Nvidia’s Cloud Strategy Evolves, Factory-Cognition Drama, Instinct Pushes Into Shopping | Diet TBPN

John Coogan (0:02)

It was just like last year that Nvidia was like, we're going to layer Coreweave, or we're gonna layer all the Neo clouds, we're gonna be the front door to compute with DGX Cloud Lepton. I don't know if you remember any of this, but this started back in like 2023, post the Chachapiti moment, Jensen comes out on stage at one of the keynotes he gives, and he says like, this is the iPhone moment, and we're going to aggregate all of the underlying compute, and you're gonna be able to come to us and deploy whatever resources you are, and we're gonna compete with like AWS and Azure and GCP, and we're gonna be a one-stop shop for this stuff, and it's gonna be really important. And at the time, I feel like people were sort of like, whoa, okay, this is bold, but this could be really, really powerful because they have this odd tension between the Neo clouds, where if you go in, you lose your connection to your customer effectively, and the margins could get compressed, you're in a much more competitive environment. And at the same time, if you don't go in, then Nvidia can just say, hey, like all the DGX Cloud leptin customers, well, we're gonna send them to the clouds that did sign up. And so you have this like tension. They really did, really did, it really did feel like Nvidia had some leverage, biggest company in the world very shortly after, and obviously a very aggressive business, like they're not just gonna be like, oh, that would be improper to compete. But it played out very differently, and Nvidia has become very friendly with CoreWeave, and they both found their places where there's still little competition here and there, but I was just reflecting on like, how that journey came together, where that landed. And the long story short is that like, the DGX product, it came out, and the reviews weren't like, okay, yeah, this is actually better than working with CoreWeave. And there was still so many benefits from going to a Neo Cloud or even a true cloud provider. There was just more to do.

Nvidia made an acquisition. They bought Lepton, where the name came from, in April of 2025, which sort of kicked most of that discourse off. The rumor was that they paid between $300 and $900 million for that.

And Jensen was like so gassed up on this. I loved his quote. He said, NVIDIA DJX Cloud Lepton connects our network of global GPU cloud providers with AI developers, together with our NCPs, which is NVIDIA Cloud Providers, we're building a planetary scale AI factory. That's what his tagline was, which makes sense. There is a world where that would have played out. That's not what happened.

Jordi Hays (2:41)

Give me one second to talk about how we have failed so far. To convince a lot of our friends not to call supercomputers AI factories.

John Coogan (2:49)

Yeah.

Jordi Hays (2:50)

Right?

John Coogan (2:50)

If we continue to fail, they're going to be like, it's a refinery.

It's an AI offshore decker. It's an oil decker for AI.

Jordi Hays (3:01)

Gabe says it's a data ranch, brother.

John Coogan (3:03)

Data ranch. There we go.

Jordi Hays (3:04)

You're putting a data ranch in your backyard.

John Coogan (3:07)

But yeah, I mean, at the time, there was a bunch of reporting of people familiar with the matter. The information got a scoop from a bunch of the NeoCloud compute providers saying Leptun was a potential competitive threat.

And if you refuse to participate, that creates a risk that Nvidia could direct workloads toward competitors. It was a bit of this tense time, but the end result was more balanced. So instead of layering their customers and trying to control aggregate demand, which would be really powerful if it happened, the product evolved into this unified AI platform, which is their words. Basically, it's between an actual Expedia for GPUs and just a set of standards. So it's a bunch of proprietary software. It is what makes up the CUDA moat. So everything from handling GPU node groups, interactive dev pods, distributed batch jobs, inference endpoints, storage, observability, reservations more. So all the things that you need to actually schedule, manage GPUs, you can run that, but you can still use your own hardware, do whatever you need to have the relationship with the underlying owner of the GPUs. And so the rest of the Neo Clouds are generally pretty safe.

And so it was interesting because it was heralded as this like, oh, okay, things could get spicy. You could see all the Neo Clouds really chirping at Jensen, Jensen really coming for the throat. Didn't really play out that way, but it does sort of go to show that Jensen does not take lightly to any whispers that the Cuda moat is going away anytime soon. Like the Ace is still up the sleeve, in my opinion. Like they still have to do a lot of work because when SemiAnalysis actually demoed it, they were like, it's not working as well as some of the other systems. Like if you were to put, I don't even know if Nvidia Cloud Lepton is on Cluster Max, but they would, at the time that they tested it, they were like, yeah, you're gonna have a better experience if you just go direct core weave Nebius, the other ones that are in the platinum tier.