In this theCUBE + NYSE Wired segment from “AI Factories – Data Centers of the Future,” Nebius co-founder and CBO Roman Chernin sits down with theCUBE’s John Furrier at the New York Stock Exchange to unpack how AI factories are reshaping enterprise infrastructure and the future of data centers. Chernin outlines Nebius’ two-track strategy: a multi-tenant cloud built for developer experience and managed services, and large-scale, mostly bare-metal deployments for hyperscalers and AI labs. He discusses the significance of Nebius’ Microsoft deal (described as “up to $20B” and set to become one of the largest single-site GB300 deployments) as both an engineering milestone and a way to feed scale and cash flow back into the core cloud business. The conversation explores why enterprises want “the baby of supercomputer in the cloud,” marrying cloud flexibility with supercomputing efficiency to minimize time-to-value without sacrificing performance.
Chernin details Nebius’ specialization in AI-centric workloads (large distributed training and inference at scale), a platform roadmap that moves beyond infrastructure into inference, fine-tuning and reinforcement learning as services, and a commitment to helping customers build on open-source models for control, cost and data leverage. He traces customer waves from foundational model builders to vertical AI companies and tech-forward enterprises, noting early traction with firms like Shopify and momentum in regulated sectors such as healthcare following Nebius’ compliance milestones. With roots in Yandex’s large-scale engineering culture and meaningful exposure to ClickHouse, Chernin also weighs in on the economics of AI-scale infrastructure (power and capacity as gating factors), hybrid orchestration and sovereignty, and why latency priorities vary by use case – from reasoning models to voice agents – as AI factories become the new unit of value in modern enterprise compute.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: AI Factories - Data Centers 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: AI Factories - Data Centers 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: AI Factories - Data Centers of the Future
Please sign in with LinkedIn to continue to theCUBE + NYSE Wired: AI Factories - Data Centers 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
Jason Hoffman, Switch
In this theCUBE + NYSE Wired segment from “AI Factories – Data Centers of the Future,” Nebius co-founder and CBO Roman Chernin sits down with theCUBE’s John Furrier at the New York Stock Exchange to unpack how AI factories are reshaping enterprise infrastructure and the future of data centers. Chernin outlines Nebius’ two-track strategy: a multi-tenant cloud built for developer experience and managed services, and large-scale, mostly bare-metal deployments for hyperscalers and AI labs. He discusses the significance of Nebius’ Microsoft deal (described as “up to $20B” and set to become one of the largest single-site GB300 deployments) as both an engineering milestone and a way to feed scale and cash flow back into the core cloud business. The conversation explores why enterprises want “the baby of supercomputer in the cloud,” marrying cloud flexibility with supercomputing efficiency to minimize time-to-value without sacrificing performance.
Chernin details Nebius’ specialization in AI-centric workloads (large distributed training and inference at scale), a platform roadmap that moves beyond infrastructure into inference, fine-tuning and reinforcement learning as services, and a commitment to helping customers build on open-source models for control, cost and data leverage. He traces customer waves from foundational model builders to vertical AI companies and tech-forward enterprises, noting early traction with firms like Shopify and momentum in regulated sectors such as healthcare following Nebius’ compliance milestones. With roots in Yandex’s large-scale engineering culture and meaningful exposure to ClickHouse, Chernin also weighs in on the economics of AI-scale infrastructure (power and capacity as gating factors), hybrid orchestration and sovereignty, and why latency priorities vary by use case – from reasoning models to voice agents – as AI factories become the new unit of value in modern enterprise compute.
>> Welcome back everyone to theCUBE studios here. The New York Stock Exchange is theCUBE East Coast studio partnering with the NYSC Wired community. This is our wired program. Of course, the AI factories is an ongoing series. We feature leaders in large-scale infrastructure. Obviously supercomputing is now the data center. Data centers connect to each other, they form more compute power. This is the key trend that is really fueling a generational shift. We've kind of been seeing it coming, but it's now hitting mainstream. Infrastructure is the key investment in a naval area. We've got a guest here, friend of theCUBE going back before the founding of theCUBE actually, Jason Hoffman, chief strategist at Switch. Really early days on a lot of technology cloud. Former co-founder of Joyent, well-documented company that essentially brought in the era of DevOps, part of what they call Cloudarati. Gone on to do many, many things. Jason, we need you to come in and help explain to the real world what these AI factories are. First of all, great to see you. Thanks for coming in.
Jason Hoffman
>> Yeah.>> Nice rig you got going there. Thanks for coming in.
Jason Hoffman
>> Thanks for having me, John. I really appreciate it. And yeah, as Wired described me, I'm the creator of Clouds Baymax, so that's something.>> Yeah, pioneers take the bullets, but early days, remember those days in 2008 the meetups, in the early days, a lot's changed, but now the AI factories are here. Seems like a similar movie with the data gravity is data centers. Switch, a big part of that. Remember when you guys were formed. Now more than ever, data centers are in demand, but unpack this AI factory kind of large-scale computing trend. You've got the big build-outs, but then you got the enterprises trying to figure out how to do AI when they've had cloud or on-premises activities. What does an AI factory in your mind mean for the industry?
Jason Hoffman
>> Oh yeah, that's a great question. I mean in the case of Switch, we've pioneered a certain data center design, which is largely the encapsulated hot aisles that go up to a plenum roof and going up to a pretty good scale density on a per-rack basis in there. And that's been the history of the company for the last 25 years is Rob Roy is still the CEO of the company, is the founder of the company in there. And I'd say that in the previous generation, a lot of the conversation around data centers, the building itself was part of the overall design and you'd go and put servers inside of the data centers and then sort of call today. It was a lot of ways a hotel business for these type of compute elements and the like.
In the case of an AI factory, it's a lot of what's in the name, the internals of the building start mattering quite a bit and they start mattering quite a bit because of the shift of the kind of equipment that goes into them. When we sort of kicked off computing 50 years ago, it was a lot of scientific computing and these big sort of vertically scaled systems and we've always had these kinds of supercomputers that are present in scientific computing and other sort of areas in there. And then what sort of kicked off in parallel to that was the use of mainframes and personal computers and cloud and everything that's sort of being used in the enterprise space in there. And then with AI taking the whole swing back, what we're really seeing is that kind of commoditization of these supercomputers. So things that need to vertically scale, not sort of scale out in a lot of places like what we have with cloud and enterprise. And so on these AI things, the more dense these racks can be, the better. And we're seeing these kinds of fund densities like going from an industry average of 20 KW to two megawatts sort of rack from a roadmap perspective. And it means that all of a sudden the internals of these buildings become almost like the internals of a semiconductor fab, which we don't even talk about the shell of a semiconductor fab. We talk about the internals and the machinery that's in there and the building of the chips. And so in a lot of ways for the data center industry as it sort of shifts into this kind of factory mode, it's really about the internals and what goes inside of it and a totally different set of metrics and everything.>> I'm really glad to do this interview with you because I remember the pioneering days when you guys started Switch. I think we covered, I think they called the SUPERNAP or something. Is that still around that name or that was an early name.
Jason Hoffman
>> Yeah. Yeah.>> Love the word super, super cloud, SUPERNAP. I mean at that point that was pioneering. I mean you guys were early. What were some of the learnings that you could look back now and say what Switch did? Obviously smart to think about it that way as a self-contained system. It's a systems design. Systems thinking is now kind of new. It's a nouveau thing. You guys have been doing it for a while. So as companies figure out, one, who they partner with for say large-scale supercomputing building/system, as well as maybe an existing data center that might not have a power envelope that could be designed around it. I mean it's bounded by power in the smaller, well, small regular data centers, but then how do I evaluate this? Take us through what you've learned at Switch, what the learnings were because you guys had to do that module design, you thought about the big picture and you had to deal with the power, you had to deal with the density, you had to look at the substrate of the building. How do I get more copper? I mean all these things are now being talked about mainstream.
Jason Hoffman
>> Yeah, I mean what was super clear for us is one of the... I mean when the NVIDIA H100 systems came out, meaning the equivalent of sort of a chat like a GPT-2, 2.5, 3, these kind of things we're trained on. We housed what was one of the first large clusters of that, so roughly about 7,000 H100 systems inside of there. And it was all air cooled because we could do those kinds of densities in there. And Switch had this design which has become the industry standard now of here's a building and air handlers are outside and everything exits out to a plenum and goes out to these external air handlers. And just this whole idea that let's go ahead and build a hundred megawatt sized facility 17, 18 years ago, which was 18 years ago, it was sort of crazy that you'd go build a building that size. And as you mentioned the first of the SUPERNAPs, Las Vegas seven was exactly that kind of a hundred megawatt size like facility. And so when we went and looked at these first kind of air cooled cluster deployments, it was very clear that->> It won't last long.
Jason Hoffman
>> Yeah, the balancing, it's very difficult to, as you sort of go to these higher densities and you deploy more power, you can strand space and if you want to fill up the space, then you're basically not building a cluster as compact as you can be and so on like that. So it was clear to us a couple years ago, pretty soon after the first ChatGPT took off, so we're really talking the first half of 2023 that we would have to basically introduce liquid cooling into it. And so the design that Rob came up with was taking the air handlers that are outside that were sort of in the megawatt range, minimizing them down and then sort of going from this macro climate perspective to a microclimate perspective. And we started encapsulating each and every rack inside of its own effectively like refrigerator. As we brought those air handlers inside, it meant we actually brought those kind of liquid and plumbing inside of it and everything else and all the supporting structures to do that. And so we started thinking that what we really had to develop at Switch was sort of an equivalent type of system roadmap that would line up with an NVIDIA system roadmap that would line up with a TSMC semiconductor process node and packaging roadmap. That just like how you have these nanometers flow up here and you have the packaging that comes up into it and it comes up into these system sizes. In some ways there's sort of this first era of systems that are now showing up that are sort of the 150 to 300 KW size. There's going to be sort of the second era that's 600, 700 KW and sort of this third era that's like a megawatt plus type rack. And the technical challenge that we gave ourselves a couple years ago then was to figure out how do we go and design for a two megawatt per rack kind of future. But do it in a way that minimizes the number of parts and makes it relatively straightforward to still sort of manufacturer, it gives us a lot of optionality. Because we don't know exactly how these clusters are going to work out or the like. And so Rob came up with his EVO design that's now been where the world first GB200s and GB300s and B100s were deployed in our design in that. And we have these concepts largely of what you think of a density node where it's a known amount of cooling and power delivery in it. And then these sort of cluster packages where we have known our own sort of abstractions around what cluster topology is going to look like because not everything in there is liquid cooled right now and there's still networking and storage and alike and so on. And so we just managed to come up with a design that flows from a TSMC type of roadmap, flows into a NVIDIA system roadmap, flows into a long-term sort of twenty-year design of our facility roadmaps in it and took that approach.>> First of all, I love hearing you talk about this because the system's thinking is a new thing. I mean at the large scale you guys are just pioneering, you're doing it. What's interesting is that back in the old days, remember the old drop ship me a server and we'd load Linux on it or something, actually load software on it. I mean people want drop shipped data centers as supercomputers. I mean, so when you talk about a density node, what pops in my head is, "Okay, I might have a power constraint or I have a location, I need a fully encapsulated system that just plugs in." I mean that's like a boatload of servers, some network fabric, full stack. All those things have to be built in. That kind of sounds a little bit like just drop shipping a machine. I remember Amazon came out with, what were they called? Outposts, remember those days? I had that concept of just having replica of a cloud. We're kind of seeing what you guys are doing and seeing, okay, the edge is going to get smarter, there'll be some form factor, some may be bigger, maybe the size of a old telco box at the side of the road or the size of a camera on a pole where training and inference will have to happen. So as you look at that distributed computing paradigm expanding, what do you guys see as the practitioner angle? Because you have infrastructure people, there's not a lot of talent out there. I mean they claim there's IT talent, but that's IT. That's not system designed like what you guys do. Then you got the chief AI officers out there going, "Let's go get some GenAI, give me some magic dust and build me some CapEx." Not that easy. So take me through how you see the evolution here because it's going to get distributed, there'll be big mega centers, SUPERNAPs, and then there'll be like mediums and then smalls, density nodes, whatever you call it. Take me through that.
Jason Hoffman
>> Yeah, the question really is what is small and is small really achievable in that kind of sense? And I say this again as somebody who's done a cloud computing company, I did an edge computing company, the Google acquired, now we're sort of building these large scale AI factories. I build mobile networks. I've gone as->> You've done a lot, that's why I asked.
Jason Hoffman
>> Yeah, so in the case of having worked on 4G and 5G and the sort of overall architecture of it, I mean I would say that the comparable CapEx from what's being deployed right now from an AI perspective was really done building out 4G and 5G networks. Meaning the deployment of 5G networks and sort of going into LTE advanced globally was spending on the order of one and a quarter, one and a half trillion dollars to sort of go deploy that across every country on the planet. We see sort of a similar thing on the more non-technical side of it that occurred over decades was the deployment of air conditioning and refrigeration globally. That's roughly about 220 gigawatts of things. In the US, it's about sort of 80 gigawatts, if you will. If you look at the AI footprint that people are building into the US, they're building into a footprint that's about the size of air conditioning and refrigeration. Air conditioning and refrigeration is why the grid was really built out post World War II in a lot of spots in sort of the world. And these kind of build outs, I think people sort of underestimate the size of them in a lot of ways and what's sort of required of it. And we'll get down to the distributed part of there. In the case of Switch, we're headquartered in Nevada and we have a very large footprint in Nevada. As we continue building out that footprint because in Nevada we also can function as a behind the meter unregulated utility and the like in there. I mean I think fully built out in Nevada we're something like a third of southern Nevada's power, two-third of Northern Nevada's power and we have a total power footprint that's like the equivalent of Los Angeles. And so if you think about what it takes to go and build total infrastructure like that->> So you're comparing it to the 4G, 5G build out CapEx equivalent, but at data center scale.
Jason Hoffman
>> I mean it's kind of thing where these AI campuses we're developing are right now a top 10 industrial site historically. And then I think as we start making them even more dense over the next five to 10 years, they'll surpass anything that we ever saw from large scale oil fields, refineries, smelters, sort a number of things like that. There are in the US for example, outside of a Gulf based refinery site and so on like that, if you go and you look in various states, the AI workloads are the largest industrial always on power sites. That's basically sitting in those kind of areas right now. And so when you start thinking about minimizing those, the question always about how much is required in the middle mile if you will, is going to come down to how powerful a device like this is and then how powerful the centralized compute is going to be. And so even the sort of, quote, unquote, "edge" conversations and where inference is going to live, I would say that it's going to live on the device and it's going to live in very centralized data centers. The things that are in between are still going to stay relatively lightweight. Because if right now it's already a challenge to get 700, 800 kilowatts into a tower around sort of base stations being present there, we're not going to be able to stick a two megawatt rack at the bottom of that tower or do that kind of thing. These density around these systems are technically very challenging to go and distribute. I mean a single rack has the same power footprint as most traditional enterprise data centers have in total. And so a lot of people have a four megawatt or eight megawatt data center. Well that's just four racks in the future of some of this gear. So how exactly is that going to work in a very distributed footprint? And so increasingly it's like a lot of the inference ends up happen on the device and ends up happen in very centralized data centers. It ends up following I think a lot of the same curves that you see in gaming and sort of other stuff. You have a very powerful PC and then very centralized computing doing these types of stadia in the middle is just not something that economically penciled out over time.>> Explains why Texas likes the AI. They love their oil too. As you said, it's oil fields down there. I love that point because I think to me right now at the edge, inference is how it can infer at the edge makes a lot of sense, but training is coming in. Edge devices get new data all the time. So this comes back down to your point about the architecture of the middle mile or distributed nature of the mega-centers, which I do buy that with you. I would see that happening at mega centers.
Jason Hoffman
>> Well, if you look at the largest footprint of neural processors, it's on device. It's not in a base station or in an old central office or in a sort of regional data center and the like. And most of the experience around these things is such that you're typing into a ChatGPT and it's printing out. It's not something that has to occur within 10 milliseconds or 20 milliseconds or that type of->> Well talk about networking. Obviously you have a networking background as well. Obviously at Switch it's highly interconnected. Interconnects with the new thing. Jensen Wong was on stage last year at GDC and he said, "KV Cache is the operating system for the AI factory." And I'm like, "Oh, he said it, someone finally said." I've been saying on theCUBE, what runs on an AI factory? What OS? What is the OS? In the very elementary computer science definition, you had an operating system, it does a function, it runs hardware and it makes stuff work. I'm oversimplifying obviously, but the question is what is the OS for these clusters, these large-scale AI factories? And it's interesting, he said that and I'm like, "Oh, it's networking what brings systems together." So, one, what's your reaction to that and is networking the OS of the AI factory or everything?
Jason Hoffman
>> No, but I'd say that in general, and it goes to a AI factory, even if we anchor it in traditional data center metrics like PUE or sort of other things in there, it's going to come down fundamentally to token per watt. And then if you basically go to the next sort of thing of, "Okay, we have these systems that are sitting there." A data center is responsible at the end of the day for powering and cooling, say a given rack of these systems and we're trying to accomplish a certain token output per watt. When you start looking at the data packets through a network stack, then on the model layer you're trying to accomplish sort of token flow through these attention blocks. When you go into the data center analogy and you think of virtual memory management and paging and putting things in and out of a register, which is what typically an operating system does. So if you think of what an operating system normally does is it boots up a piece of hardware and it's fundamentally doing virtual memory management and taking things in and out of there, putting things on and off the CPU. Well you would say that a model layer equivalent to that is KV Cache allocation of eviction. You could absolutely say that. And then when you look at operating system address space expansion and what sort of occurs there. That would be context window extensions present inside of a model. When you start thinking about distributed operating system kernels, that would be the sort of multi-node sort of cache sharding and those kinds of things that you'd see in these models. And then these sort of long-lived processes with state full context and stuff starts being these kind of streaming inference that you get with sort of persistent key value stores and that types of things. And so I'd say that->> I think you just summarized it beautifully. I think one of the things I was talking about, I want to get your thoughts on the, I used to call it the holy trinity of computing, storage, networking and compute, but I want to add a fourth pillar in their database. Database is now key part, you mentioned key value store. The role of the database isn't just the software abstraction now, it's actually part of the data piece of the networking, storage and compute. It's almost a fourth pillar of that OS dynamic you mentioned. Some of the things there, but this is distributed computing in the modern era, this is the science behind that and this is what people are trying to figure out. What is the ideal science for that environment and then how do you optimize it, design it, design optimize, run. Any reaction to that? Any thoughts?
Jason Hoffman
>> No, databases at the end of the day sit on top of storage. It's sort of in there. It's just another abstraction layer. I mean a key value cache is technically a type of database that's in that. And so the analogies on these hold, I mean I'd say that in the case of AI, the sort of vector databases are particularly important of course for keeping state and then these retrieval caches and feature stores and other things that are in there, these are essentially the semantic memory of these kinds of systems. They actually give a model any sort of persistence of knowledge, grounding, recall anything beyond the weights of the model. And so if you go from this kind of weights only model you're interacting with that knows nothing and you need to start using it from an enterprise perspective, then yeah, it's where that sort of knowledge sits.>> All right. Let's talk about Switch. You're the chief strategy officer, by the way, you have a great pedigree of experience. As you talk to customers, what are some of the things you guys are doing that you would say illustrates where the wave is right now? I mean obviously you got the data centers, a super center, SUPERNAPs, but also there's a business where you guys are actually deploying on behalf of customers too to run operations. Is that part of it? Where are you getting pulled into? What does your 3D chessboard look like?
Jason Hoffman
>> Yeah, I mean, Switch is unique in the data center space in that we design, build and operate large scale campuses where enterprises can come and get the benefit of a scale, though it would be very difficult for any enterprise to get. And campuses that are large enough that even hyperscalers go, "Oh, we could build a region there." So for us when we talk about exascale campuses, these are things that are typically designed to start out in the three to four gigawatt type of range and are capable of scaling beyond that. And where we can go and do literally terraforming utility development, build out the network, sort of do everything that's in there. And so we function a lot like the nuclear-powered aircraft carrier at the data center space as how we think of us from sort of the fleet in that. And then I would say that we are founder run. Our founder is a very good structural mechanical electrical engineer and has been doing that in the data center space for the last 26 years. We've gone and added a lot to the team of myself and others that are experts in distributed systems and software and bringing new sensors and so on sort of like that. And so a lot of what we've been doing is how do we go and digitize all of these types of things, our own design process, the construction process, those sorts of things in there. How do we do a lot of prefabrication and manufacturing beforehand? How do we introduce a lot of optionality? How do we make this more modular? And so what we function with in our largest partners is it is our job to go ahead and have a very future-proof long-term roadmap and design around what we're doing there. And making sure that those are aligned with folks like Dell that are doing the sort of total systems, people like NVIDIA, even sort of the semiconductor manufacturers and people that are sort of going and doing that. And making sure that if somebody goes and deploys these kind of leading edge systems right now, that they'll still continue doing that in our facilities 10, 15, 20 years from now just has been the case historically. And so we try to be that exceptionally knowledgeable design partner and people that are then capable of building and operating and having those things continuously flow together.>> I just had Michael Dell on, he'd be pleased to hear that. I mean you guys are like a colo on the nuclear-powered aircraft carrier. Can I have some of that?
Jason Hoffman
>> Well, our Austin campus, that's part of his old campus, so our data centers are nestled within those buildings sort of in there. And Dell is who we get our own GPU systems from.>> Yeah, Dell's kicking ass. And he said on theCUBE that they deliver the goods and that's maxed out immediately. So you're seeing the demand cycle heavily. Interesting view on the campus, I think that's almost a real estate play and look at data centers that start out as real estate investment trusts. Now they're full on supercomputing, so you can almost see the dots connecting as you lay that out. Large campuses have optionality. If they have power and utility, why not turn that into a factory?
Jason Hoffman
>> Yeah, yeah. Look, and we do this thing where I would say that, go ahead and be maybe a tad arrogant for a moment, but we're probably the most advanced user of NVIDIA's Omniverse and sort of digital twin efforts that are there. We have some of the largest models that have ever been done in that system and we sort of look at how we've gone and digitized aspects of the power chain and so on in there. My goal is to really go and support Rob and the broader team so that when we engage with our customers, we can basically take our team and put it together with any sort of technical infrastructure team in any company in the world and basically be a credible technology partner.>> I mean you guys are so smart. I mean you guys pioneered and been following obviously Switch from the beginning. I remember when you first got the high bandwidth there. I think we got a tour in Vegas early on. You guys are thinking material scientists as well as not just computer scientists. It's a lot of engineering and the upside potential's huge because now you are solving the problems. Now NVIDIA's Omniverse is interesting because you're getting real data, they're trying to supplement with synthetic data. You're a huge advantage for NVIDIA's Omniverse, you're pulling in massive data and your digital twin.
Jason Hoffman
>> Yeah, no, absolutely. Yeah. Well, on a means to that our own compute sophistication and footprint and our own deployment of these GPU systems internally for our entire design chain. How these things get commissioned, how they're operated on a go-forward basis, how do we do everything from sort of a capacity reliability to so on has become really using every possible modern tool that one can use.>> Well, Jason, great to catch up with you. I'm so glad we connected. Definitely want to follow up with you and put you on the roster for hitting you back up because, one, the work you guys are doing there, what you've done, your unique experience. You're seeing all the theaters that you've done body work in your career, telecom, networking, cloud, Switch now. So you got a unique perspective. So we really appreciate you spending the time with us here in theCUBE and the NYC Wired community. And of course we're in our New York studio and NVIDIA stock prices looking good or Tesla.
Jason Hoffman
>> Thank you NVIDIA for being up 16.8 x since 2020. My only portfolio regret is that you weren't 100% of it.>> Absolutely. Everyone's feeling good. It's a great wave right now. It's almost like for us as we look at this next generation, it's a generational moment. It really is. Just your thoughts on that as you look at the industry right now, it's like, "Wow, this is highly accelerated roads of super intelligence," whatever you want to call it.
Jason Hoffman
>> Yeah, well it's a part of a continuation in many ways. I mean, if we go and we think of... it just depends on how far back do you want to go. I mean, if we go all the way back to literally the printing press and the first Gutenberg Bible and the Reformation, well, large language models are a distillation of language themselves. It's interacting with language and utilizing that in sort of different ways. So you can think of it as maybe the culmination of the last 500 years of technology efforts or the like. But I think for me, look, as you and I first met 20 years ago as cloud was basically kicking off. When cloud was taking off as a trend, it wasn't like something that impacted the world's largest companies. So recall back then, the world's largest companies were oil and gas. You had telcos in the top 10, that kind of thing. Then when the iPhone comes out and mobile phones come out and you start getting social media on mobile phones, the sort of effort there around mobile, it's like, "Yeah, mobile impacted consumers." And you had the consumerization of the enterprise and so on like that. But now we literally in the case of AI, we have a technology, a sort of or however we want to call it, but we basically have a thing going on right now that directly impacts the world's largest companies. And so we're sort of seeing is honestly something we haven't seen before in that. Because you have sort of like, "Hey, these are the top 10 or 15 companies in the world, and then this is a new technology that it's potentially disruptive for them. It's also the biggest opportunity they've ever had.">> They could win too. They get paid to focus on it.
Jason Hoffman
>> So everyone's going. And that was not the case with mobile networks. It was not the case with cloud when it rolled up. It was not the case with even sort of the PC. It didn't have this type of thing of like, "Oh, we must do this in order to there." And so you're seeing an unprecedented amount of money being spent for sort of this >> And the intellectual capital, I mean I've observed that on that same note where the generations of old and young are interacting. It's not like the young beat the old, there's so much because it impacts everyone. It's the connections. Everyone's in. Everyone's in the game from day one.
Jason Hoffman
>> No, no. And again, being here in the Bay Area where I live in Los Gatos, it's like we got a Google head of something that lives on the corner and another one sort of there and another one's sort of there. And the local barbecues, if you will, include old and young and sort of this and that. But it's from all the places you'd think. And it is an area of just sort of like a tremendous intellectual interest right now.>> Now's a positive interaction and relationship. Jason, thank you so much for coming on theCUBE. Again, great to see. You keep alumni many times on. Really appreciate your friendship and thanks for taking the time.
Jason Hoffman
>> Thanks for having me.>> All right, cool. I'm John Furrier here at the New York Stock Exchange, theCUBE studios part of our NYC Wired program. Of course, we've got our Palo Alto studios, we've got all the events. As AI goes mainstream, it affects the entire society at large, the planet, and even in space. This data center is going to be in space soon. Who knows. Thanks for watching. We're doing our part to bring you the linguistics and the data. Thanks for watching.
>> Welcome back everyone to theCUBE studios here. The New York Stock Exchange is theCUBE East Coast studio partnering with the NYSC Wired community. This is our wired program. Of course, the AI factories is an ongoing series. We feature leaders in large-scale infrastructure. Obviously supercomputing is now the data center. Data centers connect to each other, they form more compute power. This is the key trend that is really fueling a generational shift. We've kind of been seeing it coming, but it's now hitting mainstream. Infrastructure is the key investment in a naval area. We've got a guest here, friend of theCUBE going back before the founding of theCUBE actually, Jason Hoffman, chief strategist at Switch. Really early days on a lot of technology cloud. Former co-founder of Joyent, well-documented company that essentially brought in the era of DevOps, part of what they call Cloudarati. Gone on to do many, many things. Jason, we need you to come in and help explain to the real world what these AI factories are. First of all, great to see you. Thanks for coming in.
Jason Hoffman
>> Yeah.>> Nice rig you got going there. Thanks for coming in.
Jason Hoffman
>> Thanks for having me, John. I really appreciate it. And yeah, as Wired described me, I'm the creator of Clouds Baymax, so that's something.>> Yeah, pioneers take the bullets, but early days, remember those days in 2008 the meetups, in the early days, a lot's changed, but now the AI factories are here. Seems like a similar movie with the data gravity is data centers. Switch, a big part of that. Remember when you guys were formed. Now more than ever, data centers are in demand, but unpack this AI factory kind of large-scale computing trend. You've got the big build-outs, but then you got the enterprises trying to figure out how to do AI when they've had cloud or on-premises activities. What does an AI factory in your mind mean for the industry?
Jason Hoffman
>> Oh yeah, that's a great question. I mean in the case of Switch, we've pioneered a certain data center design, which is largely the encapsulated hot aisles that go up to a plenum roof and going up to a pretty good scale density on a per-rack basis in there. And that's been the history of the company for the last 25 years is Rob Roy is still the CEO of the company, is the founder of the company in there. And I'd say that in the previous generation, a lot of the conversation around data centers, the building itself was part of the overall design and you'd go and put servers inside of the data centers and then sort of call today. It was a lot of ways a hotel business for these type of compute elements and the like.
In the case of an AI factory, it's a lot of what's in the name, the internals of the building start mattering quite a bit and they start mattering quite a bit because of the shift of the kind of equipment that goes into them. When we sort of kicked off computing 50 years ago, it was a lot of scientific computing and these big sort of vertically scaled systems and we've always had these kinds of supercomputers that are present in scientific computing and other sort of areas in there. And then what sort of kicked off in parallel to that was the use of mainframes and personal computers and cloud and everything that's sort of being used in the enterprise space in there. And then with AI taking the whole swing back, what we're really seeing is that kind of commoditization of these supercomputers. So things that need to vertically scale, not sort of scale out in a lot of places like what we have with cloud and enterprise. And so on these AI things, the more dense these racks can be, the better. And we're seeing these kinds of fund densities like going from an industry average of 20 KW to two megawatts sort of rack from a roadmap perspective. And it means that all of a sudden the internals of these buildings become almost like the internals of a semiconductor fab, which we don't even talk about the shell of a semiconductor fab. We talk about the internals and the machinery that's in there and the building of the chips. And so in a lot of ways for the data center industry as it sort of shifts into this kind of factory mode, it's really about the internals and what goes inside of it and a totally different set of metrics and everything.>> I'm really glad to do this interview with you because I remember the pioneering days when you guys started Switch. I think we covered, I think they called the SUPERNAP or something. Is that still around that name or that was an early name.
Jason Hoffman
>> Yeah. Yeah.>> Love the word super, super cloud, SUPERNAP. I mean at that point that was pioneering. I mean you guys were early. What were some of the learnings that you could look back now and say what Switch did? Obviously smart to think about it that way as a self-contained system. It's a systems design. Systems thinking is now kind of new. It's a nouveau thing. You guys have been doing it for a while. So as companies figure out, one, who they partner with for say large-scale supercomputing building/system, as well as maybe an existing data center that might not have a power envelope that could be designed around it. I mean it's bounded by power in the smaller, well, small regular data centers, but then how do I evaluate this? Take us through what you've learned at Switch, what the learnings were because you guys had to do that module design, you thought about the big picture and you had to deal with the power, you had to deal with the density, you had to look at the substrate of the building. How do I get more copper? I mean all these things are now being talked about mainstream.
Jason Hoffman
>> Yeah, I mean what was super clear for us is one of the... I mean when the NVIDIA H100 systems came out, meaning the equivalent of sort of a chat like a GPT-2, 2.5, 3, these kind of things we're trained on. We housed what was one of the first large clusters of that, so roughly about 7,000 H100 systems inside of there. And it was all air cooled because we could do those kinds of densities in there. And Switch had this design which has become the industry standard now of here's a building and air handlers are outside and everything exits out to a plenum and goes out to these external air handlers. And just this whole idea that let's go ahead and build a hundred megawatt sized facility 17, 18 years ago, which was 18 years ago, it was sort of crazy that you'd go build a building that size. And as you mentioned the first of the SUPERNAPs, Las Vegas seven was exactly that kind of a hundred megawatt size like facility. And so when we went and looked at these first kind of air cooled cluster deployments, it was very clear that->> It won't last long.
Jason Hoffman
>> Yeah, the balancing, it's very difficult to, as you sort of go to these higher densities and you deploy more power, you can strand space and if you want to fill up the space, then you're basically not building a cluster as compact as you can be and so on like that. So it was clear to us a couple years ago, pretty soon after the first ChatGPT took off, so we're really talking the first half of 2023 that we would have to basically introduce liquid cooling into it. And so the design that Rob came up with was taking the air handlers that are outside that were sort of in the megawatt range, minimizing them down and then sort of going from this macro climate perspective to a microclimate perspective. And we started encapsulating each and every rack inside of its own effectively like refrigerator. As we brought those air handlers inside, it meant we actually brought those kind of liquid and plumbing inside of it and everything else and all the supporting structures to do that. And so we started thinking that what we really had to develop at Switch was sort of an equivalent type of system roadmap that would line up with an NVIDIA system roadmap that would line up with a TSMC semiconductor process node and packaging roadmap. That just like how you have these nanometers flow up here and you have the packaging that comes up into it and it comes up into these system sizes. In some ways there's sort of this first era of systems that are now showing up that are sort of the 150 to 300 KW size. There's going to be sort of the second era that's 600, 700 KW and sort of this third era that's like a megawatt plus type rack. And the technical challenge that we gave ourselves a couple years ago then was to figure out how do we go and design for a two megawatt per rack kind of future. But do it in a way that minimizes the number of parts and makes it relatively straightforward to still sort of manufacturer, it gives us a lot of optionality. Because we don't know exactly how these clusters are going to work out or the like. And so Rob came up with his EVO design that's now been where the world first GB200s and GB300s and B100s were deployed in our design in that. And we have these concepts largely of what you think of a density node where it's a known amount of cooling and power delivery in it. And then these sort of cluster packages where we have known our own sort of abstractions around what cluster topology is going to look like because not everything in there is liquid cooled right now and there's still networking and storage and alike and so on. And so we just managed to come up with a design that flows from a TSMC type of roadmap, flows into a NVIDIA system roadmap, flows into a long-term sort of twenty-year design of our facility roadmaps in it and took that approach.>> First of all, I love hearing you talk about this because the system's thinking is a new thing. I mean at the large scale you guys are just pioneering, you're doing it. What's interesting is that back in the old days, remember the old drop ship me a server and we'd load Linux on it or something, actually load software on it. I mean people want drop shipped data centers as supercomputers. I mean, so when you talk about a density node, what pops in my head is, "Okay, I might have a power constraint or I have a location, I need a fully encapsulated system that just plugs in." I mean that's like a boatload of servers, some network fabric, full stack. All those things have to be built in. That kind of sounds a little bit like just drop shipping a machine. I remember Amazon came out with, what were they called? Outposts, remember those days? I had that concept of just having replica of a cloud. We're kind of seeing what you guys are doing and seeing, okay, the edge is going to get smarter, there'll be some form factor, some may be bigger, maybe the size of a old telco box at the side of the road or the size of a camera on a pole where training and inference will have to happen. So as you look at that distributed computing paradigm expanding, what do you guys see as the practitioner angle? Because you have infrastructure people, there's not a lot of talent out there. I mean they claim there's IT talent, but that's IT. That's not system designed like what you guys do. Then you got the chief AI officers out there going, "Let's go get some GenAI, give me some magic dust and build me some CapEx." Not that easy. So take me through how you see the evolution here because it's going to get distributed, there'll be big mega centers, SUPERNAPs, and then there'll be like mediums and then smalls, density nodes, whatever you call it. Take me through that.
Jason Hoffman
>> Yeah, the question really is what is small and is small really achievable in that kind of sense? And I say this again as somebody who's done a cloud computing company, I did an edge computing company, the Google acquired, now we're sort of building these large scale AI factories. I build mobile networks. I've gone as->> You've done a lot, that's why I asked.
Jason Hoffman
>> Yeah, so in the case of having worked on 4G and 5G and the sort of overall architecture of it, I mean I would say that the comparable CapEx from what's being deployed right now from an AI perspective was really done building out 4G and 5G networks. Meaning the deployment of 5G networks and sort of going into LTE advanced globally was spending on the order of one and a quarter, one and a half trillion dollars to sort of go deploy that across every country on the planet. We see sort of a similar thing on the more non-technical side of it that occurred over decades was the deployment of air conditioning and refrigeration globally. That's roughly about 220 gigawatts of things. In the US, it's about sort of 80 gigawatts, if you will. If you look at the AI footprint that people are building into the US, they're building into a footprint that's about the size of air conditioning and refrigeration. Air conditioning and refrigeration is why the grid was really built out post World War II in a lot of spots in sort of the world. And these kind of build outs, I think people sort of underestimate the size of them in a lot of ways and what's sort of required of it. And we'll get down to the distributed part of there. In the case of Switch, we're headquartered in Nevada and we have a very large footprint in Nevada. As we continue building out that footprint because in Nevada we also can function as a behind the meter unregulated utility and the like in there. I mean I think fully built out in Nevada we're something like a third of southern Nevada's power, two-third of Northern Nevada's power and we have a total power footprint that's like the equivalent of Los Angeles. And so if you think about what it takes to go and build total infrastructure like that->> So you're comparing it to the 4G, 5G build out CapEx equivalent, but at data center scale.
Jason Hoffman
>> I mean it's kind of thing where these AI campuses we're developing are right now a top 10 industrial site historically. And then I think as we start making them even more dense over the next five to 10 years, they'll surpass anything that we ever saw from large scale oil fields, refineries, smelters, sort a number of things like that. There are in the US for example, outside of a Gulf based refinery site and so on like that, if you go and you look in various states, the AI workloads are the largest industrial always on power sites. That's basically sitting in those kind of areas right now. And so when you start thinking about minimizing those, the question always about how much is required in the middle mile if you will, is going to come down to how powerful a device like this is and then how powerful the centralized compute is going to be. And so even the sort of, quote, unquote, "edge" conversations and where inference is going to live, I would say that it's going to live on the device and it's going to live in very centralized data centers. The things that are in between are still going to stay relatively lightweight. Because if right now it's already a challenge to get 700, 800 kilowatts into a tower around sort of base stations being present there, we're not going to be able to stick a two megawatt rack at the bottom of that tower or do that kind of thing. These density around these systems are technically very challenging to go and distribute. I mean a single rack has the same power footprint as most traditional enterprise data centers have in total. And so a lot of people have a four megawatt or eight megawatt data center. Well that's just four racks in the future of some of this gear. So how exactly is that going to work in a very distributed footprint? And so increasingly it's like a lot of the inference ends up happen on the device and ends up happen in very centralized data centers. It ends up following I think a lot of the same curves that you see in gaming and sort of other stuff. You have a very powerful PC and then very centralized computing doing these types of stadia in the middle is just not something that economically penciled out over time.>> Explains why Texas likes the AI. They love their oil too. As you said, it's oil fields down there. I love that point because I think to me right now at the edge, inference is how it can infer at the edge makes a lot of sense, but training is coming in. Edge devices get new data all the time. So this comes back down to your point about the architecture of the middle mile or distributed nature of the mega-centers, which I do buy that with you. I would see that happening at mega centers.
Jason Hoffman
>> Well, if you look at the largest footprint of neural processors, it's on device. It's not in a base station or in an old central office or in a sort of regional data center and the like. And most of the experience around these things is such that you're typing into a ChatGPT and it's printing out. It's not something that has to occur within 10 milliseconds or 20 milliseconds or that type of->> Well talk about networking. Obviously you have a networking background as well. Obviously at Switch it's highly interconnected. Interconnects with the new thing. Jensen Wong was on stage last year at GDC and he said, "KV Cache is the operating system for the AI factory." And I'm like, "Oh, he said it, someone finally said." I've been saying on theCUBE, what runs on an AI factory? What OS? What is the OS? In the very elementary computer science definition, you had an operating system, it does a function, it runs hardware and it makes stuff work. I'm oversimplifying obviously, but the question is what is the OS for these clusters, these large-scale AI factories? And it's interesting, he said that and I'm like, "Oh, it's networking what brings systems together." So, one, what's your reaction to that and is networking the OS of the AI factory or everything?
Jason Hoffman
>> No, but I'd say that in general, and it goes to a AI factory, even if we anchor it in traditional data center metrics like PUE or sort of other things in there, it's going to come down fundamentally to token per watt. And then if you basically go to the next sort of thing of, "Okay, we have these systems that are sitting there." A data center is responsible at the end of the day for powering and cooling, say a given rack of these systems and we're trying to accomplish a certain token output per watt. When you start looking at the data packets through a network stack, then on the model layer you're trying to accomplish sort of token flow through these attention blocks. When you go into the data center analogy and you think of virtual memory management and paging and putting things in and out of a register, which is what typically an operating system does. So if you think of what an operating system normally does is it boots up a piece of hardware and it's fundamentally doing virtual memory management and taking things in and out of there, putting things on and off the CPU. Well you would say that a model layer equivalent to that is KV Cache allocation of eviction. You could absolutely say that. And then when you look at operating system address space expansion and what sort of occurs there. That would be context window extensions present inside of a model. When you start thinking about distributed operating system kernels, that would be the sort of multi-node sort of cache sharding and those kinds of things that you'd see in these models. And then these sort of long-lived processes with state full context and stuff starts being these kind of streaming inference that you get with sort of persistent key value stores and that types of things. And so I'd say that->> I think you just summarized it beautifully. I think one of the things I was talking about, I want to get your thoughts on the, I used to call it the holy trinity of computing, storage, networking and compute, but I want to add a fourth pillar in their database. Database is now key part, you mentioned key value store. The role of the database isn't just the software abstraction now, it's actually part of the data piece of the networking, storage and compute. It's almost a fourth pillar of that OS dynamic you mentioned. Some of the things there, but this is distributed computing in the modern era, this is the science behind that and this is what people are trying to figure out. What is the ideal science for that environment and then how do you optimize it, design it, design optimize, run. Any reaction to that? Any thoughts?
Jason Hoffman
>> No, databases at the end of the day sit on top of storage. It's sort of in there. It's just another abstraction layer. I mean a key value cache is technically a type of database that's in that. And so the analogies on these hold, I mean I'd say that in the case of AI, the sort of vector databases are particularly important of course for keeping state and then these retrieval caches and feature stores and other things that are in there, these are essentially the semantic memory of these kinds of systems. They actually give a model any sort of persistence of knowledge, grounding, recall anything beyond the weights of the model. And so if you go from this kind of weights only model you're interacting with that knows nothing and you need to start using it from an enterprise perspective, then yeah, it's where that sort of knowledge sits.>> All right. Let's talk about Switch. You're the chief strategy officer, by the way, you have a great pedigree of experience. As you talk to customers, what are some of the things you guys are doing that you would say illustrates where the wave is right now? I mean obviously you got the data centers, a super center, SUPERNAPs, but also there's a business where you guys are actually deploying on behalf of customers too to run operations. Is that part of it? Where are you getting pulled into? What does your 3D chessboard look like?
Jason Hoffman
>> Yeah, I mean, Switch is unique in the data center space in that we design, build and operate large scale campuses where enterprises can come and get the benefit of a scale, though it would be very difficult for any enterprise to get. And campuses that are large enough that even hyperscalers go, "Oh, we could build a region there." So for us when we talk about exascale campuses, these are things that are typically designed to start out in the three to four gigawatt type of range and are capable of scaling beyond that. And where we can go and do literally terraforming utility development, build out the network, sort of do everything that's in there. And so we function a lot like the nuclear-powered aircraft carrier at the data center space as how we think of us from sort of the fleet in that. And then I would say that we are founder run. Our founder is a very good structural mechanical electrical engineer and has been doing that in the data center space for the last 26 years. We've gone and added a lot to the team of myself and others that are experts in distributed systems and software and bringing new sensors and so on sort of like that. And so a lot of what we've been doing is how do we go and digitize all of these types of things, our own design process, the construction process, those sorts of things in there. How do we do a lot of prefabrication and manufacturing beforehand? How do we introduce a lot of optionality? How do we make this more modular? And so what we function with in our largest partners is it is our job to go ahead and have a very future-proof long-term roadmap and design around what we're doing there. And making sure that those are aligned with folks like Dell that are doing the sort of total systems, people like NVIDIA, even sort of the semiconductor manufacturers and people that are sort of going and doing that. And making sure that if somebody goes and deploys these kind of leading edge systems right now, that they'll still continue doing that in our facilities 10, 15, 20 years from now just has been the case historically. And so we try to be that exceptionally knowledgeable design partner and people that are then capable of building and operating and having those things continuously flow together.>> I just had Michael Dell on, he'd be pleased to hear that. I mean you guys are like a colo on the nuclear-powered aircraft carrier. Can I have some of that?
Jason Hoffman
>> Well, our Austin campus, that's part of his old campus, so our data centers are nestled within those buildings sort of in there. And Dell is who we get our own GPU systems from.>> Yeah, Dell's kicking ass. And he said on theCUBE that they deliver the goods and that's maxed out immediately. So you're seeing the demand cycle heavily. Interesting view on the campus, I think that's almost a real estate play and look at data centers that start out as real estate investment trusts. Now they're full on supercomputing, so you can almost see the dots connecting as you lay that out. Large campuses have optionality. If they have power and utility, why not turn that into a factory?
Jason Hoffman
>> Yeah, yeah. Look, and we do this thing where I would say that, go ahead and be maybe a tad arrogant for a moment, but we're probably the most advanced user of NVIDIA's Omniverse and sort of digital twin efforts that are there. We have some of the largest models that have ever been done in that system and we sort of look at how we've gone and digitized aspects of the power chain and so on in there. My goal is to really go and support Rob and the broader team so that when we engage with our customers, we can basically take our team and put it together with any sort of technical infrastructure team in any company in the world and basically be a credible technology partner.>> I mean you guys are so smart. I mean you guys pioneered and been following obviously Switch from the beginning. I remember when you first got the high bandwidth there. I think we got a tour in Vegas early on. You guys are thinking material scientists as well as not just computer scientists. It's a lot of engineering and the upside potential's huge because now you are solving the problems. Now NVIDIA's Omniverse is interesting because you're getting real data, they're trying to supplement with synthetic data. You're a huge advantage for NVIDIA's Omniverse, you're pulling in massive data and your digital twin.
Jason Hoffman
>> Yeah, no, absolutely. Yeah. Well, on a means to that our own compute sophistication and footprint and our own deployment of these GPU systems internally for our entire design chain. How these things get commissioned, how they're operated on a go-forward basis, how do we do everything from sort of a capacity reliability to so on has become really using every possible modern tool that one can use.>> Well, Jason, great to catch up with you. I'm so glad we connected. Definitely want to follow up with you and put you on the roster for hitting you back up because, one, the work you guys are doing there, what you've done, your unique experience. You're seeing all the theaters that you've done body work in your career, telecom, networking, cloud, Switch now. So you got a unique perspective. So we really appreciate you spending the time with us here in theCUBE and the NYC Wired community. And of course we're in our New York studio and NVIDIA stock prices looking good or Tesla.
Jason Hoffman
>> Thank you NVIDIA for being up 16.8 x since 2020. My only portfolio regret is that you weren't 100% of it.>> Absolutely. Everyone's feeling good. It's a great wave right now. It's almost like for us as we look at this next generation, it's a generational moment. It really is. Just your thoughts on that as you look at the industry right now, it's like, "Wow, this is highly accelerated roads of super intelligence," whatever you want to call it.
Jason Hoffman
>> Yeah, well it's a part of a continuation in many ways. I mean, if we go and we think of... it just depends on how far back do you want to go. I mean, if we go all the way back to literally the printing press and the first Gutenberg Bible and the Reformation, well, large language models are a distillation of language themselves. It's interacting with language and utilizing that in sort of different ways. So you can think of it as maybe the culmination of the last 500 years of technology efforts or the like. But I think for me, look, as you and I first met 20 years ago as cloud was basically kicking off. When cloud was taking off as a trend, it wasn't like something that impacted the world's largest companies. So recall back then, the world's largest companies were oil and gas. You had telcos in the top 10, that kind of thing. Then when the iPhone comes out and mobile phones come out and you start getting social media on mobile phones, the sort of effort there around mobile, it's like, "Yeah, mobile impacted consumers." And you had the consumerization of the enterprise and so on like that. But now we literally in the case of AI, we have a technology, a sort of or however we want to call it, but we basically have a thing going on right now that directly impacts the world's largest companies. And so we're sort of seeing is honestly something we haven't seen before in that. Because you have sort of like, "Hey, these are the top 10 or 15 companies in the world, and then this is a new technology that it's potentially disruptive for them. It's also the biggest opportunity they've ever had.">> They could win too. They get paid to focus on it.
Jason Hoffman
>> So everyone's going. And that was not the case with mobile networks. It was not the case with cloud when it rolled up. It was not the case with even sort of the PC. It didn't have this type of thing of like, "Oh, we must do this in order to there." And so you're seeing an unprecedented amount of money being spent for sort of this >> And the intellectual capital, I mean I've observed that on that same note where the generations of old and young are interacting. It's not like the young beat the old, there's so much because it impacts everyone. It's the connections. Everyone's in. Everyone's in the game from day one.
Jason Hoffman
>> No, no. And again, being here in the Bay Area where I live in Los Gatos, it's like we got a Google head of something that lives on the corner and another one sort of there and another one's sort of there. And the local barbecues, if you will, include old and young and sort of this and that. But it's from all the places you'd think. And it is an area of just sort of like a tremendous intellectual interest right now.>> Now's a positive interaction and relationship. Jason, thank you so much for coming on theCUBE. Again, great to see. You keep alumni many times on. Really appreciate your friendship and thanks for taking the time.
Jason Hoffman
>> Thanks for having me.>> All right, cool. I'm John Furrier here at the New York Stock Exchange, theCUBE studios part of our NYC Wired program. Of course, we've got our Palo Alto studios, we've got all the events. As AI goes mainstream, it affects the entire society at large, the planet, and even in space. This data center is going to be in space soon. Who knows. Thanks for watching. We're doing our part to bring you the linguistics and the data. Thanks for watching.
