Nvidia CEO 1-on-1: Software Like OpenAI's Sora Will Drive Demand

I spoke with Nvidia CEO Jensen Huang minutes after the earnings call on Wednesday, with the stock up 9% after hours on an all-around beat and strong guidance. I wanted to get beyond the quarterly numbers and understand Nvidia's strategy for continuing to grow in the inferencing stage of AI, and using software to build a moat. Jensen delivered.

Here's a transcript:

Jon

Well, you just got done with the call. You got beats on the top and bottom lines. Strong guide led, of course, by the data center. And you got GTC, which has turned into an A.I. conference coming up in about four weeks. So I want to pick up where the call left off and get more color and detail on where you see things going forward.

So start off, where do you expect consumer applications, and I guess to some degree professional applications, like OpenAI’s Sora, to affect demand for your technology from here?

Jensen

Well, Sora is just an extraordinary, extraordinary breakthrough. The ability to go from text to video is used by Runway. You might know about this video editing, video creation service. Incredible company. By automotive simulation, autonomous vehicle simulation systems like Wave, like what Nvidia uses. And the ability to go from words and scenarios to describe videos that are generated — these models all require enormous, enormous infrastructure to train. Because you're training not from words, but you're training by watching video. Literally. You're going to see a lot more modalities of these. You're going to see text with images, with sound, with video, all being trained at the same time. And so today, notice most of it most of the videos are silent.

But in the future, the words and the video will be registered perfectly. And not only will you generate video, you will generate associated and appropriate sounds that go along with it and vice versa. And so you're going to see very large models, multi modality models being able to be trained on a large amount of content that will scale over the next several years.

And so this will of course make a very big difference in the way games, content, movies, consumer oriented content, digital media content of all different types. For enthusiasts, for consumers, it's going to make a pretty big impact. However, some of the things that I'm most excited about is using the fundamental technology we just described in industrial robotics, industrial automation. In how we simulate proteins, how we simulate the weather.

You're going to see many of these type of examples at GTC. And it's about moving beyond words, moving beyond text into all of the world's modalities for generative AI. Big opportunities.

Jon

Wow. Okay. Well, so now alongside that and I, of course, connected with you recently at ServiceNow’s gathering back in April. And on the earnings call, I believe you mentioned how their Vancouver rollout and other enterprise software rollouts drove demand. So unpack how the partnerships and investments that you've been making over the past several quarters are going to affect your pipeline.

Jensen

Really exciting. I really appreciate you asking it. ServiceNow is just a great example.

What's going to happen is, this. The world's enterprise software platforms represent approximately a trillion dollars. That trillion dollars represents rules platforms like ServiceNow. It represents data platforms like Snowflake, Dropbox, Box, SAP, Oracle. These application oriented, tools oriented platforms and data oriented platforms are all going to be revolutionized with these AI agents that sit on top of it. And the way to think about that is very simple.

Whereas these platforms used to be tools that experts would learn to use, in the future these tools companies will also offer AI agents that you can hire to help you use these tools, or to help you reduce the barrier of using these tools. And so imagine instead of a million I.T. professionals that know how to use ServiceNow around the world, I'm making up the number, Bill McDermott would know better. Millions of I.T. professionals that use ServiceNow. In the future they'll rent, on top of ServiceNow, billions of digital I.T. professionals to augment the millions. So that it will be easier to use ServiceNow. So that ServiceNow will be able to do more things. And so this is the new opportunity for the trillion-dollar enterprise software platform companies.

I'm super excited. I think they're all sitting on gold mines, frankly. You're going to see all kinds of incredible AI agents coming from ServiceNows, and let's just hypothesize. I'm sure Cadence will have them. Ansys will have them. Autodesk will have them. You’ll see, Adobe will have them. These are all AI agents that you can hire that will come and perform tasks for you. And we’ll hire, you know, and in the case of many companies, you’ll hire hundreds of thousands of them in the background doing tasks for you.

Jon

Now, one of the most interesting parts of the call to me was your response to a question about how your platform adapts and how you'll fare as demand in the market shifts from developing the models to using them from from training to inferencing. I believe you pointed out that Nvidia's platform is programmable. So unpack for me how practically that plays out.

Jensen

The fundamental difference between accelerator and accelerated computing is accelerated computing is programable. Accelerators. They aren’t configurable. They're programable in a very, very specific set of algorithms. For example, an accelerator is like a DVD decoder. That's an accelerator. An Ethernet chip is an accelerator. An Ethernet chip never becomes a DVD decoder and a DVD decoder never becomes a GPU.

And our GPUs, because of CUDA, is both an acceleration platform for many domains of applications from image processing to particle physics to, you know, quantum. You've probably seen that we announced recently some work that we've been doing in quantum. We're doing incredible work on quantum computing as a quantum emulator. All the way to an intelligence emulator we call artificial intelligence. And so our platform is programable. And also because we have the discipline of ensuring our entire installed base of GPUs are CUDA compatible. Every new algorithm that is developed on any of our GPUs will run on all of our GPUs. And that gives the developers reach. While it gives the new developers the capability to solve problems they couldn’t solve before.

And so simultaneously, we can solve new algorithms like transformers that came along, and multi modality transformers and new algorithms like SSMs and all kinds of buzzwords that most people haven't heard about. We have 100% confidence. They all work on CUDA. We have 100% confidence they all work on Nvidia GPUs. And once they work on Nvidia GPUs, because our installed base is so large, millions of GPUs in the cloud, we're in every single datacenter everywhere. We can therefore take that innovation and benefit everybody. Maybe the performance is not as fast as our latest generation, but it works on everybody’s. That is just a wonderful benefit. And you were asking about inference. The goal of inference is application reach. The goal of inference is application reach. It’s no different than the goal of writing an application for a mobile device is application reach. You prefer a phone with the largest installed base. It is fundamentally the reason why Apple's so successful. In our case, inference: Anybody who's developing an application to run inference is going to prefer NVIDIA first. And the reason for that is because the CUDA installed base is so large, it is the only acceleration platform that has a giant installed base, and is in every single cloud.

It is growing incredibly fast. And it’s available on prem or in the cloud. It's available all the way out to the edge for robotic systems. And so this architecture being so pervasive, and because we have the discipline of protecting it and maintaining it for 30 years, this architecture is now literally everywhere. And if you are somebody who is developing an application for inference, targeting Nvidia gives you the largest possible reach.

Jon

Just like iPhones. So that leads me — you mentioned Apple. It's funny. That leads me to my last question, which is Nvidia AI Enterprise. I believe you mentioned that you're charging by the GPU per year and this is about management, optimization and patching. It's really a software play. Now, I believe you said at a billion-dollar run rate. This reminds me — and it's growing quickly.

It reminds you of when Apple had to get the market used to the idea of not just looking at iPhone unit sales as a number that matters, but also at the services business. Are we entering into that phase where people who want to understand your business are going to have to look at Nvidia AI Enterprise and that services component, right? Of doing that software based management, optimization and patching for the customers who are now mostly just trying to find chips they can buy from you?

Jensen

Yeah, Jon, that’s spot on. I think at the core of it, accelerated computing is not just about the chip, it's about the software. You can't accelerate something with just the chip because if you could just be called chips, you know. Just build chips to make the computers go faster. Accelerated computing requires a full stack. Whenever we do something in quantum chemistry, it requires a quantum chemistry stack.

Whenever we do something in weather simulation it needs a weather simulation stack. Whenever we do something for AI training, it requires an AI training stack. Each one of these domains of applications, whenever we do something with SQL, we accelerate data processing, like nobody accelerates data processing. And that requires a very specialized stack. And so each one of these stacks — robotics stacks, AV stacks — each one of these tasks require a great deal of engineering.

In the case of CSPs, they can manage it themselves, they can patch it themselves, they could optimize it for themselves along with us. But for most of the enterprises and enterprise software companies, they simply don't have the large and deep expertise in accelerated computing at this point. And so we'll do it for everybody. We’ll optimize it for everybody.

We'll create these stacks for everybody and with everybody, and we'll make it run in every single cloud for everybody. And the way that we monetize it is through this engine, this Nvidia AI Enterprise engine, which is essentially an operating system for Nvidia’s AIs, Nvidia’s enterprise and acceleration algorithms. And you pay for it per GPU per year, just like an operating system.

And you can run everything that NVIDIA creates and enables. And so that, as we grow into enterprise, as we grow into enterprise software, as we grow out to the edge, this is going to be a very, very significant opportunity for us.

Jon

Well, I could talk about this for a lot longer, but I know you've got other stuff to do, so Jensen, I appreciate the time.

Jensen

Thank you, Jon. It's great to talk to you. I see you. I'll see you at GTC, I hope.