From KW to GW Scaling the Infrastructure of the Global AI Economy

having the complete sovereignty in terms of AI and not just the platform. I think India being so aspirational and ready to adopt new technology for the welfare of themselves and the welfare of the businesses, I think we would be the hub of AI development for the world. You will start seeing that happening in a few months, not years.

It’s actually heartwarming to hear that from someone who’s actually fronting India’s AI story at the moment. Nitin, as someone who is at Google, how do you see this for India? Do you think India has the right infrastructure, the right resources to build its own sovereign AI at the moment?

First of all, thank you, Corvo team, Ankush, for inviting me here. And, you know, I’ll be very happy to share my views. from Google perspective and from my personal perspective I feel yes sovereignty is very important but at the time with the sovereignty it is not a question between sovereignty or innovation it is sovereignty and innovation they have to run together they can’t be one choice versus the other one and with that Google while we have our entire data centers in India you have heard three months back we announced that we are going to be building big data centers in Vizag the announcement happened so we are ensuring that let’s say if any innovation and any data residency things are there they are being kept within the boundaries of India But then those data centers are definitely empowering the lot of AI, but they are for everyone, for all type of personas, whether they’re government, enterprises, startups, students, colleges, universities.

We understand that, you know, sometime there are going to be critical data which needs to stay even more secure. And for that, Google has created a completely indigenous data box which completely stays inside the customer premise and is fully powered by AI. So imagine that you have the full potential to run what you’re running in a Google data center, but inside your own premise. And that has full Google Gemini AI services. And that’s the definition we have for sovereignty, where you are. Also. Also controlling the hardware, not only what’s running on that hardware.

All right. So this IRCTC is one of the most heavily used websites in India. my data research says close to 50 million users visiting every month on an average correct me if I am wrong but how are you incorporating or leveraging the use of AI especially in peak periods when you look at say tatkal booking time when the traffic actually dramatically peaks up

yeah so we had tremendous mismatch of demand and supply as far as railway ticketing is concerned so we have the peak the morning 8 o ‘clock when the ticket is opened for the 60 days hence travel then 10 o ‘clock the AC tatkal and 11 o ‘clock the sleeper tatkal so there is a lot of huge demand and there is a demand supply mismatch as of today so people try to misuse and use the automated tools for accessing it So this is a constant, I would say, cat and mouse game we are sort of playing. And we are using AI also. We are using AI of very advanced AI solution. Maybe they are said to be the best in the world solution we are using

Any indigenous models are used?

Indigenous, of course, we have a layer of indigenous. There is a startup also who are doing the analysis, data analysis, and they constantly monitor the social media and see what is happening, what is the strategy. So it is basically a collaboration between the Indian startups and the global technology strength of a global company. So we are using AI and ML -based model. The model constantly learns and tries to… mitigate those automated…

Okay. Ankush, what differentiates Bharat GPT in terms of its vision when you compare it to say global models like ChatGPT or even Gemini and especially how is it curated for Indian citizens and enterprises? What is that differentiating factor?

Yeah, see, our tagline is AI with purpose and trust, right? So whatever we are doing, so I had read that book, Seven Habits of Highly Effective People very early on in my career. So begin with end in the mind. We always think what’s the use case? What’s the problem you’re going to solve? And then see what kind of model you need, tiny, small, medium, large. And then you see, okay, from where the data would come. See, the Bharat GPT family of models, right? It’s not the large language model. It’s not ready for consumers yet, right? So we work with our partners, get their data and train the model for their users because we believe we…

is easy for us to solve the problem of enterprises because the enterprises say like IRCTC, they already know their domain. We cannot learn, right? And if you say, hey, I can create travel AI solutions, very, very difficult, right? So they know travel, they know railways. So it would be, I think, much better to work with them, learn from them. They already know, they are already solving a lot of problems and they also know the problem, the real problem. They don’t have existential crisis, right? So they are not just in the game of valuation. So they are solving the real world problem.

That’s why we have him on stage with you today. He’ll share those precious tips. My last question, since we are running short of time, Nitin, I think this is also not to highlight the achievements, it’s also to perhaps highlight the concerns. And right now, one concern which the Indian Prime Minister has also highlighted is that of inclusivity. How is Google trying to bridge that divide as far as you can see? As far as digital divide is concerned, how do you make Google more accessible for the underprivileged, for those in rural areas? Nathan, before you answer, it should be shortened. I have my colleagues from other team, Vertex. I would like to apologize to them for this delay, but allow us just to wind this up.

Yeah, I’ll take a minute. Okay. So, great question. And, you know, Google has always been, you know, in the forefront of inclusivity, whether you call it Gmail, whether you call it search. You know, it is empowering billions of users every day. And just to summarize and give a recent example, we have very recently Sundar Pichai has announced that the JEE main exams, mock exams are available on Gemini free of cost for any student to try. That’s the inclusivity we want. We want to make sure that student at his home can keep on trying the mock test at free.

All right. Amazing. Amazing. Which is inclusive. Inclusive and democratic. Many thanks to you three gentlemen. It was a pleasure having you all over here. Thank you so much.

Good morning to all of you. As Rakesh has already introduced, two companies are planning for a lot of things together. As I said, I am part of strategy for India and managing Vortiv strategy and market development. The important thing, what we are bringing today for you is, as we see a lot of gigawatt infrastructures are getting announced, and that poses a lot of challenges for us. Till this time, data centers are getting built from outside in approach. and then now time is there or time has come where data centers are getting filled from inside out approach. So it’s first GPU gets decided or the workloads get decided and then the whole infrastructure gamut comes into the picture.

To discuss this, I have two friends, two industry veterans from Vertiv and NVIDIA to discuss the Fireside Chat. So we have a Jigar, I think by this time Jigar is already known to the industry because immense contribution Jigar has done for the AI ecosystem in India working with all the ecosystems, all the layers, infrastructure application, use cases and so on and so forth. He managed solution and engineering for India in NVIDIA and I have another my friend Peter Panfil . Peter is Encyclopedia in Vertiv. He is based in US. He is our senior vice president for technical business development. And he’s the one who’s involved into many designs, a large scale data centers and gigawatt designs.

I would request Jigar and Peter, please come on the stage. Let’s have a round of applause for Jigar and Peter. So, Jigar and Peter, it’s all yours now. Go to

Thank you. Thank you. Thank you. So, my friend, we got our introductions. Let’s see. Are we on? Are we on? You guys can all hear us? Yeah? Can you hear us? Good? We’re good? Okay. All right. So, my friend, great to see you. Great to see you. So, I got to start with how we would normally end. I believe that any discussion like this should start. with us telling you what we think you’re going to get out of this. So what key message or messages do you think this audience needs to hear before we get started? And then we can spin off of that and go into the kinds of details we really need to.

So where do you think, what do you think this audience is the most interested in?

Okay. Am I audible? Okay, great. So I think as the topic suggests as well, my view, what you will get to hear us next 30 -35 minutes is about why this AI is becoming so much of notion for every country. What is that is the building blocks of these AI factories and the sovereignty aspect of it? What is it two of us are trying to contribute in this journey for everyone for that matter? And how do we scale? and make it work for everyone, to make AI for all, how India wants to call it as, AI for all, is what I feel we should be discussing about here. Because that will be most relevant for the conference, for the audience, and what we can contribute back to the humanity as well.

What are your thoughts?

I agree with you completely. So the three things I feel are most relevant are speed at scale. Now, it’s not just the speed of the compute. It’s the speed of deployment. The faster we can get the GPU structures in place, the faster we can benefit from it. And scale, you and I talked about the scale. And you’re going to quote some numbers, I think, that the tops of their heads are going to blow off. But speed at scale. The second thing is we’ve got to stop. We’re not thinking the way we thought in the cloud world. In the cloud world, we were thinking a high -density rack was 10 kilowatts. And that we would start at the source, at the grid, and work our way to the chip.

What I’m here to advocate for you to do is start at the GPU. Start at the chip. Let’s start at the chip, define the most economical, most efficient, fastest from a compute perspective, and figure out how to deploy that as a pod, then replicate that pod, and achieve the speed. And the third is, don’t be scared. We got it covered. We got you covered. We know how to do this. This, we made a big, I got to just tell you, I told you this in the hallway. Vertiv made a big bet. with NVIDIA. We made a big bet. I actually reassigned myself. I was leading what we call a GSA, Global Strategic Account Pursuit Team, and I said, if we’re going to do this right, we’ve got to immerse ourselves in GPUs, understand how to deploy them, understand what drives our customers, and how we’re going to make them successful.

And I think that that has worked to both of our benefits.

Absolutely. And through the humanity as well, right? We are fundamentally changing everything that has been pursued so far, and you bring out the cloud part of it. I was just thinking while putting my hand on my beard that only a few hairs were white back then. It’s not that far that I’ve seen the retrieval clouds. We store the information and we’re just retrieving it to process the application to get us the information out, right? to the world of now generating every single time a new data and processing it right there to give you all the time a new input and a new output, right? Because prompts are new, the outputs are new, and thereby the world sees every time something different which is getting processed and being delivered to the customers, right?

So such an amazing and a fastest -paced change of how these clouds have emerged and what are your thoughts in terms of what this space is all about, how our customers are keeping up with this, and what are we contributing in that journey, if you can throw some light towards that.

Sure, that’s great. So first of all, it comes with understanding and having a transparent provider that says, here is what I’m producing today, here’s what I think I’m going to be producing a year from now, here’s what I think I’m going to be producing two years from now. Now, our goal is to make every deployment that you take on an AI factory. We all know what an AI factory is, right? An AI factory, think of it as a car factory, washing machine factory. Just, it’s a data factory, okay? And so our goal, I will just tell you, our goal along with your team is start as an AI factory. Yes, you might want to have mixed mode CPU and GPU workloads in your facility, but you’ve got to pilot the GPU configurations, at least pilot them.

When I say I reassign myself, I was working primarily with cloud providers, mostly hyperscalers, and they had a prescriptive formula. You know, they had their hacks, their number of racks. They would deploy them. We all knew which ones they were. Now, we can take a GPU pod, design it once, build it many, and apply it to the GPU that we need from that generation. It’s a complete change in the way we think about how to deploy the IT.

That’s so true. By the way, did you notice, every time we are talking about GPU, the screen is blinking. There you go. I think that’s a good message.

I think it’s because I owe somebody a nickel every time I use the letters GPU. It must be trademarked somewhere, all right? So I owe them a nickel. Okay, all right.

No, so I think the transition that we see because it’s generating something new every single time, the compute demand because of which is just exploding, right? And thereby, the possibility of what we could do more and new is every time becoming bigger and better, essentially. Right? And with that, I think the journey of data center is also evolving much more faster than what we have thought, right? You mentioned it, 10 kilowatt to 15 kilowatt, not that far. We were talking about this about four or five years back. To 40 kilowatt, what we transitioned it to it, to now to 120, 130 kilowatts. And as we announced it in January, we are now talking about 240, 230, 210 kilowatt per rack, which means this size hall could probably run a great portion of India with so many services that is probably never imagined before.

So I think it’s interesting that you comment about that, because one of the things that we’ve heard back from our customers who first do a lot of research, how do they take their critical infrastructure from CPU -based to GPU -based? And I think that’s something that we’re seeing a lot of growth in. First, there’s that transition to liquid. Don’t worry about it. We’ve been doing liquid cooling for 40 years. We know exactly how to manage it. Then there’s the density of the compute itself. I’m amazed at how quickly and easily our customers understood the move from a 10 -kilowatt rack to a 130 -kilowatt rack. I credit you all. So if you’ve already made that transition, I credit you.

You’re doing a spectacular job. Our job is to prepare you to have that go up by an order of magnitude. Not right away, but in future generations of compute. And so what we try to do is we try to prepare you for future -ready thinking. I know you don’t want to think three years down the road. You can do it. You can do it. You can do it. You can do it. You can do it. You can do it. let’s at least think three years down the road in three years based on the rate of what you’re seeing what we’re seeing both here in India and around the world

my perspective is I think all reports are talking about 5, 6 gigawatts kind of a number over the next three years my personal understanding from the lens I look at it both from NVIDIA as well as what industry and government is trying to do my anticipation is we will cross 10 to 12 gigawatts in the next three years and that’s not far and I’m not going by any of the announcements that has been made in the last three years. I know where the reality stands in terms of what inferencing and training workloads. I repeat, I started with inferencing. I did not start with training.

Yep. I noticed that.

The reason is we are a consumer country. Make a note of that. Yes. Right? He started with inferencing, not learning. Yes. All right? Because we are a consumer country. We have always been in the mode of first to consume, then to build. And thereby, we are the largest chat GPT consumer base for the globe. We are the largest for public city. We are the largest for even for Gemini as well. Right? I think we were number two about a month or so back. But my view is with this geo announcement, we should have crossed by now number one position. But the delta was pretty small. Right? What does that mean? That means. If that entire compute capacity.

that is currently not getting processed in the country should come back to India because of the DPDP law that has got enforced last month or so, then this number will be even higher. And we are very democratic that way. You know, we are not closing the doors for any businesses. We have never done that. I’m sure, knowing the country, we will never do that with this leadership that we have from Prime Minister Modi. That means we will still allow these processing to happen outside of India, but at the same time, we will do the regulatory reasons of some of the verticals, say, fintech, say, healthcare, defence and so on and so forth, or some of the government, you know, bodies.

Even if they start to do influencing locally, this number will easily touch 10 plus. And I’ve not included the industry at scale yet, which is what Anthropic and J &J, Gemini and others are trying to even capture it from that market perspective. So my understanding, it should cross 10. while all the reports are talking 5, but India will

So it’s amazing. We didn’t compare this note before we got on this stage. What was the number you gave me just 20 minutes ago? 10, right? So let’s think about that just for a second. We’re at 1 .5 now. We’re going to get to 10. So to get to 10 in that three – to five -year horizon, we’re going to have to scale pretty far, pretty fast. We’re going to have to draw on our shared expertise. And by drawing on our shared expertise, we’re going to help be a trusted advisor to you. Who’s your trusted advisor? I’ve got my trusted advisors. Somebody. Somebody I can always go to, and they’re always going to give me. the right answer. It might not be the answer I like, but they give me the right answer.

So what we want to do is make sure that you know, we understand how to scale. You understand how to scale. We understand how to scale. I think if we’re talking about doubling, getting to 10 in three years says we double every year starting this year. My one and a half goes to three, my three goes to six, my six goes to 12. We’re doubling every year. Now, if I was to take you outside of India now, North America market, when the North America market first started becoming aware of GPUs, there was a wide variety of acceptance. There were the folks that said, yep, I want to be there. And I want to be there, and I want to do a pilot with you.

I want to design a pilot that I can replicate into all of my either hyperscale or multi -tenant data center environments. The other thing they wanted to do was no data center left behind. They didn’t want to leave behind any capacity because they knew capacity was going to be the currency. They knew power and land and GPUs, that’s where they needed to be. The third thing was their project scales moved. We used to live in the cloud world at project scales of 18 months. We live now in the GPU world of project scales of between four and six months. So a dramatic compression of schedules, a dramatic increase in capacity, what does that mean? We’ve got to build capacity at a faster rate.

Than we ever have before. and I know we’re up to it. We’ve added the capacity we need to be able to support that kind of demand.

Peter, that actually brings to a very good question. When we talk about this at scale, and you said that in the U .S., you guys have already started to build this at scale because you see this as a great opportunity, essentially, and India is yet to build, right? In all fairness, I think some of the largest clusters are in 10Ks of GPU, essentially, right? While in the U .S., we’re talking about millions of GPU in a single data center, essentially. Would you like to throw some light on some of the learnings and, Kodi Kakar, a quick bite for the audience to know what are those quick things that India could do in terms of having these things done in, let’s say, three to eight months’ time frame, not just the project planning, not just the understanding of BOQs, not just the understanding of…

who is going to deploy my project and how does the project look like and the 3D version of that. How do I get the entire project done in, let’s say, six to eight months’ time frame? Including from land is what I have, and from there onwards, GPUs running and hugging and making the production environment happen.

shifted to 250. Now, along the way, we said, okay, let’s take these 10s and put them together and make a 50, and let’s take the 50s and put them together and make 100, and let’s put the 100s together and make a 220. Shoot me now. What we found is, let’s pick an optimum building block that supports the number of GPUs that is, I’ll call it, reasonable at scale, don’t take a design that has never been created before. Let’s take a design that we have a good basis on. For example, the pod. You just published some standards on pods. Reference designs.

Reference designs, okay.

We worked closely with your team on reference designs. We came up with the magic numbers that are reference designs that minimize the amount of, of underutilization, so maximize the utilization and make them the most efficient. I’ve been an advocate for efficiency within the data center space my entire career if you save a watt that’s a watt you don’t have to generate at the source you don’t have to distribute it you don’t have to reject it so the fewer watts you lose and the more watts you can put into the compute the more tokens I can generate and so our goal our goal in working with the GPU I’ll call it the AI factory mentality is how much power can we deliver from the source to the GPU as much power as we possibly can and how can we deploy that physically as quickly as we possibly can and it boils down to take the reference designs we’re not saying all the designs are going to be the same we know that’s not going to be the case so but I could show you a pod design it’s part of the reference design I could show you a pod design that supports three generations of GPUs so this year, next year, next year after that three generations of GPUs just by changing the way those pods are populated on the compute side and in fact we’ve got one customer who wants to be able to seamlessly mix GPU platforms within a pod he says I’m going to have one compute line up number one as one generation of GPUs, pod two is another generation of GPUs third generation of GPUs, so they want to be able to seamlessly move between GPU generations because at some point they’re going to optimize particular functions and particular outputs and services against a GPU platform.

You just brought up a perfect point, right? So a few things are why it’s important to follow the reference design. Just to bring everybody on the same page, CPU world was very different. Having a node down means few hundred dollars getting downtime. A GPU node down translates to few thousands of dollars going into a downtime, right? And the fortunate or unfortunate part is if your training workload is running, if a node fails, you start from the checkpoint that you have done it. Assume that your checkpoint was done eight hours before. For eight hours of, say, 4 ,000 GPUs of time multiplied by that much is what you have lost the compute time in the cloud. Unfortunately, that translates to…

Real money.

Hundreds of thousands of dollars. Real money. Right? Real money. So while you as a cloud provider might be thinking, and I’m talking about both the sides, that, hey, let me do a little bit of cut corners, do something here, something there, and I’m still making the cluster up and running. But you know what? That’s going to cause a lot. And customer may not have SLAs with you in that direction because these are not the standard SLAs we’re talking about, right? What the world has seen in the typical cloud world. These are different type of SLAs that customer signs with you. And, you know, if it’s an inferencing workload and if it’s critical with the enterprises, we’re talking about down times, which is, again, by all law of cloud, is not acceptable.

But the key question could come in that, hey, why do I need these large -scale clusters only for training? Is that the only thing I do it? The answer is no. I don’t know, but I’m sure most of you might be following what Jensen talks about it. The three scaling laws that we have it. We’ll not go. We’ll go into detailing of it. I think Jensen has mentioned it like at least 100 times of his keynote. But in a simple term, if I have to tell you, let me take one or two good examples from the country itself, right, and what we announced in the last three days. So, taking a very simple example, as everybody knows, we are 1 .4 billion people, right?

Half of the audience is associated with farming in the country. Half of the audience or the, you know, citizen base is associated with the, you know, farming, and thereby one -third of the families of the country are completely aligned to the farming aspect of the story, right? They contribute just 15 % of our GDP, but half of the population does and associates with the farming, right? Now, government of India has two simple applications that has been launched, right? One is to check the subsidized… Food, you know, that government gives it to these half of the, you know, audience, half of the citizens in the country today. subsidized to the level which is a cent or two. In Indian rupees, it is one rupee to five rupee, how government gives it.

And a feedback call goes to all these citizens, asking how was the quality, did you get the right quantity, have they done any kind of fraud, and so on and so forth. A call per day, if government has been able to scale in the last one month to about 50 ,000 calls a day to citizens through a bot, which is talking in a local language, has been able to save a fraud work of around per day, and I’m talking about per day, in the range of a couple of millions of dollars. Okay, take that fraud. Talk about financial fraud. This would be another one, right? Because we are the world’s largest online payment transaction country. We contribute 50 % of the digital, and that’s…

by the NPCI data, globally acceptable, and it does at free of cost in this country. We call it as UPI, right? Most of the Indian people would know about it, right? And imagine the innovation that takes place in the fraud, you know, when the UPI transactions are taking place, right? And I do this transaction using mobile, from your mobile to your mobile, in a fraction of milliseconds. That data is in hundreds of millions, right? To prevent this fraud is where the AI is getting used. Now, if I’m putting a couple of hundreds of millions of dollars for five years as an initial investment, think of the economic benefits and the money back that I’m giving it to the citizens by not having these frauds.

And thereby for each of these applications, right? I have another good example, which is we have 22 official languages speaking in 500 dialects in the country, unofficial languages. So, officially, we have over 100 plus languages in the country. right government of India has an application called bashingi which does basically the translation you know and ASR and TTS in all different languages of India and government of India and state government has about 10 ,000 websites that government runs it we have only touched 1 ,000 and we are already hitting 100 million requests per hour right and this translates to in a simple term roughly about 2 million 2 megawatts of data center consumption per minute right in in 2 megawatt if I’m able to cater to 100 million requests a minute that’s massive

yeah look at the productivity improvement and I’m bringing it to the nation’s and this is just thousands of those websites of government of India so we yeah so we we talked we started with scale at speed Okay. That’s where we began. It’s not just the scale of the data center environment. It’s the scale of the applications and the benefit that they’re going to bring when they get fully populated.

Absolutely.

So I’m going to put you on the spot. How much do you think, on the journey right now, where are we? Are we at 3%, 5%, 10 %? I will tell you, I cannot wait for AI to take every mundane task I have to do in every day of my life and just do it. Okay? And then once those mundane tasks are out of the way, I can use every gray cell up here for productive work.

Absolutely. Absolutely.

So where do you think we are in terms of that scale?

But you touched upon a good point, how Meta calls it as personalized AI for everyone. So we are getting there, right? But in terms of data, and I think even Minister made that announcement yesterday when he was talking at the inaugural. He gave a nice statistics talking about we as India generate 20 % of world’s data, and what data center capacity the country has today is 3 % of the world’s data center capacity. So which means even if I don’t assume data generation speed over the next 3 to 5 years, even if I keep the data rate at 20 % only, and we are a young population, so we are bound to generate more data, and ours is the cheapest data rate in the world for 5G that we have, but assume that we don’t.

We don’t generate much of data, right, and we restrict it. We still have a long, long way to go in building the… large -scale data centers just to make sure our own data we process it by ourselves. Right? And that’s where the whole theme of this sovereignty, what government is talking about at least let’s protect our data. It’s more critical, not the general data. And that’s where the gigascale is more important.

But I don’t look at the sovereign data center approaches so much as a protection. I look at it as, where’s the most efficient place to process the data? It’s where the data is generated. The most efficient and effective place to process data is at the source of the data. Absolutely. And we are limited by energy. So we want to protect that layer as much as possible. And so I see a world where the data gets generated, it gets processed as closely and as quickly after the generation of the data. It’s used to further improve the performance and generation of subsequent data. So that data gets cleaned up as it goes. It gets more refined and more accurate.

We all know we make good decisions with good data. We know that. We make bad decisions with bad data. So the real issue here is we’ve got to take the data. I won’t say that our data is not clean now, but it’s not. Okay? I mean, you spend, give the audience an idea when a model is being put together. How much of the time is actually in cleaning the data and pre -processing of it, and how much of it actually goes into the language model itself?

So just to build, because India just announced 10 of their foundation models, cleaning of data typically is three to six months of journey on thousands of GPUs. for a language model that we are trying to build it in, right? If it’s a specific model for a particular task or a vertical that we are trying to build it in, and if the data is more notorious in terms of having more videos and images and stuff, it could be even longer.

Got it.

Right? And then comes the foundation model building itself and, you know, conversion of the model. That’s another 6 to 12 months of journey. It depends on the model size and type that you are trying to build.

So it could be a third of the time of realizing my language, my large language model, is cleaning of the data.

That’s correct.

Processing that data. Now, once it’s there, I’ve got a solid foundation of data to use for future models.

That’s correct.

Okay. So, again, I think if we’re talking about it in terms of percentage, are we 5 % there? Are we 10 % there?

So I would not put percentage. The reason is what type of model we are trying to build it in. depends on that. If it’s a language model, I think specifically to India per se, I will not comment about other countries because it depends on where their journey is in their data building. But India, in my view, has already nailed the data creation for a mid -sized model, how we call it, as a small to a mid -sized model in play. And they’re going to make it open source as well, how it has been announced. So, I will not claim it that we have a very large data set for a very large model, but a small to mid -sized, I think in the last one, one and a half years due to this India AI mission, we have been able to generate a pretty good amount of data and a pretty amazing clean data.

Perfect. Alright. So, I’m getting a hook from the guys in the front row. Okay. Yeah. I run long. I’ll always run long.

Sorry, I’m pausing you there. And I want to diverge a little bit right, asking as an India person. and I’m an Indian first, then I’m an Indian, what is that Vertiv is trying to contribute in this journey for, let’s say, India to begin with, and for the globe as well, you know, in the building blocks of these things that we are trying to do for these gigawatt -scale data centers? If you can throw some light.

Sure.

I know it’s a little bit silly question.

No, it’s not a silly question.

No, no, it’s not. We want to push manufacturing. We want to push, you know, India ecosystem, be more indigenized as much as possible, be more self -reliant. I want to know what Vertiv is trying to do.

So Vertiv is investing in people, in process, in production capacity.

Amazing.

Our goal, our goal is to build as much of the critical infrastructure here in India as we possibly can. And what that, it starts with. Working with our partners and our customers on first pilots and then production. So that production, you’re going to benefit. I will just tell you, India, you’re going to benefit from the mistakes that have been made in other regions for the last 12 to 18 months. So you’re going to benefit from that. You’re going to be able to jump right to it, all right? All right, so here’s the sum up. I asked you to give what you thought the panel should, this discussion, they should get out of it. What should they have heard from us that you want them to keep in their minds for the rest of the day?

For the rest of the day?

Rest of the day.

My view is, and I know it’s going to be a mix of audience here, you should be listening to it. How are these building blocks of AI factories that are getting to learn from the globe that India could adopt it fast, followed by? Followed by what’s happening in the modern world, because that’s the most and the fastest. building things that is happening in the world and the most fascinating because changing the world is so fast, followed by how are these models getting deployed, right? And what are the applications which are changing our world on a day -to -day basis? And fundamentally, the businesses are getting challenged on how they have been operating it for decades or centuries, right?

Versus how they could do this business today. If I would be you as audience, and that’s what I’m trying to do, being the audience as well, trying to constantly learn from this conference, what’s that the people who have done this at scale, what can I learn from there that I can deploy back in my country, in my profession, in my day -to -day life as a learning, is what I’m trying to do. And that’s what I would recommend everyone else to do as well.

Perfect. So let me add on top of that. It’s scale at speed. And it’s not just speed of build, it’s speed of compute, it’s speed of adoption. Yes. Second, stop thinking grid to chip and start thinking chip to grid and let the chip help us define what that critical infrastructure needs to look like and the third is we’re going to make it as sustainable as we possibly can because a lot that I don’t waste is one I don’t have to generate transmit or reject alright I think I think you’re up next Any questions? Can we have time to take questions from this? Okay, we have. Can we? Okay. Okay, we have one end up. She’s going to run a mic over to you.

Yes.

Hi, my name is Ani. I have a question. As I can see…

Use your outside voice. That’s what my family always says.

As I can see, everywhere is AI. And in today’s era, it is totally about AI. So, as you also said that this is a… AI whereas everywhere industry and company and education in every sector using AI. So the day is not getting far once AI humans is totally dependent on the AI and once AI is in the subconsciousness as humans thinking as humans. There is any chance where humans and AI both are in the same niche?

So I think that early on AI got a bad rap. It was going to be the computers were going to take over and blow up the earth. That’s not what we’re finding. What we’re finding is that AI makes our life easier. life better every single day. I know that traffic systems in the city that I’m in now use AI to look at traffic congestion and traffic patterning. And they actually time the lights to improve the throughput on particular roads at particular times a day. Now, that’s where AI is going to really benefit society. It’s going to benefit it in transportation, in medicine, in research. I’m not so worried about the data being used for evil.

I’m really excited about the data being used for good because that’s where I think we’re going to get the most benefit.

True. But what if AI get their own subconsciousness? They don’t need humans to just act.

I wish you see that day. Somebody told me when I started my journey with the phone. That’s what is going to happen. You will lose touch with your family. You will always be busy with the phone. And so I don’t think so. We have even touched that level as a surface, even after having this phone with me for 20 years.

Here’s the example I like to give. Do you think about breathing and blinking? No. You do it automatically. So let’s let AI take those autonomous functions and do them for you automatically so that you don’t have to think about them. And then if I don’t have to think about breathing and blinking, then all of a sudden I can use my brain matter to do other things. So many things. So. I look at it as it’s going to free us. It’s going to free us from the mundane tasks that breathing and blinking. Come on, you’re laughing at me. But do you think about breathing? No. You only think about breathing when you’re trying to hold your breath.

Okay? So I think what’s going to happen is AI is going to become to us like breathing and blinking. It’s going to become an autonomous function that just runs in the background of our lives constantly and makes it better. It’s going to learn what we do and how we do it and how to improve that performance and give us more freedom to do what we really should be doing, and that is making the world better.

Thank you.

Thank you. That’s a good question. I’m glad you asked that question. We have one more. We have one more? Yeah. Hey, hi. We’re going to that side. Hello. Big one.

This is Shlom. I was watching the interview of Mr. Jensen Hong from NVIDIA, and he explained AI as a five -player stack, you know, energy -cheap infrastructure model and application. Which layer do you think, he also explained how US and China are working on different layer and how they are, you know, ahead of us many years in different layers. Which layer do you think India can excel with them or match with them in upcoming years?

So, I think we are already doing that, right? It’s a great question. When we talk about sovereignty, these are the layers we should be sovereign, essentially, right? We cannot be importing energy from anybody. We need to generate by ourselves. Otherwise, how will we run these lights and so many functions and how will we power these data centers, right? So, the good news and I think Prime Minister gave this answer so nicely yesterday in his keynote. He explained, no, the Minister said about it, sorry. He explained these five layer cake once again and I am proud to say he made that statement which we all know it. Half of our energy today is generated which is a green energy, right?

So, that layer is sorted. And I and you have a lesson to learn. We have to contribute more of the companies who have to contribute by having solar, hydro and air and other methods, right? Where NVIDIA is trying to contribute to the nation today is on the top three layers, right? We are helping the nation with AI factories. building it with all the learnings what Peter also mentioned. You don’t have to learn from all our mistakes of last 18 months that we have undergone in other regions because they were ahead. India was slightly delayed by at least 12 months or so. But we have put in all those learnings and the factories have come up way too faster than anywhere else in the world, right?

By all means. The second layer is one of the layer is the serving layer when you build these applications. How do you do inferencing? You’ll be surprised to hear Indian cloud providers never had a control plane, right? We were dependent on other nations to give us a control plane to run these cloud inferencing stack. NVIDIA has open sourced that work and shared that with government of India and that was the announcement that Sarvam did with the product named as how we call it Prava, if I’m not mistaken. I hope I’m pronouncing it correctly. And that layer is now completely owned by government of India and an Indian company to do the entire inferencing locally. right?

And the last piece, which is the application, right? I’m sure you would have visited the booth downstairs on the Hall 5. I don’t think that we have left any booth. Every booth is powered by NVIDIA open source stack that we have given it to build the agent -DIG AI platforms and formulation models. That’s the contribution we have done it for the nation. And India is right there. I think what’s missing, and I will fully agree with that, is we are missing with our own chips, right? And that’s the autonomy that every country is trying to drive across. I’m again proud to say that NVIDIA is fabulous. We don’t produce, right? We outsource that to Taiwan and a few other countries, essentially, right?

We have opened up partnerships in many countries, and we are very open to partner with India as well to give away our technology. Thank you. We will do the modifications and do the manufacturing by themselves. That’s the last piece which is left, and I’m very confident with this Semicon mission. this is going to happen very soon, even if we NVIDIA with somebody else.

Thank you so much. The past year time, we’ll have to get into our next session. 10 megawatt, 12 megawatt, and today we have…

Gigawatts. Gigawatts, baby. Gigawatts. Gigawatts.

I just wanted to leave important information. It took about 8 years time to build one 5 gigawatt. And another 10 gigawatts is going to happen next year. So look at the speed and scale. We both have to work together. And as Jigar rightly mentioned, all 5 layers will have a tremendous opportunity to work. Energy, infrastructure, compute, models, application, and so on and so forth. Huge amount of resources required. Huge amount of support required. And very exciting time ahead. Thank you so much.

And it’s going to be a system approach. System. Systems. Think systems. we as an industry have thought boxes for too long. We think, I got this compute box or that compute box. It’s now a system. It’s a platform. And that platform generates tokens. The new measure should be tokens for watt per dollar.

Absolutely. Absolutely. Very well said. Thank you so much.

He’s one of the guiding principles to implement a lot of large -scale data centers for Vertiv or all the entire ecosystems. Let me welcome Srikanth on the stage. A good round of applause for Srikanth. And another gentleman we have from Vertiv. He’s about 35 years of experience in leadership roles in Europe, Middle East, Africa, India, Southeast Asia, Asia, you name the region. He’s been there for many years. His name is Sanjay Sainani. He joined us as a senior vice president, technical business development. He’s the one strategizing all technical strategies for Vertiv and a business development area. Let me welcome Sanjay on the stage. A good round of applause for Sanjay. And I’ll be asking some questions on behalf of you.

I would also open the floor maybe sometime later. Welcome. Yeah, am I audible? Okay, so let me start, Srikanth, from you. Last question first. First, what is the one learning you want to give it to the audience from your experience of implementation when you build a large AI -scale factories? That was my last question, but I want to ask you first. One piece of advice or experience? Experience, out of your experience, because you already have good hands -on on implementation. So from a sustainability standpoint, implementation standpoint, what is one learning you want to give it to us in India when we’re building a scale of the factories and things like that?

Yeah, it’s an interesting question, right? Like when… One year back or one and a half year back, I came to India to review some data centers. And when I was asked to do that, one of the first things that crossed my mind is, wow, India is building data centers at scale? Because when we were growing up, power used to be a big issue. The reliability of the power used to be a big issue. The availability of power used to be an issue. And when I came here, I was amazed at how far, you know, I have been away from the ecosystem for a little bit, but I was amazed at how far things have come in terms of availability of power and the reliability of that power.

And the second thing I was amazed at is also just the knowledge here in the ecosystem as well as everything related to everything from safety to speed of light construction and the product ecosystem has come such a long way. I think the next step in terms of where India is going in this AI factory build -out is if you look at the U .S., it’s a little further ahead in terms of gigawatt scale and high -density racks, deploying high -density racks, high -density liquid -cooled racks. There’s a lot more experience over there. And I think our combined companies have created that experience. Like I’ve been working with WordUp for the last four to five years in the R &D work, engineering work, and then eventually the deployment work.

So now we have actually matured a lot in what we consider AI factories versus data centers. So there is a lot of advantage for India to draw from that experience. Our combined knowledge pool, again, it’s the same company. Whether you go to Europe or… US or India, it’s still Word of an NVIDIA. It has to be a strong cross -pollination between the ecosystem in the US and here, a strong knowledge sharing. And we are in year two or year three of this AI factory build -out worldwide. And as India is picking up pace in this journey, there’s a huge opportunity to not relearn all those hard lessons, or the hard way, but instead share that knowledge, share, you know, our combined teams share that knowledge and build it much faster here.

That’s first, as a thought leader, both sides we need to do that, we need to equip the market for those kind of things. And let me also tell you, on Vertiv’s side, whatever innovations we are doing in the US, we are real -time bringing to India, so that there’s no latency here, and absolutely whatever is going to happen in the US, we want to bring it to India. That takes me to our next question to Sanjay. Sanjay, we have heard about speed, and so far we have heard about speed of clock. Now, Peter, sometime back, and Jigar spoke about speed at scale. what is your thought process about speed at scale or ramp up of infrastructure happening at the speed level what’s your thought process

I mean most of us who are in the space of mission critical applications and then within IT and if you’re dealing with semiconductors we all know Moore’s law and that was pretty much a 10x almost a couple of years in terms of performance and while performance was 10x the energy required to reach that performance was probably 2 -2 .5x every generation so you were getting amazing efficiency in terms of performance because you were getting a 10x performance with 2 -2 .5x kind of additional energy usage and that’s what you saw for the past many many decades and we all thought that Moore’s law is kind of now reached a plateau there’s not much happening … and this is where companies like NVIDIA, working with other semiconductor ecosystem, came up with multi -tiered chip structures.

When you look at today, some of the chipsets, these are three -story, four -story, six -story buildings. If you had to look under a microscope, there are layers and layers of transistors, billions of transistors layered together. And the innovation that is happening now or that has kick -started now is again kind of retracing Moore’s law. So if you look at what NVIDIA is announcing in terms of the new generation of chipset, there’s a humongous amount of performance improvement every generation. While the performance generation is 10x, 20x, 50x, the energy consumption is also jumping up. It’s not 10x, but it’s 2x, it’s 2 .5x. So like Jigar and Peter mentioned a little while ago, you have the current generation.

The current generation at 130, 140 kilowatt per cabinet, while the next one is 250, 260. and the one down the road is 400, 500 kilowatt per rack. And while I don’t want to give away a bit too much, but one megawatt rack is not too far away. People are already testing it. So now think about it, one megawatt of rack. A few years ago, the whole data center was one megawatt. The white space would have 200 racks of five kilowatt each, and you had generators, chillers, transformers, facilities supporting that one megawatt. So the white space was 80 % of your footprint. The rest of this stuff was 20 -30 % of your footprint. Now this has flipped. You have only one cabinet. But you still need all of that.

You still need one megawatt worth of power, generators, chillers, transformers, everything. So in that context, if you see, we are innovating at tremendous speed. If you invest anything, today it’s outdated two years down the road. So that’s number one. That’s a challenge. The second challenge is that it costs a lot of money. Jigar mentioned the cost of a data center may be a billion dollars, or let’s make the numbers a bit more reasonable, $100 million. But the amount of GPUs sitting inside is probably $2 billion worth of GPUs. So now if I place an order today with $2 billion of GPUs, I want to monetize this project very, very quickly. If you build a project, and in olden days we used to build a home in India, not just in India, in most other parts of the not -so -developed or developing countries, would have people carrying bricks on their heads and building a house.

It takes two years to build a home. Now as a homeowner, you don’t see that as a problem. You’re trying to save $5 a year and $2 a year. You’d rather have a person taking a brick on the head rather than bringing a cement mixer because you thought you were saving money. In this world, you’re losing money. because the money you are spending is still going to be the same, probably 10 % cheaper, but your return will start after two years because you will monetize that investment after two years, after three years, because only when you turn on the switch, only when you turn on the tokens, that’s when you make money on your investment. So it’s speed to token.

Whether you spend $100 million or $1 billion, you need to spend it fast, get the factory up and running very fast, so that the token comes out very fast, so that you can get your return on your capital employed. So if you are anyone here who is from the finance industry, Rocky, return on capital employed is a seriously important KPI for money. So that’s speed. And the third is scale. The demand is so heavy. Jigar and Peter in their conversation talked about a few kind of areas where they have high applications. Think of agentic AI as what it can do for you and in how many areas of your daily life it can affect you. The scales are crazy.

And so not only we need to work on the degree of difficulty in terms of density, we need to deploy it tomorrow morning and we want to deploy it at massive amount of scale. And that’s the kind of problem statement or opportunity that we have.

So Sanjay, when you say speed at scale and that’s an idea which you have given because every month or a week save to deployment is going to be a go -to -market fast, right? And generally when you have to speed at scale, you also have to design for scale. And that’s where the blueprint discussion starts. Now why, Srikant, when it is a blueprint, it starts from a GPU architecture or GPU cluster architecture. What is your thought process? When you say… Scale for design, you not have to scale for it first which GPU you want to go with today and then you scale for that. What’s your thought process when you talk about why the GPU has to start with the… Why the blueprint of any data center has to start with a GPU cluster?

Could you repeat the last part again?

Okay, when I say when we have to speed at scale, we have to design for the scale. And that’s where the blueprint of GPU starts. GPU is the first thing we need to start with. And why is that?

Yeah, I think there’s a couple of things, right? When we first started designing, you know, the early phase of AI factories, we were relying on, you know, general purpose built data centers. And we were changing them rapidly into what… They weren’t even really AI factories, but we were trying to figure out how to make it work, right? It was not designed at scale. It wasn’t designed for… It was not purpose built designs. But I think the moment came on us so quickly. And again, NVIDIA and WordUp together foresaw that moment. We didn’t foresee the scale. We foresaw the moment. And we went… We went from very quickly from 10 megawatts to… Now we’re talking about gigawatts.

And so infrastructure doesn’t move at that speed. Infrastructure moves, you know, the design can move at that speed, but someone has to actually build out the AI factory. Someone has to build out the data centers. We have to make so many CDUs. So we were in a phase where we made it work, but we made it work in a very, we had to make it work way, right? If we had to do it all over again, that’s not how we would do it. So now we have a moment where we say, okay, if we were to do it the right way, now we know what the future looks like. We know, that’s why we’ve redefined the data center as an AI factory, which is a fully integrated, you know, where you go from a chip design to system design to the liquid cooling design or the power design.

It’s all, in fact, even the shell and the campus is all purpose -built as an AI factory. So we have to start thinking both in terms of design as well as manufacturing, as well as delivery, as well as operation. We have to think and start thinking about it at that scale. and I think we’ve already started doing that at the design. Like, you know, NVIDIA has a DSX reference design now, which is actually based out of word of, you know, smart run products and large -scale CDUs. So now we have to start deploying it that scale. That is one of the things that NBIS’s focus is, is how do we deploy it speed of light.

Everything from logistics to operations, everything is being redefined. So that’s why we have to, like, you know, you have to think of it as an end -to -end integrative product.

So you say about we have to design for the future. That means every design what we do has to have a future proof. What are two important ingredients you want to suggest to our audience or all of us when you talk about future proof from a design standpoint?

Yeah, I think the biggest one that I still have to repeat sometimes because it hasn’t caught on is we used to think of, I mean, Jigar and others, have spoken so much about rack density. we have to stop thinking about rack density we have to start thinking about row and data hole level density because how we almost are slowly retrofitting the entire footprint to match an AI factor design we will not be doing that generation to generation that’s just very expensive if we keep changing the technology it’s going to be very expensive for you’re not only spending a lot on building it you’re spending a lot on retrofitting it we don’t want that because that’s going to eat into the ROI so we have to start thinking about I’m at 30 today, I’m going to do this 40 tomorrow, I’m going to do this 100 tomorrow, I have to do something else 200 or 1 megawatt, I have to do something completely different we have to stop that mindset we have to start thinking about it in bounding boxes data hole level or row level bounding boxes and that’s what our latest reference designs do which is start looking at the entire pod as one big block don’t change the technology optimize it with a future proof mindset, right, will this work for that one megawatt rack and today with the digital twins you don’t need to actually build it to do it, you can actually simulate that so that’s number one I would say is take those bounding boxes and take that bounding box mentality now map that technology wise map that right up from the chip to the utility which is same redundancy this redundancy for compute this redundancy for network and have that cluster mindset where you say you map the cluster to the power and thermal perfectly so that every watt goes into maximizing tokens versus going into redundancy and your old school way of thinking so I think if you combine both of those elements you would get into a future proof data center again it’s the hyperscalers have have mastered over the last 10 -15 years.

Again, we pretend like AI is the first time we’re doing infrastructure build -out, but it’s not. The hyperscalers have been doing since the late 2000s, right? So they have mastered the concept of a reference design, a global reference design, where you once lock in that design, you generation -wise you stay in consistency. You build a template and you just feed it out.

I would like to ask the same question to you, Sanjay. From your perspective, what are two things you would like to offer from a design standpoint, infrastructure standpoint, when you want to give a future -proof design for at least for two or three generations, which Peter spoke about?

I think whether we like it or not, the speed of change in the semiconductor IT AI world is very different than the speed of change on the physical world in terms of power and cooling. And even the life cycles and depreciation cycles are very different. So, for example, compute storage or, you know, in the IT world is depreciated every three to five years because that’s the pace of evolution. generators, chillers, transformers, UPS batteries are deprecated 10 to 15 year cycle. So you got to figure out a way how do you run 2 to 3 cycles of IT within one cycle of infrastructure. This is a requirement. If you don’t do that to the point that was just made, Srikanth made that you would be keeping on investing and that’s not good business at all.

So now how do you do that? And in the cloud world again, we mastered that to the sense that in very simple English, how are we doing it today in the cloud space? We have a 30 megawatt data center. We have 2 to 3 to 4, 5 megawatt per data hall. Then we don’t worry about what’s inside the data hall. How does it matter? I have a 5 megawatt power, 5 megawatt cooling capacity. Bring whatever you want as long as it’s 5 megawatt, you’re good to run. The only thing that you are probably retrofitting, if at all you have a generational change, is the final mile of cable or connectors. Now, that becomes a slightly more complicated in the AI world because your densities are much higher while providing power is relatively easy.

Pumping a lot of air or now pumping a lot of liquid is not as simple. There’s much more piping happening. In fact, I joke with the people, the future is of electricians and plumbers, believe me. There’s so much plumbing now in a data center, you will need plumbers in the data center. So, the only way to do it is to again look at what was mentioned in the previous discussion also, is look at certain capacity pods, 2 .4 megawatt pod, 6 megawatt pod. So, now you have a pod. It fits certain number of GPUs of today’s generation. It has certain power capability and liquid capability and it’s done. All the upstream to that in terms of transomers, generators, utility connections is designed for 6 .2, 6 .4, whatever the case.

Now let’s say over the next three years, generations change. Well, all you have to do is reconfigure the cabinets, nothing else, everything else stays the same. Precisely what we are doing in the cloud world. It took us a couple of years to figure this out because this was all being done for the first time. But now this will definitely be the way to go going forward.

So, Sanjay, let me bring to a very different topic now on energy efficiency. When we are talking about gigawatt scale, energy conservation or energy saving is the most important piece. Now, we as a country are a tropical, right? We have a temperature right from 10 degree to 48 degree. So in such a span, what do you think the right approach to improve the PUE? Okay, maybe water usage or what are the important best practices you would like to suggest to the market when it comes to saving energy efficiency or improvising the PUE? Of course, because of liquid adoption, it is anyway has scaled down to an extent of what was there for the normal. But what would be the next stage of best practices you would like to suggest through your experience?

I think the word PUE is, I don’t know if this is the right word, but probably a very abused word in the industry. It’s used so commonly, thrown out there so easily that everyone believes, well, I have a lower PUE. Well, first of all, I can give you better PUE without doing anything. I can increase the air temperature. Suddenly your PUE is much better. You think your PUE is better, but right now your computer fans, your server fans will speed up. The temperature is more. They need to throw more. So the IT load increases. but you increase the temperature so your electrical load reduces, I mean your cooling load reduces, you suddenly have a better calculation.

But in reality, your total power increase, which you don’t realize, so the PUE is better. So PUE is a bit of a, you know, thrown out word, but here is how I look at this. I think the PUE in the data hall in the white space, irrespective where you build it, is the same. Because I need liquid at a certain temperature, I need air at a certain temperature, it needs to enter the rack. The rack is doing what it is doing, it doesn’t matter whether you build in Mumbai, Singapore, I live in Dubai, you build in Timbuktu, it’s exactly the same. The question is, how do you throw the heat out? Because now that depends on the environment outside.

So are you in Singapore? Rains all the time. Are you in Iceland? It’s never more than 20 degrees any time of the year. or are you in, you know, Dubai where it reaches 52 degrees in summer. At least that’s what we designed for 52 degrees. And that’s where the different technologies need to be adopted. Now, whether it is, you know, air -cooled chillers, whether in some markets you can have, you know, water -cooled chillers. One of the unique solution sets that we have started to see is that, especially in India, you have our cities and the way we are located in between the latitudes, our thermal variation or our temperature variation during the year is different. We have very hot in the summer and we have reasonably good weather in the winter.

So there are some entitlements that you can get in the winter. So, for example, we can use chiller technologies where during the winter months we are able to use a bit more free cooling. And in the summer months or during demand months, we add a bit more of, you know, chillers. Chiller technologies, I mean DX technologies, comparator elements that come in and help us to add that extra cooling factor when required. and so what we could do is optimize the way we cool across the thermal cycles of the year and bring down the annual PUEs of the year because at the highest point of temperature you will need that cooling whether you like it or not and so it’s this management of PUE through thermal cycles and some optimization also through load cycles because load also especially in the AI world may not be like a cloud business uniform throughout the year, throughout the day, through every month and so again certain optimizations in how you use your CDUs or fan wall units to bring that energy down will help us to improve the PUE.

One thing I would say about that is the design is there, right? Based on whether it’s the water temperatures, we’re all designing to the same targets. The design is there. Where it becomes extremely manual is again we’re still, in the traditional mode of operation in data centers where we have a large control room and we are optimizing for… for uptime and safety, and safety in the sense there’s no risk of downtime. We’re very risk -averse. But we haven’t, even if we have to do what Sanjay just suggested, which is optimize that, there is no automated way of doing that because the chip -level telemetry doesn’t talk to the data center -level telemetry. And that’s what NVIDIA’s reference design is looking to change today, is, again, if you were to retrofit a brownfield facility, this will be harder.

But if you were to build a purpose, of course, this is an opportunity for India, if you’re building an AI factory today, there is no reason why you can’t integrate telemetry from chip to chip. There’s no reason why you cannot simulate how to optimize that and simulate a traditional sample workload and see how you save energy. I’m sure that simulation will tell you that you’ll save a ton of energy without any human intervention.

You spoke about retrofit. So there have been normal cloud services or normal workloads have been working, let’s say about 5, 10, 15, 15 kilowatt of load. what do you see when it comes to AI augmentation or a GP augmentation in a same platform or a same aisle how the retrofit will be easy or difficult or what could be your one or two tips to do that like if you are talking about AI optimization for telemetry specifically there is already an existing workload which is working with a very small medium to small densities but in that row you want to put a GPU or a liquid could GPU or air GPU which means you are retrofitting some amount of passive infrastructure how difficult or easy would be that actually?

I think again if you go back to that journey even the design and the retrofit was extremely commercial even today I think at enterprise level it is extremely difficult if I was an enterprise CTO looking to deploy AI compute I might actually and I look at our experience in the last one year I might actually be a little you’re looking at a very cumbersome everywhere from design to following local regulations for the high power and liquid cooling having secondary loop built out that’s going to be that could be pretty scary at the end of the day but I think what Verte was doing for example with smart runs you know fully integrated mechanical electrical system that can be purpose built for any pod size that can track our you know our most scalable reference designs I think that would be the way to go right like that’s the importance why even Jigar mentioned that you know our following our reference design as closely as possible all these innovative designs and offerings will improve the adaptability part for the future change is what I can say

my last question to you the future of the the future of the the future of the there seems to have some NVIDIA -ready design offerings or certification offerings. Would you like to say, talk, or would you like to give some insight about that? Certification programs of NVIDIA -ready data centers. But NVIDIA -ready designs.

Yeah, I think whether it’s a colo or whether it’s at a cloud scale, an NCP scale, what we’ve been doing from the beginning is, you know, just like we’ve been enabling other partners, we’ve been enabling a lot of colo partners to build NVIDIA -ready data centers. Okay. And that optimizes for, you know, the water temperatures that we’re recommending, the port sizes that we’re recommending, the redundancy that we’re recommending, the integration between telemetry that we’re recommending. So for the partners that have followed that design, we have, you know, whether it’s DGX -ready or NVIDIA -ready. Now, the only thing I would encourage these partners and also those who are looking forward to this vertical, who is actually doing that?

at speed of light, in a sense. Like, you know, a lot of the data center industry is still in the mode of, you know, they’re thinking more like real estate developers, you know, waiting for, for example, you know, you have these tranches of data centers that you’re purpose -building for everyone. That is a traditional way of thinking and saying, I’m giving this space, this cage to you, and I’m going to build it out the way you want it. But you can’t wait. Like, the way the industry is operating, no one can wait for that, right? So the partners who are building purpose -built AI factories, they are part of, or want to be part of that future, building at large scale, and then whether they give those tranches or not, but they’re built on NVIDIA design, so when the customer comes, you already have built basically according to the specs.

That’s really insightful. Many of our Colo customers will take good insight from that. With this, I would come to an audience for any other question for them.

Hi, I’m Dal Bhanushali. Thanks for the talks, this one and the previous one. We have been talking about how we will scale India in the future. We also need to scale the talent. I wanted to get some viewpoints from you, from your experiences as we double capacities. You also need those people to run the data centers. We need DC ops as special. We can run the NVIDIA optimized containers in our laptops, but those word -to -you chillers, those skills are not common and cannot be easily taught in schools today. So what’s the plan? How do you think we should be going in the future? Especially double every year. is a huge challenge, right?

So I’ll just take this question for a while. So at Vati, we realized this challenge much ahead of time, and we started with a lot of skill development program. So the first thing first is about operation and management of the infrastructure, okay? That’s something which we have started with in collaboration with Indian Institute of Technology, Chennai, where Diploma and BTEC are graduate engineers. We train them for managing how to manage operation and maintenance of data centers. That’s about eight to 12 weeks program, extensive programs, off -site as well as on -site. So this is one part. And there are many other programs which are on the cards to develop design, engineering, and many other things, actually.

That’s what I can tell. And these programs are already available on the web. Anybody can have a look and enroll for that, okay? Any other thing which maybe anybody would like to say about skill development or any other thing which maybe anybody would like to say about skill development or any other development activity which NVIDIA would like to do with the need of ours when we are scaling so high?

I think that’s a question you also want to have. Could you repeat the last part of the question, if you don’t mind?

So he’s asking about how the scale is going up. There’s a lot of resources required, and the skill development is also a big challenge. So while NVIDIA is taking care of an operation and management piece, we are developing a lot of people through colleges and engineering institutions. What are the initiatives NVIDIA also…

managing how to manage operation and maintenance of data centers. That’s about 8 to 12 weeks program, extensive programs, off -site as well as on -site. So this is one part. And there are many other programs which are on the cards to develop design, engineering, and many other things actually. That’s what I can tell. And these programs are already available on the web. Anybody can have a look and enroll for that. Any other thing which maybe NVIDIA would like to say about skill development activity which NVIDIA would like to do with the need of ours when we are scaling so high? I think that’s a question you also want to have.

Could you repeat the last part of the question if you don’t mind?

So he’s talking about scale is going up. There’s a lot of resources required. And the skill development is also a big challenge. So while NVIDIA is taking care of an operation and management piece, we are developing a lot of people through colleges and engineering institutions. What are the initiatives NVIDIA is also taking to develop the skills within the ecosystem?

Yeah. I think a couple of things I would say. One is, you know, as you keep going up on the scale, the prefab systems that Vertebra is developing are going to be absolutely critical because, like, when I was talking about the enterprise -level difficulties right now, all that can be solved with. But it’s, you know, a lot of times you’re waiting for the data center, you’re waiting for the data hole to get ready before you can deploy the compute systems. Yeah. And each of them have dependencies on each other that all are centered around that space, right? When you’re doing off -site prefab integration, you’re doing prefab, you know, manufacturing, you can do that all in parallel.

You can do that all at scale in parallel and then bring it all into one place. And in the meantime, you could do the testing off on the factory. A lot of the testing is done today in the data hall. So you could avoid all that, move it all to the left by bringing it all outside of the data hall and then bring it all into the data hall once the data hall is ready, once the shell is built up, and you could really condense that build -out.

Srikant, as you say it rightly, as we are taking a lot of activity, it is supposed to happen on -site, taking to off -site, which by means of pre -engineering it, developing and building at the scale and getting deployed at the site. So that’s a way forward. Any more questions? Otherwise, we can hold it here.