Jensen Huang (NVIDIA CEO) | NVIDIA Keynotes Summary

NVIDIA Keynotes

1. Key Themes

Agentic AI Has Crossed the Threshold from Experimental to Profitable

Jensen marks a decisive inflection: agentic AI is no longer a future concept but an operational, revenue-generating reality. The proof is in GitHub commit velocity — commits nearly tripled in the first months of 2026 alone, implying that AI-assisted software development is producing measurable economic output at scale.

"In 2023, the number of commits was 300 million. 2024, 400 million. 2025, 500 million commits. In the first few months of 2026, it has nearly tripled." 00:06:46

"Tokens are now profitable units of revenues. Because it is now profitable, the AI companies want to build a lot more tokens, generate a lot more tokens, build more AI factories." 00:09:34

Compute Has Become a Revenue-Generating Asset, Not a Cost Center

Jensen reframes the entire economics of data centers: every watt of compute is now a direct revenue generator, and performance-per-watt is the true unit of profitability. This fundamentally changes how AI infrastructure should be evaluated and purchased.

"Compute is revenue now. Compute is profit. The absence of revenues and profit is loss... If you have one gigawatt of power, then throughput per watt is revenues. Because every token is profitable. Every token is revenues." 00:34:37

"Choosing the wrong architecture just because the chips are cheaper doesn't translate, doesn't make sense. You need to make sure that your revenues per watt, the more you buy, the more you make." 00:36:34

Vera Rubin Is a Paradigm Shift: From GPU to Full Agentic Infrastructure

Vera Rubin is explicitly not a GPU upgrade — it is a five-rack-scale, disaggregated, distributed computing system purpose-built for the agentic workload pattern. Jensen calls it the most ambitious endeavor in NVIDIA's history.

"Vera Rubin is not one chip. Vera Rubin is not a GPU only... This entire thing is Vera Rubin from end to end. It has GPUs, Vera Rubin NVLink 72... the storage systems, revolutionary... everything across this is secure because the AI model is so precious." 00:23:21

"Vera Rubin was built to run agents. This is an agentic system... Vera Rubin, the first multi-rack pod-scale supercomputer, built for the agentic age." 00:46:10

The CPU Is Being Reinvented for Agents, Not Humans — A Massive New Market

Jensen argues that every CPU ever built was designed for human interaction patterns (seconds-scale latency, core-per-rental economics). Agents operate at nanosecond scale and are deeply impatient. NVIDIA's Vera CPU is built from scratch around this new requirement, and Jensen explicitly calls it the fastest product launch in NVIDIA's history.

"All of the CPUs until now were created for people. We were the users... Agents are impatient. They don't live in a world that is in seconds. They live in a world that's in nanoseconds." 00:51:53

"The orders already is going to make it the fastest and the most successful product launch in our company's history." 01:52:35

AI Factory Infrastructure Is NVIDIA's New Identity — Beyond Chips and Systems

NVIDIA has redefined itself as a full-stack AI factory infrastructure company. A 1-gigawatt AI factory now costs $50–100 billion, and the complexity of making it work profitably from day one justifies NVIDIA's entire DSX stack.

"Each one of these at one gigawatt level started at $30, $20, $30 billion. It is at 50, 60 billion dollars and soon it will be 80, 100 billion dollars per gigawatt... It must work the first time and it must work right away." 00:29:48

"NVIDIA has really become an infrastructure company, not just a GPU company, not just a systems company, but an infrastructure company to help you generate the maximum revenues, the maximum profit." 01:52:35

The PC Is Being Reinvented for the First Time in 40 Years

NVIDIA and Microsoft have co-designed a new PC platform (RTX Spark) built around local agentic workloads. Jensen draws a direct parallel to the smartphone revolution and predicts household AI supercomputers will become as common as home theaters.

"Microsoft and NVIDIA over the last three years, it took this long to completely reinvent how the PC is going to work so that we could be ready for this moment." 01:21:14

"I could totally imagine that someday there's actually an AI supercomputer in your house, and it's running all of your agents... you want to assist AI agent computers running in your house. And these in time becomes a lot more like R2-D2 to you." 01:35:22

Physical AI and Robotics Are NVIDIA's Next Frontier — Cosmos 3 Is the Foundation Model

Jensen positions physical AI (robotics, autonomous vehicles, humanoids) as the next wave, with Cosmos 3 serving as the world model equivalent of what large language models are to digital AI. The data scarcity problem for physical AI is being solved by using compute itself to generate synthetic data.

"In the case of language models, all the English and all the language that we have on the internet that we trained on was from the perspective of us... However, in order to create data for AI, robotics, it has to be in the perception, the perspective of the robot." 01:38:01

"Now that we have AI, compute is data... Cosmos 3 is the frontier of physical AI. We are at the frontier with language models... However, in physical AI, we are absolutely the world's best." 01:42:46

The Same Agentic Computing Pattern Repeats Across Every Device Category

Jensen's deepest architectural insight: the model + harness + tools + runtime pattern is universal — identical whether running in a hyperscale data center, an enterprise, a PC, an autonomous vehicle, or a humanoid robot. This means NVIDIA's software and silicon investment compounds across every category simultaneously.

"The computing pattern will repeat over and over and over again. This computing pattern of an agent that's a model, a harness, harness that uses tools with skills and runs in a runtime. That runtime depends on whether it's in the cloud or on-prem, on a PC or in a robot. But the computing pattern is exactly the same for all of them." 01:50:48

NVIDIA's CUDA Library Moat Deepens as Agents Become Expert Tool Users

CUDA libraries (CUDAX) are transitioning from human-facing developer tools to machine-consumed agent tools. Jensen argues agents will use these libraries more effectively than humans ever could, and the moat compounds because NVIDIA is now shipping "skills" — essentially manuals for agents to self-learn tool usage.

"Today, we're able to now present these CUDAX libraries to agents who can use it much more effectively than even humans... The CUDAX library, some skills, basically a manual. The AI reads it and goes, aha, that's how you use it." 00:15:22

Taiwan's Supply Chain GDP Is Directly Coupled to AI Token Demand

Jensen explicitly ties Taiwan's projected ~10% GDP growth directly to AI compute demand, framing Taiwan's entire supply chain ecosystem as the upstream engine of the global AI factory buildout.

"Somebody told me last night that the annual GDP of Taiwan is going to grow almost 10%... The supply chain we created for Vera Rubin is twice as large as Grace Blackwell." 00:05:53 / 00:40:26

2. Contrarian Perspectives

AI Creates More Jobs, It Does Not Destroy Them

Most mainstream discourse assumes AI will reduce software engineering employment. Jensen argues the productivity multiplier creates an incentive to hire more engineers, not fewer.

"The number of engineers, software engineers, is actually increasing. People talk about AI reducing jobs, complete nonsense. It's causing more software engineers to be hired. And the reason for that is very simple. If you can hire a software engineer and you could generate $9 trillion worth of productive work, why wouldn't you want to hire more software engineers?" 00:08:42

Agentic AI Saves Enterprise Software Companies — It Doesn't Disrupt Them

The conventional narrative is that AI agents will cannibalize SaaS incumbents. Jensen argues the opposite: agents will consume more software tools than humans ever could, making this the best time ever to be a software company — provided tools are presented in agent-consumable formats.

"A lot of people have said, you know, Jensen, AI is coming, agentic AI is coming, therefore all of the software companies are going to go out of business. I said it's exactly the opposite. Because there are going to be so many agents, the world is no longer limited by the number of people. Therefore, those agents are going to use more tools than ever." 00:14:27

Cheaper Chips Are a False Economy in AI Factories

The prevailing buyer instinct in commodity hardware is to minimize chip cost. Jensen directly inverts this: in a power-constrained factory where every token is revenue, performance-per-watt is what matters, and buying cheaper, less efficient chips actually destroys revenue.

"Choosing the wrong architecture just because the chips are cheaper doesn't translate, doesn't make sense. You need to make sure that your revenues per watt, the more you buy, the more you make." 00:36:34

Inference Was Never "Easy" — It Was Always the Hard Money Problem

There was a period when the industry consensus was that inference was a solved, commoditized problem. Jensen rejects this entirely, and Grace Blackwell's commercial dominance in inference validates the view.

"Everybody said, Jensen, you know, NVIDIA is really good at pre-training. Inference is so easy. Do you remember that? People used to say inference is so easy. We could do that too. But as you know, inference equals money." 00:48:04

The Personal Computer Will Be Unrecognizable in 10 Years — Household AI Supercomputers Are Inevitable

The mainstream view is that AI is an app or a feature on existing PCs. Jensen's prediction is more radical: the PC form factor will be replaced by household AI supercomputers running continuous autonomous agents, analogous to how phones were replaced by smartphones.

"There is no question this reinvention of the computer is as big of a deal as the reinvention of the phone into what we now know as the smartphone... I could totally imagine that someday there's actually an AI supercomputer in your house, and it's running all of your agents." 01:35:22

3. Companies Identified

CoreWeave Cloud computing company specializing in GPU infrastructure. Cited as a primary example of how small companies, empowered by NVIDIA's full stack, have scaled to extraordinary valuations rapidly.

"CoreWeave is worth 50, 60, 70 billion dollars and growing incredibly fast." 00:31:46

Nebius AI cloud company. Cited alongside CoreWeave as a fast-growing regional AI cloud built on NVIDIA infrastructure, serving frontier AI customers.

"Recently, we worked with Nebius. And again, they're growing incredibly fast." 00:31:46

Cadence EDA (Electronic Design Automation) software company. Highlighted as the first major enterprise super-agent deployment — partnering with NVIDIA to build chip design verification agents that compress weeks of work into hours.

"Cadence and NVIDIA built a design verification agent... What once took weeks now takes hours. Verification cycles over 40 times faster. Together, NVIDIA and Cadence are reinventing chip design with AI agents." 01:13:30

Microsoft Technology giant. Co-designer of the new RTX Spark PC platform and Windows agent runtime; first to stand up an operational Vera Rubin NVLink 72 engineering rack.

"Congratulations to Microsoft for their operational Vera Rubin NVLink 72 engineering rack... Microsoft and NVIDIA over the last three years, it took this long to completely reinvent how the PC is going to work." 00:43:17 / 01:22:14

Cursor AI-powered software coding company. Named as a key customer of NVIDIA-powered AI clouds, representing the leading edge of agentic software development adoption.

"Cursor, the software coding company, Black Mountain Labs, Image Generation, World Labs, World Foundation Model, Revolut, the leading financial services AI company, and Shopify." 00:31:46

World Labs World foundation model company. Cited as a customer of NVIDIA AI clouds building at the frontier of physical world modeling.

"World Labs, World Foundation Model." 00:31:46

Revolut Fintech company. Called "the leading financial services AI company" — notable endorsement of AI adoption in regulated financial services.

"Revolut, the leading financial services AI company." 00:31:46

Shopify E-commerce platform. Named as an NVIDIA AI cloud customer, signaling deep AI infrastructure investment by a major consumer commerce company.

"Revolut, the leading financial services AI company, and Shopify." 00:31:46

Nscale AI cloud company. Cited for having Google as a customer — a notable signal that even Google uses third-party AI clouds running on NVIDIA infrastructure.

"This is Nscale, and their customers are British Telecom, Google. Google is using one of our AI clouds." 00:31:46

Thinking Machines Frontier AI lab. Called "super exciting" by Jensen, named as a customer of Nscale's AI cloud.

"Thinking machines, a frontier labs company, which is super exciting." 00:32:38

Naver Cloud Korean cloud platform. Cited for serving Bank of Korea and Hyundai as AI cloud customers — representing NVIDIA's penetration into national and industrial AI infrastructure in Korea.

"Here's Naver Cloud in Korea, Bank of Korea, Hyundai, so many incredible companies." 00:32:38

Together AI / AI Singapore AI cloud company based in Singapore, building in Australia. Cited as part of NVIDIA's global regional AI cloud ecosystem.

"Here's one based in Singapore, building in Australia, Together AI. AI Singapore." 00:32:38

Endosat AI cloud company based in Indonesia. Named as part of NVIDIA's expanding Southeast Asian AI infrastructure ecosystem.

"This is one in Indonesia. Each one of these companies are serving regional as well as global customers." 00:32:38

GMI AI cloud company based in Taiwan. Named by Jensen with an explicit prompt for applause — signaling strong partnership.

"Here in Taiwan, GMI. It's okay to clap." 00:32:38

CrowdStrike Cybersecurity company. Named as a partner deploying NVIDIA's enterprise AI agent toolkit for their workflows.

"We're working with so many companies, Cadence and CrowdStrike and DeSole and Palantir, SAP and ServiceNow." 01:18:18

Palantir Data analytics and AI platform. Named as an enterprise agent toolkit partner.

"Cadence and CrowdStrike and DeSole and Palantir, SAP and ServiceNow." 01:18:18

SAP Enterprise software company. Named as an enterprise agent toolkit partner — significant signal that SAP is building agents on NVIDIA's platform rather than being disrupted by it.

"Cadence and CrowdStrike and DeSole and Palantir, SAP and ServiceNow." 01:18:18

ServiceNow Enterprise workflow automation company. Named as an enterprise agent toolkit partner.

"Cadence and CrowdStrike and DeSole and Palantir, SAP and ServiceNow." 01:18:18

MediaTek Taiwanese semiconductor company. Co-designed the N1X chip powering RTX Spark laptops — a landmark partnership between NVIDIA and a major ARM-ecosystem chip designer.

"This is the N1X that we built in partnership with MediaTek... This is a beautiful chip. This is a chip that, frankly, would take 33 years to build." 01:25:13

Adobe Creative software company. Re-engineered the core of Photoshop and Premiere for RTX Spark, delivering 2x performance and adding MCP server agent-interoperability.

"They have re-engineered the architecture, the core of Adobe Photoshop and Premiere, and they'll release it for RTX Spark. It is twice as fast... With its MCP server, it can now interact with agents on your laptop." 01:30:02

Foxconn / Quanta Taiwanese electronics manufacturers. Named as the production sites where the Grok 3 LPX system for Vera Rubin takes shape.

"At Foxconn and Quanta, Grok 3 LPX takes shape." 00:44:15

Micron / SK Hynix / Samsung Memory manufacturers. Named as HBM4 memory suppliers for Vera Rubin — representing the full global memory supply chain mobilized for NVIDIA's flagship system.

"HBM4 memory from Micron, SK Hynix, and Samsung." 00:42:22

TSMC Semiconductor foundry. Named as the manufacturer of Vera Rubin chips using 3nm process with CoWoS-R and CoWoS-L advanced packaging.

"It starts at TSMC. The seven new chips that make up Vera Rubin take shape through hundreds of processing steps. Three nanometer process, CoWoS-R and CoWoS-L packaging." 00:42:22

Dell Technology company. Named alongside Microsoft and CoreWeave as an early Vera Rubin NVLink 72 engineering rack deployer.

"Congratulations to Dell and CoreWeave as well for standing up their Vera Rubin NVLink 72 engineering rack." 00:43:17

Red Hat / Canonical Linux distribution companies. Named as early adopters of NVIDIA OpenShell, the open-source enterprise agent runtime.

"NVIDIA OpenShell is open source. You can see so many companies adopted, Red Hat, Canonical, Microsoft." 01:10:48

New York Stock Exchange Financial exchange. Named as a live production user of Vera CPU for real-time stream processing, achieving 6x performance improvement.

"This is Vera CPU running real-time stream processing for New York Stock Exchange. Lynn Martin, the president of New York Stock Exchange, has been so gracious to partner with us." 01:05:08

Black Mountain Labs AI image generation company. Named as a customer of NVIDIA AI clouds at the frontier of generative image AI.

"Black Mountain Labs, Image Generation." 00:31:46

Sharpa Robotics hardware company. Named as the manufacturer of the 25-degrees-of-freedom hands on NVIDIA's Isaac Groot reference humanoid robot.

"25 degrees of freedom on each hand, made by Sharpa." 01:47:50

4. People Identified

Lynn Martin President of the New York Stock Exchange. Highlighted for personally partnering with NVIDIA to deploy Vera CPU for real-time stream processing — a remarkable signal of AI infrastructure adoption at the world's most important financial exchange.

"Lynn Martin, the president of New York Stock Exchange, has been so gracious to partner with us. This system is run all over the world in real-time stream processing." 01:05:08

Satya Nadella CEO of Microsoft. Referenced as Jensen's partner in the multi-year PC reinvention project, with a joint keynote announced for the following day.

"Tomorrow night, I think it's tomorrow night our time, but I'm going to be with Satya. We're going to talk a lot more about the work that we're doing together." 01:21:14

Rick (MediaTek representative) Unspecified executive at MediaTek. Acknowledged by Jensen on stage during the RTX Spark N1X chip unveiling, signaling a high-level personal partnership.

"I think I saw Rick earlier. This is N1X. This is a beautiful chip." 01:25:13

5. Operating Insights

AI Factories Are Over-Provisioning Power by 40% — DSX Max LPS Recovers That as Revenue

This is an immediately actionable operational insight for anyone building or operating AI infrastructure. Current AI factories waste up to 40% of their power budget, which directly translates to stranded revenue. NVIDIA's DSX Max LPS software layer recovers this without adding hardware.

"Today's AI factories over-provision power by up to 40%. DSX Max LPS lets operators safely deploy more GPUs inside the same power budget, adding billions in annual revenue." 00:27:56

Digital Twin Simulation Before Physical Build Is Now Mandatory at Gigawatt Scale

For any operator planning large-scale AI infrastructure, simulating the entire factory in Omniverse before breaking ground is not a nice-to-have — at $50–100 billion per gigawatt, the cost of a misconfigured rack is catastrophic. NVIDIA's DSX-SIM blueprint makes this a standard practice.

"With the DSX-SIM Omniverse Blueprint, partners design and validate an NVIDIA Vera Rubin AI factory before a single rack lands. They plan the layout, simulate the power and cooling, design the network, validate every integration, test every change in the digital twin." 00:26:57

Assembly Time as a Supply Chain KPI: Vera Rubin Reduced Rack Assembly from 2 Hours to 5 Minutes

Jensen uses rack assembly time as a proxy for supply chain maturity and throughput capacity. This 24x improvement is a direct driver of how fast AI capacity can come online globally — and a benchmark for evaluating any hardware partner's manufacturability.

"What used to take two hours to assemble one Grace Blackwell rack now only takes five minutes. So not only is the capacity higher, the throughput is a lot faster, and we need it all to support the demand." 00:40:26

The Enterprise Agent Toolkit Framework: Four Non-Negotiables Before Deploying Agents

Jensen distills the enterprise agent deployment requirement into four components. Any enterprise evaluating AI agent readiness can use this as a checklist: models (modifiable), harness (orchestration), tools with skills (domain-specific libraries), and a secure runtime.

"There are four things that companies need in order to build agents as a service or build agents to operate. The first thing you need is you need models... The second is you need a harness to orchestrate the whole thing. The third, these models want to use tools... And then lastly, you need a runtime." 01:08:58

Chip Design Verification Agents: A 40x Cycle Compression Playbook Applicable Across Engineering

The Cadence partnership produced a replicable template: take a multi-week expert engineering verification cycle, decompose it into sub-agent specialists (RTL generation, test bench creation, regression testing, debug), orchestrate with Codex/Claude Code, and compress the cycle 40x. Any engineering-intensive company should be mapping their own verification and QA loops to this pattern.

"What once took weeks now takes hours. Verification cycles over 40 times faster... NVIDIA has thousands of chip designers. We are going to hire hundreds of thousands of Cadence super agents that work with us so that we can accelerate our company." 01:14:24

6. Overlooked Insights

NVLink Is Now an Open Ecosystem Standard — NVIDIA Just Became a Platform for Competitor Silicon

Jensen mentioned "NVLink Fusion" and explicitly said "Everyone's welcome to the NVLink party" in the closing rap — but buried in the keynote is the implication that NVLink chip-to-chip connectivity is being opened as a fabric for third-party ASICs to connect into NVIDIA's ecosystem. This is profoundly significant: it means NVIDIA is not just competing with custom silicon, it is positioning itself as the interconnect fabric that custom ASICs must plug into. Companies building their own AI chips (hyperscalers, startups) may find it more practical to integrate via NVLink than to build competing end-to-end stacks.

"Memory-coherent NVLink chip-to-chip connects GPUs directly to the fabric. Beyond GPUs, NVLink chip-to-chip can scale Vera up to multiple sockets, enabling massive bandwidth between CPUs." 01:01:15

The KV Cache / Memory Management Layer Is the Hidden Bottleneck — and an Unaddressed Investment Opportunity

Jensen spent considerable time explaining that the memory management system of agentic AI — specifically KV cache compaction, retrieval of structured vs. unstructured data, and ontology management — is "incredibly complicated" and will "cause the storage system to be completely revolutionized." He named it as one of the hardest parts of the entire agentic stack. Yet no specific company or product was named to address it. This is a conspicuous gap in the ecosystem map Jensen drew — a critical infrastructure layer without a dominant winner, representing a significant investment and building opportunity.

"What to remember? Compaction, not just compression, but how to retrieve. Do you retrieve structured data? Do you retrieve unstructured data? What is the ontology, the relationship of all of these different data to itself? That entire processing is incredibly complicated. The memory system of AIs is going to cause the storage system to be completely revolutionized." 00:22:26