- 13 minutes ago
At the India Today AI Summit 2026, industry leaders unpacked what truly powers AI, i.e., from advanced chips and data centres to voice tools and open-source platforms.
Category
🗞
NewsTranscript
00:00Good evening, behind every breakthrough model, every intelligent agent, every transformative
00:09application lies an invisible foundation and this invisible foundation is of massive, massive
00:17chips, sprawling data centers and mounting energy demands.
00:24That's going to be the panel discussion.
00:26Let me introduce my panelists, Preet Banerjee is from Synopsys, he is pushing the frontiers
00:32of AI, driven chip design, Samhita, from Resilience AI, mapping climate risks to build truly resilient
00:41systems.
00:41We have Keshav Reddy from Equal AI, he is trying to democratize powerful AI for India's diverse
00:50realities and Martin Tasne is from AI Collaborative and the current AI championing open, equitable,
00:59non-aligned infrastructure.
01:01I'm going to begin with you, Preet.
01:04With AI now designing what is being seen as power AI, are we then heading towards what
01:12will be seen invariably in future, some kind of unbreakable physical limit of scaling?
01:20That's a great question.
01:24So, the current AI models, as you said, right, are being powered by AI chips from AMD, NVIDIA,
01:33et cetera, and they are going into the data centers from the cloud companies, the Microsofts
01:38and others.
01:40And Synopsys is the leading provider of electronic design automation tools to design those chips
01:46and design the systems that use those chips.
01:51What is going on is now, those things used to be designed by humans, but the pace of innovation
01:58in these chips is increasing, the complexity is increasing, and so something that an AI chip
02:07for NVIDIA would take maybe two years to design is now being designed in a year.
02:11something that had a hundred million transistors now has a trillion transistors in it.
02:18So, it's impossible for a human designer using EDA tools to design those chips.
02:23So, what we are working on is using what is called agentic engineers, agents that are doing
02:31the job of the human engineers to automate that process and accelerate the pace of innovation.
02:37Because these AI data centers are incredibly complex systems, because you have to manage
02:43the power, the thermal, and all of those things, and it is impossible for a human designer to
02:50take care of these multidisciplinary optimizations.
02:53Those are things that are now being done with the Synopsys tools.
02:57Okay, Keshav, on one hand is what is being called by our executive editor-in-chief, Kalipuri,
03:04digital imperialism. And on the other hand is what India aims for, which is AI that listens,
03:13understands, and acts. How do you make AI inclusive then?
03:18Thanks, Maria, for the wonderful introduction. I've never heard models in AI models be introduced
03:25like that before. So, I'm on a company called Equal AI. It's a call assistant. It picks up your phone,
03:32speaks on your behalf, and then in the future will do tasks for you, help them direct people
03:38to your use cases. So, it's a consumer app. For our use case, you know, we're seeing insane amount
03:44of need, want, because you get a lot of calls actually nowadays. Even with the advent of like
03:49quick commerce, you get a lot of OTPs, you get addresses, which is much more frequent,
03:53apart from telemarketing and other aspects. One of the biggest challenges that I see in AI infra
04:01is that actually India is a DAO farm for most global companies. There are millions and hundreds
04:07of millions of active daily users. But the infrastructure is not there. Even for like we
04:12scaled 4x last week, and we peaked out on all of our partners. And the good thing or the bad
04:20thing
04:20is that they were elastic. But after a point, you have to buy dedicated DPUs, you have to ensure
04:25scale up the concurrency needs, because we are a real time use case. So essentially, we would need
04:32real time streaming part of the infrastructure. And that's not there, actually. And I know a lot of
04:39people have made statements of building assets for India. But those come live in two years, 18 months,
04:4424 months, 24 months. And so demand is not going to wait, as a consumer company. So I think there
04:50are challenges in being able to deliver the demand. But people using AI in India, I mean,
04:56it's it's blowing up, actually, people are loving using it.
04:58Yes, of course, you know, many would say that it's a little messy right now. But the data centers are
05:05being
05:06built. There is infrastructure in terms of mobility. And the focus is also on what was there,
05:14earlier. We are just trying to build on it. So Martin, we are racing to close the AI empires.
05:23You know, how do you actually then build an open, non-aligned infra backbone?
05:29Thank you. Thanks for having me. And thanks for asking. I think what's important to think about when
05:35we think of the infrastructure is not to only focus on the bricks and mortar. So the computer is
05:40important. The data centers is important. We'll talk about the environmental impacts, I'm sure,
05:45in a second. I'd encourage all of us to make a link to the session just before this one, and
05:50to also
05:51think about data as infrastructure, and to think about open source as infrastructure. So I think we're
05:57making a mistake. There's a lot of conversations about AI sovereignty. I think AI sovereignty is critical,
06:02but is often misunderstood, but very few countries around the world. I would even put a question mark
06:09as to whether the US and China can actually be fully sovereign when it comes to AI. And I actually
06:14don't
06:14think they should be. I don't think anyone really should be. What we want, I would say what the public
06:19wants in our understanding and from surveys, is agency over how AI is used in their lives or not used
06:27in
06:27their lives. And I can think of no better way to have that agency to have as many layers of
06:32the stack
06:33as open source as possible so they can be used, shared, and modified so people can really sort of
06:40like get stuck in and understand it. So I think that's critical. So on the infrastructure piece, I would
06:46really encourage us to think of open source as a critical part of the infrastructure that needs
06:51investment. The good news is that open source doesn't have the environmental externalities that
06:56data centers do. It's also a lot cheaper than buying hundreds of thousands in video GPUs. And at the same
07:03time to think of data as part of the critical infrastructure, ways in which, and I'll just finish
07:08with that, we've had fantastic innovation around compute, but very little innovation really over the past
07:1310 years when it comes to data. If we had a fraction of the world's brains and money and power
07:18focused on data governance, on technical innovations around data, so we can use personal data in safe
07:24ways, on governance innovations around data, I think that would be fantastic. And I can think of no better
07:29place to do it than India. Since you brought in the environmental aspect of it, let me bring in
07:33Samitha now. Samitha, AI data centers are seen as units that guzzle energy and the heat of cities as well.
07:44How do we make AI infra then resilient to fight that crisis?
07:52Thank you so much.
07:55So data centers are important. We can't minus it. It's like as much as you have power, there is lights.
08:00We are talking, you know, it's become a very critical part of our lives.
08:05The growth of power today to have electricity in every mile, in every last mile has taken its own years
08:11and it has been driven by a need.
08:14So is the growth of data centers. So that fundamental understanding of having data centers is absolutely important when we
08:21think of energy.
08:23Now, as development happens and India is going to invest one billion in infrastructure, it's going to contribute to that
08:30energy consumption and energy release at the same time, in the same way that data centers do.
08:34Now, what data centers can do differently are two things. Number one, at the time of construction of a data
08:40center itself, if there are ways of using AI to understand the environmental nuances of that location.
08:49So if you can understand that there is a water network or a water system already flowing in there, there
08:55is a large patch of vegetation, there is a road there that has already been there, there is a settlement.
09:00And key factor that when you are constructing a data center, that's the first level of resilience, which we call
09:08as a practitioner in the terms of climate adaptation.
09:11The other one, which is, of course, the suffix and prefix of climate today is carbon emission, which also is
09:18as critical as climate adaptation.
09:21And for that, the data centers that are getting built today are very modernized.
09:26Even at a rack rate, they're able to optimize the energy they consume to be able to churn that level
09:33of intelligence and data.
09:35And this panel is equipped to answer that the data centers that used to get constructed even five years back
09:41and the data centers that are getting constructed today are more architecturally sound so that the energy consumption is minimized.
09:49It's also built with sustainable materials and building materials so that it cannot forfeit the entire amount of energy consumed
09:59and the carbon emission, but it can minimize.
10:01But considering the fact that data centers are going to be absolutely essential in the ecosystem of things that we
10:07are talking about, and especially with, we are part of the AI summit, which is computing, generating, and agentifying more
10:15than petabytes of data at the same time to be able to give us the intelligence we need across sectors
10:20of life.
10:20So that would be my understanding of data center, and I truly believe the data centers are modernized today, much,
10:26much, than where we used to be even five years back.
10:28Okay, so let's talk about one big idea, which is on display at Bharat Mandipam.
10:33What do you think will be the idea of tomorrow?
10:36Preet.
10:38So I would say there's a lot of discussion today on AI models, but those are AI models for words
10:46as tokens, right?
10:47I mean, people have been working on AI for many, many years, but until chat GPT happened in 2022, nobody
10:55really knew what was going on.
10:56And what chat GPT did was bring AI to the average person, right?
11:02So that you could write poetry, you could write essays, and so on.
11:05But all of those are using words as tokens.
11:08But then, after that, there was DALI that came in with using images as tokens.
11:14And so now you could create all kinds of things with images.
11:17You could create new image.
11:19You could say, I want to create a picture of a Boeing airplane hitting a Golden Gate Bridge, which is
11:26a bad example.
11:27But you don't have to have an actual picture like that.
11:30It would generate that picture.
11:32Sora created, sort of did it with videos as tokens.
11:37But the future is what we call physical AI.
11:40The world around us is governed by the laws of physics, which is, say, fluids physics, which is Navier-Stokes
11:47equations,
11:47or structural physics, which is Euler equations, or electromagnetic physics, which is Maxwell's equations.
11:54So, Synopsys just recently acquired ANSYS.
11:57I used to be CTO at ANSYS before.
11:59So, we believe that the real future in AI infrastructure will be using physical AI,
12:06where the real world around us has world models, those different physics that I talked about,
12:13and it will understand that physics, and be able to reason and say, here is an airplane.
12:19I mean, you were just talking about the data center and understand the environment.
12:24It would actually look at the physics of here is the water, here is something,
12:28and because of the water, you could actually do liquid cooling,
12:31understanding the world around you, building those world physics models, physical AI.
12:37That is actually the future.
12:38So, physical artificial intelligence, how will it really work?
12:43So, what it works is, as I said, with words as tokens, you are using tokenized words.
12:49You are saying, Mary has a little lamb, right?
12:52So, basically, you have scraped all the words in the English language and says,
12:56when you have Mary has a little, you predict the last word would be lamb,
13:00or I wish you a merry Christmas.
13:03That's how words as tokens are used in foundational models.
13:07Now, if you, instead of those words, you use images, then you can essentially say,
13:13from this image to this image to this image, what is the next image?
13:15That's gen AI applied to images.
13:18Then you do the same with videos.
13:20You have a video screen here, et cetera.
13:21You will predict the next video.
13:23So, the physical AI is saying, just like, suppose I were to increase my temperature here,
13:29at some point I'll burn, I'll do whatever.
13:32That is the physics.
13:34It will predict the physics of the world around us,
13:37and that is incredibly computationally difficult, right?
13:44But it will be very, very powerful, because it will actually, I mean,
13:49Artemis mission is sending NASA's thing to the space, right, to the moon.
13:56Imagine, those things are being designed by NASA engineers.
13:59In the future, physical AI will be able to design a spacecraft with AI,
14:04which will understand the physics of an aircraft going through,
14:08the physics of gravity, and how it comes back.
14:10That's where the world is headed.
14:11Okay, since India is aiming for making AI more accessible,
14:16making it for every citizen,
14:19so, Keshav, which tool do you think is powerful enough right now
14:24to ensure that AI is democratized?
14:27See, the way Indians will use AI at scale,
14:32so there are studies saying that over a billion people in India have never used AI.
14:36I mean, we are sitting in a room which is a cohort who uses it very often,
14:41but I'm saying, like, look at rural India, consumers, they don't know.
14:46A white screen, when they can't type very well, and they can't read very well,
14:52is not a mode for them to actually use it.
14:55So it has to be conversational.
14:56That's the first thing.
14:58Second, conversational, multilingual, with dialects, that is still improving.
15:03With every model, with every model company, they are improving.
15:07Actually, you know, one point is on open source.
15:09I really hope, and I know that we're working on a lot of stuff.
15:12The government is working on a lot of stuff, and that will be very exciting.
15:15And then the next feature, which I think,
15:18the next application that will be wide-scale used in India will be voice,
15:22very simple, and it doesn't need to be labeled as AI.
15:26It needs to be so simple.
15:28Indians love options, but love simplicity.
15:31So if you look at any UPI app, it's cluttered, but has very simple actions.
15:37I think an AI assistant or anything that works at scale in India
15:41will be something which is speech-oriented and very simple.
15:45I don't think that exists at scale, what Indians can use at scale.
15:48And that's kind of the direction which a lot of people are trying,
15:51including Equalist, trying to do that with voice first.
15:54There's a reason we're doing calls first and voice first,
15:57because it's very natural.
15:58And then we move into doing tasks for you,
16:00and then we do end-to-end tasks for you,
16:03for everything that you want to do.
16:04So earlier today we had Professor Raskar,
16:07who is with MIT Media Lab.
16:09He had a very interesting idea,
16:11and which is already in motion, which is about doot.
16:16Every Indian, just as an Aadhaar,
16:18will have an AI agent of their own,
16:22so perhaps a Keshav AI agent.
16:24And that AI agent will be communicating with a Maria AI agent
16:29or a Sametam AI agent and functioning at that level.
16:33That is finally perhaps the aim.
16:36Do you think, how far are we to that level of,
16:40that kind of scale where each and every person sitting in this room
16:44will have their own AI agent?
16:45Very soon, 12 to 18 months.
16:48It's actually just a modality.
16:4912 to 18 months?
16:50Yeah, easily.
16:50Because if you're using any platform that is using natively AI,
16:54and you call me and I call you and our agents speak,
16:57it's actually not that far away on basic communication.
17:02But doing a lot more complex tasks will still take some more time,
17:06because obviously you have more nuances that come around it.
17:09But communication should be there very soon.
17:11Today, we do millions of calls.
17:13A lot of those calls are from AI agents,
17:16which your AI agent is speaking back.
17:18So that's already happening at a business-to-consumer level.
17:22Consumer-to-consumer level will happen.
17:24When UPI started,
17:25nobody thought UPI will be a peer-to-peer money-sharing platform.
17:28They thought it will be meant for businesses to customers,
17:30or customers to businesses.
17:32But now 90% is largely peer-to-peer.
17:35So you'll see a lot of innovation there, I feel.
17:37Yes, so Martin, the fact is that India has adopted something called the UPI,
17:43and the scale at which, and the pace at which it was adopted.
17:47Every person right now on the street is perhaps having a UPI app on their phone,
17:52and they're using it for basic shopping of vegetables and others.
17:56Do you see AI becoming so accessible at the same time?
18:02Thanks for asking.
18:04It's very linked to the question you were just asking, Keshav.
18:06I completely agree on the voice piece.
18:10I think that I find it hard to reconcile the fact that,
18:14and I'm certainly not an expert in India, to be clear,
18:17but I find it hard to reconcile the two things you were saying earlier.
18:20One of them is that a billion people don't use AI,
18:23and at the same time that every Indian will be interacting with an agent.
18:27At Current AI, we've been working with Bashni,
18:30the translation platform, which is fantastic,
18:32and we've been developing, together with them,
18:36a very simple open hardware product that we're launching on Friday at the summit,
18:41and what that product does, and that I can see it used by everyone,
18:45is a voice translation mechanism, right?
18:50Very simple.
18:51You can be in any use case.
18:52You could be a farmer looking in a field at a particular sort of pesticide or other,
18:58and it will not only translate what someone tells you,
19:02but it will also analyze the text and either speak it back to you
19:06or have it back to you, whichever you wish,
19:07and this is a very basic open hardware tool.
19:10Anyone can make it.
19:11We're not designing it to sell it,
19:13but for anyone to be able to mess with it in the spirit of the maker movement,
19:17and it will also be able to tell what's in front of you,
19:21so if you cannot see well, if you're blind,
19:23it will be able to say, right,
19:25and it will speak to you in your language and say,
19:27right, here are a few stairs.
19:28You might need to watch out.
19:29So that type of an approach,
19:31especially something that's available to all to make and to sort of mess with,
19:36I can see that being really sort of taken up at quite an interesting scale.
19:40Just to clarify, I think I was mentioning that apps today at India scale,
19:45those are not there,
19:46but your application use case is very relevant.
19:49It doesn't need to look like AI,
19:51but it works at that intelligence layer, so I agree with you.
19:55Okay, Samit.
19:55I would like to add just one thing here,
19:57that all these,
19:59first of all, the UPI is an absolutely amazing innovation that happened in India,
20:03and I live in the U.S.,
20:05and I mean, people in the U.S. have often wondered,
20:09had U.S. been able to adopt a thing like UPI,
20:11wonderful things that have happened,
20:13but the UPI is a unique identify for the humans.
20:17There are one billion people in India, right?
20:20But there are a hundred billion machines.
20:24There will be a trillion machines,
20:25like the world of Internet of Things, IoT.
20:29So, and I was talking about physical AI.
20:31So, imagine if all the humans and the machines,
20:35all of it had that unique ID,
20:37and then, and these machines are generating data,
20:41like just the humans are generating data, right?
20:43You can gather all that information.
20:45That's the physical AI I'm talking about.
20:47That combined,
20:49the human interacting with the machines
20:51would be the absolute end state of AI,
20:54in my opinion.
20:55So, hyper-local risk tools.
21:00I'm just saying that, you know,
21:01the question almost fell on the lap.
21:05Okay, no, fantastic.
21:07I love the conversation of having,
21:10mimicking AI.
21:11Again, always have human assistance in the loop.
21:14That's fundamental of science for us,
21:17because we deal in disasters, natural disasters.
21:19We can't just let everything on a machine.
21:24Hyper-local AI is just taking your conversation.
21:28It's happening today.
21:29If you test a product called Resilience 360,
21:3430 minutes,
21:35give us any location anywhere in the world.
21:38We are able to tell you that that structure
21:41is at what exposure, if flood happens.
21:45It's not forecasting flood.
21:46It's saying, go with the open premise
21:48that the flood will happen.
21:50It's modeled in such a way,
21:52considering environmental factors,
21:54considering the construction factors,
21:56considering built environment,
21:57which is physical,
21:59and considering all the climatic factors together.
22:03Think of it,
22:03a machine is doing that at an abnormal speed,
22:06in 30 minutes,
22:07able to give you a risk score.
22:09Explainable to an ordinary person like me.
22:13Heavily computed.
22:15Computed so much,
22:16and trained that in 30 minutes,
22:18is able to give you a result.
22:19But at the same time,
22:20it's physical infrastructures
22:22having a unique ID of its own.
22:25It's just started with a risk score
22:28of a natural disaster.
22:29We can start adding things like
22:31risk around healthcare.
22:33Can it command,
22:34or become an incident command center?
22:37Can it tomorrow become
22:38a three-story building,
22:40and being able to visualize it
22:41at a rapid scale,
22:42and start doing the construction itself,
22:44visualizing the construction
22:45before the construction happens?
22:47Those are the advanced cases,
22:48but completely agree with the entire panel
22:50that having human assistance in the loop
22:52when we do hyper-local
22:54is essential.
22:56Otherwise,
22:56there will be errors,
22:58and those errors
22:59are going to be expensive.
23:00That's the only caveat
23:02we use
23:02when we build our models
23:03on physical AI.
23:05So,
23:05since we are talking about physical AI,
23:07and our next panel
23:08is, of course,
23:09on healthcare,
23:10and we have Dr. Trehan
23:11in the,
23:13right now with us,
23:14so he'll be speaking about it.
23:16We'll shift focus on that in a bit.
23:18So, let me begin with you, Martin,
23:19and start wrapping up
23:20this conversation.
23:22We are looking at the scale,
23:23and this is the scale of AI
23:26which will be,
23:28which is being built
23:29for,
23:30in a country such as ours
23:32for 1.4 billion people.
23:35So, not just physical AI,
23:36we are looking at how,
23:38making AI accessible
23:40and ensuring that there is no digital divide.
23:43The challenges in a country such as ours
23:45is very, very different.
23:48So, where do you see
23:50current AI fitting in?
23:54I think, okay,
23:55so, two pieces,
23:56and I might try and link it
23:57to the panel afterwards.
23:59I think,
24:00like,
24:00one of the incredible,
24:02of the many incredible things in India
24:03is the diversity of languages.
24:05So, just to go back to earlier point,
24:07I gave that example
24:08of the translation device.
24:09You were talking about voice.
24:11You know,
24:11the fact that a platform like Bashni
24:12is already working in 26,
24:14soon to be,
24:1422,
24:15soon to be 36 different languages.
24:17I think that type of scale
24:19and leaning into it
24:20is really important.
24:21My concern for any country,
24:25be it one on the huge scale
24:27of India or others,
24:28is that AI risks
24:30flattening our culture
24:32into a monoculture
24:34or a few,
24:35small few monocultures.
24:37So, I think of linguistic diversity
24:40really as culture
24:41and as cultural preservation.
24:43So, one of the first places to start
24:45is not only having the AI working
24:48and being able to talk to it
24:49in different languages,
24:50but for the data
24:51to be available
24:52and accessible
24:53with the wealth of cultures
24:55that you have
24:56in a country like India,
24:57that would be the first place.
24:58I think one of the hardest areas
25:01to work in
25:01and one of them
25:03which is so close
25:04to everyone's hearts
25:05is around health.
25:06And the health,
25:07I think the real,
25:08the challenge there
25:09is around,
25:10back to the earlier panel,
25:12around privacy protections
25:13for people to feel comfortable
25:15sharing patient treatment
25:16outcome data
25:17and sharing genetic data.
25:19And that's why I think
25:19India as well as any country
25:21really has a lot to do
25:23to innovate in that area.
25:24Okay, Kishav.
25:27I'm super excited
25:28with the AI summit.
25:31You know, we were there,
25:32we presented in a booth
25:33and it was outstanding.
25:35I think if you look
25:35at the sheer scale
25:37and the quality of people
25:38and every single serious leader
25:41of AI,
25:42barring a few,
25:42were there,
25:43I think we're on to something big.
25:45The government is doing
25:46a really good job
25:47bringing all the resources together.
25:50I'm a firm believer
25:51that if India will use
25:53AI at scale,
25:54it has to be simple,
25:56it has to be reasonable
25:58and cheap priced,
25:59so it has to be open sourced,
26:01and it has to be available
26:04at a lot of languages
26:06and available very freely.
26:08So I think I would imagine
26:11that in the next 24 months
26:12there will be a lot of this
26:13will start opening up
26:14and we are ready
26:16for an AI world
26:17where India,
26:18everyone uses AI
26:19for everything.
26:20Okay.
26:2230 seconds to both of you.
26:24No, fantastic.
26:24I think,
26:26and I believe in it,
26:27the fact that we are talking
26:29about technology,
26:30think of your WhatsApp,
26:31it doesn't have a digital divide.
26:32I don't think the AI
26:33will ever have a digital divide
26:35because it's more focused
26:36on the applications of AI
26:38and less on just the stack itself.
26:40And when an application
26:41becomes the subject
26:43of the conversation,
26:44then that application,
26:46be it health sector
26:47or be it climate
26:48or be it environment
26:49or for that matter,
26:50chip design,
26:51it is democratized
26:53and it will always be democratized
26:55for language,
26:56for any other models
26:57it will be democratized.
26:58I don't believe
26:59that AI is going to create
27:00a digital divide.
27:01Let's wrap this up.
27:02And what I would share
27:04with the audience
27:05is that the future
27:06is about intelligent product design,
27:10products in every field,
27:11automotive, aerospace,
27:13healthcare, energy
27:14and high tech.
27:16And in that world,
27:17we kind of have three pillars.
27:19One is it is silicon design.
27:20There will be a chip
27:21inside that system
27:23which will be software driven.
27:25There will be a software
27:25that runs on it
27:26but it will be AI powered.
27:28So it's AI powered,
27:30silicon designed,
27:31software designed
27:32intelligent systems
27:33with which you can build
27:34digital twins
27:35of everything in the world
27:37and automobile
27:38and aerospace,
27:39airplane,
27:40data center,
27:41whatever.
27:41And it's that
27:42incredibly exciting world
27:44that we are working towards.
27:46All right.
27:46Thank you so much
27:47for this conversation,
27:48Martin, Keshav,
27:49Samitha and Preet.
27:51Thank you so much.
27:51Thank you so much.
Comments