
The Man Who Built India’s First Data Centre
Sharad Sanghi On Power, AI & The Future

He built India's first data centre when the term didn't even exist. Today, he's powering the AI revolution.
In this episode of The Core Report Special Edition, Govindraj Ethiraj speaks with Sharad Sanghi, founder of NetMagic and CEO of NEYSA, about the evolution of India’s data centre industry — from server closets in the early 2000s to AI-powered hyperscale campuses today.
A must-watch for anyone interested in tech infrastructure, enterprise IT, AI adoption, and the future of cloud services in India.
NOTE: This transcript contains the host's monologue and includes interview transcripts by a machine. Human eyes have gone through the script but there might still be errors in some of the text, so please refer to the audio in case you need to clarify any part. If you want to get in touch regarding any feedback, you can drop us a message on [email protected].
—
TRANSCRIPT
Govindraj Ethiraj: Hello and welcome to the Core Reports Special Edition. We're going to talk about data centres today and more importantly, how they started and what really is a data centre and to answer all those questions, I have with me Sharath Sanghi who started what is conceivably the first data centre in India as NetMagic, a company that is subsequently sold to NTT and today runs a company called NEYSA, that's in the artificial intelligence over data centres, but we'll understand all of that, but the objective of this conversation is two or threefold. First to understand what are data centres, where do they fit into the whole IT and systems and information architecture, including in a country like India and then of course, what's the growth and opportunities ahead. So Sharath, thank you so much for joining me.
Sharad Sanghi: My pleasure. Thank you for having me.
Govindraj Ethiraj: So you built a data centre where that maybe the term was not known, maybe it was called something else, maybe it was called a server farm. So tell us about that early journey as NetMagic and then we will try and fast forward to today.
Sharad Sanghi: Sure. So I came back from the US in 95 and I was consulting with various companies. In 1998, the government opened up the internet sector to private companies.
So private internet licences were given out. I was fortunate to have met the founders of Exodus Communications, Mr. B.V. Jagdish and K.B. Chandrasekhar, who started the first data centre company in the world, which is called Exodus Communications in the US. So when I started in India in 98, when the licences were given out, I was consulting for most of the private ISPs, like Satyam Infoway and others, now they're called SIFI.
And what I saw was that most of them were focussing on retail customers, because at that time valuation were dependent on number of subscribers or number of eyeballs on the content. And very few people were focussing on how internet could be used for mission critical businesses. And since I was exposed to data centres, thanks to the founders of Exodus, I decided to start data centres in India.
It took me a little bit of time to raise money. And so I started the first data centre in India in October 2000. And the reason data centres are important is most enterprises at that time, what they were doing was they were setting up these private closets, if you will.
So in their office, they would designate a room and set up a little bit of IT infrastructure. And the only data centre in India at that time was Videsh Sanchar Nigam, which is actually an internet service provider, POP, point of presence. But they didn't have the concept of precision air conditioning or round the clock electricity and things like that.
So it was actually a room facing the sea with windows where even pigeons could fly in with no UPSs or anything like that. Is this the one at Prabhadevi? Prabhadevi.
Now it's, of course, a world class facility.
Govindraj Ethiraj: Which is where the tables also land.
Sharad Sanghi: And I'll come to that later. Yeah. So after Tata took over, now that's become a world class facility.
But at that time, I'm talking about in the mid 90s, late 90s, it was actually not a data centre in today's sense of the word. But that was what India started with. And we saw this opportunity that, look, and I'd seen world class data centres in the US, which Exodus started.
So we said, look, this is an opportunity that we need to tap into India. Now what happened was, basically, it's an environment where you put servers. So you have more servers in these data centres than people.
And you have very high densities. And these densities have started increasing with AI workloads. So you have a lot of server infrastructure.
So you have to make sure that it's cool around the clock, because it generates a lot of heat, that there's electricity fed to it around the clock. Imagine a mission critical banking site or a media website, for example, not being accessible sometime on the day, or Amazon.com not being able to reach any time of the day. So you have to make sure that there's power supplied all the time.
You have to make sure connectivity. The two lifelines are power and fibre. You have to make sure that there's connectivity all the time.
So you have to have redundant fibres, redundant fibre parts, because if you can have a fibre cut, you lose fibre. So you have this facility which has redundant power, redundant fibre, around the clock cooling, and a lot of security, because there's mission critical data residing on these servers. Initially, when I started, many of the enterprises were sceptical about third party data centres, because they were not comfortable.
Govindraj Ethiraj: And they used to be called third party data centres.
Sharad Sanghi: Yeah, I mean, they were called various things, including third party data. They were not very comfortable, because they were used to having data in their own premises. And they felt that they were letting go control.
And so there was a lot of education required. And over a period of time, people realised that it's better to focus on their core competence and outsource this. And of course, the security measures also in data centres started improving.
And I think there were a couple of major incidents like, you know, an earthquake in Bhuj, as well as the floods in Mumbai, where people realise the value of data centres even more. So it actually took some, you know, natural disasters.
Govindraj Ethiraj: So one reason why people might have held back is also the speed of transmission of data, right? So if you say that I'm my data sitting in some data centre somewhere, as opposed to my own office, yeah, the one thing that I would wonder about is apart from the fact that whether some line breaks down and what happens, second is that the speed at which it was that an issue in the early days, so, you know, it's actually a hybrid.
Sharad Sanghi: So like if you have a local file server in your office that you need to access, I would still recommend I mean, there are many organisations still put it in their private office data centre for two reasons. One is speed and one is also reliability. So you know, you may have redundant internet access, but what if it breaks, right?
So you don't want to be out of data. So normal, you know, so a lot of organisations even put that on third party data centres, but many organisations want to keep their local file server local, right? So that's still happening.
But what I'm talking about is a mission critical data that then let's say you have a bank and you need all the branches to access right now you if you're off if you put this in your office in your central headquarters, and that loses power or that loses connectivity, then all the branches are isolated. And since data centres specialised to handle that, you know, now even banks are outsourcing to data centres. So you know, so that that issue of local reliability now, there are technologies like Citrix and others where people use virtual desktops.
So those are put in third party data centres. But a lot of people who don't use those technologies or want at least a local file service not to be dependent on you know, an internet connection or a point to point link, they put those in their own facility.
Govindraj Ethiraj: So who was the if you remember who was the first sort of major client and which sort of reflected that shift in thinking and confidence in data centres.
Sharad Sanghi: So one of the first clients for us was Nirmal Jain of India Infoline. And he's still a client. One of the first clients also from was Bazi.com, which then got acquired by eBay. Avneesh was with us right in the beginning. And he's still he's an investor in Nesa now. So these are some of the earlier clients.
And what happened at that time, you know, within a year or two of us starting some of the larger companies started offering data centres. Now you don't lose your job if you outsource to IBM. That's the famous saying, right?
So when you outsource to a Tata or Reliance, you know, people were okay, we were a startup. So it took a while before we started getting mission critical enterprise, really large enterprise customers coming to us. But once that started happening, then life gets open.
Govindraj Ethiraj: And where did the banks start coming?
Sharad Sanghi: I think banks started coming around 2000, I think around 2007, 2008. Most of the banks, you know, were doing it themselves. And then over a period of time, they started moving one of their sites, say near DR site.
And then they started moving even their core to near disaster recovery site.
Govindraj Ethiraj: So once the bank started coming, would you say that was like a turning point in the growth of data centres?
Sharad Sanghi: Yeah. So I think there were two major turning points. So one was clearly from an enterprise adoption perspective, it was when some of the banking and financial institutions started coming.
I think the brokerages came earlier, as I said, India Infline was earlier. But then when the bank started coming, maybe 2007, 2008, that's when enterprise adoptions really kicked off. But I think the bigger turning point for data centres and why you see this mad rush right now is when the hyperscalers started opening offices in India.
So you have the Google, Amazon, Microsoft, Oracle, Meta, all these guys have set up server farms or data, you know, they've either set up their own or outsource to third parties like NTT in India. This happened from 2015 onwards. So from 2015, data centres growth just exploded.
And that's continuing. Now, there's a new wave, the third wave, and that is AI. So now all these guys are deploying AI workloads, and even enterprises are using AI, AI native startups are using AI, government is using AI.
So because of that, the energy density per rack has gone up from 10 kilowatts per rack to 20 kilowatts per rack, now to as high in the Blackwell chips, Nvidia Blackwell chips will use 132 kilowatts per rack.
Govindraj Ethiraj: So I'm going to come back to the AI part, which is the most current part. But let me sort of take you back again, as you were scaling up, so let's say you had India Infoline, you had Bazzi, you had a bunch of others, and you said the first bank started coming in around 2007, 8. What did it mean in terms of the size of the data centre?
I mean, what were we looking at now in terms of size and chips, sorry, size and servers and computers?
Sharad Sanghi: Great question. My first data centre was just a 5000 square foot server house. Where was that?
That was in Nirlan Complex in Goregaon, that was the first data centre we set up. Currently, we're building 300,000 square foot data centre building. So that is how it evolved.
The first customer, first customers, we used to take like a one rack or half a rack. So we had actually done some customisation for India where we designed quarter racks with locked cabinets, half racks, etc. Then we had our first cage customer, which was Bazzi, we took a four rack cage.
Today, we are selling not only, then we started selling server halls to the hyperscalers. Then we started selling entire buildings to the hyperscalers. Now we're starting selling to entire campuses to hyperscalers.
So that's how it's grown. From one or two servers, if we got a one full rack customer, a rack is a 42 space of unit, a unit of space, to now full campuses being sold to hyperscalers, which they take over a period of time, right? So that's how it has evolved.
Govindraj Ethiraj: Right. And when you say cage, the cage is a concept of security or is it?
Sharad Sanghi: Security. So what, since a data centre hall has multiple customers, some client wants to have a little enclosure, so that where the racks are set up, so that no other customer who's accessing that server hall can access their racks. So some clients prefer that, some clients don't.
Govindraj Ethiraj: Right.
Sharad Sanghi: But especially the banking and finance people.
Govindraj Ethiraj: First one was 5,000 square feet in Erlan in North Mumbai. And today you're saying we're typically talking about millions of.
Sharad Sanghi: Yeah. So we are now more than four and a half million square feet.
Govindraj Ethiraj: When you say we, you mean? NTT. Okay.
Sharad Sanghi: So India is much larger, NTT has about 22% of the market.
Govindraj Ethiraj: Right. And what's the average size of a data centre today?
Sharad Sanghi: So NTT is typically building about 300,000 square foot, three to 400,000 square foot. But it varies. I would say average size would be somewhere in the two to 300,000 square foot range.
Govindraj Ethiraj: So when, whether it was 5,000 or whether it's 300,000, what goes into connecting a data centre into the rest of the world, including the power and the.
Sharad Sanghi: So as I said, two lifelines of data centres are power and fibre. So you need, how do you connect, these data centres need to be connected to internet, they need to be connected to, let's say, there are two kinds of customer requirements. One is internet access.
So let's say you have a media website like moneycontrol.com. So you need people to come over the internet. So you need your data to be connected to the internet, data centre to be connected to the internet.
The other is a private application, like a bank, as I said, where you have, let's say a bank, which has a thousand branches. Now only these branches need to connect to the data centre. They don't need to necessarily come over the internet.
They could come over a secure VPN if they're using the internet, or they could come over a private MPLS connection. So, so that's the, that's what we see. That's how the data centres are connected.
From a power perspective, since you need round the clock power, you need a very good, reliable power, which is why Bombay happens to be the capital, between 40 to 60%, depending on various analysts. You have really roughly half of data centre capacity of India is in Bombay. Three, four reasons for that.
Number one reason is power. So there's the most reliable power, Tata, Adani, and even MSCDCL is the most reliable power. Second reason is connectivity.
So we are a fibre undersea cable lands in Bombay, in Prabhadevi, in VSNL, and also in, and now in Malad and other places.
Govindraj Ethiraj: So these are, these are the cables coming globally.
Sharad Sanghi: Global, to connect India to the rest of the world. There is fibre, there's very good connectivity of fibre within the city. So you have multiple service providers like Tata, Airtel, Reliance, Jio, and others, MTNL, VSNL, MTNL, whatever.
And so you have very good fibre. You have very good power. You have the financial capital.
So a lot of local domestic demand. And typically the financial industry wants to be in a data centre that's close to their headquarters. The one reason is also because of the, if you want a point to point link, that will be cheaper to do it locally, but they, you know, psychologically world over, you want to be close.
So if there's something go wrong, you can send your engineer there, even though you don't need to, but psychologically that's the case. And the fourth reason is also good talent. So, I mean, there could be another city in the middle of nowhere in India, which may have good power, may have good, you know, may have connectivity also if it's in the fibre path, but may not have talent, for example.
So we have good talent as well. I think the only area where Bombay doesn't score as well is the climate because it's humid and hot. So from a cooling perspective, you don't get, unless you use the latest liquid cooling technologies, the power usage efficiency, you can get better power usage efficiency in a colder place than in Mumbai.
Govindraj Ethiraj: So what would be a colder place today in India?
Sharad Sanghi: So for example, even Bangalore, we get even better PUEs at least in some parts of the year than in Mumbai. Of course you have, you know, people are talking about building data centres in Himachal and all the others, but then you have to look at all the other parameters of connectivity. Also it needs to be earthquake resistant zone, Bombay comes in zone three.
So it's not too bad. Bangalore is actually better, Bangalore is in zone two, but then Bangalore doesn't have the undersea cable and doesn't have reliable power like Bombay. So which is why Bombay is the capital.
Govindraj Ethiraj: Right. And I'm going to come to the talent point because that's quite interesting. So you're saying that a city like Bombay, for example, is the reason why, or at least it scores on three of the four factors.
Actually four out of the five. Actually four out of the five factors. So let me then ask you about talent.
So this is not IT talent, right? Or is it sort of engineering IT or is it?
Sharad Sanghi: So there are two kinds of talent required. So there are some data centres that do just co-location. Co-location means they provide you a space where they give you good connectivity, reliable connectivity, reliable power, reliable cooling, and reliable security.
And then you bring in your, the enterprise brings in their servers and manages the IT themselves. So this is called pure vanilla co-location. For that data centres don't need any IT talent.
They need people with electrical and mechanical skills, security skills, and things like that. But a lot of data centre providers also provide IT services. So for example, in NetMagic, we provided cloud services.
We provided, we managed the infrastructure, IT infrastructure within the data centres for some of our clients. So for that, you need IT talent. So both are now, you know, Bombay is one of the top places to get both these talents.
Govindraj Ethiraj: Okay. So now let's come to the third phase. And we talked about the artificial intelligence, these things.
So I have a couple of sort of starting questions. One is artificial intelligence and the need for it at this kind of speed, scale, and all that is more global at this point than in India, but it may come to India very soon, or is it, it's coming to India already. So do data centres in general supply global needs or is it more local?
Because you, for example, said that in the case of finance, it's very local.
Sharad Sanghi: Great question. So this is data centres typically will serve local audiences typically because of latency issues, right?
Govindraj Ethiraj: So now latency is still an issue today.
Sharad Sanghi: Yeah. Latency issues. So, you know, so for example, if there is good data centres available everywhere, all things being equal, I'd rather do well, latency is lesser, right?
That's why it is local. So that's why typical Indian audience will want data centres within India. And of course, then within banking and finance, they would want it even within their city.
Enterprises preferred within the same city. But that, that is true. So nobody outsources data centre capacity.
Now a global multinational will outsource to India if they have to put the data in India or if they're targeting Indian audience. So with data localisation, data sovereignty with the DPDP act, et cetera, many of the global financial services companies started moving some data centre, some of the data into India, and they started using data centre capacity to India. And that is another cause for the data centre growth, right?
So that, that is one, but, but for many, many, many kinds of use cases, it doesn't matter where the data centre is. So for example, you're using a cloud. You could use a cloud anywhere, right?
If you're an internet centric company, it could, if latency is important, you like it to be local. If latency is not that important, you could technically use it anywhere. Now, most of the cloud providers realising that there's going to be data localisation requirements and latency issues, latency requirements has started moving their cloud instances in India.
So the cloud, the cloud of all the major hyperscalers now has presence in India. Some they've built themselves, some Azure, Microsoft Azure, AWS, GCP, Oracle, Meta, all of their infrastructure now is available in India, right? Because sometimes they, let's say financial services companies using them, they can't put that data outside.
So they have to cater to them locally. And also otherwise, you know, they've got a very large target audience in India. I mean, GPay, for example, has a very, very large percentage of Indian market.
I think the second only to PhonePay. So GPay has to be hosted in India because of regulatory requirements. And so therefore they do it partly, some of these hyperscalers build their own and some of them, a lot of the workloads are outsourced to third party data centres.
Govindraj Ethiraj: So I'm going to come back to the hyperscalers in a moment. But since you brought up regulation, let me spend a minute on that. So first is how much is regulation driving the hosting and the growth of data centres?
Because we clearly see, let's say, Reserve Bank asking banks, credit card companies to co-locate, host locally and all that. So that's one part. Secondly, how big is banking itself and financial services itself in the pie, at least in the pie that we are measuring in India and here?
Sharad Sanghi: So, so let me answer this question. I would have straight away given an answer, but I'm going to give you a slightly nuanced answer. As a data centre provider, my biggest client is actually a hyperscaler, right?
Now within the hyperscaler, there will be some banking, et cetera. So if I look at the actual end user, about roughly 20 to 25 percent, 25 percent would be banking and finance. We call banking, finance, insurance, broking, all of that one sector, BFSI, would be in the between 20 to 25 percent.
Government is a very large user, maybe around 20 percent or so would be government.
Govindraj Ethiraj: Of private data centres?
Sharad Sanghi: Of private data centres. Then there's, you know, there's media, there's AI, born in the cloud companies, there's manufacturing, but the largest clearly is the banking and finance.
Govindraj Ethiraj: Single chunk. And is that a global phenomenon or?
Sharad Sanghi: Pretty much, pretty much. In India, definitely. Okay.
Govindraj Ethiraj: So let's come back to the AI. So you were saying that many of these companies are obviously now, have now over the years built local capacity, are hosting locally and so on, but that they're doing predominantly for the local market and audience. Yes, yes.
But what is the cross-border opportunity, if so?
Sharad Sanghi: So, you know, if you're targeting audiences overseas, right? So there are two issues, right? Now, there are a lot of that.
Let me answer one more question. A lot of training of data, of overseas data can also happen in India because training is not latency sensitive. Inferencing, when you're doing predictions, that's what becomes latency sensitive.
So you'll have the inferencing servers to be close to your clients. Training can be happening anywhere, wherever you get the lowest cost, wherever, for example, power is cheap, et cetera, right? You can do training anywhere.
The only factor that kind of, you know, governs where you do the training is if there's a large amount of data. So now, if you have petabytes of data, then migrating that data, let's say that's in the US, that data is very difficult to move to Indialytics several days, right, even with today's internet speeds. So, but having said that, there is, you know, a lot of local demand.
We are seeing demand from four kinds of, when it comes specifically to AI, four kinds of enterprise, if you will. First is the research institutes doing a lot of training and fine tuning. Second is government itself, the centres of excellence, like a centre of excellence for healthcare, for smart cities, for agriculture.
Government has just set up three centres. So they are looking at AI in a big way. Third is AI born in the cloud companies.
So born in AI native startups, right? There are a lot of companies that have got funded that are coming out with some solution for enterprises for AI. And so they need a lot of GPU capacity, if you will, and data centre capacity because of that.
And last but not the least, if I take all the enterprises in one bucket, then I would say banking, insurance is where I've seen the maximum adoption. I've seen healthcare. Healthcare is another big vertical.
We've seen in logistics. We've seen in retail manufacturing. And yeah, so these are some of the segments.
Govindraj Ethiraj: Yeah. But you said that roughly a quarter is government in any case. And about 20% even today.
And of potential future, which includes AI, government continues to be a similar size.
Sharad Sanghi: Yeah. So government has two... In different ways, I guess.
So in different, I'm saying overall market. Now, a lot of the government workloads actually reside in NIC. They don't necessarily come to a third party data centre.
So the third party data centre for government departments by default is NIC. But a lot of the government organisations are not bound to put that data centre only in NIC. They can also use third party data centres.
But yeah, government is a very large spender.
Govindraj Ethiraj: So, I mean, it seems to be like it's an encouraging sign that the government is using 25% of AI, potential AI data centre capacity.
Sharad Sanghi: Yeah.
Govindraj Ethiraj: I don't know if it's the same number elsewhere in the world.
Sharad Sanghi: Yeah, I'm not sure about that for the rest of the world. I can find out. But in India, it's a very encouraging capacity.
I think the government has played a very positive role. They've set up the AI mission, not only these three centres of excellence, but they've also had the AI mission where they're not only looking at subsidising GPU capacity in a public-private partnership, but also funding research.
Govindraj Ethiraj: Yeah, which was recently announced.
Sharad Sanghi: Which was recently announced, yes. Yeah.
Govindraj Ethiraj: So as you look ahead, I mean, there are some very...
Sharad Sanghi: And of course, they're consumers. They're large consumers of AI. So government departments want to use AI.
Govindraj Ethiraj: So as you look ahead, there are some very encouraging projections and predictions by many of the people in the data centre space in terms of actual growth, growth into other cities and so on. How are you seeing it from your perspective?
Sharad Sanghi: No, absolutely. I think whatever estimates, we always end up beating the estimates. There's one industry where we've ended up beating the analyst estimates for the last few years consistently.
Give me an example. People are talking about doubling in the next three, four years. I won't be surprised if we double faster.
We'll probably double in two years, if not three years. We were five data centre companies till about 2015, 2016. Today, there are more than 20 data centre companies.
I'm talking about companies with scale.
Govindraj Ethiraj: Including Ambani's and Adani's.
Sharad Sanghi: Ambani's, Adani's, NTT, STT, Ambani, Adani, and many others. So that's one. Secondly, the AI workload, as I said, from 20, 30 kilowatts per act, now we're going to 100 plus kilowatts per act.
So that requires a lot more power density. And today, the way data centres build customers is based, even the space is built on the number of megawatts. Because the amount of megawatts you consume is what is the kind of infrastructure data centres need to deploy.
And so the benchmark is X dollars per kilowatt per hour. Or kilowatt per month, whatever.
Govindraj Ethiraj: I'm actually quite curious about that. I mean, maybe not now anymore, but maybe it's useful to explain that a little more. When people say, oh, Adani is setting up a 100 megawatt data centre or Reliance is setting up 150 megawatt.
So the use of the word megawatt, which one would traditionally sort of associate with power generation is interesting.
Sharad Sanghi: That's the IT load. So two things. Some people use, what the conservative people like NTT and typically what people should use is the IT load.
The actual load will be depending on the power usage efficiency, depending on how much of energy goes in wastage and cooling. That's the total load. So some people use total load.
Some people use IT load. IT load is, you take the total number of servers that you have in a data centre. Let's say an average consumption is 20, or total number of racks.
Let's say average consumption is 20 kilowatts per rack. You multiply by total number of racks, you get the total IT load. So that is the total number of, that's it.
Govindraj Ethiraj: So how much would that be in a 300,000 square feet?
Sharad Sanghi: About close to 30 megawatts.
Govindraj Ethiraj: Per? 30 megawatts.
Sharad Sanghi: 30 megawatts being used at any time. Now, the 30 megawatt of power, but to cool it, you'll probably waste another 15 megawatts or so. So you have, that will be a 45 megawatt data centre.
So total power consumed will be 45 megawatts total. Now with AI data centres that in 300,000 square foot, you'll probably end up having higher density. So because you're using liquid cooling.
So now you'll have much higher capacities in the same footprint.
Govindraj Ethiraj: So much more cooling as well. So you're saying cooling.
Sharad Sanghi: Yeah, so what, so conventional cooling, air cooled designs. Now what has happened, there's direct to chip cooling. So water actually, there's a manifold and the water inlet takes the heat from the chip itself.
It's called direct to chip cooling. Or you have liquid immersion cooling where you immerse servers in the liquid. And that gives you, so this term called PUE is called power usage efficiency, which is the total load divided by the IT load.
So for example, let's say I have a server that requires 100 kilowatts. Okay, and I use 150 kilowatts to cool it, my PUE is 1.5, right? If my total power is 150, which includes the cooling, et cetera, and the IT load of this, and actually the useful load is 100, then my PUE is 1.5. So typical data centre designs with when you're not using direct to chip cooling or liquid immersion cooling, any form of liquid cooling, is typically in a climate like Bombay is between 1.35 to 1.4. These are good designs. The traditional designs are more like 1.7, 1.8. But now you can, you know, become much better, much more energy efficient. With liquid cooling, we've achieved, with direct to chip cooling, we've achieved 1.2. And with liquid immersion cooling, we've achieved below 1.1, 1.08, right? So that means only 8% of energy is lost in the case of liquid immersion cooling.
Now, the problem with liquid immersion cooling is that the servers need to be something that can be immersed. If the servers cannot be immersed, then you cannot use liquid immersion cooling. So the most popular technology right now, direct to chip cooling for AI workloads.
Govindraj Ethiraj: Okay, so, I mean, this obviously brings us to the whole environment and power consumption. So in the US, for example, the hyperscalers are saying, or rather they're buying nuclear plants, or they want to buy nuclear plants. They want to set up microgrids.
Micro, this thing, and small. Modular. Modular, yeah, SMRs, and so on and so forth.
So in India, we are nowhere near the nuclear part right now, at least. Maybe it'll come. But how are you seeing the whole power?
Because this is, when you say 30 plus 15, 45 megawatt, this is really an entire town.
Sharad Sanghi: Yeah, yeah, yeah. If not more. I mean, total capacity that NTD uses right now is about 300, what is in production is close to 300 megawatts.
And what is, the total power will be more like 450 or 440 megawatts, which is, yeah. So, and the one campus in Navi Mumbai that we've recently taken live, that has a capacity of 14 buildings, four of which are live, and that campus alone will be about 500 megawatts. Reliance has talked about a gigawatt in Gujarat, or wherever, right?
So that requires huge, but so I want to clarify one thing. While data centres consume a lot of energy, but data centres are actually good, because if I were to just imagine this workload was just distributed in everybody's offices, the PUs in these typical office data centres is two point something. So for every one unit, you're wasting one unit.
Whereas here, the PUs are sub 1.4, 1.3, or maybe in liquid cooling, for AI data centres, it'll be sub 1.2. So you're actually, there's a study done by some folks in the industry where third party data centres and well-designed energy efficient data centres are actually saving almost 70 to 80% of energy. So it's a necessary- So for what it is you're saying, data centres are more efficient than putting it in their own- It's a necessary evil, right? You can't be in this day and age without, everything runs on the cloud, and you can't be without data centres.
And given that that's a necessary evil, so while yes, they consume energy, but then data centre operators today, at least in NTT, we have announced a sustainability goal of being carbon neutral by 2030. Today, about 45 to 50% of energy is green. So we've set up captive solar, hybrid wind solar plants, that where we feed energy to the grid and get wheeled back.
Govindraj Ethiraj: And that's how we save, that's how we've- So my question actually is a little different. It's more to do with the availability of electricity. Yes, it is a high sort of power-hungry beast.
So my question is, in countries like India, more in countries like India, of course, even in the US now, because AI was never there in the reckoning. Are we, do we have a, I mean, so we have ambitions, saying that we're going to double or triple data centre capacity.
Sharad Sanghi: So I think currently, there's no immediate, at least in a city like, in the Mumbai metropolitan area, there doesn't seem to be a constraint in power right now. Of course, the government needs to constantly upgrade the, not only distribute transmission, but also distribution, right? So right now, but I was running, before I left Entity from an executive role, I was running the global data centre business for Entity.
We had constraints in Singapore, we had constraints in Germany, we had constraints in Virginia, which is a, North Virginia, the data centre capital of the world, right? And Mumbai, like a mini version of Virginia. There was constraints in that particular county.
So there, it's only inevitable that the data, which is growing, that there could be some constraints, which is why some of the global companies are looking at microgrids. Some people are looking at gas-based, you know, there's a fuel cells, gas-based fuel cells, there's a company called Bloom Energy that is doing that for some data centres in the US. There's also these company called Oklor, I'm not sure if I'm getting the name right, that does these smart nuclear modular reactors.
India is exploring that, but I think it's too early. I think in the recent budget, there was some incentives and announcements given for that. I think it'll come, it'll take a few years, but it'll come to India as well.
And the reason is that you don't want to be dependent on the grid. What if, you know, the growth is so fast that the grid becomes unstable, so you want to be able to cater to this locally itself. So yes, people are exploring it, but it's not, I mean, I don't see it in production for another two to three years.
Right.
Govindraj Ethiraj: And the other big development, obviously, in this space, or that has affected this space is DeepSeek, which seems to demonstrate that you don't actually need as much computing power that you thought you did, or you could achieve more for less, at least in terms of infrastructure and chips. What's your reading?
Sharad Sanghi: So I think DeepSeek is a good thing because, you know, there'll be more companies now where, so earlier these, you know, these really large scale, 100,000 GPU kind of deployments were left to the big four or five or six, the, you know, the Metas of the world, or the Elon Musk, and, you know, Amazon, Google, Microsoft, and these kinds of companies. And none of the other companies were really training to that scale, because that requires hundreds of millions of dollars, if not billions of dollars. But now DeepSeek has shown that you can do achieve, as you rightly said, you know, with the constraints that were imposed because GPUs cannot be sent to China.
So they use some older generation GPUs and they use a lot of well-known methodologies like mixture of experts and reducing the precision of the matrix multiplication and a lot of other techniques that they used to actually achieve similar accuracy results with one 10th, as they say, of the cost of doing that. I think it's a good thing because there'll be many more companies now attempting to do that, right? And so the volume, you know, when you, you may, and the funny thing is that the big boys have not cut down their budgets.
If you look at the earnings report for any of the big companies, whether it's, $25 billion, $62 billion, $70 billion, whatever, they're all talking about big. So they're continuing to do what they're doing. And now the many more companies will attempt to do these foundational models, right?
And so I think it's a good thing for the industry.
Govindraj Ethiraj: In your own journey, you started and sold NetMagic to NTT and NTT is, of course, in India in many ways and continues to expand. I just read that they've commissioned a new subsea cable into India. And you're also the chairman of the company in India.
NTT's venture capital arm has invested in NYSA. So tell us a little bit about NYSA before we go.
Sharad Sanghi: Yeah, sure. So the intent of NYSA, so one of the things I realised, so at NetMagic NTT, we had set up India's first indigenous cloud, right? The compute cloud that we had set up.
That time there were no GPUs, I mean, without GPUs, that was, and that was doing really well. Now, obviously not the same scale as the hyperscaler, but that did really well for, we had a known niche of clients in India. And I saw this opportunity when the, in late 22, early 23, when Gen AI came, but my role in NTT was different.
So I was now running the global COLO business of data centres and not the IT piece. And so I decided to, you know, quit, start this, still have great relationship with NTT. They kept me on the board and of the data centre business because our business, in fact, I'm a very large client, a decently sized client for NTT because all my GPUs are in NTT data centre.
And the opportunity we saw was that, look, most enterprises today are, so many enterprises are mature. So AI is not new, right? Machine learning has been there for some time.
So banks have been using machine learning for fraud detection for a while, right? And they've got data scientists and machine learning engineers now for the last few years, right? Gen AI is obviously more recent.
but many companies today, many enterprises were not familiar with AI or wanted to do just proof of concept, but not really take it into production. And we felt that there was an opportunity to democratise AI, to be able to make enterprises use AI as easily as possible, as cost effectively as possible and as often as possible. And that's how we set up NESA.
So what we've done at NESA is we've got not only the infrastructure as a service, which is provides GPU and other infrastructure as a service in three flavours. One is on bare metal for research institutes for people who want to do training and fine tuning and also using virtual machines or containers. We also have two platforms on top, which is one is a developer platform, which is again, open source based.
And one is a low code, no code platform that allows people to, without writing code, build Gen AI apps. And then, and last but not the least, we provide inference as a service. And we layer this with both security.
So we have our own security tools plus some third party security tools and observability. So as you scale, you want to be able to manage this infrastructure. So that's what we set out to do.
And, you know, it's, we've gone live about six, seven months back and we're seeing very good traction. Right.
Govindraj Ethiraj: So last question, you've worked, as you said, you spotted the data centre opportunity very early and you worked with people and BV Jagdish is on your current board as well.
Sharad Sanghi: Yes.
Govindraj Ethiraj: And an investor as I understand. Yes. They saw the opportunity for data centres.
They were the first to see it in the world with Exodus. Yes. So how do you spot something like this?
Sharad Sanghi: I think it just, you know, as you're in the business, right? So I was, so let's take the example of NetMagic. I was consulting to ISPs.
They were only focussing on retail users, consume, you know, there were a number of broad, it became retail users, then later on became broadband subscribers. And, you know, so at that time, when internet just started, when 98 to 99 period, nobody was looking at enterprise, mission critical enterprises. So given that I had worked at the NSF net and given my US background for six years, I was in the US and the fact that I knew the Exodus founders, I was lucky if you will, that they were the ones who actually spotted that opportunity in the world.
And I kind of benefited because I saw what they had done. And I saw that India needed something like that. So that wasn't the first time.
The second time also, I'm not the first to set up GPU. I mean, Corvi set up GPU as a service. They're just going IPO soon.
And we note the, initially when I was running the NetMagic business and the NetMagic cloud, there was not so much demand for AI related services. But the moment when everybody started using Corvi, that was something that everybody started using, right? So it was only a matter of time.
And then we started getting requests from customers saying that, we'd like to do some AI related stuff on the cloud. And we started losing these customers to the hyperscalers. And that's when I said, this is a great opportunity.
And we saw the success of Corvi. So I would not claim to be the first to do something in the world, I mean, either of these, but kind of taking what trends that you saw somewhere else, seeing the local demand and then adapting it to the Indian market is what.
Govindraj Ethiraj: Right. So absolutely last question. Why did you call your company NetMagic?
Sharad Sanghi: So it was, my wife chose that name. So I'll have to ask her, I just liked the name. And Nisa was actually chosen by my CA.
I couldn't get a company these days. You try to register a new company, either whichever you have the domain name, or you have a trademark. So getting both the domain name and trademark not taken is very difficult.
So I asked my CA to give me a bunch of names that were the trademark and the domain name was available. Nisa was one of them. And what I liked was it means pure in Greek.
So that's why we chose Nisa. Right.
Govindraj Ethiraj: Good note to end on. All the best. Thank you so much for visiting us.
Sharad Sanghi: Thank you so much. My pleasure. Thank you.

Sharad Sanghi On Power, AI & The Future

Sharad Sanghi On Power, AI & The Future