Jensen Huang (NVIDIA CEO) | NVIDIA Keynotes Summary

1. Key Themes

AI Factories as Revenue-Generating Infrastructure, Not Data Centers

Jensen reframes the entire mental model of AI compute infrastructure. These are not data centers in the traditional sense — they are factories that manufacture tokens, and tokens are the building blocks of productive intelligence.

"Nobody really thinks about their data center as a revenue generating facility. But they think of their factories, their car factories as revenue generating facilities and they can't wait to build another factory because whenever you build a factory, revenue grows shortly after... Those ideas are exactly the same ideas in these AI factories." 00:39:49

Reasoning Models as the Core Driver of Exponential Compute Demand

The shift from one-shot generative AI to reasoning/agentic AI is not incremental — it multiplies token generation by orders of magnitude, which directly justifies the leap from Hopper to Blackwell architecture.

"One prompt like that on the original chatbot would have generated a few hundred tokens. But now with that one single prompt into an agent to solve a problem, it must have generated 10,000 times more tokens. This is the reason why Grace Blackwell is necessary." 00:59:02

The Third Wave of AI: Agentic and Embodied Intelligence

Jensen identifies a clear three-wave structure — perception, generative AI, and now agentic + physical AI — and argues that both information robots (agents) and physical robots (humanoids) are now technically feasible and entering deployment.

"These capabilities, the fundamental technology to enable agents which are basically information robots and embodied AI, physical robots, these two fundamental capabilities are now upon us." 00:16:25

Quantum Computing Is Hitting a Real Inflection Point

Jensen makes a serious, non-hype claim about quantum computing timelines, predicting that every next-generation supercomputer will have a QPU, and that CUDA Q now bridges classical and quantum workloads on Grace Blackwell.

"It is clear now we are within reach of being able to apply quantum computing, quantum classical computing in areas that can solve some interesting problems in the coming years... every single one of them will have a QPU assigned and QPU connected to GPUs." 00:10:03

Europe as a Strategic AI Infrastructure Buildout

Europe is not just a market — it is becoming a sovereign AI infrastructure zone. Jensen announces a 10x increase in AI compute capacity in Europe within two years, with indigenous cloud providers, supercomputing centers, and national AI technology centers across seven countries.

"In total, in just two years, we will increase the amount of AI computing capacity in Europe by a factor of 10." 00:43:18

Digital Twins as the New Design and Operations Standard

Jensen argues that everything physical — factories, warehouses, train stations, fusion reactors, cars — will first be built, optimized, and operated digitally. This is not a future state; it is happening now with BMW, Mercedes-Benz, Toyota, and others.

"Everything physical will be built digitally. Everything that's built magnificently will be built digitally. Everything that's operated at gigantic scale will be first built digitally and there will be digital twins that operate it." 00:20:04

Humanoid Robotics as the Next Massive Industrial Wave

Jensen ties together the AI software stack, digital twin training environments, and manufacturing heritage of Europe to argue that humanoid robotics may become one of the largest industries in history — and that the key bottleneck (programming difficulty) is now being solved.

"Humanoid robots are going to be a thing. We now know how to build these things, train these things, and operate these things. Humanoid robotics is going to potentially be one of the largest industries ever. And it requires companies who know how to manufacture things of extraordinary capabilities. This speaks to the European countries." 01:25:44

Open Models Are Only Months Behind Proprietary Ones — and NVIDIA Enhances Them

Jensen explicitly states open models like Mistral, Llama, and DeepSeek R1 are exceptional and only months behind closed models, and that NVIDIA's Neotron post-training program systematically improves them — making open models top leaderboard performers.

"The open models are also moving at light speed. Only a few months behind. Whether it's Mistral, Llama, DeepSeek, R1, R2, coming Q1, these models are all exceptional. Every single one of them exceptional." 00:48:50

NVIDIA's Software Stack as the True Moat

The real defensibility of NVIDIA is not just the GPU hardware — it is the 400+ domain-specific libraries (CUDA X), the developer ecosystem, the NIM deployment containers, the Nemo training framework, and the fact that NVIDIA is the only AI architecture available in every cloud.

"NVIDIA is the only AI architecture that's available in every cloud. It's the only computing architecture aside from x86 that's available everywhere." 00:44:20

Inference Demand Is Going Exponential

Jensen cites a 100x growth in inference users in just a couple of years and argues that agentic AI will push this further. The combination of more users, more tokens per prompt, and more frequent use creates a compounding demand curve.

"The number of people that are using inference has gone from 8 million to 800 million — a 100 times in just a couple of years." 01:34:30

2. Contrarian Perspectives

AI Data Centers Should Be Valued Like Factories, Not IT Cost Centers

Most CFOs and investors still model data centers as cost infrastructure. Jensen argues the opposite framing entirely — and the investment and operational implications are completely different.

"These AI factories are now part of a country's infrastructure... They are revenue generating facilities and they are designed to manufacture tokens." 00:40:50

Scaling Up (Not Out) Is the Hard Problem Everyone Is Getting Wrong

The industry conversation focuses on scaling out (adding more nodes via Ethernet). Jensen argues the real breakthrough is scaling up — building one giant virtual GPU — which requires memory-semantics interconnects, not networking, and is fundamentally harder.

"Scaling out computing is not that hard. Just connect more CPUs with Ethernet. Scaling out is not hard. Scaling up is incredibly hard... And so what we decided to do was we created a new interconnect called MVLink. MVLink is a memory semantics interconnect. It's a compute fabric, not a network." 00:23:53

Quantum Computing Is Not Overhyped — It Is at a Real Inflection Now

The prevailing view in tech investing is that quantum is perpetually 10 years away. Jensen takes a specific, time-bound position with a Moore's Law analogy applied to logical qubits, and commits NVIDIA's full Grace Blackwell stack to it.

"I could totally expect 10 times more logical cubits every 5 years, a 100 times more logical cubits every 10 years... Quantum computing is reaching an inflection point." 00:10:03

Everything Physical Will Be Built Digitally First — Full Stop

This is not a niche industrial use case. Jensen argues this is the universal future of manufacturing, infrastructure, and operations — and that digital twins must be photorealistic and physics-accurate because robots learn from photons.

"The idea that we would build everything in software is now upon us. Everything physical will be built digitally... these digital twins have to look real and behave realistically." 00:20:04

The First Customer of the DGX1 Was OpenAI — and Jensen Drove It There Himself

This is a deeply non-obvious historical fact that reframes how OpenAI got started computationally, and also illustrates Jensen's personal sales instinct in the early days of a product with "no customers, no interest, 100% confusion."

"A startup, a nonprofit startup in San Francisco was so delighted to see the computer. They said, 'Can we have one?'... I put a DGX1 in my car and I drove it up to San Francisco and the name of that company is OpenAI." 01:04:14

3. Companies Identified

NVIDIA Semiconductor and AI computing platform company. Central to the entire keynote — Jensen describes NVIDIA's full stack from chips (Blackwell, Thor) to software (CUDA, Nemo, NIM, Lepton, Omniverse) to ecosystem partnerships spanning every cloud and major industry.

"NVIDIA is the only AI architecture that's available in every cloud. It's the only computing architecture aside from x86 that's available everywhere." 00:44:20

Mistral French open-weight AI model company. NVIDIA is building a joint AI cloud with Mistral to serve European AI startups and enterprise customers with regional models.

"Today, we're announcing that we're going to build an AI cloud together here to deliver their models as well as deliver AI applications for the ecosystem of other AI startups." 00:48:50

OpenAI AI research company. Named as the first-ever customer of the DGX1, received personally by Jensen in 2016 when the company was a nonprofit.

"I put a DGX1 in my car and I drove it up to San Francisco and the name of that company is OpenAI." 01:05:15

Perplexity AI-powered reasoning search engine. NVIDIA is partnering with Perplexity to integrate regional European language models directly into Perplexity's platform.

"We're announcing today that Perplexity will take these regional models and connect it right into Perplexity so that you could now ask and get questions in the language, in the culture, in the sensibility of your country." 00:53:20

Siemens German industrial conglomerate. Deep partnership with NVIDIA on industrial AI, digital twins, and factory simulation. CEO Roland Buch highlighted personally. Siemens also built the Synapse One in 1992 — described as the world's first neural network computer.

"Siemens 1992. We have a great partnership with Siemens and Roland Buch, the CEO, is supercharging the company so that they could leap completely the last IT industrial revolution and fuse the industrial capabilities of Europe with artificial intelligence." 01:11:40

BMW German automaker. Building next-generation factory digital twins in Omniverse.

"This is BMW doing building their next generation factory in Omniverse." 01:15:38

Mercedes-Benz German automaker. Building digital twins of their factories in Omniverse.

"This is Mercedes-Benz and their digital twins of their factories built in Omniverse." 01:16:38

Toyota Japanese automaker. Building digital twin of their warehouse in Omniverse.

"This is Toyota building a digital twin of their warehouse in Omniverse." 01:16:38

Cisco Networking and enterprise technology company. Announced AI security platform built jointly with NVIDIA; demonstrated a multi-agent AI architecture for security.

"This is Cisco. They announced it yesterday. We're building AI platforms together for security." 01:00:41

SAP German enterprise software company. Building AI business application automation on NVIDIA.

"This is the way SAP — they're building an AI platform in Nvidia. SAP is building an AI business application automation on NVIDIA." 01:08:01

Deepl Language translation platform. Building their language framework and platform on NVIDIA AI.

"Deepl is building their language framework and platform on NVIDIA AI." 01:08:01

Photoroom AI video and photo editing platform. Building on NVIDIA.

"Photoroom, a video editing and AI editing platform, building their platform on NVIDIA." 01:08:01

Kodu (formerly Kodium) AI coding agent. Building on NVIDIA.

"This is Kodu, used to be I think Kodium — incredible coding agent built on Nvidia." 01:08:01

Iola Voice AI platform. Building on NVIDIA.

"This is Iola, a voice platform built on Nvidia." 01:09:03

Hugging Face AI model repository and developer platform. Integrated with NVIDIA's Lepton so developers can deploy models from Hugging Face to any cloud with one click.

"Hugging Face and NVIDIA has connected Lepton together. And so whenever you're training a model on Hugging Face, if you would like to deploy it into Lepton and directly into Spark, no problem. It's just one click." 01:06:14

Schneider Electric French energy and industrial automation company. Partnering with NVIDIA on physically and digitally building AI factories.

"One particular one I want to highlight — our partnership with Schneider, building the — even building these AI factories. We build them digitally now." 00:46:55

TSMC / Samsung Semiconductor fabs. Named as the production sites for computational lithography using NVIDIA's cuLitho library.

"Computational lithography... runs in a factory at TSMC, Samsung, large semiconductor fabs." 00:05:56

Google Named for demonstrating the world's first logical qubit in 2023, a milestone Jensen uses to anchor the quantum computing inflection point narrative.

"In 2023, almost 30 years later, the world's first logical qubit was demonstrated by Google." 00:09:03

Stargate Massive AI infrastructure project in the US. Cited by Jensen as a 1-gigawatt AI factory holding approximately 500,000 GPU dies — a visual example of what AI infrastructure at national scale looks like.

"This is Stargate. This doesn't look like a data center. It looks like a factory. This is 1 gigawatt. It will hold about 500,000 GPU dies." 00:42:31

Disney Research R&D arm of Disney. Partnering with NVIDIA and DeepMind to create the world's most sophisticated physics simulation for robot training.

"We announced a big partnership with Disney Research and DeepMind. And we're going to work together to create the world's most sophisticated physics simulation." 01:29:32

DeepMind Google's AI research lab. Co-developing physics simulation with NVIDIA and Disney Research for humanoid robot training.

"We announced a big partnership with Disney Research and DeepMind." 01:29:32

Red Hat Enterprise Linux provider. Named as part of the enterprise software stack that NVIDIA's new RTX Pro Server supports natively.

"Some data centers required enterprise stacks, the ability to run Linux, Red Hat or Nutanix or VMware." 00:34:19

Dell Technology company. Named as a key NVIDIA partnership for enterprise AI system integration.

"Great partnership with Cisco... Dell, great partnerships, NetApp, Nutanix." 00:45:18

NetApp / Vast / Weka Storage companies. Named as storage partners integrated with NVIDIA's enterprise AI stacks.

"Storage systems from Dell EMC, Hitachi, NetApp, Vast, Weka, so many different storage systems." 00:34:19

Lambda Cloud / AWS / GCP Cloud providers. Named as deployment targets in NVIDIA's Lepton multi-cloud orchestration platform.

"Here's the Lambda cloud, the AWS cloud... It could be AWS. It could be GCP." 01:02:29

Yoda / Nebus / Scale Named as cloud/compute partners integrated into DGX Cloud Lepton.

"DGX Cloud Lepton provides on-demand access to a global network of GPUs across clouds, regions, and partners like Yoda and Nebus." 01:07:05

Barcelona Supercomputing Center European supercomputing institution. Named as a key partner in quantum computing and supercomputing advancement.

"I saw Barcelona Supercomputing last night. It is clear now we are within reach of being able to apply quantum computing." 00:10:03

Kion Company building digital twin for warehouse logistics in Omniverse.

"This is Kion, their digital twin for warehouse logistics." 01:16:38

Schaefer Company building digital twin of their warehouse in Omniverse.

"This is Schaefer and their digital twin of their warehouse built in Omniverse." 01:16:38

4. People Identified

Jensen Huang Founder and CEO of NVIDIA. Delivered the entire keynote, demonstrating encyclopedic command of NVIDIA's full hardware, software, and ecosystem strategy across compute, AI, robotics, quantum, and industrial applications.

"We are driven to enable the geniuses of our time to do their life's work, and we can't wait to see the breakthroughs you create." 00:32:23

Roland Buch CEO of Siemens. Personally called out by Jensen for championing the industrial AI transformation at Siemens and for bringing historical context — Siemens built the Synapse One in 1992, the world's first neural network accelerator.

"Roland Buch, the CEO, is supercharging the company so that they could leap completely the last IT industrial revolution and fuse the industrial capabilities of Europe with artificial intelligence." 01:11:40

President Emmanuel Macron President of France. Named as a keynote attendee at VivaTech and as a participant in announcing further AI infrastructure commitments for France, including the Mistral partnership.

"President Macron's going to be here later on. We're going to talk about some new announcements." 00:46:55

5. Operating Insights

Treat AI Agents Like Digital Employees with a Full HR/IT Lifecycle

Jensen frames the operational management of AI agents explicitly as an HR and IT process — onboarding, fine-tuning, training, evaluation, guardrails, security, and continuous improvement. Companies that build this lifecycle infrastructure early will have a durable operational edge over those treating AI as one-off deployments.

"The way you treat an AI agent is a little bit like a digital employee. So your IT department would have to onboard them, fine-tune them, train them, evaluate them, keep them guardrailed, keep them secure and continuously improve them. And that entire framework platform is called Nemo." 01:09:03

Deploy One Agentic Architecture Across All Environments via Multi-Cloud Orchestration

Rather than rebuilding AI pipelines for each cloud or on-prem environment, Jensen advocates a single Helm chart deployment via Lepton that routes to wherever compute is cheapest or most appropriate — reducing operational fragmentation and accelerating time to production.

"You have this Helm chart, an AI agent that you've developed, and you want to run it here and parts of it you want to run in AWS and parts of it you want to run in a regional cloud somewhere — you use Lepton, you deploy your Helm chart and it magically shows up here." 01:05:15

Build and Operate AI Factories as Revenue-Per-Token Businesses, Not Cost Centers

Operators of AI infrastructure should model their systems on factory economics: maximize throughput (tokens per second across users) and intelligence quality (tokens per query), because the product of those two axes equals factory revenue. Underutilization in a $50B AI factory is catastrophically expensive.

"You want the throughput of the factory to be as high as possible so you could support as many people as possible so that the revenues of your factory is as high as possible... These AI factories are so $50 billion sometimes — if the utilization of that factory is not at its fullest, the cost to the factory owner is going to be incredible." 00:28:32

Use Open Regional Models + Neotron Post-Training to Own Your Enterprise Data Advantage

Rather than relying entirely on closed proprietary models, Jensen's playbook is: take an open model, enhance it with your proprietary enterprise data (which belongs to you and is not on the internet), extend its context window, and deploy it via NIM. This is a replicable, defensible enterprise AI moat.

"Your data belongs to you. It is the history of your people, the knowledge of your people, the culture of your people... You should use that data, use an open model like Neotron and all the tool suites that we provide so that you can enhance it for your own use." 00:52:19

When a Developer Asks for a GPU, the Answer Is Always Yes

Jensen's personal origin story with OpenAI contains a direct operating principle he states explicitly — the developer relationship is the most valuable one to protect, and the cost of saying no compounds in ways that are impossible to predict.

"If a developer reaches out and needs a GPU, the answer is yes." 01:05:15

6. Overlooked Insights

The Clinical Trial Automation Platform — an Invisible Giant

In a single passing mention buried in a list of NVIDIA ecosystem partners, Jensen names "the world's largest automation platform for clinical trials" as being built on NVIDIA. This is a massive, largely invisible vertical — clinical trials are a multi-hundred-billion-dollar industry with enormous inefficiency, and the company building the dominant AI automation layer for it is almost certainly a category-defining company. Jensen mentions no name, which makes it even more worth investigating.

"This one is a clinical trial platform — the world's largest automation platform for clinical trials built on Nvidia." 01:09:03

MVLink Spine Bandwidth Exceeds the Entire Global Internet's Peak Traffic

This is stated almost as a throwaway technical boast, but the implication is profound for understanding the trajectory of AI model scale. If a single rack's internal bandwidth already exceeds peak global internet traffic, the architectural gap between AI-native infrastructure and conventional distributed computing is not a matter of degree — it is a different category of machine. Any investor or operator still thinking about AI inference in terms of traditional distributed systems is reasoning from the wrong model entirely.

"The bandwidth of this is about 130 terabytes per second... it is more than the data rate of the peak traffic of the world's entire internet traffic on this back plane. And so this is how you shrink the internet into 60 pounds." 00:25:52