Jensen Huang on The Future of Computing | AI, Infrastructure & What’s Next

Jensen Huang
2 followers
NVIDIA CEO Jensen Huang shares his vision for the future of computing at #DellTechWorld. From AI factories and accelerated infrastructure to the next evolution of intelligent systems, the architecture of computing itself is being reinvented in real time.
AI Generated·May contain errors
Video Summary
AI GeneratedJensen Huang discusses the transition from generative AI to \"useful AI\" through agentic systems. He explains how AI agents—capable of reasoning, planning, and tool use—are exponentially increasing computational demand and productivity. He also details NVIDIA's hardware architecture, including the Vera CPU and GB200 systems, designed to support hybrid AI and large-scale agent orchestration.
Please join me in welcoming a great partner and friend,
a true leader and visionary of the AI age, NVIDIA's Jensen Wang.
Thank you, Jensen.
Jensen, it's great to have you back here at Dell Tech World.
You know, we've been talking about agents.
I'm here every year selling Dell.
We appreciate it, man.
We appreciate it.
So give us your perspective on where we are in this
agentic recursive self-improvement world.
It looks like every time, you know, we wake up,
there's been some leap in the model and the capabilities.
And sort of, it doesn't seem to be plateauing.
Give us your perspective, you know, from the front lines.
Well, two years ago when I was here,
we had just started with the agentic, excuse me, generative AI journey.
Generative AI can, of course, generate content.
But remember, it can also generate content to think with, generate thoughts,
which led us to reasoning, which led us to planning, which led us to agentic systems.
So now we have, we now have, for the very first time, useful AI,
which is the reason why your demand, my demand is going parabolic, utterly parabolic.
Because, right?
It's going parabolic because agentic systems, the AI has to understand, has to reason,
has to think, it has to plan, use tools, look at the results of the tools,
think about it some more, come up with maybe an improved plan.
And so it iterates until it can get the job done using a whole bunch of different tools.
Well, the amount of computation necessary,
because it's running autonomously for so long,
instead of just responding to a query,
the amount of computation has grown 100x, 1000x.
And depending on the work that you're doing,
sometimes will kick off a software programming job.
It doesn't finish for a week.
Of course, it did a week, what would have taken a whole team a month to do.
And so big deal in productivity, but gigantic leap in computation requirements.
Let's just say that computing went up by 100 or 1000 times.
Meanwhile, because it's so useful,
the number of people using agents is now all over the place.
Every enterprise, our company, your company,
everybody's using agents all over the place to do software development, DevOps, SRE,
all of our CI, CD work, QA, testing.
The amount of software work that we do in the company now supported by agents is incredible.
One engineer, a really good engineer today is working with an agent,
but a really great engineer in the future is going to be orchestrating a whole bunch of agents,
who are going to be orchestrating a whole bunch of sub-agents to do work.
Well, between the amount of computation and the amount of demand use going up,
the product of that, that's our demand.
And so we've now arrived at the era of useful AI,
which is really exciting for all of us,
because until now, it's been novel, interesting, incredibly exciting.
But really, at the enterprise level, many of you have said before,
last couple of years, the impact of AI is, the potential is incredible,
but the actual use was minimal.
Now it's taken off.
And we're starting to see now, a couple of years ago,
we didn't have 5,000 enterprise customers.
Now we have the biggest companies in the world,
like you saw with Lilly and Samsung, Honeywell,
they're piling into this in a big way.
But really, that's just starting, right?
And to reimagine their workflows,
to be able to understand the trajectory of how this is improving
and how it can affect ultimately what their companies can become,
that really is just an idea.
We haven't seen that in any scale.
I mean, there are some companies that are doing it,
but it's a very small number today.
No doubt.
And well, you're experiencing and I'm experiencing,
our company has always gone fast, but it's gone.
Yeah, it's really going fast now.
And just the amount of content that's being created internally,
the progress that's being made.
You know, people said that AI is going to make us more productive.
There's no question about that.
What took months now takes weeks.
What took weeks now takes days.
And what takes days now takes hours.
And things that would take an hour,
you know, you and I pretty much expect it instantly now.
And so what has really changed is that our ambition has changed.
There's no question my ambition has changed.
You know, I wanted to be somebody to do something, make a contribution.
But that was the old Jensen.
You know, the new Jensen, I got big ambitions now.
You got it.
We all have to ask ourselves the question, how high is up?
Well, it's pretty high, right?
Yeah, yeah.
So we've got all these great products.
You know, we're embedding AI in everything and enabling,
you know, this distributed inference and intelligence.
This is the best room I've ever been in.
Which is your favorite?
Well, you know, it's hard to, I love all my kids.
How about the one on the end?
You got to love that one most though.
There's, yeah.
Well, this is incredible.
And so we, Michael and I have been working on a whole new line of computers to run agents.
The way to think about agents is this.
There's a large language model.
It's gigantic.
It's the most computationally intensive piece of software the world's ever known.
And that would run in that system towards the end.
That NVLink 72, the world's largest scale up single domain computer.
Okay.
It's just one giant system operating as one computer.
And that system has a large language model in it.
It could, you know, one terabyte, 10 terabyte of parameters, no problem.
And so that's one giant system.
That's the brain.
However, an agent starts with the harness.
That harness has to sit in a secure and governed container.
We call it a sandbox.
And so the NVIDIA open shell, open sourced sandbox is the security system that just about
everybody in the industry is using.
Inside, we have a reference harness we call Nemo Cloud.
Nemo Cloud runs on a CPU.
And that CPU could be a CPU in there, a CPU in here, a CPU in there.
It can also run Nemotron or any of the models, open source models that you would create your
own specialized agents for your own companies, trained for your own special domain of data
or skills.
And that would run locally if you like.
And then the large language models could run in the cloud.
And so you have this hybrid AI.
The thing that's really cool is that NVIDIA's architecture is the only architecture in the
world that runs every frontier AI model.
And lately, recently, in the last year or so, Anthropic has been really leaning into
the NVIDIA architecture.
So now we support every single frontier model.
We support every open source model.
And we support them in the cloud or locally.
As you can see, our computer is the first one in the world that runs in every cloud.
But it also runs locally.
And if you have your models that you're quite sensitive about, our systems are built with
confidential computing.
This way, you don't have to trust the operator that's operating the data center with your
secure data.
And so all of these architectures now run in every cloud, runs every model, runs hybrid
AI, and runs agentic systems.
And so you have your harness.
It runs on a CPU.
The CPU also uses tools.
And the CPU we created, I think you mentioned it earlier, called Vera.
Yes.
The CPUs of the past were built for hyperscale clouds.
And so you're renting the CPU cores.
And so you're optimizing for as many CPU cores as possible.
Well, agents in this new world, you're generating tokens.
You're not renting CPU cores anymore.
You're generating tokens and marketing that.
That's the economics of this AI era.
And so the AI wants to generate, run its work, and generate as many tokens as possible, as
quickly as possible, because that's the output of the intelligence.
Vera CPU has the highest single-threaded performance of any CPU in the world.
It has three times the memory bandwidth of the fastest CPU in the world.
And as a result, Starburst, DuckDB, all these databases run incredibly fast because the
agents are pounding on the databases, so the CPUs better be super fast.
The agents want to get through its work, so the CPU has got to be super fast.
Otherwise, that big machine down there is waiting for the agent to get its work done.
And so now you have the harness running on the systems here, local AI models running
on the systems here, and giant models running in the cloud or in your own data center on
the big machine down there.
Well, let's go check it out.
But as we do that, what's exciting is to think about, in the past, as humans, we would do
work, right?
And we would pass it on to the next human.
But now we have all these agents that we're managing, right?
And an individual can manage, I don't know, 1,000, 100 agents.
And so the possibilities that that unlocks in terms of human creativity and what humans
are going to be able to do.
And I love this idea of the unmetered intelligence, right?
Where you've got the power in your own PC, in your own data center, and you can use it
with your own data.
You know, that's sort of super cool.
That's a thing.
You don't have to struggle with token anxiety.
Yep, yeah.
You tell your employees.
Run out of tokens or get that dreaded bill.
But Jensen, since you're here, would you do us the honor of autographing this latest
plaque?
All right, well, we've got to climb up here.
I'm not as tall as you are.
All right, all right.
This is, what's the date today?
18th.
May 18th.
We're not going to sell this one.
This one is.
You got to sign it too.
I already signed it.
You did?
Yeah.
Well, you got it right here.
There you go.
Michael's already signed it right here.
And so this is, Michael, this is a hundred times larger than this one.
The station.
Right.
Exactly the same architecture the GB 300 GB 300 in here. That's a hundred times larger than this one
This is the only computer that side computer in the world that can run a terabyte a one trillion parameter AI model
Okay, so this is what it would have been unimaginable even a year or two ago. I know it
One trillion, I mean this this would basically be a cloud just two days ago
You know and so so this is this is a station and this this is a hundred times bigger
so yeah this one and that's five six times larger than this one and
This is one of my favorites
this is a
the smallest gb10
Same architecture is that one is that one? It's incredible right one architecture incredible. Well Jensen
We really appreciate you being here. We treasure the incredible partnership with NVIDIA and all we've been able to do
for customers around the world and
Look forward to and we grew up together practically. Yes, we did about 31 years. We've been doing this
Guys keep up the great work. Thank you everybody. Thank you Jensen by Dell
