial intelligence ecosystem, providing over nearly three million open models spanning language, computer vision, digital biology, physics, and autonomous systems. To anchor this sprawling landscape, NVIDIA designs and releases six dedicated families of open frontier models, bundled with the training datasets, recipes, and optimization frameworks required for domain-specific fine-tuning:

```mermaid flowchart TD accTitle: NVIDIA Open Frontier Model Families accDescr: Six specialized open model families spanning reasoning, physical AI, robotics, mobility, digital biology, and AI physics. subgraph LanguageAndPhysicalAI ["Language, Vision & Physical AI"] direction TB NEMO["Nemotron<br/>Reasoning, Language, Vision, RAG & Speech"] COSMOS["Cosmos<br/>Physical AI & World Generation"] ALPA["Alpamayo<br/>Reasoning Autonomous Vehicle AI"] end subgraph SpecializedDomains ["Scientific & Embodied Systems"] direction TB GR00T["GR00T<br/>General-Purpose Humanoid Robotics"] BIONEMO["BioNeMo<br/>Digital Biology & Molecular Design"] EARTH2["Earth-2<br/>Climate Forecasting & AI Physics"] end ```

Huang stresses NVIDIA's dual philosophy: combining vertical integration across hardware and software with horizontal openness across the entire ecosystem. Unlike proprietary walled gardens, NVIDIA guarantees long-term iterative development on each family. As Huang promises the audience: “Nemotron 3 is going to be followed by Nemotron 4, Cosmos 1 was followed by Cosmos 2, GR00T, GR00T generation two—each and every one of these, we're going to continue to advance these models.” On open leaderboards for voice, visual comprehension, physical world simulation, and general robotics, these models consistently claim top ranks.

- **Nemotron**: Frontier reasoning models powering multi-turn language, multimodal vision, speech, safety guardrails, and retrieval-augmented generation (RAG). - **Cosmos**: World foundation models for physical AI simulation, synthetic world generation, and spatial dynamics. - **Alpamayo**: The industry's first thinking and reasoning AI model specifically architected for autonomous driving decisions. - **GR00T**: Foundation models driving general-purpose humanoid robots and articulated robotic manipulation. - **BioNeMo**: Computational open models targeted at biochemistry, drug discovery, and molecular design. - **Earth-2**: High-resolution weather prediction and climate modeling driven by accelerated AI physics engines.

---

[02:00:23](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7223s) Sovereign AI and the Global Nemotron Coalition

Emphasizing the geopolitical and economic need for national self-reliance, Huang explains that global enterprises and nation-states require the freedom to post-train and fine-tune foundation models into sovereign AI tailored to localized languages, cultures, and proprietary industrial data. To serve as this international baseline, NVIDIA announces **Nemotron 3 Ultra**, designed to be the world's premier customizable base model.

To accelerate the upcoming **Nemotron 4** generation, Huang unveils the **Nemotron Coalition**, an alliance uniting leading AI model builders, application designers, and tooling platforms. NVIDIA has poured billions of dollars into AI training infrastructure not merely to power inference acceleration libraries, but to supply open-source architectural engines for global developers.

- **Black Forest Labs**: Frontier visual synthesis and image generation architectures. - **Cursor**: Accelerated developer environments and AI-assisted code generation. - **LangChain**: Widely deployed framework for chaining agentic workflows, counting over a billion downloads. - **Mistral AI**: Led by co-founder Arthur Mensch, pioneering efficient open-weight reasoning models. - **Perplexity**: Conversational search and agentic research engine. - **Sarvam AI**: Foundational language systems built specifically for sovereign Indian language models. - **Thinking Machines Lab**: Advanced AI research lab founded by Mira Murati. - **Reflection**: Advanced multi-step reasoning systems and agentic workflows.

---

[02:02:37](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7357s) Enterprise Token Budgets and the NemoClaw Reference Architecture

Huang argues that enterprise software is undergoing a generational shift comparable to the arrival of the World Wide Web or Linux. The global enterprise IT sector is transforming from a traditional $2 trillion tooling business into a multi-trillion-dollar labor market of specialized, rentable digital agents. Every modern software company will soon evolve into a dual-role enterprise: a token consumer that augments its workforce, and a token manufacturer that serves domain expertise to its customers.

Illustrating this shift, Huang forecasts how engineering compensation and organizational productivity will be restructured around raw compute:

> “I could totally imagine in the future, every single engineer in our company will need an annual token budget. They're going to make a few hundred thousand dollars a year their base pay. I'm going to give them probably half of that on top of it as tokens so that they could be amplified ten X.”

Huang notes that compute allocation has already transformed into a primary recruiting perk across Silicon Valley, where top engineers negotiate the volume of tokens bundled with their job offers. To anchor this shift, NVIDIA offers **NemoClaw**, a fully optimized, enterprise-grade reference architecture built on top of the open-source **OpenClaw** standard. Designed to run performant, secure, and authenticated long-running agents, NemoClaw integrates directly into the NVIDIA Agentic AI toolkit and open model families.

---

[02:05:26](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7526s) Physical AI and the Expansion of Robotics Ecosystems

Moving beyond purely digital software agents, Huang turns to physically embodied agents: robots that perceive, reason, and act in physical space. Over 110 physical robots are showcased on-site, and Huang asserts that virtually every major robotics enterprise worldwide now builds on NVIDIA hardware and software. The underlying development pipeline rests on a synchronized three-computer framework:

```mermaid flowchart LR accTitle: NVIDIA Three-Computer Physical AI Architecture accDescr: Interplay between the AI training supercomputer, Omniverse simulation engine, and onboard edge robotics computer. A["AI Training Supercomputer<br/>(DGX / Vera Rubin)"] -->|"Pre-trains policies & foundation models"| B["Simulation Supercomputer<br/>(OVX / Omniverse)"] B -->|"Generates synthetic data & digital twins"| A B -->|"Validated policies deployed to edge"| C["Robotic Computer<br/>(NVIDIA DRIVE Thor / Jetson)"] C -.->|"Edge telemetry & corner cases"| B ```

Huang announces that autonomous vehicles have finally crossed their “ChatGPT inflection point,” proving that reliable autonomous vehicle operations can be solved at scale. NVIDIA reveals four new manufacturing partners adopting its **Robotaxi-Ready** platform: **BYD**, **Hyundai**, **Nissan**, and **Geely**. Together, these four automotive giants manufacture 18 million passenger vehicles annually, joining established partners such as Mercedes-Benz, Toyota, and General Motors. Furthermore, NVIDIA is partnering with **Uber** to connect fleets of Robotaxi-ready vehicles directly into its ridesharing network across multiple metropolitan centers.

The physical AI ecosystem extends beyond automotive: - **Industrial Automation & Manipulation**: Core integration with ABB, Universal Robots, and Kuka to embed physical AI foundation models and simulation pipelines directly onto manufacturing and assembly floors. - **Heavy Machinery**: Collaborative deployments with Caterpillar for semi-autonomous earth-moving and construction equipment. - **Autonomous Telecommunications**: T-Mobile is converting traditional cellular radio towers into **NVIDIA Aerial AI-RAN** nodes. These software-defined radio masts function as intelligent robotic towers, reasoning about localized cellular traffic and dynamically adjusting beamforming parameters in real time to minimize energy consumption and optimize data fidelity.

---

[02:08:41](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7721s) Video Showcase: Alpamayo AV Reasoning and Differentiable Physics Simulation

The presentation transitions to a video demonstration showcasing the rollout of NVIDIA's physical AI technologies. The demonstration features **Alpamayo**, NVIDIA’s reasoning model for autonomous driving. Unlike conventional autonomous driving stacks that rely on opaque heuristic networks, Alpamayo runs an explicit, real-time internal chain-of-thought, explaining its navigational decisions and interacting naturally with cabin occupants:

- The autonomous vehicle vocalizes lane changes: *“I'm changing lanes to the right to follow my route.”* - Confronted with unexpected street blockages, it reasons out loud: *“There's a double-parked vehicle in my lane. I'm going around it.”* - When an occupant issues a conversational request—*“Hey Mercedes, can we speed up?”*—the system confirms and adjusts dynamically: *“Sure, I'll speed up.”*

The narration highlights that the real world contains far too many chaotic edge cases to train physical robots using real-world operational logs alone. To overcome physical data scarcity, NVIDIA treats compute as data, employing massive synthetic data generation to pre-train World Foundation models using internet-scale video and human demonstrations.

To bridge this simulation-to-reality gap, NVIDIA provides **Isaac Lab**, an open-source framework dedicated to robot policy training and evaluation at scale, alongside **Newton**, a GPU-accelerated differentiable physics engine that allows developers to simulate complex contact dynamics, material deformations, and closed-loop control with extreme physical fidelity.

[02:10:27](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7827s) Closing the Physical AI Data Gap Across the Industry

The presentation moves into the rapid adoption of NVIDIA's physical AI stack across robotics developers, highlighting how NVIDIA Cosmos world models provide neural simulation while Project GR00T foundation models power open robot reasoning and action generation. The narrator explains that with massive compute capabilities, developers worldwide are systematically eliminating the physical AI data gap. In operating rooms, Peritas AI is leveraging NVIDIA Isaac Lab to train assistant robots, utilizing Cosmos world models to dramatically multiply their available synthetic training data.

Similarly, Skild AI relies on Isaac Lab combined with Cosmos to generate post-training datasets for the Skild AI brain, applying reinforcement learning to stress-test and harden the policy across thousands of environmental variations. Humanoid robotics companies and industrial leaders are following suit: Humanoid uses Isaac Lab to train full-body control and fine-grained manipulation policies, Hexagon Robotics deploys it for high-volume simulation and data generation, and both Foxconn and Noble Machines fine-tune GR00T foundation models directly inside the framework. Furthermore, Disney Research links its proprietary Kamino physics simulator with the Newton physics solver and Isaac Lab, allowing dynamic motion and expressive behavior policies to be trained across character robots inhabiting diverse entertainment universes.

- **Peritas AI:** Trains surgical and OR assistant robots inside Isaac Lab, leveraging Cosmos world models for data multiplication. - **Skild AI:** Hardens the "Skild AI brain" using Isaac Lab, Cosmos post-training synthetic generation, and reinforcement learning over thousands of variations. - **Humanoid & Hexagon Robotics:** Employ Isaac Lab for whole-body control, fine manipulation policies, and synthetic simulation data generation. - **Foxconn & Noble Machines:** Fine-tune open GR00T robotics models within Isaac Lab pipelines. - **Disney Research:** Integrates their Kamino physics engine with the Newton solver inside Isaac Lab to train robotic character behaviors across Disney properties.

```mermaid flowchart TD accTitle: Physical AI Data Pipeline Architecture accDescr: Diagram illustrating Isaac Lab, Cosmos, and GR00T foundation models powering robotics training across multiple companies.

subgraph CorePlatform["NVIDIA Physical AI Platform"] Cosmos["NVIDIA Cosmos World Models<br/>(Neural Simulation)"] IsaacLab["NVIDIA Isaac Lab<br/>(Reinforcement Learning)"] GR00T["Project GR00T Models<br/>(Reasoning & Action)"] end

subgraph Partners["Industry Deployments"] Peritas["Peritas AI<br/>(OR Assistant)"] Skild["Skild AI<br/>(Skild AI Brain)"] Humanoid["Humanoid & Hexagon<br/>(Whole-Body & Manipulation)"] Foxconn["Foxconn & Noble Machines<br/>(GR00T Fine-Tuning)"] Disney["Disney Research<br/>(Kamino + Newton Policies)"] end

Cosmos --> IsaacLab IsaacLab --> Peritas IsaacLab --> Skild IsaacLab --> Humanoid GR00T --> Foxconn IsaacLab --> Disney ```

---

[02:12:45](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=7965s) Olaf Takes the Stage With Jensen Huang

Jensen Huang invites a surprise guest to the stage: a bipedal robotic incarnation of the character Olaf from Disney's *Frozen*. Olaf walks steadily onto the stage calling out, “Coming through,” prompting Jensen to marvel at the demonstration: “Newton works. Wow. Omniverse works. Olaf, how are you?” Olaf responds in his cheerful signature voice, declaring, “I'm so happy now that I'm meeting you.” Jensen playfully reminds Olaf of his origin: “I know because I gave you your computer. Jetson.” When Olaf asks what that is, Jensen replies with a laugh, “Well, It's in your tummy,” to which Olaf enthusiastically responds, “That's going to be amazing.”

Jensen recounts how Olaf learned to balance and walk natively inside Omniverse. Olaf expresses sheer delight at his newfound mobility: “I just love walking. This is so much better than riding on a reindeer. Gazing up at the beautiful sky.” Jensen explains that the breakthrough stems directly from physics modeling using the Newton solver running on NVIDIA Warp—an engine co-developed by NVIDIA, Disney Research, and Google DeepMind to let robots adapt dynamically to real-world physical constraints. Praising Olaf's intelligence, Jensen is met with a sharp, witty retort: “I'm a snowman, not a snowclopedia.” Jensen then looks forward to what this technology represents for the theme park industry: “Could you imagine this? The future of Disneyland. All these. All these robots. All these characters wandering around.” Before transitioning to the conclusion, Jensen playfully teases the snowman's stature: “Oh, you know, I have to admit, though. I thought you were going to be taller. I've never seen such a short snowman, to be honest,” prompting a dry “Nope” from Olaf.

---

[02:14:24](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=8064s) Keynote Summary and Closing Montage

Jensen Huang turns to Olaf to help wrap up the event, breaking from the usual solo recap routine. “Usually I close the keynote by telling you what I told you,” Jensen explains, running through the defining pillars of the 2026 keynote: the inflection point of real-time AI inference, the industrial architecture of AI factories, the rapid expansion of the open-core agent revolution, and the emergence of physical AI and robotics.

Instead of delivering a long monologue, Jensen decides to bring in friends to help conclude the presentation: “Tell you what, why don't we get some friends to help us close it out?” With Olaf’s enthusiastic “Of course,” Jensen cues the final video. A musical montage showcases the breadth of robotics, accelerated computing systems, and AI models introduced throughout the day. At the conclusion of the video, Jensen returns to offer a brief, warm farewell to the live and virtual audience: “All right. Have a great GTC.”

## Notable quotes > [01:00:34](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3634s) If you have the wrong architecture, even if it's free, it's not cheap enough. — Jensen Huang

> [01:42:02](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6122s) AI factory revenues are equal to tokens per watt, so with power constraints every unused watt is revenue lost. — Jensen Huang

> [01:05:31](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=3931s) Your data center, it used to be a data center for files, it's now a factory to generate tokens. — Jensen Huang

> [13:50](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=830s) Structured data is the foundation of trustworthy AI. — Jensen Huang

> [01:52:48](https://www.youtube.com/watch?v=jw_o0xr8MWU&t=6768s) Every single SaaS company will become Agaas company, an Agentic-as-a-service company. — Jensen Huang

## Mentioned Jensen Huang, Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton, Andrew Ng, Dylan Patel, Peter Steinberger, Satya Nadella, NVIDIA, CUDA, CUDA-X, GeForce, RTX, DLSS 5, cuDF, cuVS, IBM, IBM Watsonx.data, Nestlé, Dell, Google Cloud, Amazon Web Services (AWS), Microsoft Azure, CoreWeave, Palantir, OpenAI, Anthropic, Meta, Groq, TSMC, Samsung, DGX-1, Hopper, Blackwell, Grace Blackwell NVLink 72, Vera Rubin, Rubin Ultra, Vera CPU, NVLink, Spectrum-X, Kyber, Groq LP 30, NVIDIA DSX, NVIDIA Omniverse, OpenClaw, NVIDIA NemoClaw, Nemotron, Cosmos, Alpamayo, Project GR00T, BioNeMo, Earth-2, Isaac Lab, Newton, Disney Research, Olaf, Jetson, Google DeepMind

## Video description Watch NVIDIA Founder and CEO Jensen Huang’s GTC keynote as he unveils the latest breakthroughs in AI and accelerated computing. See how agentic AI, AI factories, and physical AI are powering the next generation of intelligent systems.

0:00 The Economics of Tokens: Scaling Intelligence via High-Throughput Tokenomics 3:15 Welcome to GTC 2026 6:02 20 Years of CUDA 10:23 GeForce 13:50 DLSS 5 16:19 Structured Data is the Ground Truth of AI 19:45 NVIDIA's Vertical Integration 20:16 IBM Reinvents Data Processing With NVIDIA 38:26 CUDA-X 38:31 NVIDIA Foundational Technology Montage 43:15 AI Natives 46:01 Inference Inflection Arrives 50:43 "The inflection point for inference has arrived." 52:23 Inference Inflection Drives Strong Growth 1:00:53 How Nvidia delivers industry’s lowest token cost - 50x perf/watt leads to 35x lower cost 1:04:50 InferenceMAX King 1:06:13 NVIDIA is the Global Standard for AI Inference at Scale 1:06:46 AI Factories are the Industrial Infrastructure of the AI Era 1:07:56 A Decade of AI Infrastructure Innovation: From NVIDIA DGX-1 to NVIDIA Vera Rubin 1:11:26 NVIDIA Vera Rubin 1:13:53 NVIDIA Vera Rubin, NVLink and Groq 1:15:29 NVIDIA Spectrum-X Switch, Co-Packaged Optics, Vera and NVIDIA BlueField-4 1:17:38 Rubin Ultra 1:19:41 Inference Performance and Efficiency Drive Company Results 1:29:16 Uniting Processors of Extreme Performances 1:32:52 NVIDIA Groq 3 LPX 1:33:30 Announcing NVIDIA Launch Partners 1:35:26 NVIDIA Vera Rubin: 7 Chips – 5 Rack Systems 1:36:28 NVIDIA Extreme Co-Design Delivering X-Factors Every Year 1:40:05 NVIDIA DSX AI Factory Platform 1:42:15 How AI Factories Maximize Tokens, Power, and Profit With NVIDIA DSX 1:45:40 Space-1 Vera Rubin Module 1:46:23 NemoClaw for OpenClaw 1:47:47 OpenClaw: The ChatGPT Moment for Long-Running, Autonomous Agents 1:57:01 NVIDIA Nemotron and Open Models 1:57:29 How NVIDIA Open Models Power Every Industry’s AI 2:01:46 Announcing Global AI Leaders Join NVIDIA Nemotron Coalition to Advance Open Frontier Models 2:04:43 Announcing NVIDIA NemoClaw Reference OpenClaw 2:05:22 Physical AI and Robotics 2:08:33 The Age of Physical AI and Robotics 2:13:00 Olaf Takes the Stage With Jensen Huang 2:14:55 Official Keynote Closing Video | GTC 2026

[Source](https://www.youtube.com/watch?v=jw_o0xr8MWU)

Reading summary

At GTC 2026, NVIDIA CEO Jensen Huang framed the company's evolution around the concept of the 'AI factory,' projecting at least $1 trillion in hardware and infrastructure demand through 2027. Huang argued that physical power limits make full-stack accelerated computing and neural simulation essential to sustain the transition from generative software toward autonomous, embodied agents. The Economics of AI Factories and Inference (60:34) Huang presented modern data centers as industrial factories whose primary output is tokens rather than static data storage. Because modern facilities operate under strict physical power limits—such as a fixed one-gigawatt utility ceiling—operators can only expand revenue and model performance by maximizing output per watt. NVIDIA claims that full-stack hardware and software optimization increased token delivery from an average of 700 tokens per second to roughly 5,000 tokens per second across production fleets at inference providers like Together AI and Fireworks without altering baseline silicon. Full-Stack Hardware and Modular Rack Architecture (71:29) Addressing performance demands beyond the Moore's law slowdown, NVIDIA outlined architectural shifts spanning Grace Blackwell and the newly detailed Vera Rubin system. Vera Rubin incorporates the Vera CPU—designed with LPDDR5 memory to accelerate single-threaded agent orchestration loops—and a fully liquid-cooled rack that eliminates external backplane copper cabling. For higher-density deployments, the Kyber rack introduces a vertical midplane interconnecting 144 GPUs via NVLink, while third-generation LP 30 chips developed with Groq handle deterministic decode phases. Accelerating Enterprise Data and Transitioning to Agents (17:51) NVIDIA emphasized that artificial intelligence requires grounding in enterprise data to prevent hallucinations and maintain auditability. Libraries such as cuDF for relational tables and cuVS for semantic vector search are aimed at speeding up data lakes: Nestlé reported a fivefold execution speedup and an 83% drop in compute costs after applying GPU acceleration to IBM Watsonx.data pipelines. Huang stated that enterprise software is shifting from traditional SaaS models to autonomous agent architectures, requiring open tools like OpenClaw and secure runtime layers like NemoClaw. Physical AI, World Models, and Robotics (125:26) Extending artificial intelligence from digital environments into the physical world, NVIDIA outlined a three-computer workflow that links training supercomputers, Omniverse simulation engines, and edge computers such as Jetson or DRIVE Thor. To bridge real-world physical data scarcity, the company relies on synthetic generation through Cosmos world models, Isaac Lab, and the Newton physics solver. Huang highlighted automotive partnerships adopting NVIDIA's Robotaxi-Ready platform—including BYD, Hyundai, Nissan, and Geely—and showcased Disney Research's robotic Olaf demonstrating real-time bipedal locomotion developed natively inside Omniverse simulation. While Huang forecast sustained demand exceeding $1 trillion through 2027, the operational feasibility of this expansion relies on overcoming severe gigawatt-scale grid constraints, manufacturing complex liquid-cooled rack infrastructure, and proving that autonomous physical robots can safely transition from simulation to open-world reality.

Source quotes

[60:34] Jensen Huang: If you have the wrong architecture, even if it's free, it's not cheap enough. [102:02] Jensen Huang: AI factory revenues are equal to tokens per watt, so with power constraints every unused watt is revenue lost. [65:31] Jensen Huang: Your data center, it used to be a data center for files, it's now a factory to generate tokens. [13:50] Jensen Huang: Structured data is the foundation of trustworthy AI. [112:48] Jensen Huang: Every single SaaS company will become Agaas company, an Agentic-as-a-service company.

Original transcript

[0:08] [INTRO VIDEO]: This is how intelligence is made. A [0:14] new kind of factory generator of tokens. The building blocks of AI. [0:26] Tokens have opened a new frontier, turning  data into knowledge and drawing on all we have [0:33] learned. Tokens are harnessing a new wave of clean  energy. And unlocking the secrets of the stars. [0:53] In virtual worlds. They help robots  learn. And in the physical world, [0:59] perfect. Forging new paths. And clearing the  way for a Bountiful harvest. In the moments [1:19] that matter. Tokens are already there. And in  the miles between they never stop. They work [1:34] where human hands cannot. So we may all breathe  easier. And the smallest hearts beat stronger. [2:05] Tokens are helping us break new ground. On a scale never attempted. To empower the [2:16] world. So we can reach Star Cloud one. Separation  confirmed well beyond it. Together we take the [2:43] next great leap into a bright new future. Built  for all mankind. And here is where it all begins. [3:18] [JENSEN HUANG] Welcome to GTC! [3:38] I just want to remind you this is a tech  conference. All these people lining up so early in [3:46] the morning. All of you in here. It's great to see  you, GTC. We're going to talk about technology. [3:56] We're going to talk about platforms. NVIDIA has three platforms. [4:01] You think that we mostly talk about. One of them  is related to CUDA-X. Our systems is another [4:08] platform. And now we have a new platform  called AI factories. We're going to talk [4:12] about all of them. And most importantly  we're going to talk about ecosystems. [4:15] But before I start. Let me thank our pre game  show hosts. I thought they did a great job. [4:25] Sarah Guo of conviction. Alfred Lin. Sequoia  capital. NVIDIA's first venture capitalist. [4:34] Gavin Baker NVIDIA's first major institutional  investor. These three people are deep in [4:42] technology, deep in what's going on. And of course, they have just a really [4:47] broad range of technology ecosystem. And  then, of course, all of the VIPs that are [4:52] hand-selected to join us today, all star team,  I want to thank all of you for that. [5:03] I also want to thank all the companies  that are here. NVIDIA, as you know, [5:09] is a platform company. We have technology. We have  our platforms. We have rich ecosystem. And today [5:18] there are probably 100% of the $100 trillion  of industry here. 450 companies sponsored this [5:27] event. I want to thank you. A thousand technical  sessions, 2000 speakers. This is this conference [5:37] is going to cover every single layer of the five  layer cake of artificial intelligence from land, [5:43] power, and shell to infrastructure to  chips to the platforms, the models. [5:48] And of course, the most important and ultimately  what's going to take get this industry taken [5:54] off is all of the applications. But it all began. It all began here. [6:02] This is the 20th anniversary of CUDA. We've  been working on CUDA for 20 years. For 20 years [6:14] we've been dedicated to this architecture, this  revolutionary invention. SIMT single instruction, [6:21] multithreaded writing scalar code could spun off  into multithreaded application. Much, much easier [6:29] to program than SIMD. We recently added tiles so  that we could help people program tensor cores [6:38] and the structures of mathematics that are so  foundational to artificial intelligence today. [6:45] Thousands of tools and compilers and frameworks  and libraries in open source. There's a couple of [6:53] hundred thousand public projects. CUDA literally  is integrated into every single ecosystem. [7:01] This chart basically describes 100% of NVIDIA's  strategies. You've been watching me talk about [7:09] this slide from the very beginning. And  ultimately, the single hardest thing to [7:15] achieve is the thing on the bottom installed  base. It has taken us 20 years to now have built [7:23] up hundreds of millions of GPUs and computing  systems around the world that run CUDA. We are [7:29] in every cloud. We're in every computer company.  We serve just about every single industry. [7:36] The installed base of CUDA is the reason why the  flywheel is accelerating. The installed base is [7:44] what attracts developers, who then creates new  algorithms that achieves a breakthrough, for [7:49] example, deep learning. There are so many others.  Those breakthroughs lead to entirely new markets, [7:57] which builds new ecosystems around them with  other companies that join, which creates a [8:02] larger installed base. This flywheel. This  flywheel is now accelerating. The number of [8:08] downloads of NVIDIA libraries is incredibly  accelerating. It's at a very large scale and [8:14] growing faster than ever. This flywheel is what  makes this computing platform able to sustain so [8:22] much applications, so many new breakthroughs. But most importantly, it also enables these [8:31] infrastructures to have extraordinarily useful  life. And the reason for that is very obvious. [8:36] There are so many applications that you can run  on NVIDIA CUDA. We support the entire every single [8:42] phase of the AI lifecycle. We address every single  data processing platform. We accelerate scientific [8:50] principal solvers of all different kinds. And so  the application reach is so great that once you [8:57] install NVIDIA GPUs, the useful life of it is  incredibly high. It is also one of the reasons [9:03] why Ampere that we shipped some six years ago.  The pricing of Ampere in the cloud is going up. [9:11] And so all of that is made possible fundamentally  because the installed base is high, the flywheel [9:17] is high, the developer reach is great. And when  all of that happens and we continuously update [9:23] our software, the computing cost declines, the  combination of accelerated computing speeding [9:31] up applications tremendously. Meanwhile,  as we continue to nurture and continue to [9:36] update software over its life, not only do you  get the first time pop, you get the continuous [9:42] cost reduction of accelerated computing over time,  and we're willing to nurture willing to support [9:49] every single one of these GPUs in the world  because they're all architecturally compatible. [9:54] We're willing to do so because the installed base  is so large. If we release a new optimization, [9:59] it benefits millions. This applies to everybody  in the world. This combination of dynamics is what [10:08] makes the NVIDIA architecture expand its reach,  accelerating its growth. At the same time driving [10:15] down computing costs, which ultimately encourages  new growth. So CUDA is at the center of it. [10:23] But our journey to CUDA actually started  25 years ago. GeForce. I know how many [10:36] of you grew up with GeForce. GeForce is  NVIDIA's greatest marketing campaign. We [10:45] attract future customers starting long before  you could afford to pay for it yourself. [10:52] Your parents paid. Your parents paid your  parents. Your parents paid for you to be [10:58] NVIDIA customers. And every single year they  paid up year after year after year until someday [11:05] you became an amazing computer scientist and  became a proper customer, a proper developer. [11:12] But this is this is the house that GeForce  made. 25 years ago, we started our journey, [11:19] which led to CUDA. 25 years ago, we invented  the programmable shader, a perfectly unobvious [11:28] invention to make an accelerator programmable  the world's first programmable accelerator, [11:35] the pixel shader, 25 years ago. That led us to  explore further and further 20 years later. [11:43] Five years later, the invention of CUDA, one  of the biggest investments that we made, and we [11:48] couldn't afford it at the time, and it consumed  the vast majority of our company's profits, [11:54] was to take CUDA on the backs of GeForce to every  single computer We dedicated ourselves to creating [12:01] this platform because we felt so much—we  felt so strongly about its potential. But [12:07] ultimately, the company's dedication to  it, despite the hardships in the beginning, [12:13] believing in every single day for. For 13  generations of 20 years, we now have CUDA [12:20] installed everywhere. The pixel shader led  to, of course, the revolution of GeForce. [12:28] And then ten years ago we introduced about  ten years ago. What is it? Eight years ago, [12:33] we introduced RTX, a complete redesign of our  architecture for the modern era of computer [12:40] graphics. GeForce GTX CUDA to the world.  GeForce therefore enabled Alex Krizhevsky [12:47] and Ilya Sutskever and Geoff Hinton, Andrew  Ng, and so many others to discover that the [12:54] GPU could be their friend in accelerating deep  learning. It started the big bang of AI ten years [13:02] ago. We decided that we would fuse programmable  shading and introduced two new ideas ray tracing, [13:12] hardware, ray tracing, which is incredibly  hard to do, and a new idea at the time. [13:17] Imagine about ten years ago, we thought that AI  would revolutionize computer graphics. Just as [13:24] GeForce brought AI to the world, AI is now going  to go back and revolutionize how computer graphics [13:30] is done all together. Well, today I'm going to  show you something of the future. This is our [13:35] next generation of graphics technology.  We call it neural rendering, the fusion, [13:42] the fusion of 3D graphics and artificial  intelligence. This is DLSS 5. Take a look at it. [14:19] [VIDEO - NO [14:37] NARRATION] [JENSEN] [14:59] Is that incredible? Computer graphics comes to  life. Now what did we do? We fused controllable 3D [15:11] graphics. The ground truth of virtual worlds. The structured data. Remember this word? The [15:17] structured data of virtual worlds of generative  generated worlds. We combine 3D graphics, [15:24] structured data with generative AI probabilistic  computing. One of them is completely predictive, [15:32] the other one probabilistic, yet highly realistic.  We combine these two ideas, combine these two [15:39] ideas, control through structured data controlled  perfectly and yet generating at the same time. And [15:47] as a result, the content is beautiful. amazing  as well as controllable. This concept of fusing [15:57] structured information and generative AI will  repeat itself in one industry after another, [16:03] industry after another industry. Structured  data is the foundation of trustworthy AI. [16:11] Well. This is going to scare you a little bit.  I'm going to flip the slide and don't gasp. [16:18] ________________ So we're going to go through [16:21] this schematic for the rest of the time. This  is my best slide. Every time I asked my I asked [16:30] the team, what's my best slide? Repeatedly. This  was it. They said, don't do it, Jensen. Don't do [16:35] it. I said, no. These seats are free, for some  of you. So this is your price of admission. So [16:48] this is this is structured data. You've heard  of it. SQL, Spark. Pandas, velox. Some of these [16:56] really really important very large platforms. Snow. Snow. Snowflake. Databricks EMR, Amazon EMR, [17:07] um Azure fabric. Google cloud BigQuery. All  of these platforms are processing data frames. [17:16] These data frames are giant spreadsheets and  they hold all of life's information. This is [17:23] the structured data, the ground truth of business.  This is the ground truth of enterprise computing. [17:30] Well, now we're going to have AI use  structured data and we better accelerate [17:36] the living daylights out of it. It used to  be okay. And we would, you know, of course, [17:41] we would accelerate structured data so that we  could do more. We could do it more cheaply. [17:47] We could do it more frequently per day  and keep the company running at a much [17:51] more synchronized way. However, in the future,  what's going to happen is these data structures [17:56] are going to be used by AI, and AI is going to  be much, much faster than us. Future agents are [18:02] going to use structured databases as well. And then of course the unstructured database, the [18:08] generative database this database is represents  the vast majority of the world vector databases, [18:15] unstructured data, PDFs, videos, speeches, all  of the world's information. About 90% of what's [18:22] generated every single year is unstructured data. Until now, this data has been completely useless [18:29] to the world. We read it, we put it into our file  system, and that's it. Unfortunately, we can't [18:34] query it. We can't search for it. It's hard to do  that. And the reason for that is because there's [18:40] no easy indexing of unstructured data. You have  to understand its meaning, its purpose. And so [18:46] now we have AI do that. Just as AI was able  to solve multi-modality perception, you can [18:54] and understanding you can use that same technology  multi-modality perception and understanding to go [19:00] read a PDF to understand its meaning, and from  that meaning embedded into a larger structure [19:07] that we can search into, we can query into. NVIDIA created two foundational libraries, [19:14] just like we created RTX for 3D graphics. We  created cuDF for DataFrames, structured data, [19:20] we created cuVS for vector stores,  semantic data, unstructured data, [19:27] AI data. These two platforms are going to be two  of the most important platforms in the future. [19:34] Super excited to see its adoption throughout the  network. That's complicated network of the world's [19:38] data processing systems. And the reason for that  is because data processing has been around a long [19:43] time and therefore so many different companies  and platforms and services. It has taken us a long [19:50] time to integrate deeply into this ecosystem. I'm super proud of the work that we're doing here. [19:57] And then today we're announcing several of  them. IBM, the inventor of SQL, one of the [20:06] most important domain specific languages of all  time, of all time is accelerating Watsonx.data [20:12] with cuDF. Let's take a look at it. [VIDEO NARRATION - JENSEN] [20:19] 60 years ago, IBM introduced the system 360,  the first modern platform for general purpose [20:26] computing. Launching the computing era, then  SQL, a declarative language to query data without [20:33] requiring the computer to be instructed step by  step and the data warehouse. Each. The foundations [20:40] of modern enterprise computing today. IBM and  NVIDIA are reinventing data processing for the [20:48] era of AI by accelerating IBM Watsonx.data data  SQL engines with NVIDIA GPU computing libraries. [20:56] Data is the ground truth that gives AI context  and meaning. AI needs rapid access to massive data [21:03] sets. Today's CPU data processing systems can't  keep up. Nestle makes thousands of supply chain [21:11] decisions every day. Their order to cash Data  Mart aggregates every supply order and delivery [21:18] event across global operations in 185 countries.  On CPUs, Nestle refreshed the Data Mart a few [21:26] times a day with accelerated Watsonx.data data  running on NVIDIA GPUs. Nestle can run the same [21:34] workload five times faster at 83% lower cost. The  next computing platform has arrived. Accelerated [21:43] computing for the era of AI. [JENSEN] [21:53] NVIDIA accelerates data processing in the  cloud. We also accelerate data processing on [21:59] prem. As you know, Dell is the world's leading  computer systems maker. And they also are one [22:06] of the world's leading storage providers.  And they worked with us to create the Dell [22:10] AI data platform that integrates cuDF and cuVS  to create an accelerated data platform. Well, [22:18] for the era of AI, and this is an example of  what they did with NTT data. Huge speed up. [22:25] This is cloud, Google Cloud and Google  Cloud. As you know we've been working [22:30] with Google Cloud for a very long time. We accelerate Google's Vertex AI. We now [22:36] accelerate BigQuery. really important framework  and really important platform. And this is an [22:42] example of our work together with Snapchat,  where we reduce their cost of computing [22:47] by nearly 80%. When you accelerate data  processing, when you accelerate computing, [22:55] you get the benefit of speed, you get the benefit  of scale, but most importantly, you also get [23:01] the benefit of cost. And so all of those come  together as one. It was originally called Moore's [23:08] Law. Moore's law was about getting performance  doubling every couple of years. That's another way [23:13] of saying so long as the price remains about the  same and most computers remained about the same. [23:19] You're also getting twice the performance every  year, or you're reducing the cost of computing [23:24] every single year. Well, Moore's Law has run out  of steam. We need a new approach. Accelerated [23:31] computing allows us to take these giant leaps  forward. And as you will see later, because we [23:37] continue to optimize the algorithms. And NVIDIA  is an algorithm company as we continue to optimize [23:43] the algorithms. And because our reach is so  large and our installed base is so large, we can [23:49] reduce the computing cost, increasing the scale,  increasing the speed for everybody continuously. [23:55] This is Google Cloud. You could  see this pattern I just mentioned. [24:00] I just wanted to show you three versions of it.  NVIDIA built the accelerated computing platform, [24:07] has a bunch of libraries on top. I gave you  three examples. RTX is one of them. cuDF is [24:12] another cuVS and we'll show you a few more.  These libraries sit on top of our platform, [24:18] but ultimately we integrate into the world's  cloud services, into the world's OEMs and [24:27] together and other platforms that I'll show  you. Together, we're able to reach the world. [24:33] This pattern. NVIDIA, Google Cloud. Snapchat  will repeat over and over again and it kind [24:40] of looks like this. And so this is one example. NVIDIA with Google Cloud we accelerate vertex AI. [24:46] We accelerate BigQuery. We accelerate. I'm super  proud of the work that we've done with Jackson [24:51] XLA. We are incredible on PyTorch. We're the  only accelerator in the world that's incredible [24:57] on PyTorch and incredible on Jackson, XLA and  the customers that we support, the Basetens, [25:03] the CrowdStrike, Puma, Salesforce, they're not our  customers, but their customers, developers of ours [25:11] that we've integrated the NVIDIA technologies  into that we can then land on the clouds. Our [25:18] relationship with cloud service providers are  essentially us bringing customers to them. We [25:25] integrate our libraries, we accelerate workloads,  and we land those customers in the clouds. [25:32] And so as you could see, most of our cloud  service providers love working with us, and [25:38] they're always asking us to land the next customer  on their cloud. And I just want to let you know, [25:44] there are a lot of customers. We're going to  accelerate everybody. And so there'll be lots [25:50] and lots of customers. We'll be able to land in  your cloud. Just be patient with us. [25:54] And so this is Google Cloud. This is AWS.  We've been working with AWS a long time. [25:59] And one of the areas, one of the, one of  the things I'm super excited about this [26:03] year is we're going to bring Open AI to AWS. And so it's going to drive enormous consumption [26:09] of cloud computing at AWS. It's going to expand  the reach and expand the compute of Open AI. [26:15] And as you know, they are completely compute  constrained. And so AWS, we accelerate EMR, [26:21] we accelerate SageMaker, we accelerate bedrock.  NVIDIA's integrated really deeply into AWS. They [26:29] were our first cloud partner. ________________ [26:31] Microsoft Azure. NVIDIA's A100 supercomputer  was the first one we built was for NVIDIA. [26:42] The first one we installed was at Azure,  and that led to the the the big successful [26:49] partnership with OpenAI. But we've been  working with Azure for quite a long time. [26:53] We accelerate Azure Cloud. Now their AI foundry,  we partner deeply with we accelerate Bing search. [27:02] We work with them on Azure regions. This is  one of the areas that is incredibly important [27:10] as we continue to expand AI throughout the  world. One of the capabilities that we offer [27:16] is confidential computing that in confidential  computing, you want to make sure that even the [27:24] operator cannot see your data, even the operator  cannot touch or see your model's confidential [27:32] computing NVIDIA's GPUs as the first ones  in the world to do that. It's now able [27:37] to support confidential computing and protected  deployment of these very valuable OpenAI models. [27:43] And and Anthropic models throughout clouds  and different regions, and all because of [27:49] our confidential computing. Confidential  computing is super important. And here's [27:53] an example where we have different customers that  we work with. Synopsis. A great partner of ours, [27:59] we're accelerating all of their  EDA and C workflows. And then we [28:02] landed at Microsoft Azure. ________________ [28:06] We were Oracle's first AI customer. Most people  would have thought we were their first supplier. [28:12] We were their first supplier also, but we were  their first AI customer. I'm quite proud of the [28:18] fact that I explained AI clouds to Oracle for the  first time, and we were their first customer. [28:25] Since then, they've really taken  off. We've landed a whole bunch of [28:29] our partners there. Coherent fireworks.  And of course, very famously, Open AI. [28:35] ________________ A great partnership with [28:38] CoreWeave. They're the world's first AI native  cloud, a company that was built with only one [28:48] singular purpose to provision to host GPUs. As  the era of accelerated computing showed up and [28:55] the host for AI clouds. They've got some fantastic  customers and they're growing incredibly. [29:00] ________________ One of the platforms that I'm [29:01] quite excited about is Palantir and Dell. The  three of our companies have made it possible to [29:08] stand up a brand new type of AI platform, the  Palantir ontology platform and AI platform. [29:14] And we could stand up these platforms in any  country, in any air gapped region, completely on [29:21] prem, completely on site, completely in the field.  AI could be deployed literally everywhere without [29:27] our confidential computing capability, without  our ability to build the end to end system, [29:33] as well as offer the entire accelerated computing  and AI stack from data processing, whether it's [29:40] vectors or structures all the way to AI. It  wouldn't have been possible. I wanted to show [29:45] you these examples. This is our special working  relationship with the world's cloud service [29:53] providers. And while all of them are here and I  get the benefit of seeing them during booth tour, [29:58] and it's just so incredibly exciting, I just  want to thank all of you for the hard work. [30:03] What NVIDIA has done is this, and you're  going to see this theme over and over again. [30:09] NVIDIA is vertically integrated  the world's first vertically [30:14] integrated but horizontally open company. And the reason that's necessary is very simple. [30:24] Accelerated computing is not a chip problem.  Accelerated computing is not a systems problem. [30:31] Accelerated computing has a missing word. We just  never say it anymore. Application acceleration. [30:40] If I could make a computer run everything faster,  that's called a CPU, but that's run out of steam. [30:47] The only way for us to accelerate applications  going forward and continue to bring tremendous [30:53] speed up, tremendous cost reduction is through  application or domain specific acceleration. [31:00] I dropped that phrase in the in the front.  And therefore it just became “accelerated [31:05] computing.” And that is the reason why  NVIDIA has to be library after library, [31:10] domain after domain, vertical after vertical. We are a vertically integrated computing [31:16] company. There is no other way. We have to understand the applications. [31:20] We have to understand the domain. We have  to understand fundamentally the algorithms, [31:24] and we have to figure out how to deploy the  algorithm in whatever scenario it wants to [31:31] be deployed, whether it's a data center, cloud  on prem at the edge, or in a robotic system. [31:37] All of those computing systems are different. And  finally, the systems and chips, we are vertically [31:44] integrated. What makes it incredibly powerful,  and the reason why you saw all the slides is [31:49] because NVIDIA's horizontally open will work  and integrate NVIDIA's technology into whatever [31:55] platform you would like us to integrate into. We  offer you the software, we offer you libraries, [32:00] we integrate with your technology so that  we can bring accelerated computing to [32:05] everybody in the world. ________________ [32:08] Well, this GTC is really a great demonstration of  that. You know, most of the time, most of the time [32:16] you'll see me talk about these verticals and I'll  use them some examples, but in every single case, [32:21] whether it's automotive, by the way, financial  services, the largest percentage of attendees at [32:27] this GTC is from the financial services industry. I know. I'm hoping it's developers, not traders, [32:37] guys. Here's, here's, here's one thing I wanted to  say. And so. In the audience represents NVIDIA's [32:48] ecosystem upstream of our supply chain and  downstream of our supply chain. And we work, [32:54] we think about our supply chain upstream and  downstream. And it's just so exciting that our [33:02] entire upstream supply chain this last year,  irrespective of whether you're a 50 year old [33:08] company, we have 70 year old companies. We  have 150 year old company who are now part [33:15] of NVIDIA supply chain and partnering with  us either upstream or downstream. And last [33:19] year you had your record year, did you not? Congratulations. We're on to something here. [33:33] This is the beginning of something very, very  big. And so if you look at accelerated computing, [33:38] we've now set the computing platform. But in order  for us to activate those computing platforms, [33:45] we need to have domain specific libraries that  solve very important problems in each one of the [33:52] verticals that we address. You see us addressing  every single one of these autonomous vehicles, [33:58] our reach, our breadth, our impact. Incredible.  We have a track on that. Financial services I [34:05] just mentioned, algorithmic trading is  going from classical machine learning [34:10] with human feature engineering called quant. The quants did that too. Now, supercomputers [34:17] studying massive amounts of data, discovering  insight and discovering patterns by itself. And [34:23] so this is going through its deep learning and its  transformer moment. Healthcare is going is going [34:28] through their ChatGPT moment. Some really exciting  work that we're there we have we have a great [34:34] keynote track here. We have a great keynote track,  Kimberly, a great keynote track. For healthcare, [34:40] we're talking about AI physics or AI biology for  drug discovery, AI agents for customer service [34:49] and support of diagnosis, diagnosis, and of  course, physical AI, robotic systems. All [34:56] these different vectors of AI have different  platforms that NVIDIA provides industrial. [35:02] We are completely resetting and starting the  largest build out of human history and most [35:10] of the world's industries. Building AI factories.  Building chip plants, building computer plants are [35:16] represented here today. Media and entertainment.  Gaming, of course. Real time AI platform so that [35:23] we could translation and broadcast support and  live live games and live video. Enormous amount [35:32] of it will be augmented with AI. We have a we have  a platform called Holoscan quantum. There are 35 [35:38] different companies here building with us the next  generation of quantum GPU hybrid systems. Retail [35:45] and CPG using NVIDIA for supply chain using. Creating a genetic shopping systems. AI agents [35:54] for customer support. A lot of work being  done here. $35 trillion industry robotics, [36:01] $50 trillion industry in manufacturing. NVIDIA  has been working in this area for a decade now, [36:07] building three computers, the fundamental  computers necessary to build robotic systems. We [36:12] are integrated with working with literally every  single company that we know of building robots. [36:18] We have 110 robots here at the show. And then  telecommunications, about as large as the world's [36:25] IT industry, about $2 trillion. We see. Of course,  base stations everywhere. It's one of the world's [36:32] infrastructures. It was the infrastructure  of the last generation of computing. [36:37] That infrastructure is going to get completely  reinvented. And the reason for that is very [36:42] simple. That base station, which is. It does one  thing, which is base station. It's going to be an [36:51] AI infrastructure platform in the future. AI will  run at the edge. And so lots of, lots of great, [36:58] um, great, uh, discussion there. And our  platform there is called Ariel. Our AI [37:03] ran big partnership with Nokia. Big partnership  with T-Mobile and many others. At the core of our [37:10] business, everything that I just mentioned,  computing platforms, but very importantly, [37:17] our CUDA libraries, our CUDA libraries is the  algorithm, the algorithms that NVIDIA invents. [37:24] We are an algorithm company. That's what  makes us special. That's what that's what [37:28] makes it possible for me to be able to go  into every single one of these industries, [37:32] imagine the future and have the world's best  computer scientists describe and solve problems, [37:38] refactor it, re-express it, and turn it into  a library. We have so many. I think we have. [37:46] At this show, we are announcing 100 100  libraries, 70 libraries, maybe 40 models, [37:54] and that's just at the show. We're updating these  all the time. We're updating them all the time. [37:59] The libraries is the crown jewels of our company. It is what makes it possible for that platform, [38:06] the computing platform, to be activated in service  of solving a problem making impact. One of the [38:13] biggest, one of the most important libraries that  we ever created, cuDNN CUDA deep neural networks, [38:21] it completely revolutionized artificial  intelligence, caused a big bang of modern AI. [38:26] Let me show you a short video about CUDA-X. ________________ [38:32] [VIDEO NARRATION - JENSEN] 20 years ago we built CUDA, [38:35] a single architecture for accelerated computing.  Today we've reinvented computing a thousand CUDA-X [38:43] libraries helped developers make breakthroughs  in every field of science and engineering. [38:49] cuOpt opt for decision optimization.  cuLitho for computational lithography, [38:57] cuDSS for direct sparse solvers to cuEquivariance  for geometry aware neural networks. Aerial for AI [39:09] ran. Warp for differentiable physics. Parabricks  for genomics. At their foundation are algorithms. [39:21] And they are beautiful. [JENSEN] [41:56] Everything you saw was a simulation. Some  of it was principal solvers, fundamental [42:02] physics solvers. Some of it was AI surrogates,  AI physical models, and some of it was physical [42:10] AI robotics models. Everything was simulated.  Nothing was animated. Nothing was articulated. [42:18] Everything was completely simulated. That is  what fundamentally NVIDIA does. It is through [42:25] the connection of understanding of the algorithms  with our computing platforms that we're able [42:32] to open up to unlock these opportunities. NVIDIA is a vertically integrated computing [42:39] company with open horizontal integration with the  world. So that's CUDA x. Well, just now you saw [42:49] a whole bunch of companies. You saw Walmart. And  you know there's L'Oreal and incredible companies, [42:56] established companies JP Morgan and Roche.  And these are companies in companies that have [43:02] defined society to today. Toyota is here. These  are some of the largest companies in the world. [43:09] ________________ It is also true that there's a whole bunch [43:13] of companies you've never heard of. These are  companies. We call them AI natives. A whole bunch [43:20] of small companies. It's the list is gigantic. I couldn't this is just a little tiny, tiny bit [43:25] of it. And, um, I couldn't decide whether to show  you more or show you less. And so I made it so [43:32] that you couldn't see any. And nobody's feelings  are hurt. However, inside this list are a bunch of [43:41] brand new companies. They're companies. Like, for  example, you might have heard a couple of them, [43:45] OpenAI, Anthropic, but there's a whole bunch of  others. There's a whole bunch of others and they [43:50] serve different verticals. Something happened in  the last two years, particularly this last year. [43:57] We've been working with the AI natives for a long  time, and this last year it just skyrocketed. [44:02] I'll explain to you why it happened. This  this industry has skyrocketed $150 billion [44:08] of investment into venture investment into  startups, the largest in human history. This [44:15] is also the first time that the scale of the  investments went from millions of dollars, [44:22] tens of millions of dollars to hundreds of  millions of dollars and billions of dollars. [44:26] And the reason for that is this is the  first time in history that every single [44:32] one of these companies needs compute, and  lots and lots of it. They need tokens, [44:39] lots and lots of it. They're either they're either  going to create and build and create tokens and [44:45] generate tokens, or they're going to Integrate. Add value to tokens that are available created [44:53] by Anthropic and OpenAI and others. And so this  industry is different in so many different ways. [44:59] But the one thing that is very clear, the impact  that they're making this, the incredible value [45:05] that they're delivering already is quite tangible.  AI natives, all because we reinvented computing, [45:14] just like during the PC revolution, a whole bunch  of new companies were created just as just as [45:20] during the internet revolution, a whole bunch  of companies were created and mobile cloud, [45:23] a whole bunch of companies were created.  Each one of them had their own standards. [45:26] And we're talking about one of the major standards  that just happened. Incredibly important. And this [45:33] generation, we also have our own large number  of very, very special companies. We reinvented [45:40] computing. It stands to reason there's going  to be a whole new crop of really important [45:45] companies Consequential companies for the future  of the world the Googles, the Amazons, the Meta's [45:53] consequential companies that have come as a result  of the last computing platform shift. We are now [45:59] at the beginning of a new platform shift. But  what happened in the last couple of years? Well, [46:04] we've been watching as you know, we've been  working on deep learning and working on AI. [46:08] The big bang of modern AI. We were right there at  the spot, and we've been advancing this field for [46:13] quite some time. But why the last two years?  What happened in the last two years? Well, [46:18] three things. ________________ [46:20] ChatGPT, of course, started the generative  AI era. It's able to not just understand, [46:28] perceive and understand. It's able to  also translate and generate generation [46:33] of unique content. I showed you the fusion  of generative AI with computer graphics and [46:39] brought computer graphics to life. You guys  just everybody in the world should be using [46:44] ChatGPT. I know I use it every single morning. Used it plenty this morning. And so ChatGPT was [46:49] the generative AI, the era, the second. By the  way, generative generative computing versus [46:56] the way we used to do computing. It's not. It's  generative. AI is a capability of software, but [47:02] it has profoundly changed how computing is done.  Computing used to be retrieval based. Now it's [47:10] generative. Keep that thought in mind when I talk  about certain things, and you'll realize why it [47:16] is that everything that we do is going to change  how computers are architected, how computers are [47:21] provided, how computers are going to be built out,  and what is the meaning of computing altogether. [47:28] Generative AI 2023. End of 22, 2023. The next  reasoning AI O1, which and then took off with [47:38] O3. Reasoning allowed it to reflect, allows it to  think to itself, allows it to plan, break down, [47:46] break down problems, and decompose a problem it  couldn't understand into steps or parts that it [47:53] could understand. It could ground itself on  research O1 made generative AI trustworthy [48:01] and grounded on truth that caused ChatGPT  to simply took off. And that was a very, [48:09] very big moment. The amount of input tokens  that was necessary in order to produce, and the [48:15] amount of output tokens it generated in order  to reason the model was a little bit larger. [48:21] It, you know, of course, you could have much  larger models. The model O1 was a little bit [48:25] larger, not much larger, but its input token usage  for context and its output token for thinking [48:35] increased the amount of computation tremendously.  Then came Claude Code, the first Agentic model, [48:41] it was able to read files code, compile it, test  it, evaluate it, go back and iterate on it. Claude [48:50] Code has revolutionized software engineering.  As all of you know, 100% of NVIDIA is using a [48:56] combination of or oftentimes all three of them  Claude Code, Codex and cursor all over NVIDIA. [49:04] There's not one software engineer today who is  not assisted by one or many AI agents helping [49:11] them code. Claude Code completely revolutionizes  the the new inflection and for the first time, [49:20] you don't ask the AI what, where, when, how you  ask it. Create do build. You ask it to use tools. [49:37] Take your context, read files. It's able to break  down a problem. Reason about it. Reflect on it. [49:45] It's able to solve problems and actually perform  tasks. An AI that was able to perceive became an [49:54] AI that could generate an AI that could  generate, became an AI that could reason, [49:58] an AI that could reason now became an AI that  can actually do work, very productive work. [50:04] The amount of computation in the last two years,  we know that everybody in this room knows the [50:10] computing demand for NVIDIA GPUs off the charts.  Spot pricing is skyrocketing. You couldn't find a [50:18] GPU if you tried. And yet in the meantime, we're  shipping GPUs out incredible amounts of it, and [50:25] demand just keeps on going up. There's a reason  for that, this fundamental inflection. Finally, [50:34] AI is able to do productive work and therefore  the inflection point of inference has arrived. [50:43] AI now has to think in order to think, it has  to inference. AI now has to do. In order to [50:50] do it has to inference. AI has to read. In  order to do so. It has to inference. It has [50:55] to reason. It has inference. Every part of AI,  every time. It has to think, it has to reason, [51:02] it has to do. It has to generate tokens. It  has to inference. It's way past training. [51:09] Now it's in the in the field of inference. So  the the inference inflection has arrived. [51:16] At the time when the amount of tokens, the  amount of compute necessary increased by [51:22] roughly 10,000 times. Now, when I combine these  two, the fact that since in the last two years [51:29] the computing demand, computing demand  of the work has gone up by 10,000 times, [51:34] and the amount of usage, the amount of  usage has probably gone up by 100 times. [51:44] People have heard me say, I believe that computing  demand has increased by 1 million times in the [51:52] last two years. It is the feeling that we all  have. It is the feeling every startup has. It's [51:58] the feeling that OpenAI has. It's the feeling  that Anthropic has. If they could just get more [52:03] capacity, they could generate more tokens. Their  revenues would go up. More people could use it. [52:08] The more advanced, the smarter the AI could  become. We are now at that positive flywheel [52:16] system. We have. We have reached that moment. The  inflection, the inference inflection has arrived. [52:23] Last year at this time, I said. That where I stood  at that moment in time, we saw about $500 billion. [52:41] We saw $500 billion of very high confidence  demand and purchase orders for Blackwell and [52:51] Rubin through 2026. I said that last year. Now,  I don't know if you guys feel the same way, [53:02] but $500 billion is an enormous amount of revenue.  Not one impressed. I know why you're not impressed [53:20] because all of you had record years. Well,  I'm here to tell you that right now where [53:30] I stand a few short months after GT, CDC. One  year after last GTC right here where I stand. [53:42] I see through 2027. At least $1 trillion. ________________ [54:00] Now. Does it make any sense? And that's what I'm  going to spend the rest of the time talking about. [54:06] In fact, we are going to be short. I am certain  computing demand will be much higher than that. [54:15] And there's a reason for that. So the first  thing is we did a lot of work in the last year. [54:21] Of course, as you know, 2025 was NVIDIA's year of  inference. We wanted to make sure that not only [54:27] were we good at training and post training that we  were incredibly good at every single phase of AI. [54:33] So that the investments that were made,  investments made in our infrastructure could [54:38] scale out for as long as they would like to use  it. And the useful life of NVIDIA's infrastructure [54:44] would be long, and therefore the cost would be  incredibly low. The longer you could use it, [54:48] the lower the cost. There's no question in  my mind, NVIDIA systems are the lowest cost [54:54] infrastructure you could get for AI infrastructure  in the world. And so the first part was last year [55:00] was all about AI for inference. And it drove  this inflection point. Simultaneously, we were [55:08] very pleased last year that Anthropic has come  to NVIDIA, that MSL Meta SL has chosen NVIDIA. [55:15] And meanwhile, meanwhile, and as a collection as  a group, this represents one third of the world's [55:25] AI compute. open source models. Open source models  have reached near the frontier and it is literally [55:32] everywhere. And NVIDIA, as you know today were  the only platform in the world today that runs [55:38] every single domain of AI across every single  one of these AI models in language and biology [55:48] and computer graphics, computer vision and  speech. Proteins and chemicals, robotics and [55:56] otherwise edge or cloud, any language. NVIDIA's  architecture is fungible for all of that, [56:03] and we're incredible for all of that. That allows us to be the lowest cost, [56:08] the highest confidence platform. Because when  you're building these systems, as I mentioned, $1 [56:14] trillion is an enormous amount of infrastructure.  You have to have complete confidence that the [56:20] trillion dollars you're putting down will  be utilized, would be performant, would be [56:26] incredibly cost effective, and have useful life  for as long as you could see that infrastructure [56:32] investment you could make on NVIDIA. You could  make with complete confidence. We have now [56:38] proven that it is the only infrastructure  in the world that you could go anywhere in [56:43] the world and build with complete confidence. You want to put it in any of the clouds. We're [56:48] delighted by that. You want to put it on prem.  We're happy about that. You want to put it in [56:53] any country, anywhere. We're delighted to support  you. We are now a computing platform that runs [57:01] all of AI. Now, our business already starting to  show that 60% of our business is hyperscalers, [57:11] the top five hyperscalers. However, even  within that top five hyperscalers, some of [57:17] it is internal AI consumption. The internal AI  consumption. Really important work, like recsys [57:23] is moving from recommender systems of tables and  collaborative filtering and content filtering. [57:29] It's moving towards deep learning and large  language models. Search. Moving to deep learning. [57:35] Large language models. Almost all of these  different hyperscale workloads are now moving, [57:40] shifting towards a workload that NVIDIA GPUs  are incredibly good at. But on top of that, [57:45] because we work with every AI lab, because  we work with every AI, we accelerate every [57:50] AI model. And because we have a large ecosystem of  AI natives that we work with that we can bring to [57:57] the clouds. That investment, no matter how large,  no matter how quick that compute will be consumed. [58:05] And that represents 60% of our business. The other 40% is just everywhere. Regional [58:11] clouds. Sovereign clouds. Enterprise. Industrial.  Robotics. Edge. Big systems. Supercomputing [58:19] systems. Small servers. Enterprise servers. The  number of systems. Incredible. The diversity of AI [58:30] is also its resilience. The span of reach of AI is  its resilience. There is no question this is not a [58:39] one app technology. This is now fundamental.  This is absolutely a new computing platform [58:46] shift. Well, our job is to continue to advance the  technology. And one of the most important things [58:53] that I mentioned last year was last year was our  year of inference. We dedicated everything. We [59:01] took a giant chance and reinvented while Harper  was at its prime and it was just cooking, we [59:09] decided that the Hopper architecture, the NVLink  by eight, had to be taken to the next level. [59:16] We completely re-architected the system,  disaggregated to the computing system [59:20] altogether and created NVLink 72. The way  that it's built, the way it's manufactured, [59:26] the way it's programmed, completely changed. Grace  Blackwell NVLink 72 was a giant bet, and it wasn't [59:34] easy for anybody. And many of my partners here  in the room. I want to thank all of you for the [59:38] hard work that you guys did. Thank you. NVLink 72 NVFP4 for not just FP4 precision. [59:50] FP4 is a whole different type of  tensor core and computational unit. [59:55] We've demonstrated now that we can inference NVFP4  for without loss of precision, but gigantic boost [60:03] in performance and energy efficiency. We've also been able to use NVFP4 for [60:08] training. So NVLink 72 NVFP4 the invention  of Dynamo TensorRT-LLM, a whole bunch of new [60:16] algorithms. We even built a supercomputer to  help us optimize kernels and help us optimize [60:22] our complete stack. We call it DGX cloud. We  invested billions of dollars of supercomputing [60:28] capability help us create the kernels, the  software that made inference possible. [60:34] Well. The results all came together and  people told. People used to tell me. But, [60:42] Jensen, inference is so easy. Inference is  the ultimate hard. Inference is ultimate [60:48] hard. It is also ultimate important  because it drives your revenues. [60:51] ________________ And so this is the outcome. This [60:53] is from SemiAnalysis. This is the largest, most  comprehensive sweep of AI that has AI inference [61:01] that has ever been done. And what you see here on  the left, on this side, on this side is tokens per [61:09] watt tokens per watt is important because every  data center, every single factory, by definition, [61:16] is power constrained. A one gigawatt factory will  never become two. It's physically constrained. The [61:22] laws of atoms, the laws of physicality. And so  that one gigawatt of data center. You want to [61:29] drive the maximum number of tokens, which is  the production, the product of that factory. [61:36] So you want that. You want to be on top of that  curve as high as you want this. The x axis is the [61:43] interactivity, the speed of inference, the speed  of each inference. The faster you can inference, [61:51] the faster you could of course, respond. But  very importantly, the faster you can influence. [61:57] The larger the models, the more context you  can process, the more tokens you can think [62:02] through. This axis is the same as smartness  of the AI, and so this is the throughput of [62:09] the AI. This is the smartness of the AI. I notice  the smarter the AI, the lower your throughput. [62:17] Makes sense. You're thinking longer. Okay. And  so this axis is the speed. And I'm going to come [62:23] back to this. This is important. This is where  I torture all of you. But it's too important. [62:29] Every CEO in the world you watch, every CEO in  the world will study their business from now on [62:35] in the way I'm about to describe, because this  is your token factory. This is your AI factory. [62:43] This is your revenues. There's no question  about that going forward. And so this is [62:47] the throughput. This is the intelligence better  per per watt for a given power of data center. [62:54] The more throughput, the more tokens you  can produce on this side is cost. Notice [63:01] NVIDIA is the highest performance in the  world. Nobody would be surprised by that. [63:06] They would be surprised by the fact that in one  generation. whereas Moore's Law would have given [63:12] us through transistors 50%. Two times. Moore's  law would probably give us one and a half times [63:22] more performance. You would have expected from  Hopper, age 201 and a half times higher. Nobody [63:28] would have expected 35 times higher. I said  last year at this time that NVIDIA's Grace [63:37] Blackwell NVLink 72 was 35 times perf per watt. Nobody believed me. And then SemiAnalysis came [63:46] out and Dylan Patel had a quote. He accused me of  sandbagging. He accused me of sandbagging. He says [63:57] Jensen sandbagged. It's actually 50 times. And  he's not wrong. He's not wrong. And so our cost [64:05] per token. Yeah. Our cost per token is the lowest  in the world. You can beat it. I've said before, [64:19] if you have the wrong architecture, even if it's  free, it's not cheap enough. And the reason for [64:25] that is because no matter what happens, you  still have to build a gigawatt data center. [64:30] You still have to build a gigawatt factory. And that gigawatt factory for 15 years amortized [64:36] across that gigawatt factory is about $40  billion. Even when you put nothing on it, [64:41] it's $40 billion in. You better make for darn  sure you put the best computer system on that [64:47] thing so that you could have the best token  cost. NVIDIA's token cost is world class, [64:54] basically untouchable at the moment. And  the reason that's true is because of extreme [64:59] co-design. And so I'm very happy that he named us.  There was a monkey King, token king. Well, we take [65:31] we take all of our softwares, as I told you. We vertically integrate, but we horizontally [65:35] open where vertical integration horizontal open.  We integrate all of our software and all of our [65:40] technology. However, we could package it up  and integrate it into the world's inference [65:45] service providers. And these these companies  are growing so fast. They're growing so fast. [65:52] Fireworks. Lin is here. Together. They're just  growing so incredibly fast. 100 times in the [65:59] last year. They are token factories.  And the effectiveness, the performance [66:06] and the token cost production capability for  their factories is everything to them. [66:11] ________________ And this is what happened. [66:14] This is we updated their software same system. And notice their token speeds. Incredible. The [66:24] difference. Before before NVIDIA updated  everything and all of our algorithms and [66:30] software and all the technology that we bring  to bear about 700 tokens per second average [66:38] went to nearly 5007 times higher. And so this  is the incredible power of extreme co-design. [66:46] ________________ I mentioned earlier the importance [66:49] of factories. This is the importance of factory  your data center. It used to be a data center for [66:53] files. It's now a factory to generate tokens. Your  factory is limited no matter what. Everybody is [67:00] looking for land, power and shell. Once you build  it, you are power limited within that power. [67:06] Limited infrastructure. You better make for darn  sure that your inference, because, you know, [67:12] inference is your workload and tokens is your  new commodity. That compute is your revenues, [67:18] that you want to make sure that the architecture  is as optimized as you can in the future. Every [67:24] single CSP, every single computer company, every  single cloud company, every single AI company, [67:32] every single company period, are going to be  thinking about their token factory effectiveness. [67:39] This is your factory in the future. And the reason  why I know that is because everybody in this room [67:45] is powered by intelligence. And in the future,  that intelligence will be augmented by tokens. [67:51] So let me show you how we got here. ________________ [67:56] [VIDEO NARRATION - JENSEN] On April 6th, 2016, a decade ago, [68:00] we introduced DGX one, the world's first  computer designed for deep learning. Eight [68:07] Pascal GPUs connected with the first generation  NVLink 170 teraflops in one computer, the world's [68:16] first computer designed for AI researchers.  With Volta. We introduced NVLink switch 16 [68:25] GPUs connected with full all to all bandwidth,  operating as one giant GPU, a giant step forward, [68:32] but model sizes continue to grow. The data center  needed to become a single unit of computing, [68:40] so Mellanox joined NVIDIA. In 2020, DGX A100  Superpod became the first GPU supercomputer [68:52] combining scale up and scale out architecture. NVLink 3 for scale Up, ConnectX 6, and Quantum [68:59] InfiniBand for scale out. Then Hopper, the first  GPU with the Fp8 transformer engine that launched [69:09] the generative AI era NVLink for ConnectX 7  Bluefield three DPUs second generation quantum [69:18] InfiniBand. It revolutionized computing. Blackwell  redefined AI supercomputing system architecture [69:29] with NVLink 7272 GPUs connected by NVLink spun  130TB per second of all, to all bandwidth compute, [69:40] trace, integrate Blackwell GPUs, Grace  CPUs ConnectX 8 and BlueField-3. Scale [69:48] out runs over Spectrum for Ethernet with three  scaling laws in full steam pre-training, post [69:55] training and inference, and now Agentic systems  compute, demand continues to grow exponentially. [70:04] And now Vera Rubin architected for every phase of  a genetic AI, advancing every pillar of computing, [70:12] including CPU storage, networking, and  security. Vera Rubin, NVLink 72, 3.6. [70:22] Exaflops of compute 260TB per second of all to  all NVLink bandwidth. The engine supercharging the [70:31] era of Agentic AI, the Vera CPU rack designed  for orchestration and agentic workflows. The [70:39] Stat rack AI native storage built with Bluefield  for scale out with Spectrum-X co-packaged optics, [70:47] increasing energy efficiency and resiliency, and  an incredible new addition, the Groq 3 LPX rack, [70:55] tightly connected to Vera Rubin Groqs LPU's  massive on chip SRAM, a token accelerator to the [71:02] already incredibly fast Vera Rubin together  35 times more throughput per megawatt. [71:09] The new Vera Rubin platform seven  chips, five rack scale computers, [71:15] one revolutionary AI supercomputer for agentic AI  40 million times more compute in just ten years. [71:29] [JENSEN] Now, in the. [71:37] In the good old days, when I would say Hopper,  I would hold up a chip. That's just adorable. [71:48] This is Vera Rubin. When we think. When we when  we think Vera Rubin, we think the entire system [72:01] vertically integrated completely with software  extended end to end optimized as one giant system. [72:10] The reason why it's designed for Agentic systems  is very clear, because agents. Of course, [72:15] the most important workload is its thinking. The large language model, the large language [72:20] models are getting larger and larger and larger.  It's going to generate more and more tokens more [72:24] quickly, so it can think more quickly. But it  also has to access memory. It's going to pound [72:30] on memory really hard. KB cache structured data,  unstructured data, CVS. It's going to be pounding [72:39] on the on the storage system really, really hard,  which is the reason why we reinvented the storage [72:46] system. It is also going to use tools and unlike  humans that are more tolerant to slower computers, [72:56] AI wants the tools to be as fast as possible. These tools. Web browsers. In the future, [73:03] they could also be virtual PCs in the cloud. Those  PCs have to be. And those computers have to be as [73:09] fast as possible. We created a brand new CPU,  a brand new CPU that's designed for extremely [73:18] high single threaded performance, incredibly high  data output, incredibly good at data processing [73:27] and extreme energy efficiency. It is the only  data center CPU in the world that uses LPDDR5, [73:36] LPDDR5, and incredible single threaded  performance and performance per watt. [73:43] That is unrivaled. And so that's why we built  that, so that it could go along with the rest [73:49] of these racks for a genetic processing. And so here it is. This is the Grace black [73:57] wall. No. Vera Rubin, where is it? Here it is.  Okay, so this is the Vera Rubin system. Notice, [74:03] since the last time 100% liquid cooled, all of  the cables gone. What used to take what used to [74:11] take two days to install now takes two hours.  Incredible. And so the manufacturing cycle [74:19] time is going to dramatically reduce. This is also  a supercomputer that is cooled by. It's cooled by [74:27] hot water 45 degrees which takes the pressure off  of the data center. Takes all of that cost and [74:33] all of that energy that's used to cool the data  center and makes it available for the system. [74:40] This is the secret sauce. It is the only we're  the only company in the world that has today [74:46] built the sixth, sixth, generation scale up  switching system. This is not Ethernet. This [74:53] is not InfiniBand. This is NVLink. This  is the sixth generation NVLink. This is [74:59] insanely hard to do well. It is insanely hard  to do, period. And I'm just super proud of the [75:04] team. NVLink completely cooled. This is the brand new Groq system [75:10] and I'll show you a little bit more about  it. This system eight Groq Chips. This is [75:16] the LP 30. The world's never seen it. Anything that the world's ever seen is [75:20] V1. This is third generation. And we're in  volume production now. And I'll show you [75:27] more about that in just a second. The world's  first CPO Spectrum-X switch. This is also in [75:37] full production co-packaged optics. Optics comes  directly onto this chip interfaces directly to [75:45] silicon. Electrons get translated to photons. And  it gets directly directly connected to this chip. [75:53] We invented the process technology with TSMC.  We're the only one in production with it today. [75:58] It's called COOP. It's completely revolutionary.  NVIDIA is in full production with Spectrum-X. [76:09] This is the Vera system, twice the performance  per watt of any, any, CPUs in the world today. [76:15] It is also in production. Well, you know, we  never we never thought we would be selling [76:22] CPUs standalone. Um, we are selling a lot of  CPU standalone. This is already for sure going [76:28] to be a multi-billion dollar business for us. So  I'm very, very pleased with our CPU architects. [76:33] We've designed a revolutionary CPU and this is  the CX9 powered with Vera CPU. The BlueField-4 [76:43] STX. Our new storage platform. Okay, so these  are the four. These are the the racks and it's [76:50] connected. Each one of these racks. The NVLink  rack. This is. I've shown you guys this before. [77:00] It's a super heavy and seems to get heavier every  year. Because I think there's just more cables in [77:06] there every year. And so. So this is the NVLink  rack. We've also taken this technology because it, [77:13] it is so efficient to create a data center with  these cabling systems, structured cables. So we [77:22] decided to do that for Ethernet. So this is  Ethernet 256 liquid cooled nodes in one rack. [77:31] And it is also connected with these incredible  connectors. You guys want to see um. Rubin Ultra. [77:58] So this is the Rubin Ultra compute node.  Unlike Rubin that slides in horizontally [78:06] Rubin Ultra goes into a whole new rack. It's called Kyber that enables us to connect 144 [78:15] GPUs in one NVLink domain. And so the Kyber rack  this I, I could lift it, I'm sure, but I won't. [78:27] It's quite heavy. This, this is one compute node  and it slides into the Kyber rack vertically. This [78:36] is where it connects into this is the midplane  the Kyber racks, those four top NVLink connectors, [78:44] slide in and connect into this, and this becomes  one of the nodes. And each one of these racks is [78:52] a different compute node. And this is the  amazing part. This is the midplane and the [78:58] back of the midplane instead of the cabling  system, which has its limits in terms of how [79:05] far we could drive cables, copper cables, we  now have this system to connect 144 GPUs. [79:13] This is the new NVLink. This sits also vertically  and it connects into the midplane on the back [79:23] compute in the front, NVLink switches in the  back one giant computer. Okay, so that is. [79:34] Reuben ultra. As I mentioned, as I mentioned. How  about we take this back down? I need the rest of [79:42] my slides. Oh, it's coming down. Okay. Thank you.  Jeanine. This is what happens when you this is [79:42] what happens when you don't practice. ________________ [79:42] Okay. All right. So, um, you saw you. Take your  time. Just don't get hurt. You saw. You saw this [79:44] slide. You know, only on NVIDIA's keynote will  you see last year's slide presented again. [79:50] And the reason for that is I just want to let you  know that last year, I told you something very, [79:54] very important and it's so important is  worthwhile to tell you again. This is [80:00] probably the single most important chart for  the future of AI factories. And every CEO, [80:06] every CEO in the world will be tracking it.  We'll be studying it very deeply. It's much, [80:11] much more complicated than this. It's  multi-dimensional. But you will be studying the [80:16] throughput and the speed of your AI factories,  the throughput token speed at ISO power, [80:25] because that's all the power you have throughput  and token speed for your factories forever. [80:32] And that that analysis is going to lead directly  to your revenues. What you do this year will show [80:39] up precisely next year as your revenues. And this  chart is what it's all about. And I said on the [80:46] vertical axis, on the vertical axis, thank you  guys. On the vertical axis is throughput. On the [80:52] horizontal axis is token rate. ________________ [80:54] Today I'm going to show you this because we're  able because we're now able to increase the token [81:02] speed and because model sizes are increasing,  because the token length, the context length, [81:09] depending on the different grades of different  application use case continues to grow from maybe [81:16] 100,000 tokens input length to maybe millions. The token input length is growing and also the [81:24] output token length is growing. And so all  of these play into ultimately. The marketing [81:32] and the pricing of future tokens. Tokens are  the new commodity and like all commodities, [81:39] once it reaches an inflection, once it becomes  mature or becomes maturing, it will segment into [81:45] different parts. The high throughput low speed  could be used for the free tier. The next tier [81:54] could be the medium tier, larger model, maybe  higher speed for sure. Larger input context. [82:00] Length that translates to a different price point.  You could see from all the different services. [82:07] This one is free. It's a free tier. The first  tier could be $3 per million tokens. The next [82:12] tier could be $6 per million tokens. You would  like to be able to keep pushing this boundary [82:18] because the larger the model smarter the more  input token context length more relevant, [82:26] the higher the speed, the longer the more  you can think and iterate smarter AI models. [82:32] So this is about smarter AI models.  And when you have smarter AI models, [82:37] each one of these clicks allows you to  increase the price. So this is $45. And [82:42] maybe one day there'll be a premium model that  allows you a premium service that allows you to [82:49] generate token speeds that are incredibly  high because you're in a critical path, [82:53] or maybe you're doing really long research and  $150 per million tokens is just not a thing. [83:00] So let's translate that. Suppose you were to  use 50 million tokens per day as a researcher at [83:06] $150 per million tokens. As it turns out, as a  research team, that's not even a thing. So we [83:13] believe that this is the future. This is where AI  wants to go. This is where it is today. It had to [83:21] start here to establish the value and establish  its usefulness and get better and better and [83:26] better in the future. You're going to see most  services encompass, encompass all of that. [83:30] ________________ This is Hopper. Hopper [83:34] started and I moved it moved the chart. This is 50. This is 100 Hopper looks like [83:39] this. And you would have expected Hopper the next  generation to be higher. But nobody would have [83:44] expected it to be that much higher. ________________ [83:46] This is Grace Blackwell. What Grace Blackwell  did is at your free tier increase your throughput [83:52] tremendously, However, where you mostly  monetize your service, it increased your [84:00] throughput by 35 times. This is no different  than any product that every company makes. [84:06] The higher the tier, the higher the quality. The  higher the performance. The lower the volume, [84:12] the lower the capacity. And so it is no  different than any other business in the world. [84:17] And so now we're able to increase this  tier by 35 x. And we introduced a whole [84:25] new tier. This. This is the benefit of Grace  Blackwell. A huge jump over Hopper. Well, [84:33] this is what we're doing with. ________________ [84:38] Okay, so this is Grace Blackwell. Okay,  let me just reset. Reset this. [84:43] ________________ And this is Vera Rubin. Okay. [84:54] Now just think. Just think what just happened  at every single tier, at every single tier, at [85:00] every single tier, we increase the throughput. And  at the tier that where your highest ASP and your [85:06] most valuable segment, we increased it by ten X. That is the hard work. This is incredibly hard [85:15] to do out here. This is the benefit of NVLink 72.  This is the benefit of extremely low latency. This [85:21] is the benefit of extreme co-design that we could  shift the entire area up. Now, what does it mean [85:26] from a customer perspective in the end? ________________ [85:28] Suppose I were to take all of that and I just  multiply it against. Suppose I took 25% of my [85:35] power, used it in a free tier, 25% of my power in  the medium tier, 25% of my power in the high tier, [85:40] and 25% of my power in the premium tier. My data center only has a gigawatt. And so [85:46] I get to decide how I want to distribute.  The free tier allows me to attract more [85:51] customers. This allows me to serve my most  valuable customers. And the combination, [85:58] the product of all that allows you basically  your revenues, the revenues you can generate. [86:04] Assuming this simplistic example allows Blackwell  to generate five times more revenues. Vera Rubin [86:14] to generate five times. Yeah. ________________ [86:22] So, Vera Rubin, you should get there as soon as  you can. And the reason for that is because your, [86:29] your cost of tokens goes down  and your throughput goes up. [86:32] Now, but we want even more. We want even  more. And so let me just show you back to [86:37] this. This is as you, as I, as I told you, this  throughput requires a ton of flops. This latency, [86:46] this interactivity requires an enormous amount of  bandwidth. Computers don't like extreme amounts [86:52] of flops. Extreme amount of bandwidth. Because  there's only so much surface area for chips that [86:58] any systems has. And so optimizing for high  throughput and optimizing for low latency [87:04] are in fact enemies of each other. And so this is  what happened when we combined with Groq. Okay. [87:11] ________________ And so we, we acquired the team [87:13] that worked on the Groq chips and licensed the  technology. And we've been working together now [87:17] to integrate the system. This is what that  looks like. So at the most valuable tier, [87:24] at the most valuable tier, we're now  going to increase performance by 35 X. [87:29] Now this very simple chart revealed to you  exactly the reason why NVIDIA is so strong in [87:38] the vast majority of the workloads so far. And  the reason for that is because up in this area [87:45] throughput matters so much. NVLink 72 is so game  changing. It is exactly the right architecture, [87:52] and it's even hard to beat. Even as you add Groq to it. [87:56] However, if you extended this chart way out  here and you said you wanted to have services, [88:03] that delivers not 400 tokens per second, but  a thousand tokens per second, all of a sudden, [88:09] NVLink 72 runs out of steam and simply can't  get there. We just don't have enough bandwidth. [88:15] And so this is where Groq comes in, and this  is what happens when we push that out. [88:20] ________________ So it goes out beyond. Thank you. [88:28] Goes out beyond even the limits of what NVLink 72  can do. And if you were to do that, translate that [88:34] into revenues relative to Blackwell. ________________ [88:40] Vera Rubin is five X if most of your workload is  high throughput. I would stick with just 100%. [88:46] Vera Rubin. If a lot of your workload wants  to be coding and very high valued engineering [88:55] token generation, I would add Groq to it. I  would add Groq to maybe 25% of my total data [89:01] center. The rest of my data center is all 100%.  Vera Rubin And so that gives you a sense of how [89:08] you would add Groq to Vera Rubin and extend its  performance and extend its value even more. [89:15] ________________ This is what happens. This is a contrast. [89:19] The reason why the reason why Groq was so  attractive to me is because their computing [89:24] system, a deterministic data flow processor. It  is statically compiled. It is compiler scheduled, [89:33] meaning the compiler figures out when the  data. When do the compute. The compute and [89:39] data arrives at the same time. All of that is done  statically in advance and scheduled completely [89:46] in software. There's no dynamic scheduling. The  architecture is designed with massive amounts of [89:53] SRAM. It is designed just for inference. This one  workload. Now this one workload, as it turns out, [90:01] is the workload of AI factories. And as the  world continues to increase the amount of [90:06] high speed tokens it wants to generate with super  smart tokens, it wants to generate, the value of [90:12] this integration is going to get even higher. And so these are two extreme processors. You [90:17] could see one chip 500MB one Vera Rubin chip,  one Rubin chip, 288GB. It would take a lot of [90:29] rock chips to be able to hold the parameter size  of Rubin, as well as all of the context that has [90:37] to go. The CVV cache that has to go along with  it. so that limited Grox ability to really reach [90:43] the mainstream, to really take off until we had a  great idea. What if we disaggregated inference all [90:49] together with a piece of software called Dynamo?  What if we re-architected the way that inference [90:55] is done in the pipeline, so that we could put  the work that makes perfect sense on Vera Rubin, [91:02] and then offload the decode generation, the low  latency, the bandwidth limited challenged part [91:10] of the workload for Groq. And so we united unified  two processors of extreme differences one for high [91:18] throughput, one for low latency. It still doesn't  change the fact that we need a lot of memory. [91:24] ________________ And so Groq, we're just going to add a whole bunch [91:27] of Groq chips, which expands the amount of memory  it has. And so if you could just imagine out of [91:35] a trillion parameter model, We have to store all  of that in rock chips. However, it sits next to. [91:43] NVIDIA Vera Rubin where we could. We could hold  the massive amounts of KV cache that's necessary [91:50] in processing all of these agentic AI systems. ________________ [91:52] It's based upon this idea of disaggregated  inference. We do the prefill. That's the easy [92:00] part. But we also tightly integrate the decode. So  the attention part of decode is done on NVIDIA's [92:07] Vera Rubin, which needs a lot of math and the  feedforward network part of it, the decode part [92:15] is done. The token generation part is done on Vera  Rubin on the Groq chip, the two of them working [92:21] tightly coupled together over today Ethernet with  a special mode to reduce latency by about half. [92:28] And so that capability allows us to integrate  these two systems. We run Dynamo, this incredible [92:34] operating system for AI factories on top of it and  you get 35 times increase. 35 times increase. [92:43] Not to mention additional new tiers of  inference performance. For token generation, [92:49] the world's never seen. ________________ [92:50] So this is it. This is Groq. The Vera Rubin systems including Groq. I [93:04] want to thank Samsung, who manufactures the Groq  LP 30 chip for us. And they're cranking as hard [93:10] as they can. I really appreciate appreciate you  guys. We're in production with the Groq chip and, [93:17] uh, you know, we'll ship it in the second half,  probably about Q3 timeframe. Okay. Groq LP. [93:28] ________________ Vera Rubin You know, it's kind of hard. [93:34] It's kind of hard to imagine any more customers. You know, and, and the really great thing is, is, [93:44] um, grace Blackwell early sampling of it was  really complicated because of coming together [93:49] with NVLink 72, but the sampling of Vera Rubin is  just going incredibly well. And in fact, Satya, [93:55] I think, texted out already that the first  Vera Rubin rack is already up and running [93:59] at Microsoft Azure. And so I'm super excited  for them. We're just going to keep cranking [94:04] these things out. We have now set up a supply  chain that could manufacture thousands a week [94:11] of these systems, essentially multi gigawatts of  AI factories per month inside our supply chain. [94:20] And so we're going to crank out these, these  Vera Rubin racks while we're cranking out the [94:24] GB 300 racks. We are in full production. The Vera CPUs incredibly successful. And the [94:31] reason for that is because AI needs CPUs for  tool use and various CPU was designed just [94:37] perfectly for that sweet spot. Incredible.  For the next generation of data processing, [94:43] Vera CPU is ideal. ________________ [94:45] The Vera CPU plus CX-9 connected  into the BlueField-4 stack. 100% [94:55] of the world's storage industry is joining  us on this system, and the reason for that [95:02] is because they see exactly the same thing. The storage system is going to get pounded. [95:07] It's going to get pounded because we used to  have humans using the storage systems. We used [95:12] to have humans using SQL. Now we're going to have  AI using these storage systems. And it's going to [95:17] store cuDF accelerated storage. cuVS accelerated  storage as well as very importantly KV caching. [95:25] ________________ Okay, so this is the Vera Rubin system. [95:28] Now what's amazing is this in just two  years time, in a one gigawatt factory, [95:36] in just two years time, in one gigawatt factory.  Using using the mathematics that I showed you [95:41] earlier. Whereas Moore's law would have given  us a couple of steps, we would have, you know, [95:47] X factor, the number of transistors we had, X  factor, the number of flops we had X factor, [95:54] the number of amount of bandwidth. But with this architecture, we're [95:59] going to take our token generation speed token  generation rate from 2 million to 700,000,350 [96:10] times increase. This is this is the power of  extreme co-design. This is what I mean when [96:21] we integrate and optimize vertically, but then we  open it horizontally for everybody to enjoy. [96:28] ________________ This is our roadmap. Very quickly. [96:32] Blackwell is here. The Oberon system. In the case  of Rubin, we have the Oberon system. We're always [96:39] backwards compatible so that if you wanted to not  change anything and just keep on moving through [96:44] with the new architecture, you could do so. The old system, the standard rack system, [96:50] Oberon still available. Oberon is copper scale up,  and with Oberon we could also use optical scale [96:58] out or excuse me, optical scale up to expand  to NVLink 576. Okay. And so there's a lot of [97:08] conversation about is NVIDIA going to copper scale  up or optical scale up? We're going to do both. So [97:16] we're going to have NVLink 144 with Kyber and then  with Oberon we're going to NVLink 72, plus optical [97:27] to get to NVLink 576 the next generation of Rubin  with Rubin Ultra. We have the Reuben ultra chip, [97:37] which is coming, which is taping out, and we  have a brand new chip LP 35, LP 35 will, for the [97:45] first time incorporate NVIDIA's NVFP4 computing  structure, give you another few X factor speedup. [97:53] Okay. And so this is Oberon NVLink 72  optical scale up. And it uses Spectrum-6, [98:01] the world's first co-packaged optical.  And um all of this is in production. [98:08] The next generation from here is Feynman. Feynman  has a new GPU, of course. It also has a new LPU, [98:19] LP 40 big step up. Incredible, incredible new  technology now uniting the scale of NVIDIA and [98:30] the Groq team building together. LP 40. It's going  to be incredible. A brand new CPU called Rosa, [98:39] short for Rosalyne. BlueField-5, which connects  the CPU with the next SuperNIC-10. We will [98:49] have Kyber, which is copper scale up. We will also have Kyber CPO scale up. [99:00] So for the first time, we will scale up with both  copper and co-packaged optics. Okay. And so a lot [99:12] of people have been asking, you know Jensen, is  copper going to still be important? The answer is [99:17] yes. Jensen are you going to scale up optical?  Yes. Are you going to scale out optical? Yes. [99:29] And so for everybody who is in our ecosystem, we  need a lot more capacity. And that's really the [99:37] key. We need a lot more capacity for for copper.  We need a lot more capacity for optics. We need [99:42] a lot more capacity for CPO. And that's the  reason why we've been working with all of you to [99:48] lay the foundation for this level of growth.  And so Feynman will have all of that. [99:53] Let me see if I missed everything.  That's it. Every single year, [99:57] brand new architecture. Very quickly. Very quickly, NVIDIA went [100:08] from a chip company to an AI factory company or  AI infrastructure company, AI computing company, [100:15] these systems. And now we're building entire  AI factories. There's so much power that is [100:24] squandered in these AI factories. We want  to make sure that these AI factories come [100:29] together designed in the best possible way. Most of these components never meet each other. [100:35] Most. Most of us technology vendors now, we  all know each other. But in the past we never [100:41] met each other until the data center. That can't  happen. We're building super complex systems and [100:47] so we have to meet each other virtually somewhere  else. And so we created Omniverse and Omniverse [100:53] DS world, a platform where all of us can meet  and design these gigafactories. Giga, you know, [101:00] gigawatt AI factories virtually in system. We have  simulation systems for the racks for mechanical [101:09] thermal, electrical networking. Those simulation  systems integrated into all of our ecosystem [101:15] partners with incredible tools companies. See also operated connected to the grid so that [101:22] we could interact with each other, send each other  information so that we could adjust grid power and [101:30] data center power accordingly, saving energy.  And then inside the data center using Max-Q [101:37] so that we could adjust the system dynamically  across power and cooling and all of the different [101:42] technologies we all work on together, so that we  leave no power squandered, so that we run at the [101:50] most optimal rate to deliver enormous amount of  token throughput. There's no question in my mind [101:55] there's a factor of two in here and a factor of  two at the scale we're talking about is gigantic. [102:02] We call this the NVIDIA DSX platform. And just as  all of our platforms, there's the hardware layer, [102:09] there's the library layer and there's the  ecosystem layer. It's exactly the same way. [102:14] Let's show it to you. [VIDEO NARRATION - JENSEN] [102:18] The greatest infrastructure build out in  history is underway. The world is racing [102:24] to build chip system and AI factories, and every  month of delay costs billions in lost revenues. [102:31] AI factory revenues are equal to tokens  per watt, so with power constraints every [102:38] unused. watt is revenue lost. NVIDIA RDS is an  Omniverse digital twin blueprint for designing [102:47] and operating AI factories for maximum token  throughput, resilience, and energy efficiency. [102:54] Developers connect through several APIs.  DSXM for physical, electrical, thermal, [103:00] and network simulation. DSS Exchange for AI  factory operational data. DSS flex for secure [103:09] dynamic power management between the grid and DSS  Max-Q to dynamically maximize token throughput. [103:18] It starts with SIM ready assets from NVIDIA  and equipment manufacturers managed by PTC [103:25] Windchill PLM. Then model based systems  engineering is done in Dassault Systemes [103:31] 3D experience. Jacobs brings the data into  their custom Omniverse app to finalize design. [103:41] It's tested with leading simulation tools using  Siemens STAR-CCM+ for external thermals. Cadence [103:50] reality for internal eTap for electrical  and NVIDIA's network simulator DSX Air, [103:59] and virtually commission through Procore  to ensure accelerated construction time. [104:05] When the site goes live, the digital twin becomes  the operator. AI agents work with DS Max-Q to [104:14] dynamically orchestrate infrastructure. Phaedra's  agent oversees cooling and electrical systems, [104:21] sending signals to Max-Q, which continuously  optimizes compute throughput and energy [104:26] efficiency. Emerald AI agents interpret  live grid demand and stress signals and [104:33] adjust power dynamically. With DSX, NVIDIA  and our ecosystem of partners are racing to [104:42] build AI infrastructure around the world, ensuring  extreme resiliency, efficiency, and throughput. [104:51] ________________ It's incredible. Right? Well. On Omniverse, [105:01] Omniverse was designed to hold the world's  digital twin starting from the earth, and [105:07] it's going to hold digital twins of all sizes. And so we have just such a great ecosystem of [105:11] partners. I want to thank all of you. All of these  companies are brand new to our world. We didn't [105:17] know many of you. Just a couple of years ago, and  now we're working so close together to work on and [105:23] build together the largest computer the world's  ever seen, and also to do it at planetary scale. [105:29] So NVIDIA DSX is our new AI factory platform. ________________ [105:40] I'll spend very little time on this  at this time. However, we're going [105:43] to space. We've already been out in space. Thor  is radiation approved and we're in satellites. [105:49] You do imaging from the from satellites. In  the future, we'll also build data centers in [105:54] the in the in the in space, obviously very  complicated to do. So we have we're working [106:00] with our partners on a new computer called  Vera Rubin Space one. And it's going to go [106:05] out to space and start data centers out, out  in space. now. Of course, in space there's no [106:13] connection, there's no convection, there's just  radiation. And so we have to figure out how to, [106:18] um, cool these systems out in space. But we've  got lots of great engineers working on it. [106:23] ________________ Let me talk to you about something new. [106:32] So, so, um, uh, Peter Steinberger is here, and,  um, he, he wrote a piece of software, uh, it's [106:42] called OpenClaw. And and, um, I don't know if he  realized, uh, how successful it was going to be. [106:49] Um, but the importance is profound. OpenClaw  is the number one. It's the most popular open [106:56] source project in the history of humanity. And  it did so in just a few weeks. It exceeded it [107:08] exceeded what Linux did in 30 years. And it's that  important. It is that important. It will do, well, [107:17] uh, this is all you do. Okay. ________________ [107:20] We're announcing our support of it. Uh,  let me let me just quickly go through [107:23] this. I want to show you a couple  of things. You simply type this, [107:27] you type it this into a into a console and it  goes out. It finds OpenClaw, it downloads it, [107:35] it builds you an AI agent and then you could  tell it whatever else you need to do. Okay, [107:41] so let's take a look. ________________ [108:30] [VIDEO]: [VIDEO SOUNDCLIPS] [109:15] An open source project just dropped.  Andrej Karpathy has just launched [109:19] something called a. Research is a huge deal. You give an AI agent a task, go to sleep. [109:23] It runs 100 experiments overnight.  Keeping what works and what doesn't. [109:38] [PETER STEINBERGER] I really love what my [109:40] stuff enables that person to do. And he had  like one guy, he told me like he installed [109:45] it as a 60 year old dad and like they made beer,  connected the machine via Bluetooth to open cloth. [109:50] And then we automated everything, including the  whole website for people to order the lobster. [110:01] Hundreds of people are queuing  up for lobsters in Shenzhen. [110:05] Do you want to build open floor with OpenClaw? Everyone is talking about OpenClaw. But what is [110:13] OpenClaw? Believe it or not,  there's already a ClawCon. [110:16] ________________ [JENSEN]: [110:29] Incredible. Incredible. Now, I illustrated  effectively what OpenClaw is in this way. And so [110:37] all of you can understand it. But let's just think  what happened. What is OpenClaw? It connects. It's [110:42] an it's a genetic system. It calls and connects  to large language models. So the first thing it [110:48] has it has resources that it manages it. It could  access tools. It could access file systems. It [110:54] could access large language models. It's able  to do scheduling. It's able to do cron jobs. [111:01] It's able to decompose a problem that a prompt  that you gave it into step by step by step. [111:08] It could spawn off and call upon other subagents.  It has IO. You could talk to it in any modality [111:16] you want. You could wave at it and understand you.  You could talk to any modality you want. It sends [111:23] you messages. It texts you, sends you email. So  it's got I o. Um, what else does it have? Well, [111:34] based on that, you could, you could say, in fact,  it's an operating system. I've just used the same [111:42] syntax that I would describe in operating system  art. OpenClaw has open sourced essentially the [111:50] operating system of agentic computers. It is no different than how windows made it [111:56] possible for us to create personal computers. Now,  OpenClaw has made it possible for us to create [112:02] personal agents. The implication is incredible.  The implications are incredible. First of all, [112:10] the adoption says something, you know all in  itself. However, the most important thing is [112:16] this every single company now realizes every  single company, every single software company, [112:20] every single technology company. For the CEOs,  the question is, what's your open source strategy? [112:27] Just as we need it all have a Linux  strategy. We all need it to have HTTP, [112:32] HTML strategy, which started the internet. We all needed to have a Kubernetes strategy, [112:37] which made it possible for mobile cloud  to happen. Every company in the world [112:41] today needs to have an open strategy, an agentic  system strategy. This is the new computer. [112:48] ________________ Now this is just the exciting part. [112:52] This is enterprise IT before OpenClaw, you know,  and, and I mentioned earlier the way enterprise [112:58] it works. And the reason why it's called data  centers is because these large rooms, these large [113:04] buildings held data, held the files of people. The  structured data of business, it would pass through [113:12] software that has tools and systems of records  and all kinds of workflow that's codified into it, [113:19] and that turns into tools that humans would use. Digital workers would use. That is the old IT [113:26] industry, software companies creating  tools, saving files, and of course, [113:32] GSIs consultants that help companies figure out  how to use these tools and integrate these tools. [113:36] These these tools are incredibly valuable  for governance and security and privacy and [113:42] compliance. And all of that continues to be true.  It's just that post-OpenClaw post agentic. This is [113:50] what it's going to look like. ________________ [113:52] This is the extraordinary part. Every  single IT company, every single company, [113:57] every SaaS company, every SaaS company will  become a. Agaas company. No question about it. [114:09] Every single SaaS company will become Agaas  company, an Agentic-as-a-service company. [114:14] And what's amazing is this you now open cloud  gave us gave the industry exactly what it [114:20] needed at exactly the time, just as Linux gave the  industry exactly what it needed, exactly the time, [114:29] just as Kubernetes showed up at exactly the right  time, just as HTML showed up, it made it possible [114:35] for the entire industry to grab on to this  open source stack and go do something with it. [114:42] There's just one catch. Agentic systems in the  corporate network can have access to sensitive [114:50] information. It can execute code and it can  communicate externally. Just say that out loud. [114:57] Okay, think about it. Access sensitive  information. Execute code. Communicate [115:02] externally. You could of course access employee  information, access supply chain, access finance [115:09] information sensitive information and send it  out communicate externally. Obviously this can't [115:16] possibly be allowed. And so what we did was we  worked with Peter, we took some of the world's [115:21] best security and computing experts, and we worked  with Peter to make OpenClaw, OpenClaw, enterprise [115:31] secure and enterprise private capable. ________________ [115:36] And we call that this is our NVIDIA  OpenClaw reference for open Nemo claw, [115:43] which is a reference for OpenClaw. And  it has all these Agentic AI tool kits. [115:49] And the first part of it is technology we call  open shell that has now been integrated into [115:55] OpenClaw. Now it's enterprise ready. This stack,  this stack with a reference design we call Nemo. [116:05] NemoClaw. Okay. With a reference stack we call  NemoClaw. You could download it, play with it [116:11] and you could connect to it. The policy engine of  all of the SaaS companies in the world. And your [116:19] policy engines are super important, super  valuable. So the policy engines could be [116:24] connected. NemoClaw or OpenClaw with open shell  would be able to execute that policy engine. [116:31] It has a policy. It has a network guardrail,  it has a privacy router. And as a result, [116:38] we could protect and keep the the clause from  executing inside our company and do it safely. [116:46] We also added several things to the system. And  one of the most important things you want to do [116:52] with your own claw custom claws is so that you  can have your custom models. And this is NVIDIA's [117:01] open model initiative. ________________ [117:02] We are now at the frontier of every single  domain of AI models, whether it's Nemotron, [117:11] Cosmos, World Foundation model, GR00T, artificial  general robotics, humanoid robotics models, [117:18] Alpamayo for autonomous vehicle, BioNeMo for  digital biology, Earth-2 for AI physics. [117:26] We are at the frontier on every  single one. Take a look. [117:30] ________________ [VIDEO NARRATION - JENSEN] [117:32] The world is diverse. No single model can serve  every industry. Open models is one of the largest [117:39] and most diverse AI ecosystems in the world.  Nearly 3 million open models across language, [117:46] vision, biology, physics and autonomous systems  enable AI builds for specialized domains. NVIDIA [117:54] is one of the largest contributors to open source  AI. We build and release six families of open [117:59] frontier models, plus the training data, recipes,  and frameworks to help developers customize and [118:06] adopt new leaderboard topping models are  launching for every family. At the core, [118:11] Nemotron, reasoning models for language.  Visual understanding. RAG. Safety. And speech. [118:27] Cosmos frontier models for physical AI world  generation and understanding. Alpamayo. The [118:37] world's first thinking and reasoning. Autonomous  vehicle AI. GR00T, foundation models for general [118:45] purpose robots. BioNeMo open models for  biology, chemistry and molecular design. [118:53] Earth-2 models for weather and climate forecasting  rooted in AI physics. NVIDIA open models give [119:01] researchers and developers the foundation to build  and deploy AI for their own specialized domains. [119:08] ________________ Our models are. Thank you. [119:15] Our models are valuable to all of you because,  number one, it's on the top of the leaderboard. [119:20] It's world class. But most importantly, it's  because we are not going to give up working on it. [119:27] We're going to keep on working on it every single  day. Nemotron 3 is going to be followed by neutron [119:32] for cosmos one was followed by cosmos two. GR00T,  GR00T at generation two each and one of these. [119:39] We're going to continue to advance these models. Vertical integration. Horizontal openness. [119:45] So that we can enable everybody to join the  AI revolution. Number one on leaderboard, [119:52] across research and voice and world models and  artificial general robotics and self-driving [119:58] cars and reasoning. ________________ [120:00] And of course, one of the most important  one, this is Nemotron 3 in OpenClaw. [120:08] This is Nemotron 3 and OpenClaw. And look at the  top three. There are the three best models in the [120:15] world. Okay, so we are at the frontier. ________________ [120:23] It is also true. It is also true that we want  to create the foundation model so that all of [120:28] you can fine tune it, post train it into  exactly the intelligence you need. This [120:34] is Nemotron 3 Ultra. It is going to be the  best base model the worlds ever created. [120:40] ________________ This allows us to help every country [120:45] build their sovereign AI, and we're working  with so many different companies out there. [120:49] And one of the most exciting things that we're  doing today. I'm announcing today is a Nemotron [120:55] coalition. We are so dedicated to this. We have  invested billions of dollars of AI infrastructure [121:02] so that we could develop the core engines for  AI that's necessary for all the libraries of [121:07] inference and so on, but also to create the  AI models to activate every single industry [121:15] in the world. Large language models is really  important. Of course it's important. How can [121:19] how can human intelligence not be. However,  in different industries around the world, [121:24] in different countries, around the  world, you need to have the ability to [121:28] customize your own models and the domains. The domain of the domain of the models is [121:34] radically different from biology to physics, to  The self-driving cars to general robotics, to, [121:39] of course, human language. And we have the ability  to work with every single region to create their [121:44] domain specific, their sovereign AI. Today, we're  announcing a coalition to partner with us to make [121:51] Nemotron 4 even more amazing. ________________ [121:55] And that coalition has some amazing companies  in it. Black Forest Labs imaging company, [122:01] Cursor, the famous coding company.  We use lots of it. LangChain billion [122:06] downloads for creating custom agents. Mistral.  Arthur. Arthur Mensch. I think he's here. [122:13] Incredible. Incredible company. Perplexity.  Perplexity. His computer. Absolutely. Use it. [122:19] Everybody use it. It is so good. A multi-modal  agentic system. Reflection. Sarvam from India. [122:27] Thinking Machines. Mira Murati's lab.  Incredible companies joining us. Thank you. [122:37] I said. I said that every single enterprise  company, every single software company in the [122:43] world needs an Agentic systems need  an agent strategy. You need to have [122:48] an open source strategy. ________________ [122:50] And they all agree. And they're all  partnering with us to integrate Nemo, [122:56] the NemoClaw reference design, the NVIDIA Agentic  AI toolkit, and of course, all of our open models. [123:05] One company after another. There's so many. [123:08] ________________ And we're partnering with all of you. [123:09] I'm really grateful for that. And, um, this is our  moment. This is a reinvention. This is, this is a [123:16] renaissance, a renaissance of the enterprise. It  from what would be a $2 trillion industry. This is [123:26] going to become a multi trillion dollar industry  offering not just tools for people to use, [123:32] but agents that are specialized in very special  domains that you're expert in that we could rent. [123:39] I could totally imagine in the future, every  single engineer in our company will need an [123:45] annual token budget. They're going to make a few  hundred thousand dollars a year their base pay. [123:51] I'm going to give them probably half of that  on top of it as tokens so that they could be [123:58] amplified ten X. Of course we would. It is now  one of the recruiting tools in Silicon Valley. [124:06] How many tokens comes along with my job.  And the reason for that is very clear, [124:12] because every engineer that has access to tokens  will be more productive. And those tokens, [124:18] as you know, will be produced by AI factories  that all of you and us, we partner to build. [124:25] Okay. So every single enterprise company in today  sit on top of file systems and data centers. [124:32] Every single software company of the future will  be agentic and they will be token manufacturers. [124:39] They'll be token users for their engineers,  and there'll be token manufacturers for all [124:43] of their customers. ________________ [124:45] The open clause in event the OpenClaw  event cannot be understated. This is [124:51] as big of a deal as HTML. This is as big of  a deal as Linux. We have now a world class [124:59] open, a gigantic framework that all of us  could use to build our open core strategy, [125:07] and we've created a reference design we call Nemo.  Nemo that all of you could use that is optimized. [125:15] It's performant. It is safe and secure. ________________ [125:26] Speaking of agents, agents, as you know, perceive,  reason and act. Most of the agents in the world [125:34] today that I've spoken about are digital agents.  They act in the digital world. They reason, [125:40] they write software. It's all digital. But we also  have been working on physically embodied agents [125:48] for a long time. We call them robots. And the AIS  that they need are physical eyes. We have some big [125:54] announcements here. I'm going to just walk through  a few of them, 110 robots here, almost every [126:01] single company in the world, I can't think of one  that are building robots is working with NVIDIA. [126:06] We have three computers the training computer, the  synthetic data generation and simulation computer, [126:12] and of course, the robotics computer that  sits inside the robot itself. We have all [126:17] the software stacks necessary to do so, the AI  models to help you. And all of this is integrated [126:24] into ecosystems around the world. And all  of our partners, from Siemens to Cadence, [126:30] incredible partners everywhere. And today we're  announcing a whole bunch of new, new partners. As [126:37] you know, we've been working on self-driving  cars for a long time. The ChatGPT moment of [126:40] self-driving cars has arrived. We now know we  could successfully, autonomously drive cars. [126:46] And today, we are announcing four new partners  for NVIDIA's robotaxi ready platform. BYD, [126:58] Hyundai, Nissan, Geely. All together,  18 million cars built each year, [127:07] joining our partners from before. Mercedes.  Toyota. GM. The number of robotaxi ready cars [127:16] in the future are going to be incredible. And  we're announcing also a partnership with Uber. [127:23] Multiple cities we're going to be deploying  and connecting these robo taxi ready vehicles [127:29] into their network. And so a whole bunch of  new cars. We have ABB, Universal Robotics, [127:37] Kuka, so many robotics companies here,  and we're working with them to implement [127:43] our physical AI models integrated into  simulation systems so that we could deploy [127:47] these robots into manufacturing lines all over. We have Caterpillar here. We even have T-Mobile [127:54] here. And the reason for that is in the future,  that radio radio tower used to be a radio tower [128:01] is going to be an NVIDIA Aerial AI ram. And  so this is going to be a robotics radio tower, [128:08] meaning it can reason about the traffic,  figures out how to adjust its beamforming [128:13] so that it could save as much energy as possible  and increase the amount of fidelity as possible. [128:19] There's so many humanoid robots here, but  one of my favorites. One of my favorites [128:26] is a Disney robot. You know what? Tell you  what, let me just show you some of the videos. [128:32] Let's look at that first. ________________ [128:41] [VIDEO NARRATION - JENSEN] The first global rollout of physical AI at [128:44] scale is here. Autonomous vehicles and with NVIDIA  Alpamayo vehicles now have reasoning helping [128:52] them operate safely and intelligently across  scenarios. We ask the car to narrate its actions. [129:01] I'm changing lanes to the  right to follow my route. [129:05] Explain its thinking as it makes decisions. There's a double parked vehicle in my [129:10] lane. I'm going around it. And follow instructions. [129:16] Hey, Mercedes, can we speed up? Sure, I'll speed up. [129:24] This is the age of physical AI and robotics. Around the world. Developers are building [129:32] robots of every kind, but the real world  is massively diverse, unpredictable, [129:38] full of edge cases. Real world data will  never be enough to train for every scenario. [129:44] We need data generated from AI and simulation for  robots. Compute is data developers, pre-trained [129:53] World Foundation models on internet scale, video  and human demonstrations, and evaluate the model's [129:59] performance to prepare them for post-training.  Using classical and neural simulation. They [130:07] generate massive amounts of synthetic data  and trained policies at scale to accelerate [130:14] developers. NVIDIA built open source Isaac Lab  for robot training and evaluation and simulation. [130:21] Newton for extensible and GPU accelerated  differentiable physics simulation. [130:27] Cosmos world models for neural simulation. And GR00T open robotics foundation models [130:33] for robot reasoning and action generation.  With enough compute, developers everywhere [130:41] are closing the physical AI data gap. Peritas AI trains their operating room [130:47] assistant robot in NVIDIA Isaac Lab, multiplying  their data with NVIDIA Cosmos World models. [130:54] Skild AI uses Isaac Lab and Cosmos to generate  post-training data for their Skild AI brain. [131:01] They use reinforcement learning to harden  the model across thousands of variations. [131:07] Humanoid uses Isaac Lab to train whole  body control and manipulation policies. [131:14] Hexagon robotics uses Isaac Lab  for training and data generation. [131:20] Foxconn fine tunes Group models in  Isaac lab, as does Noble Machines. [131:27] Disney Research uses their Kamino  physics simulator in Newton and [131:31] Isaac Lab to train policies across their character [131:34] robots in every universe. ________________ [132:45] JENSEN: Ladies and gentlemen. Olaf. OLAF: Coming through. [132:54] JENSEN: Newton works. Wow.  Omniverse works. Olaf, how are you? [133:03] OLAF: I'm so happy now that I'm meeting you. JENSEN: I know because I gave you your [133:09] computer. Jetson. OLAF: What's that? [133:11] JENSEN: Well, It's in your tummy. OLAF: That's going to be amazing. [133:17] JENSEN: And you learn how  to walk inside Omniverse. [133:22] OLAF: I just love walking. This  is so much better than riding on [133:27] a reindeer. Gazing up at the beautiful sky. JENSEN: And it was because of physics. Using [133:35] this Newton solver that runs on top of NVIDIA  Warp that we jointly developed with Disney and [133:43] with DeepMind, that made it possible for you  to be able to adapt to the physical world. [133:49] Check that out. That's how smart you are. OLAF: I'm a snowman, not a snowclopedia. [133:58] JENSEN: Could you imagine this? The future  of Disneyland. All these. All these robots. [134:07] All these characters wandering around. Oh,  you know, I have to admit, though. I thought [134:11] you were going to be taller. I've never  seen such a short snowman, to be honest. [134:18] OLAF: Nope. JENSEN: Hey, tell you what, [134:22] you want to help me out. OLAF: Hooray! [134:24] JENSEN: Okay. Usually. Usually I close the  keynote by telling you what I told you. We [134:31] talked about inference, inflection. We talked  about the AI factory. We talked about the open [134:36] core agent revolution that's happening.  And of course, we talked about physical [134:41] AI and robotics. But tell you what, why don't  we get some friends to help us close it out? [134:47] OLAF: Of course. JENSEN: All right, play it. Come on. [135:32] [VIDEO - SONG PLAYS] JENSEN: All right. [138:45] Have a great GTC. ________________