Zeus Kerravala, founder and principal analyst at ZK Research, joins Dave Vellante, co-founder and co-CEO of SiliconANGLE Media, in an insightful dialogue at the New York Stock Exchange. This session, part of NYSE Wired and theCUBE's coverage of "AI Factories - Data Centers of the Future," delves into the critical role of networking infrastructure as artificial intelligence reshapes data center architectures.
In this engaging discussion, Kerravala, esteemed for deep expertise in networking, shares insights with theCUBE Research hosts. Kerravala emphasizes the increasing importance of networking in AI, illustrating how networks facilitate communication between Graphics Processing Units and enable expansion across data centers. They also provide a nuanced comparison of Ethernet and InfiniBand technologies, unpacking how both continue to be crucial in meeting evolving AI infrastructure demands.
The conversation further explores the potential disruptions in the networking industry as companies such as NVIDIA and Cisco form strategic partnerships, aiming to innovate and redefine AI-ready network solutions. According to Kerravala, enterprises must rethink network architecture to accommodate AI's expanding role, leveraging validated solutions to ensure effective deployment. They highlight how AI impacts data security, advocating for a shift toward more software-based solutions to enhance protection.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: The AI Factory - Data Center of the Future. If you don’t think you received an email check your
spam folder.
Sign in to AI Factories - Data Centers of the Future.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open the link to automatically sign into the site.
Register for AI Factories - Data Centers of the Future
Please fill out the information below. You will receive an email with a verification link confirming your registration. Click the link to automatically sign into the site.
You’re almost there!
We just sent you a verification email. Please click the verification button in the email. Once your email address is verified, you will have full access to all event content for AI Factories - Data Centers of the Future.
I want my badge and interests to be visible to all attendees.
Checking this box will display your presense on the attendees list, view your profile and allow other attendees to contact you via 1-1 chat. Read the Privacy Policy. At any time, you can choose to disable this preference.
Select your Interests!
add
Upload your photo
Uploading..
OR
Connect via Twitter
Connect via Linkedin
EDIT PASSWORD
Share
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: The AI Factory - Data Center of the Future. If you don’t think you received an email check your
spam folder.
Sign in to AI Factories - Data Centers of the Future.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open the link to automatically sign into the site.
Sign in to gain access to theCUBE + NYSE Wired: The AI Factory - Data Center of the Future
Please sign in with LinkedIn to continue to theCUBE + NYSE Wired: The AI Factory - Data Center of the Future. Signing in with LinkedIn ensures a professional environment.
Are you sure you want to remove access rights for this user?
Details
Manage Access
email address
Community Invitation
Zeus Kerravala, ZK Research
Zeus Kerravala, founder and principal analyst at ZK Research, joins Dave Vellante, co-founder and co-CEO of SiliconANGLE Media, in an insightful dialogue at the New York Stock Exchange. This session, part of NYSE Wired and theCUBE's coverage of "AI Factories - Data Centers of the Future," delves into the critical role of networking infrastructure as artificial intelligence reshapes data center architectures.
In this engaging discussion, Kerravala, esteemed for deep expertise in networking, shares insights with theCUBE Research hosts. Kerravala emphasizes the increasing importance of networking in AI, illustrating how networks facilitate communication between Graphics Processing Units and enable expansion across data centers. They also provide a nuanced comparison of Ethernet and InfiniBand technologies, unpacking how both continue to be crucial in meeting evolving AI infrastructure demands.
The conversation further explores the potential disruptions in the networking industry as companies such as NVIDIA and Cisco form strategic partnerships, aiming to innovate and redefine AI-ready network solutions. According to Kerravala, enterprises must rethink network architecture to accommodate AI's expanding role, leveraging validated solutions to ensure effective deployment. They highlight how AI impacts data security, advocating for a shift toward more software-based solutions to enhance protection.
In this theCUBE + NYSE Wired segment from the AI Factories – Data Centers of the Future series, Zeus Kerravala, principal and chief analyst at ZK Research, joins theCUBE’s Dave Vellante to unpack why networking is the heartbeat of AI factories. Kerravala explains how AI has redefined the “unit of compute” from a single server to racks, data centers and now multi–data center fabrics – making the network the connective tissue for GPU clusters and edge data access. He breaks down scale-up (NVLink), scale-out and scale-across patterns; why both InfiniBand and Eth...Read more
exploreKeep Exploring
What is the significance of networking in the context of AI and data centers?add
What are the characteristics and use cases of InfiniBand compared to Ethernet in the context of networking protocols over the past 30 years?add
What are the advantages of using turnkey systems for AI deployment in enterprises?add
What considerations should a Global 2,000 company have when re-architecting their data centers?add
What is the assessment of Cisco's position in the data center market compared to its competitors, particularly in relation to NVIDIA and Arista?add
What are the three essential components needed to ensure that AI systems function effectively?add
>> Everybody, welcome back to the New York Stock Exchange. This is Dave Vellante and you're watching myself and John Furrier's coverage of the AI Factories - Data Center of the Future. We're going to talk about the infrastructure stack generally. We're going to dig into networking specifically with Zeus Kerravala, he's the Principal and Chief Analyst at ZK Research. Good to see you, my friend. Thanks for coming in.
Zeus Kerravala
>> Hey Dave, how you been?
Dave Vellante
>> I've been really, really good, thank you.
Zeus Kerravala
>> What's that?
Dave Vellante
>> Yeah, we're super excited to be here. The NYSE Wired plus theCUBE, it's been an awesome partnership and it's great to have a super deep networking expert, but also somebody who really understands the market in general security. You and I have talked about a number of companies in that space, so let's get into it. I mean, first, let me start with people I think forget sometimes about the importance of networking in the whole AI. I know you and Bob have worked on some things in the networking, AI summit, AI for networking, networking for AI. It goes both ways, but explain to the audience why networking is increasingly important.
Zeus Kerravala
>> Yeah, it's a good question, and I agree with you, I do think people often forget about the network, and part of it is it's hard to understand, right? With the GPU, Jensen goes on stage, he shows you a GPU, you can conceptualize it. We all got used to Intel Inside, so we understand the importance of CPUs. With the network, it's a little more nebulous, shall I say. We connect things, but we don't really understand how it works, at least most people don't. They just know when they click on google.com and it doesn't work, it must be the network, and so it's hard to understand the importance of it. But I think when you think about how AI has transformed the data center, and Jensen talks a lot about this, the new unit of compute. The unit of compute was a server, then it extended to rack, and in that case, the network played an interesting role.
It connected to rack, but it was all pre-built. And so it was all one system and then it expanded across the data center. And so now your network plays an important role connected all those things in data center. And now with their ability to scale out and scale across data centers, the network's, the thing that makes all these GPUs talk to each other. So in fact, it looks like one GPU when in fact that's several GPUs scattered all over the place and increasingly all over the world. And so from that standpoint, from a back end perspective, the network connects those GPUs. Front end network perspective, you think of all the training and stuff that's being done now, and our data, in fact, I wondered about this years ago, Dave, how would we manage data in an AI world? A lot of companies tried to suck all their data back to one location that was too expensive, too complicated, too hard to manage. So now increasingly, they're just leaving the data where it is at the edge, on IoT devices, and they want their AI engines to be able to go get the data when they need it. And so how do you do that? Well, the network does that. And so front-end, back-end know wired wireless, however you want to look at the network, it's the thing that connects the data to our AI infrastructure, which are GPUs and makes AI work.
Dave Vellante
>> There's three areas that you talked about. There's the scale-up is the NVLink 72, there's the scale-out and then there's the scale-across. It was interesting in the last couple of conference calls, you heard Jensen talk about how their InfiniBand business had doubled, which was unbelievable. Everybody says InfiniBand's dead and then their Ethernet business is growing like crazy. And then they introduced this scale-across concept. And then you heard Hock Tan, which was really interesting on there in the Broadcom earnings call saying, "Yeah, the other guys, their NVLink 72, they can only scale up 72 and we've got hundreds and hundreds," but he was playing games there with the definitions, right? It's really scale-out. Can you explain that?
Zeus Kerravala
>> With the scale-up or scale-out?
Dave Vellante
>> Well, the way, I don't know if you've heard Hock's comments, you basically are trying to deposition InfiniBand-
Zeus Kerravala
>> Look-...
Dave Vellante
>> and say Ethernet scales it. Now, of course NVIDIA has Ethernet too so it's this war of-
Zeus Kerravala
>> This InfiniBand versus Ethernet war has been in place a long time actually.
Dave Vellante
>> Yeah.
Zeus Kerravala
>> In no way do I think that Ethernet's going away. In fact, if you talk to Jayshree at Arista, she always says never been against Ethernet. And I think both can be true.
Dave Vellante
>> What does she say, sorry?
Zeus Kerravala
>> Never been against Ethernet, right?
Dave Vellante
>> Oh, yeah.
Zeus Kerravala
>> It's one out of overall, if you roll back the clock 30 years, there was all these different protocols and Ethernet beat them all. InfiniBand is the last band standing, but InfiniBand has a lot of really unique characteristics. It's highly deterministic, it's lossless, and so if you need guaranteed performance, and no matter what cost, I'm connecting a back plane of GPUs in an NVLink-type scenario. So in a box, it's great for that. You can have InfiniBand grow and Ethernet grow at the same time because we're seeing AI scale-up, scale-out, scale-across. I think over time what happens is InfiniBand gets used for cases where it's a fully integrated system, something like NVLink where I'm connecting GPUs within a box. And then as I start to scale-out and across systems, Ethernet's a better system for that because it's something people are familiar with and ought to work with. And to the vendor's credit, Ethernet has come a long way. When the boom first started, Ethernet was not in a position to take over to replace InfiniBand, but it's even without the UEC, the alternate consortium vendors like Arista, Cisco, companies like that have improved the lossless nature of it and made it more deterministic to closely match InfiniBand. But I do think, in fact, World Wide Tech did a study where they compared Ethernet and InfiniBand and they said that the difference is there nominal. But if I'm running some super intense workloads and I'm Amazon and Microsoft, that nominal little bit makes a big difference. And I'm likely to go with InfiniBand. If I'm a general purpose enterprise, big banks, pharmaceutical companies, that nominal difference isn't going to make a big difference to me, and so Ethernet plays a better role there.
Dave Vellante
>> Now there's also nuance, so both can win. Okay. But there's also nuance if I'd love you to explain this within the Ethernet world, so NVIDIA would say, "Well, we do Ethernet too. We love open standards. Our Ethernet," they call it Spectrum-X, "is purpose built for AI versus generally off-the-shelf internet." I'm sure Broadcom has a response to that, but I haven't heard it, but I'd love to. How much of a difference does that make? Is a purpose built for AI Ethernet something that customers should think about? What are the trade-offs? I think they're both standard, but maybe it's standard like Unix was standard. I don't fully understand, help us parse through.
Zeus Kerravala
>> In fact, that's a great comparison because I started my career as a Unix developer and on an IBM machine, and I could not take code and port it over to an HP machine-
Dave Vellante
>> AIX versus HP-UX.
Zeus Kerravala
>> Because there were subtle differences in how it worked. So at the very base level, Ethernet's Ethernet, but you can do a lot of things to tweak the performance of it. And that's what Cisco's done with their Silicon one. That's what Broadcom's through Tomahawk, chip certainly has done that, and what Spectrum-X has done. So they all have their different flavors of it, but they are all purpose built for AI now.
Dave Vellante
>> So let's uplevel a bit, I know-
Zeus Kerravala
>> So I couldn't go grab a wiring closet switch.
Dave Vellante
>> Right. So you work with a lot of CIOs, you run IT shops. What's the blueprint for the AI, for the networking component, for the AI factory for CIOs and enterprise architects? I'm specifically interested in enterprise AI. We know the cloud guys, we know the neo clouds. They're crushing it. They're spending CapEx like crazy. They're able to raise a bunch of money, hundreds of billions of dollars. But what about the enterprise? How should CIOs and enterprise network architects be thinking about the future of AI factories and how they should be architecting their network?
Zeus Kerravala
>> I think initially, in fact the best comparison I can think of is in the early days of private clouds, in cloud build-outs for enterprises, companies like VCE offered you these fully integrated stacks. Now they were built off, VCE was VMware, Cisco EMC. It was made up of multiple vendors, but it was a reference architecture of a pre-configured validated design for private cloud.
Dave Vellante
>> And it was a SKU that made it simple to sell.
Zeus Kerravala
>> Yeah, and then over time what happened is we got good at deploying this stuff.
Dave Vellante
>> Right.
Zeus Kerravala
>> And we didn't need the turnkey nature of it anymore, and everything's opened up. And I think you'll see the same thing happen with AI is you will see companies, you see Cisco and NVIDIA working together, Pure Storage and Arista to have a partnership with NVIDIA as well and build another turnkey system. These turnkey systems to me are the right way to go for an enterprise because when you start to deploy these things, there's a lot of dials you can turn and levers you can pull in order to tweak the systems so you can spend months doing it yourself or just use these validated blueprints that all the vendors have now that'll make sure that when you deploy it's deployed in a way that you'll get the performance you need. And then over time, it'll happen in a few years, we'll open these things up and we'll be able to pull components from anywhere and build it ourselves. But in the initial term, I don't think there's enough build-outs to know what the best practices are. And so that's why the validated designs are important.
Dave Vellante
>> Okay, that's super helpful. Paint a picture just from a conceptual standpoint. I remember in 2010, Jayshree came on theCUBE and she was explaining to John and me, she's like, "Dave and John, what's historically happened is network traffic was north-south and now things are changing cloud, hybrid cloud, and it's going east-west," and that's the game, the trend that we're leaning into. So now you've got both the pendulum swinging back to scale up because of the training clusters, and you've got massive scale-out, and you've got across data centers. Will that mental model migrate into the enterprise or is that really the cloud, neocloud, giant bank framework?
Zeus Kerravala
>> I think initially it's the cloud, neocloud, but I do think enterprises need to start thinking about it too. I'm not saying your 1,000-person regional bank that's going to ever need that. They're probably going to do their stuff in a cloud. But if you're a Global 2,000, I do think that you do need to think about how to re-architect your data centers. And I do think because for a while everybody just worried about scale-out and that's why the leaf spine architecture became so popular. You can just add more boxes and you just scale your data center as wide as you want. But I think within these clusters now, you do have to worry about vertically integrated performance as well as a performance across racks. And again, it goes down to Jensen's comment about the new unit of compute. What is the new unit of compute? It is your entire data center, or multiple data centers. And that require scale-up, scale-out, and scale-across, and so from an Ethernet perspective, you got to think about the performance requirements within the rack, across racks and across data centers. And that's why even with NVIDIA's integrated optics, the co-packaged optics, that becomes important because now you don't have to worry about the little bit of performance loss from pluggables and things like that.
Dave Vellante
>> Okay.
Zeus Kerravala
>> There's been a lot of innovation.
Dave Vellante
>> I want to ask you about, I wrote a piece, it was tongue-in-cheek, but there was some data behind it, which was the new Jensen's law.
Zeus Kerravala
>> Yeah.
Dave Vellante
>> Which is buy more, make more, and the corollary there, well, buy more, make more.
Zeus Kerravala
>> Make more, right.
Dave Vellante
>> Yeah. The more you buy, the more you make is how we altered it. And then the more you buy, the more you save. And then the corollary is that the network utilization factor, if you can take your network utilization from say 65% up to 85%, NVIDIA made the statement the network is essentially free. Well, it's economically on balance. You get the network is economically neutral, let's call it that. Okay. I think that law holds up in certain cases when you're GPU constrained, oh, sorry, when you're energy constrained, when energy is really the main constraint and your network utilization is low, where that's not the case, which it may not be for a lot of enterprises, then that law may not be immutable. But from a networking standpoint, based on what you've seen so far, how underutilized are those networks? What's the opportunity? As new cycles come out, instead of Moore's law every two years, we're seeing upgrades every one year doubling or tripling the performance every year, and that has network implications. If my network is underutilized and I'm able to utilize it to a greater degree, at least for a couple waves, that's going to be economically neutral. What's the truth there?
Zeus Kerravala
>> Yeah, historical network utilization in a data center, 30%, 40% maybe really highly utilized one 60%. Part of that is once you got there, even with a lot of networks, you'd start getting a lot of congestion with Ethernet. And that's where the limitations came from, which, and like I said, with a lot of the newer systems, Tomahawk, Silicon One, that's all been cleaned up. You can push the bounds that a little bit more.
Dave Vellante
>> Will network utilization, I guess the 65%, I'm talking really efficient operators, the cloud guys, right? Okay. The neoclouds versus the enterprises are probably still not 30% just like servers were pre-VMware. Okay, so will that change?
Zeus Kerravala
>> Yeah, I think so. I think these just things get more efficient over time, and let's face it, whatever data we've got today, in fact I saw it, what was it? The Splunk event, Jeetu Patel, the chief product officer Cisco was talking about how 55% of all data generated today is machine data, created by machines. So whatever traffic we're throwing on the network today, I mean you ain't seen nothing yet. I think we're going to saturate our 800 gig networks and the vendors are all toying with one terabit networking now, which is phenomenal. I remember when the first terabit switch came out, I was talking with Andrew Feldman, he was still at Force10, that's really cool we got a terabit switch. Now we're talking about at a port level. I just think from an AI perspective, Dave, the race is on for AI to do more with the data you have. AI creates more data to work with, and I think whatever we see for volumes today, it's going to be dwarfed in a few years and that's going to create these upgrade cycles that you said. I think companies historically in the data center have waited these two, three years to upgrade. I can see 18-month cycles now, as soon as the new chips come out from Broadcom, companies now are looking at, I should be upgrading as fast as I can.
Dave Vellante
>> Let's talk about the landscape in networking. It seems like the whole market is dramatically shifting as a result of extreme parallel processing or accelerated computing, whatever you want to call it. You've got the legacy, the switch vendors, the network vendors like Cisco. You've now got HPE Juniper as a new competitor. You got companies like Dell that have tried to compete in networking in the past and really not been super successful, but now they're taking the NVIDIA stack with their AI factory strategy, so they get the NVIDIA networking so that's something that is monetizable.
And then you've got the Supermicros of the world, the Novos of the world, which are there in the infrastructure stack, of course, Arista. You've got a resurgence. It's almost a renaissance in networking for some of the reasons that you laid out before. Is the window opening? Is Cisco ripe for disruption? How do you see the HPE Juniper thing playing out? Can a company like Dell actually compete in networking now? Would that AI factory strategy, what are your thoughts on that?
Zeus Kerravala
>> That's a lot of questions. Yeah, I think-
Dave Vellante
>> Pick your favorite.
Zeus Kerravala
>> Yeah. First of all, NVIDIA, I think in the last earnings Bill said that it's $7 billion in networking revenue, right?
Dave Vellante
>> Insane, right?
Zeus Kerravala
>> They're the second-biggest networking company in the world but a lot of that is BlueField, the DPUs.
Dave Vellante
>> DPUs.
Zeus Kerravala
>> Things like that and-
Dave Vellante
>> Ask you about that.
Zeus Kerravala
>> And embedded networking. But still, even just pure Ethernet switch, they're probably selling a couple billion a quarter, which makes them a pretty big networking player. I think Cisco finally has its act together in the data center. If you had asked me this a couple years ago, I would've said they're going to get, Arista's do to them what NVIDIA did to Intel and just run over. But with Silicon One, and especially with the NVIDIA partnership, that to me is very meaningful because I think while Cisco needs NVIDIA to create some AI relevance for their stack, NVIDIA needs Cisco for its distribution channel. Nobody does enterprise distribution better than Cisco. NVIDIA does not know how to run a channel, they'll tell you that. So that is a great partnership. And so I think NVIDIA and Cisco would be there. I really like Arista. I think they're a fascinating case and a company that uses merchant silicon Broadcom, but they tend to get the Broadcom chips in and deployed in their products faster. And through their software and good hardware engineering, they actually get markedly better performance than what you might with a white box. I think from a high-performance perspective, to me those are the three: Arista, Cisco, NVIDIA. NVIDIA is going to appeal to the companies that want this fully integrated solution, I think. And then for companies that want more open, I think they'll use a mix of Cisco and NVIDIA. Juniper is interesting with HPE because neither company has been really strong in the data center for a long time, but clearly with HPE's other assets and servers and storage, there's no reason why they couldn't build their own fully integrated stack. And that's what I'm expecting to see now that they've got the Juniper networking in there, because Juniper does do high-performance well. They did service provider really well, but it's like over the last few years they get so obsessed with campus and AI that they only forgot about data center.
Dave Vellante
>> Want to ask you about just enterprise AI at GTC. Last year, I think you were there.
Zeus Kerravala
>> Yep.
Dave Vellante
>> Jensen laid out three vectors. I've talked about this a lot on Breaking Analysis and in theCUBE, three vectors of AI growth, AI in the cloud, AI and the enterprise, and AI in robotics. And here we're really focused on AI in the enterprise. And the reality is AI and the enterprise hasn't taken off. Most enterprise's data centers aren't liquid cooled. It's not like the enterprise is sucking up all the GPUs because the cloud guys are constrained and they're still trying to figure out, okay, where's my ROI on AI? Having said all that, there's a lot of proprietary data in banks and insurance companies, and they all want to keep it themselves. They all want to keep it there. They say they're not moving that data into the cloud. How do you see that playing out when you talk to your CIO colleagues? How are they thinking about bringing AI and intelligence to the data? Or are they saying, "Hey, we're just going to put it all in the cloud," what are you seeing?
Zeus Kerravala
>> No, I don't think it's going in the cloud. I think they're going to leave the data where it is, which is on their premise at their edge, and then they'll almost try and bring AI to the data. But that's a big network play. I think this is also where you'll see the rise of the neoclouds actually think that if you are going to partner with a third party like cloud provider, data residency becomes important so you're going to want it to stay in country. And so a lot of these neocloud providers are building regionally country-specific nodes. They're all doing that now. And so I think that's, in fact, you've got some, you look at what's going on in the Middle East, they're just, that part of the business is exploded, right? Building on all their different, the neoclouds there. I think that'll be the avenue to AI for a lot of enterprises. I think the biggest of the big though will build their own data centers out. In fact, I was talking to the CTO about a month ago who said that for the first time in 20 years, he's been talking about data center expansion. They've been shrinking, shrinking, shrinking, now they're trying to get their footprint bigger. I do think power's a problem and I am hopeful that SMRs maybe play a role here. It's climate weak in New York actually, and that's one of the big topics here is how do-
Dave Vellante
>> Small modular reactors which is nuclear fission for lack of a better term, versus fusion, which is going to take a little longer. There's radioactive waste that comes out of fission. There's much less or negligible, radioactive waste out of fusion, but it's probably 20, 35, maybe 32, maybe at the best where we see that.
Zeus Kerravala
>> Yeah, but I do think that's the big limiting factor today is the power. And you're right, most of them don't have liquid cool. And that's where if I can buy these integrated systems and liquid cooling built into them that maybe can be engineered to be a little more energy efficient, that's a big win.
Dave Vellante
>> What about security in this AI world? You and I have talked about security, it's an adjacency of your research area. Sometimes you really get deep into it. How does AI change the security model?
Zeus Kerravala
>> Yeah, that's a great question. And I think we have yet to see the impact of AI on security. And so there's the obvious AI can be used for security administrators to do things better, but I think securing the data requires a different model. I think this model, this heavy firewall reliance we've had of trying to place them in strategic points of the network doesn't work anymore. The whole concept of zero trust everywhere actually is the right model now and I have to be able to limit what has access to data and what doesn't. Your question has a lot of different vectors. Just from a utilization perspective, I'm a user. There got to be systems in place that limit what I can put into ChatGPT or Gemini or things like that. So if I'm an accountant and I want to create an earnings script, I shouldn't be able to drop it in Gemini and say, "Create me an earnings script based on these financials two days before the earnings call."
Limit it, but at a... And so there's a lot of work being done there and helping companies understand how to create guardrails, what to do, things like that. I think from protecting the data standpoint though, that's where you're going to see the big shift because historically in data centers, we threw up a lot of firewalls to protect the east-west traffic. That is an expensive model that's very resource-intensive and requires a lot of constant changing. I think this model, that Zscaler probably perpetuates better than anybody of to be able to deploy zero trust down to the individual server workloader. The fact they've got a cellular partnership where they can secure things at the SIEM for IoT. To me, I'm not sure how the world can evolve. The security world can evolve without fundamentally changing the model where I break away from these appliances to something that's a lot more software based.
Dave Vellante
>> All right, we're got to go. So it sounds like Ethernet with UEC or InfiniBand, the answer is yes in 2026, both.
Zeus Kerravala
>> Both. I mean, both can be true.
Dave Vellante
>> And then what's the biggest gotcha you see in AI network proofs of concept that CIOs and architects should be aware of? What should they watch out for?
Zeus Kerravala
>> I think just the performance. I think not underestimating the amount of traffic you're going to generate from AI, because these network builds take a long time and they're very expensive, and so whatever you think you're going to have for traffic, I think your estimates are conservative, yeah.
Dave Vellante
>> So GPU utilization obviously very important, but so is network-
Zeus Kerravala
>> Yeah. In fact, you require-
Dave Vellante
>> It's dollars.
Zeus Kerravala
>> Yeah. In fact, Charlie Giancarlo, the CEO of Pure said this to me when he joined Pure, because I joked, "You're going to a storage company?"
Dave Vellante
>> And he's a networking guy, right.
Zeus Kerravala
>> Yeah. He says you need three things to make sure your AI works, fast storage, fast network, and fast processor, right? Everybody understands the role of GPUs, right? If your network speeds don't match what you're trying to do to the GPU layer and your data access speeds on the storage side, don't match that. One of those as a weak link will create the whole system to not work properly, so the three of those things have to be in lockstep with one another. And again, that's why I'm a big fan of these validated solutions.
Dave Vellante
>> Zeus Kerravala, thanks so much for coming on theCUBE. Good to see you, really appreciate your contributions. All right, thank you for watching. This is Dave Vellante for John Furrier, and the NYSE Wired plus theCUBE's coverage of AI Factories - Data Centers of the Future. But right back, right after this short break from the New York Stock Exchange.
Dave Vellante