Scaling the Agentic Era with NVIDIA Vera Rubin NVL72 on CoreWeave Cloud | Ihab Tarazi, Dell Technologies & Jacob Yundt, CoreWeave

Clips
More from Scaling the Agentic Era with NVIDIA Vera Rubin NVL72 on CoreWeave Cloud

Ihab Tarazi

CTO & Senior Vice President of AI, Compute, Networking and Security

Dell Technologies

Jacob Yundt

Senior Director of Compute Architecture

CoreWeave

John Furrier

Co-Founder & Co-CEO

SiliconANGLE Media, Inc.

play_circle_outline Dell–CoreWeave co-engineering partnership enabling first-to-market Vera Rubin NVL72

play_circle_outline L11 rack concept: the rack is the computer versus traditional L10 server-based design

play_circle_outline Integrating Power, Liquid Cooling, Thermal Management, and Observability for High-Density Systems and Customer Platforms

play_circle_outline Vera Rubin NVL72 Rack-Scale Debut at CoreWeave: Joint Engineering, Onsite Validation, and Rapid Iteration

play_circle_outline Performance gains: higher HBM, NVLink bandwidth, up to 10x tokens per second

Info
Transcript

Ihab Tarazi, Dell Technologies & Jacob Yundt, CoreWeave

Ihab Tarazi

CTO & Senior Vice President of AI, Compute, Networking and Security Dell Technologies

Jacob Yundt

Senior Director of Compute Architecture CoreWeave

John Furrier

Co-Founder & Co-CEO SiliconANGLE Media, Inc.

HOST

In this interview from Scaling the Agentic Era With NVIDIA Vera Rubin NVL72 on CoreWeave Cloud 2026, Ihab Tarazi, senior vice president and chief technology officer of AI, compute and networking at Dell Technologies, joins Jacob Yundt, senior director of compute architecture at CoreWeave, to talk with theCUBE's John Furrier about how deep engineering collaboration delivered the industry's first NVIDIA Vera Rubin NVL72 deployment. Tarazi traces the partnership's track record of firsts — GB200, GB300 and now Vera Rubin NVL72 — crediting more than a year of adva... Read more

explore Keep Exploring

What came together — what was the key factor or "magical moment" — that enabled you and your partner (NVIDIA) to co-engineer and move so quickly and successfully on projects like GB200, GB300 and the Vera Rubin system? add

How has the transition from L10 (individual-server) builds to L11 rack-scale (and emerging L12) architectures — particularly with NVL 72 products — changed your engineering approach, infrastructure planning, and software/hardware design? add

What engineering and design challenges arise when compressing a data center's compute into a single high-density rack, and how are those challenges being addressed? add

How were the GB200, GB300, and Vera Rubin systems validated with CoreWeave, how does that validation fit into Dell’s overall rack-scale approach, and what lessons were learned from that iterative engineering process? add

How does CoreWeave validate and integrate new hardware (e.g., GPUs) into its platform, and what benefits does that provide to customers? add

bolt Powered by CUBE AI

Ihab Tarazi, Dell Technologies & Jacob Yundt, CoreWeave

search

John Furrier

>> Welcome back, everyone, to Scaling the Agentic Era with CoreWeave Cloud. I'm John Furrier, your host, with the engineering discussion now with Ihab from Dell. Ihab, great to see you.

Ihab Tarazi

>> Thank you, John. Love to be here.

John Furrier

>> Senior Vice President Jacob from CoreWeave. You guys are the engineering view of this. Dell and CoreWeave shipped the first Vera Rubin NVL72. Congratulations.

Jacob Yundt

>> Thank you. Thank you.

John Furrier

>> We'll see some of that product later after this segment, but let's get into the relationship because this is a major milestone for the industry. Everyone was waiting for this when they saw the announcement at Nvidia GTC. Everyone's been waiting. It happened. Talk about the relationship because this came together with the partnership with Dell.

Jacob Yundt

>> In terms of the partnership, partnership is strong and demonstrated history of success. So first with VR NVL72, first with GB300, first with GB200. And we established this relationship with GB200 when we really started to co-engineer, work together, roll up our sleeves and really think of this as a true partnership. And you can see that success. You can see that these first after first after first. And Vera Rubin's just the most recent example of that. We've started planning this year in advance. I don't know.

Ihab Tarazi

>> At least a year in advance.

Jacob Yundt

>> At least a year in advance, so ...

John Furrier

>> Yeah. Talk about this. We've had many conversations on theCUBE and also in person at events around the evolution of the AI factories. We're now seeing the kind of scale. And it's been a progression. What came together in your opinion? What was the magical moment? What was the magical piece to pull together so fast?

Ihab Tarazi

>> First of all, I just want to comment on the partnership. We have an unbelievable engineer to engineer relationship. That's why this stuff happens. This is not about talking about concepts, but we get into very detailed technical specs and we work closely with our partner Nvidia and we do that way in advance. So I count Jacob as one of the most credible advisors, technical advisors out there and we do the same. So it does take that kind of detailed technical collaboration day in, day out by all our teams to get here. The feedback we get from partners like CoreWeave, which is CoreWeave is one of our top, if not the most advanced into engaging with us, that feedback goes into our product development process. That's why this is really valuable to us. So our AI factory we announced few years ago, it's the blueprint of how you serve all the capabilities from Nvidia and from other partners, but Nvidia obviously leads the world in this case in AI, both hardware and software. And we put out the first blueprint few years ago and then we continue to expand it with more chips, more software, more capabilities. I think the biggest thing about it is that it starts from the smallest building block of on server to one gigawatt data center, same architecture, same software, and you could just scale it.

John Furrier

>> Talk about the PowerEdge relationship with Vera Rubin because these rack-scale solutions are very well received. We're seeing that response in the numbers. You see it on the business side, but in the industry it's working. What's the role? How does that fit in for the audience?

Ihab Tarazi

>> Yeah. Those are X scale solutions when you get to the L11, the whole system design, that kind of integration of compute and network has a massive multiplier. So the compute itself is accelerating both memory and compute, but when you add the network integration with the NVLink, then we're talking about 10X the number of tokens from the generation before. I'm really excited to see what Jacob and team tests in the system in real life and what numbers they will get out.

John Furrier

>> Jacob, based on what he's saying, I mean, I love the co-engineering discussion, love that piece, but scale is a huge piece. How did you guys do it at scale? Take us through some of the mechanics, some of the day in the life. What's going on behind the scenes to make that happen at scale?

Jacob Yundt

>> Yeah, let's touch a little bit on L11 or rack-scale. So if you're not familiar, L11 is an industry standard term used to define completely integrated rack. So it has all the network, maybe storage, compute, networking, all in one rack. And previously, CoreWeave, probably like a lot of other clouds, really focused on L10 builds. So we really focused on individual server, making sure that those servers were connected well with backend fabrics, front end fabrics, and that operated well at sort of an individual server level. But the NVL 72 products have really changed the landscape. They sort of change how you approach engineering. And so now the rack is the computer. The rack has this high speed interconnect that's like ultra-high bandwidth ultra-low latency connection, all of the CPUs, all the GPUs are all in sort of like one rack-level package. And that influences then all of the rest of your design for scale as well because you're no longer thinking of things as just an individual server. I have to think of them now as racks and then that sort of has a kind of cascading effect of, okay, well now I have all these racks that are working together. I need to think of them as pods or I need to think of them as rows. Now I'm thinking of these as data halls or full data centers. So the evolution of L10 to L11 to I guess L12 influences both our software design or hardware design, just holistically how we approach building our platform is designed around this sort of increasing concept of scale as we grow what we think the computer is.

John Furrier

>> Yeah, this brings up the whole systems concept we talk about here over the years. This is a system. It's not a server. There's servers in the system. And they're not just racked and stacked. They're engineered, like he's mentioned. What are some of the engineering things that you guys do? When you ship a solution, it's already tested software, everything's validated, a word that's being used a lot. Take us through the engineering side of this. This is really important.

Ihab Tarazi

>> The building block here is what Jacob said that the rack is the new system now. So you're not buying servers anymore, you're buying those L11s. First thing we do that ... and we have deployed tens of thousands of GPUs at customers' sites overall, a tremendous amount of GPUs out there. First thing is that that system, while it looks easy, has thousands of components and dozens of pieces of software inside. So we have a very extensive process to build, manufacture, design, test every component. And then for somebody like CoreWeave, they have their own architecture and IP, so we work closely to make sure they do their parts as well. And then we end up at the system level we have on our factories we do the full test of the entire system, the power, the cooling, how it performs, the network, the software test turn up and all the security. By the time we finish that, that system is fully operational and then we ship it to the site where our customers like CoreWeave would do a whole bunch of other engineering and testing. But I would say the industry change, as Jacob was saying, once the compute goes that way, we do have now a full portfolio of L11s for networking all the way to fully liquid cooled storage, compute, and then how you design that, how you put them together in a system in a scalable unit. That's where the industry is. That's where all the key innovators in that space. And most people need tremendous amount of engineering expertise, resources to be able to make that happen.

John Furrier

>> Jacob, we hear this a lot. Engineering, the word engineer, engineering, they have great engineers. I've heard terms like that with CoreWeave mentioned. Go into the details of some of the engineering around what it's producing. Just give some stats. We've heard from other segments here in this run of show where it's like, oh, all these sensors, you got power and you got cooling as an ... It's like many innovation points just on the surface would be good, but they have to work together.

Jacob Yundt

>> Absolutely. Let's touch a little bit on the design we talked about of the rack. If you think about it, if the rack is the computer, we have essentially just compressed what was previously entire rows of compute in a data center into like one rack, and that creates a tremendously complex system, very complex. So the things that you need to engineer into that in terms of observability, management, monitoring, things that you would normally have sort of like across the whole data center, you now have in a rack and then still across the whole data center, but it's about making that whole system work together. It's more than just, "Do I have the appropriate power, the cooling, the networking?" Because if any one of those things is not firing on all cylinders, you have GPUs that are throttling, you have networks that are having link flaps or something, any number of things that can impact the overall performance of that system. So in terms of engineering, yes, every single part along the way, we're working with Dell, we're working with our internal teams. We are doing things to improve or build upon liquid cooling management, we're doing things to improve upon network monitoring, but it's really focused on engineering to improve that overall solution, not just one single dimension.

John Furrier

>> You guys were the first to validate the GB200 and now Vera Rubin. How does this fit into the overall Dell piece? Because now you have, again, all this complexity. So talk about that validation process and then talk about what you guys have learned because you didn't wake up one day and say, "Hey, we're going to do rack-scale." There's been a lot of iteration. You're smiling. A lot of scar tissue.

Ihab Tarazi

>> Absolutely.

John Furrier

>> But the results are significant. 10X token performance is going to be significant on the high end, but there's also other uses of other things and now this is one big system. Talk about the validation importance and then lessons learned.

Ihab Tarazi

>> Yeah. I can't emphasize enough how important it is. It's an engineer to engineer engagement. Jointly with CoreWeave, we have been first to market with GB200, GB300 and now VR. That's not coincidental and that doesn't happen just because somebody said, "Do it." It takes tremendous amount of that. But some of the things we put in place to make it happen is that not only do we design every single component, test it, validate it, but even in our factory, CoreWeave is there with us testing every component. The whole team was with us before we shipped it and onsite at CoreWeave, we have our top engineers and the best experts. So what happens to be first, you have to have a fully working system. There's always bugs. There's so many issues that come up in this process. So once you find something, you need to change, modify, optimize manifolds or firmware or software or network. Having the whole engineering team can re-design it on the fly, tell the rest of the team in Dell as well as in CoreWeave to respond, fix it, create a new version, test it, get it done, go to the next thing, go to the next thing, so the whole system can be working. That's what it looks like.

John Furrier

>> Jacob, from your standpoint, the CoreWeave cloud gets the benefits of this as you guys continue to grow. As the learnings progress, what's your big learning now that you're here with Vera Rubin at scale with all that validation, with all that engineering? Take us through the learnings and what's it enabling CoreWeave Cloud to do?

Jacob Yundt

>> That's a good question. Let's see, some of the learnings. Yeah, to Ihab's point, these are very complex systems and the validation process is pretty extreme. We spend a ton of time making sure that any materials that we get before it lands at a CoreWeave site have been tested, validated, are running optimally, so that when we present them to our customers, our customers are getting the best experience. So I really want to emphasize that tight engineering collaboration, because truly without that, we would not have been successful for getting to market first with these, but then also that would not translate to our customer success as well. So in terms of what are the outcomes for the customers, I mean, it's huge. If you have access to the latest and greatest that has oodles more HBM, oodles more bandwidth on NVLink, that allows you to train the next generation model, that allows you to have greater tokens per second. That is a huge capability. And so the value to our customer is that, A, we have access to these very early, but, B, we also have time to integrate those into our systems so that all of our software stack, all the mission control, all of the different utilities and observability and dashboards, everything that we present to our customers, we make sure then that that is fully integrated with the platform and our customers have the best experience.

John Furrier

>> You mentioned earlier about collapsing the data center down to a rack and what that enables and creates value. Yeah, this brings up the whole density. I mean, you're talking about dense is in. It's like you have a lot of stuff going on in the system. Density metric is going to become very important. How much can you squeeze out of the density on a performance? How do you think about that? Because that becomes the new thing because it's only going to get more systems are coming in. How do you make that density equation work?

Ihab Tarazi

>> Yeah. Density is important, not just from token economics. First of all, all the new models that really matter to people are trillion-parameter models. So they no longer fit. For the most part, if you want a full performance and you want some of the use cases, they're not going to fit on an 8-way GPU standard server. They really need those 72-GPU systems. Most of them need even more memory because you have this recursive nonstop agentic work that's starting and they need enhanced storage and enhanced security. So first of all, you do need that high density to get the outcome out of these new models that everybody talks about. But then now when you talk about token economics, you're saying, "How many agents can I at the same time? What's the concurrent inferencing?" And this is where the true value of everything we talked about comes through, because at the end, that's what customers are looking for. Being able to run the latest and best at full performance and get the maximum amount of throughput.

John Furrier

>> I love the whole, "The system is the new server." Because that really does speak to where the opportunity is and certainly the scale piece. Not to date myself, but I remember back in the PC server days, it was very easy. You build it, power it up, boots it up, it runs, you ship it.

Ihab Tarazi

>> Done.

John Furrier

>> And you load an OS on it. Okay, we'll bundle in Linux or Windows NT in the old days. So PCs were even easier. This is not the same. So you really can't say that the engineering involved on the QA side or the evaluation side. What goes into that piece of it? Because there's so much more. It is literally build-to-order because you need to ship it. And one of the benefits is you're shipping basically pre-certified, I don't know what you call it, pre-certified, pre-evaluated, what, I mean, there's probably a term for it, but it's going out as a high-end piece of system work.

Ihab Tarazi

>> So I'm going to give you an example of how important that is. One of our HPC customers is using one of the latest systems now and yesterday they told me out of one of the L11 racks, they can get for one set of software and precision, they can get like five petaflops, but they can go all the way to 60 petaflops with a mixed precision or something like . This is how different it is. It could be that massive of a difference of the number of tokens and the amount of performance based on what software you use. This is a very difficult thing to navigate. This is why I say CoreWeave does a fantastic job. What our customers say is that no matter what they need, they can quickly get on it, high stability, and they can get that highest performance. So this is not Linux where you just have-

John Furrier

>> It's complicated.

Ihab Tarazi

>> It's complicated.

John Furrier

>> Jacob, talk about from your standpoint, because this is why it highlights the engineer-to-engineer peer relationship, in everything from the beginning to shipping and supporting, what's the conversation like? "Yeah, we need more of this." I'm oversimplifying it, but explain the conversation.

Ihab Tarazi

>> .

Jacob Yundt

>> Yeah, it's actually very accurate. Extremely accurate. Like, "Flash, yeah, need more flash. flash." So one of the analogies I use, which is kind of laughable because I'm not a car guy, but is F1. For F1, you need a team. It's not just who has the fastest engine. Everyone's using maybe similar tires or the same tire, but in the end it is a combined effort between the driver, the car, engineers, the crew. Literally, everything has to come together. I think where CoreWeave's been very successful is partnering with Dell on this and that you can see that throughout all of this engagement, whether it's Dell's strength in terms of manufacturing, supply chain and logistics, whether it's their L11 integration capabilities, or that direct engineer to engineer feedback, we're like, "They're great about that." We pester them all the time, we're like, "Hey, we think we see this. Well, what's going on?" And then we get our engineers together, we lock ourselves in a room, we solve tough problems. I don't think either of us would be sort of where we are today without that really tight collaboration.

John Furrier

>> Guys, talk about the relationship and specifically the pressure and the speed game of getting the first Vera Rubin NVL72 at CoreWeave, getting that up and running. Why you guys burn the midnight oil? What's the pressure? What's the reason why you hustle so hard and engineer heavily to get this out?

Jacob Yundt

>> Hustle is accurate. Not pictured. We have an army of engineers with many sleepless nights doing everything from following every single component to make sure it gets assembled and we are tracking everything to installing, burning in, validating, configuring. And everyone is just truly laser-focused with just singular goal of getting this thing online. And the reason we do that, obviously we want to be speed of light. Our customers want it. We want to get it out there as fast as we can. But that has a meaningful impact for our ability to be prepared to scale this product. So by the time that we have a bunch of these, now we're able to deploy hundreds, thousands, tens of thousands for our customers. And at that point it is a very stable, well-thought-out, well-engineered platform that our customers can use for anything. And by integrating that as early as possible, we're able to get as much runtime as possible. I'll use the F1 analogy for a minute again. If we just had our car out for the first time on race day, that would be bad. We're putting in a lot of laps. We're putting in a lot of wind tunnel time, a lot of sim time, right? We're spending lots of time and energy so that when these racks hit the ground, we're off and running and we know how to get this out there at scale.

John Furrier

>> Yeah. It's a competitive advantage too for you guys. I mean Ihab, this is a great analogy because this is not like yesterday's validation. Oh, it's running. I mean the importance and the consequences of getting it right for CoreWeave is, you can quantify it. He just mentioned we get a headstart on the competition. They have a peaked car on race day. That's basically day one. That's a differentiator for CoreWeave while everyone's just test-driving.

Ihab Tarazi

>> Yeah. I think that the demand for compute in the industry right now is tremendous. If you take compute, it translates to revenue for people instantly. So we have tremendous amount of demand in customers and these customers are waiting for this compute to show up. And obviously as soon as CoreWeave can turn it up at high performance and with the user experience, they can take advantage of it. So this is not about just an engineering milestone. This is a tremendous requirement in the industry.

John Furrier

>> The payback is significant.

Ihab Tarazi

>> Significant for our customers and they're really pushing for it.

John Furrier

>> I mean, if you think about the old way of doing things, everyone's waiting for the new system, they're bated-breath, "Oh, it's the new system." They get it and they just pound on it, pressure-test it, and go. You're ahead of the game and everyone's like, "Okay, I got to validate it." That takes time.

Jacob Yundt

>> It does.

John Furrier

>> Or maybe one workload, how can I figure out all the different use cases? You have a lot of customers, but CoreWeave's not like one thing.

Jacob Yundt

>> Exactly.

John Furrier

>> And what does that mean for you? How does that translate into the burning the midnight oil? You have to pressure-test it all?

Jacob Yundt

>> I think it just means we want to maximize our runtime as early as possible. We want to get the engineers to hammer on these systems and give them truly as much time so that when we let these loose on our customers, we give them access to this capacity, we know the ins and outs of everything. We know this car like the back of our hand, we know how it handles every single corner. We're good. There's no questions about this product.

Ihab Tarazi

>> Speed to market is really important in this case. There's tremendous amount of the industry and we talked about the big bumper performance and the additional tokens people can get out of this system and the higher performance. So there's a tremendous amount of customers who want it as soon as possible. So we have to work at speed of light jointly to make it happen. Not only we want to validate and test the first system, but we have to do many things in parallel so they can use it immediately as soon as we have it. The good thing about working with CoreWeave and the help that Jacob provides to us and his team is that we get tremendous amount of engineering feedback that we put back into our development cycle and innovation that helps us accelerate our product development. The amount of innovation we have put into our system since we've been working and collaborating with CoreWeave over multiple systems now has been tremendous. And you could see it in the industry, you could see how our success with these at-scale systems, data center design. So I really appreciate this kind of collaboration. It delivers tremendous value to us, we have fun with it, and at the same time, it does add value to us and our customers.

John Furrier

>> It speaks to the engineering mindset and this shift in the computing paradigm. This compounds for you guys. It's not like one and done, see ya next generation. Is an always-on compounding interest for both of you guys.

Jacob Yundt

>> Absolutely. Yeah. I think Ihab's underselling the iteration aspect of this. One of the things for this tight engineering collaboration speed is that we can iterate so much faster, I feel like, than we ... maybe don't want to say faster than we have previously, but just you see that even within the lifecycle of one product or the life cycle for the next product, that iteration loop, we really, really pull that in when we focus on first-to-market, time-to-market.

John Furrier

>> Well, congratulations. It's a major milestone. The Vera Rubin NVL72 is highly coveted and I won't say a race , but people want it as fast as possible. You guys got it out first. Congratulations.

Ihab Tarazi

>> Thank you.

John Furrier

>> I guess my final question for you is, you've accomplished a lot together. Does this signal the future of the AI infrastructure paradigm? Jeff Clarke and Michael Dell, and Jeff Clarke, particularly on theCUBE at Dell Tech World said it's an inverted market. Everything's changing. Is the computing paradigm change, what does this tell us about what's coming? You both have a peak into the future. Both public companies, so you can't tell me the specifics, but give us a taste of what this is signaling. From an AI infrastructure standpoint, what can we expect?

Jacob Yundt

>> Do you want to start? Do you want me to start?

Ihab Tarazi

>> I'll start. I think we said a little bit, both Jacob and I before, that this is the new system design. So I think between both companies we've done unbelievable innovation on the compute side and I have no doubt we'll continue to lead with the relationship and culture we have and brain power. But I think the opportunity now is collapsing the entire system, compute network storage. It is still three different architectures, which is an IT topic, not an AI inferencing. If you were to do AI inferencing, you wouldn't do it this way if you were starting with a white page. So we are going to continue to innovate in that area and there are different opportunities to continue to optimize and figure out how you can get more performance, more tokens per dollar, and continue to innovate. My view.

John Furrier

>> Well, that's a good point. I mean, you have the collapsing ... What you're heading at is not just the speeds and feeds and the technology of the systems, it's an economic disruption. And that's where the tokens speak to that. So awesome. Jacob, what's your take on the signaling of this? I mean, you guys are leading the pack at CoreWeave. What's it say about the direction?

Jacob Yundt

>> I think the direction is, sounds cliche, but scale, but scale also in a sense that we talked about the densification. At a certain point, we're sort of fighting the laws of physics. I keep trying to cram more stuff in. I need things to be as close as possible. I need to be as fast as possible. They're using a lot of power. They're using a lot of cooling. And so I think it's a combination of we're seeing this unprecedented scale at the same time we're trying to collapse everything to make it as dense as possible. And that has some fun, unique opportunities and challenges. I mentioned that we've moved from this L10 to L11 model. From my point of view, it's no longer like, "Hey, how am I dealing with these rack-scale devices?" It's, "How am I dealing with, again, rows, pods, whole data halls, whole data centers. How does that look like or how does that impact our power delivery? How does that impact all of our cooling?" So I think that's sort of where we're going with this now, that the system is going to continue to grow while at the same time weirdly shrinking because it's getting more dense. And I think that's where we've got some fun, unique opportunities.

John Furrier

>> It's a connected distributed competing network with super computers hanging off of it.

Jacob Yundt

>> Yeah.

Ihab Tarazi

>> And the software is changing dramatically, which basically will impact all of it. So ...

John Furrier

>> Guys, thank you so much for taking the time to unpack and congratulations on a major milestone of Vera Rubin NVL72, first to ship and get and validate. Congratulations. I know a lot of hard work went in, a lot more engineering to come, and the future's bright. AI continues to boom along and compute architecture is a coveted role. I'm sure you're very busy, Jacob.

Jacob Yundt

>> Yes. Yes.

John Furrier

>> And, yeah, it's always great to see you.

Ihab Tarazi

>> We have so much fun. Believe me, we really have a good-

John Furrier

>> It's the best time to be an engineer right now.

Jacob Yundt

>> It is.

Ihab Tarazi

>> Yeah.

Jacob Yundt

>> It's great.

John Furrier

>> There's plenty of hard problems to solve.

Jacob Yundt

>> Yes.

John Furrier

>> Well, thanks guys. I really appreciate it. I'm John Furrier, here in theCUBE in Palo Alto. Now to Harsh at the data center to take a look at the system at the CoreWeave facility.