10:02 Webinar
NVIDIA and Crusoe on AI at Scale
NVIDIA, Crusoe, and Everpure discuss what it takes to build and operate AI infrastructure at scale.
This webinar first aired on June 17, 2026
Click to View Transcript
Kevin, Omar, thanks for joining us today. Kevin, I think you got the loudest applause of everybody, so we're gonna have to bring you back every year. all right, well, let's, let's just dive right in. Kevin, you and I, we've been working together for a couple years, in and around AI.
would love you to share with the audience, what you see as, some of the biggest changes that have happened over the last couple of years, and specifically the impacts on infrastructure in the AI space. Yeah. I think what we've seen is the transition from single shot AI, where it just answered a question, to this agentic reasoning.
So we've heard a lot about that. I liken this to AI eating its own tail. So we have AI that's generating a response, and it's feeding it back to itself. It's doing tool calls. So all of those applications, I love the way you've unified all of the different data with
the data stream platform, making the data AI ready, and the networking plays a huge part in that. Been a great partnership working with you to make the infrastructure so that we can get faster time to first token, better performance, better efficiency out of the infrastructure. Absolutely. Omar, as, you know, as a, a leader in the
infrastructure space, building out AI, how does this jive with what you're seeing in, in your customer base? Yeah, I mean, I think what strikes me the most is that AI is not just a software problem anymore. It's a full stack problem, right? It starts from the frameworks that you're using, whether that's TensorFlow or PyTorch,
goes all the way down the stack, right? What version of CUDA you're using, the drivers, and eventually all the way down to the hardware. and what's really interesting for me, you know, I've been working in the cloud and building on cloud for several years, so for me, hardware is cool again.
Um- Hardware was always cool. Hard- hardware was always cool. People just didn't realize that. Yeah. I, I mean, well, from my perspective, it was always abstracted away from me, and we got to operate at a lot higher layer.
but now, you know, the, the hardware is really driving what AI is able to do. and all of those innovations are really important because AI is not gonna move forward without all of these hardware innovations that's happening across the GPU stack, the network stack, the storage stack, and all of that needs to play in concert with all of the software that's being deployed.
so I'm really, like, just seeing this, like, really interesting co-evolution of the software and the hardware, you know, kind of happening si-simultaneously, and it's really awesome. Yeah, and if hardware is cool, storage is really cool. Context is so important. We heard a lot about context, and it's really important, so storage is super important.
So, we've talked a lot of tech, got a lot of building blocks and pieces. Now, at the same time, Sean was up here a couple minutes ago, shared some pretty scary sounding statistics, right? Basically saying, "Hey, most of y'all in the room, there's a pretty big gap between what you wanna do with AI and, and kind of, reality and, and bring that to bear." so Omar, you
know, as somebody who's working with leading customers in the space, why is this so hard? What's, what are people running into? Yeah, I mean, I think there's three main factors that a lot of our customers are asking us for. it's all about reliability, performance, and operational excellence.
I mean, reliability is pretty obvious, right? Customers are making huge capital investments. They wanna make sure that that infrastructure is always operating. performance is a really interesting perspective, right? Like, if you think about an eight hundred GPU cluster of Blackwells, that's about a quarter
of a petabyte of HBME memory. if it's taking you twenty, thirty minutes to load that memory, load your model weights, load your training data, you're sitting around waiting for, for bytes to fly around, right? And so it's really about efficient data movement across of that instru- infrastructure. and that's why, that's why, you know, storage and network are equally as important as the
GPUs, right? This performance needs to be reliable 'cause you don't wanna have, you know, mi- hundreds of millions of dollars of GPUs sitting there idle. so I'm just seeing this, like, renaissance in not just the compute space, but the storage and the networking space, and our customers are asking us to really operate that for them
on their behalf. the second part of that is operational excellence. You know, I kinda tied abou- talked about, like, the, the deep coupling between hardware and software. Anytime you make a, a seemingly small change somewhere in the stack, that could have massive implications elsewhere.
So, you know, being really good at operations and being able to run that on behalf of our customers, is something that we really pride ourselves on. So, okay, so we've heard context, ontologies, CUDA drivers. there's a lot of deep tech here. Kevin, I think this, goes to the heart of what you guys are trying to boil down, but we're
working with you ar-around NVIDIA's AI data platform, to really make this, more easily, accessible and deployable. you know, how, you know, how did that come about and, and, how are you guys leaning into AI data platform to solve some of these challenges? Yeah. I think there was a realization several years
ago, and we talked with you about that and the team at Everpure and said that we need all of the data to be pre-ingested AI ready. So AI ready data, this notion of context. And as we're starting, you know, you were talking, Omar was talking about training. It's even more important for inferencing that we actually have the right data at the right
time that's delivered reliably and efficiently. Because with an agent, we're doing a harness around that. We have an agentic loop, and the AI is generating a response that's actually maybe doing a tool call and then coming back in. And so to get fast response, we need very quick access to data.
We work with you not just on, file systems, but on objects that actually curates that context. We can run it over what's called RDMA. Super important. You're a great partner to be able to build this, deliver the-
Networking is cool again, so. Networking is cool again. So it's funny. That was, you know, it's, thirteen years ago that we first heard that software is eating the world. It turns out software needs to run on something, and all of the compute and the
storage is vitally important, so maybe hardware's eating the world again. All right. so maybe, maybe switching gears, you know, it turns out infrastructure's really important, whether it's, moving data really fast into the GPUs, networking, you know, storage we've talked about. so Omar, you know, as an infrastructure leader in this space, what do you think about, what
advice would you give, you know, our audience, in terms of, evaluating vendors? what do you look for in technology choices? what led you to Everpure? Yeah, I mean, it was, you know, as we explored this partnership, it was a lot more than just the technology itself.
So I think there's three, three things that really drove us towards Everpure. one is, a really strong reputation for reliability, right? You've been serving the enterprise segment for several years, and very good ratings on that and, and the feedback has been amazing there. two is a really important piece for me is, empathy for the service provider.
so at Crusoe, we're not buying like an appliance and serving our internal needs. We're buying multiple appliances, petabytes of data to serve the needs of multiple customers, right? We're running multi-tenant infrastructure. There's variable performance and variable read and write capabilities
that our customers require. and so really, you know, I think Everpure understood what that means and the challenges that come along, serving multiple customers across, single fleets of devices. and then the last thing I would mention is, just strong understanding of the supply chain. I think both from, you know, being able to get us the capacity, but really understanding, all
of the different components that go into building an AI system, and making it very easy to deliver that for our customers. I mean, tho-those were like the kind of three shining things that, you know, helped us move forward with this partnership. Awesome. Awesome. So maybe to bring us home, lightning round, fifteen seconds, each.
a lot changing in the ecosystem. folks wanna do more with AI. Technology's changing super fast. Hard to predict the future. What advice would you leave our audience with, in terms of how to get started, how to get
successful, how to make these projects a reality? Yeah, so I think what's really important here is to get started today and that you rely on a partner that has all of the attributes that you talked about, that, reliability and the empathy and the supply chain operational excellence, but also a partner that's very nimble because this AI world is changing so fast that you have to adapt.
And what I love about the new Datastream platform is that it's built all of this compute and the ability to do the governance and the provenance and prepare it. We were talking earlier about ontology, the way we're going to organize data so that it's all part of a whole, but you're looking at the pieces with graph networks. I think this is a great engineering team that's involved and understands this, you know,
starting with Charlie, who's a deeply technical CEO and all the way down to the, the management team here. I think that's a great partnership that you wanna work with and be responsive to the customer needs. Omar- Awesome any, any closing thoughts to leave the audience with? Yeah. I mean, I, I mean, just to add on to that, I mean, I think we've seen a lot of technology
evolutions over the past several years from virtualization to cloud to containers, now AI. And so I think there's a significant step function of operating AI clusters and AI infrastructure. So I think, you know, we should get really good at that, as a community. and then lastly, you know, I think, just echoing a lot of the statements that have been
said today, I think, your data and your processes and your internal knowledge are your most valuable asset. I don't see a lot of enterprises going out and saying, "I want to build the next frontier model." They want to unlock the intelligence that they already have and leverage the capabilities that they, they already do.
So I would encourage you to look at, of ways of capitalizing on those assets. Agreed. Agreed. Awesome. So we heard AI-ready data. We heard context understanding is super important, infrastructure, super important, reliability, but also flexibility in this age.
Everpure, NVIDIA, and Crusoe on AI infrastructure at scale, from performance and data readiness to the real-world demands of building for next-generation workloads.
Personalize for Me
My Updates
No updates at this time.