Enabling Seamless AI Adoption in BFSI with High-Performance Computing | Tarun Dua | AI in Finance2.0

0:02 enabling seamless AI adoption in BFSI with high performance computing where we have Mr. Tarun Dwa founder and MD E2E networks limited which is of course uh the pioneer of AI compute infrastructure in India and the first cloud partner of Nvidia in India. Uh Tarun is a computer science engineer who founded E2 networks in 2009 and led it to become India's first homegrown hypers scale public cloud confused focused on a IML infrastructure. Under his leadership, E2 networks has focused on building scales for a IML solutions using open-source which are used by higher education research institutions, startups and enterprises globally. A IMLE transformation is a once in a generation

0:49 opportunity to leaprog businesses and organizations. E2 networks provides self-service public cloud infrastructure and platform tools to deliver this to its customers. The session will be moderated by Satic Jen who's the founder of Generational Capital and CIO and fund manager of generational capital investment ventures. So if you could uh please join me in welcoming Mr. Tarun Dua. Hi everyone. Uh thank you uh for the warm welcome and uh uh good morning to all of you. Okay. So shall we begin? Okay. So I guess like we have been talking about AI for a long time

1:36 now. So everyone has some idea about what is AI and uh how many AI users over here? Okay. Wow. Almost everyone. So, even the people who have not raised their hands, I'm pretty sure you are users of AI because your email provider is running a lot of AI on your email to show you relevant ads and the OTT provider is also running a lot of AI to show you like what to watch. So, almost everyone today is an AI user. Now let's uh uh kind of like do a brief recap. I know it may not be required although I think everyone would have heard this again and again. So what

2:22 was the old world? What's the new world? So sorry. Okay. So we do a lot of stuff. So before we talk about that um see the old world was designed uh for the CPU. Uh when I say old world I mean to say the old IT world. Now CPU when it began like I think like so many decades back there were a lot of limitations that the CPU had. So because of those limitations, all the software that was designed was to kind of like reduce the amount of

3:07 data you were sending to the CPU. So which means that when you designed your database, you put only the most relevant bits and byes into the database to kind of like make sense to the CPU and then you wrote a lot of rules for the CPU to kind of like run on that data and for that uh those IT systems to be fast, you could not like deal with like any extraneous data apart from what was there. So uh now CPU even today yes it is very powerful and while being very powerful it is also very underutilized uh but when you talk about all the a IML techniques CPU becomes like a very small piece of compute in the overall ecosystem. Now when I say small, how

3:54 small? So for comparison, typically two CPUs uh in a single server, each one of the CPUs would be 128 cores, maybe 144 cores, maybe it will go up to 192 cores, but ultimately you are looking at a sub 400 uh physical compute cores in two CPUs, which is the usual configuration. And of course there are multi- uh CPU machines as well. But the most commonly used machines like they use uh typically two CPUs. So that would be about 400 compute cores. Now if you look at the GPU world uh you are looking at like very large number of cores and very high speeded memory which is embedded onto that. So

4:40 there are of course like some specialized CPUs that use the HBM but typically like your GPU utilizes like high bandwidth memory which is typically running at several times the speed of your typical compute RAM and each GPU is about H100 for instance is about 17,000 cores uh CUDA cores. So these are not generic compute cores but like uh but equally not equally but like uh somewhat comparably powerful compared to the CPU cores. So in a single machine you're looking at 17,000 into 8 uh GPU cores and then you also have your normal 128 or uh 264 or some number of like uh CPU cores. So ultimately you're looking at

5:26 like something like thousand times more powerful compute machine at about 10x the cost. Now which is another way of saying that the cost of compute with GPU has actually come down. So when people say GPU is expensive, it's not really expensive. It is actually cheaper to run a lot of stuff on GPU. Now the question is like we say that yeah the data that CPU was doing CPU was able to handle it well enough. Why do you need a GPU? The answer to that question lies in the fact that like the limitations of the original CPUs when all the software started to be designed uh led to people writing uh suboptimal software to only

6:13 deal with 1% of a typical organization's data. So if 100 bytes of data was passing through an organization only one bite out of that was going to a database and that was being processed by the CPU. The remaining 99% of the data that passed through an organization the data in motion that was actually uh never seeing the light of the day in front of any kind of compute. So the only processors that were processing that 99% of the data was human brains. So now GPU is all about like how to utilize this 99% of the data which you break down into the fundamental constructs of data. What is visually uh

6:58 being visible what you can see now not just what your eyes can see but what a LAR camera can see or what uh webcam can see or like what any kind of like uh optical measurement equipment can see. So which means that visible spectrum, invisible spectrum, infrared, ultraviolet, uh x-ray, uh magnetic resonance data like any kind of visual data that can be visualized. Uh GPUs with the help of state-of-the-art open-source AI techniques can see that uh apart from that like language any kind of language. So the language that humans speak, the language that computers speak, uh the language of DNA that biological systems speak. So all

7:43 sorts of languages can be processed by natural language processing systems which AI can use. So all in all we are saying is that like whatever the human brain was doing although in a different fashion the AI systems can do many of those things with varying degrees of success. Now one way to judge uh AI systems is by comparing against uh the human intelligence which is I would say the wrong way of comparing uh because AI is very good at certain things which humans are not. See what do the marketers typically do? They bombard you with so much information that you can't like compare all the choices to come up

8:28 with an optimal choice of what to buy, right? Like so you just look at like uh are you trying to buy an automobile today like you look at like 20 models with same name with slight variation in the code like some of them are based on fuel some of bas them are based on injections some of them are based on number of seats being different. So and this is just like we're talking about like one uh vehicle platform of one brand. Now multiply that with number of brands, number of vehicle platforms. Suddenly the information is so much that human brain cannot process that. Now that is just an example. Now this example uh apply that to any kind of enterprise systems like talk about like simple things like nuts and bolts, right? Like

9:15 so uh the inventory management systems are like almost like several decades old now. Now but even with the best of inventory management systems mistakes would happen in the stores. So physical inventory of any kind like so it would kind of like get helped by having uh systems in place which are AI backed rather than based on human exception handling uh uh radiology data. For instance like every top every city would have like top five radiologists. So basically what we are saying is that like AI is impacting every field that we have today in a very transformational way. Whatever we were doing like certain things AI can do better than humans. Uh

10:02 especially where there is an overload of data like you and I do not have the patience of watching what to see in an MRI. So we are not experts. There would be five experts in a city. But with AI those expert systems can multiply by 50,000 and the cost can be cut down to one/10enth or maybe 100th. So uh things like uh uh uh kind of like pre-diagnostics using uh retina images. There is some medical term for that which I'm forgetting. Obviously I have a human brain not like a brain backed by that many cores. So uh so basically that gives you like okay uh what sort of diseases could the person have? Now all it needs is like one uh

10:49 kind of like an expert doctor one in a million guy to sit watch the training data and give the feedback on that okay this looks like this this looks this this and the AI learns that. Now you can replicate that across millions of devices at every opthalmologist shop for like maybe $1,000 of extra expense of a device that will interface with your smartphone. Uh as opposed to very expensive uh equipment costing like uh uh tens or 20s or 30s of lakhs or maybe even running into kores to look at like uh doing the retinopathy in that way. So uh so lot of impact every sector there is a lot of

11:35 impact. Now we are here gathered to talk about like okay what is relevant to BFSI. Now all of you I think like this is the CFA society. So I'm assuming that you are in the business of understanding business. So to understand business you need to go into the detail. The devil lies in the detail right? So now AI can help you make sense of all the detail that is coming in. So like I said that like human brain can only remember so much. The human brain is always constantly bombarded with information. The entire news cycle is trying to tell you something. Now what an AI system can help you is to determine that in the entire news cycle of bombardment of language the visuals how to make sense

12:21 of it all. So like at a very base level for any CFA for any BFSI like having reasoning systems is important. So let's uh try and see what sort of okay so like at a very base level like every corporate organization needs at least like these things. So wherever there is customer experience help required. Now if you go to any customer services oriented organization which most of BFSI is there is a lot of inconsistency. Uh what we always say is that okay your problem is not getting

13:07 solved here in this particular bank or in this insurance company. You need to find this guy. This guy will solve your problem. Why will this guy solve your problem is because this guy is an expert. So your man Friday is the expert who understands the systems better than the other people. So because like you cannot have consistency of experience spread across all the customers because the top customers are getting the man Friday. The customers with the jar are getting the man Friday. But the customers who do not have access to the jar are not getting access to the man Friday. they are the ones who are complaining on social media about yeah this company is So AI systems can help with

13:53 the consistency of uh building a user experience platform which is consistent across all customers. The dumb systems were the IVRs systems where you were trying to figure out like how how to get out of the loop and speak to a human. Now uh ultimately the expert systems would become strong enough that the system itself will determine like when to transfer the user to like a nonI human who can solve the problem and the right human find the right man Friday because you have solved all the uh normal problems for all the normal users but the pro user who has had the anomalous exceptional problem which could not be solved easily by uh an AI or an automated system or a rule-driven

14:38 system only needs to be sent to them Friday. So this is something that like almost all BFSI which is like typically dealing with like large sets of users like would require I am assuming now we talk about in a very generic fashion that like summary of like uh lengthy documents. What are the examples of lengthy documents? So passing an insurance for or figuring out like the insurance validity of like a healthcare claim uh the insurance validity of like a vehicle having suffered an accident. Now there is so much human intelligence over there that has been applied over the years which is not consistent across organizations. So for example like the user claim is

15:24 submitted is that okay uh like particular X model from a Year car suffered an accident. The picture you are seeing over there is car with a different color or the color that didn't exist in that particular model in that year or car with a user modified parts or the description says that like accident was suffered in this way. The visual aspect says that the accident uh looks like engineered in a different fashion. So now humans have gotten very good at processing those kind of claims with their own sets of false positives and false negatives. Now there would be experts in the system who end up reviewing like a lot more data. Now

16:10 whether your claim gets passed or rejected also depends on the fact that like whether that guy got his breakfast, whether that guy guy got his tea on time or not. So now with AI systems you can remove a lot of those biases, human biases. Again the man Friday problem like you need like uh be able to uh do what like a lot of like nonvalue adding uh personnel are doing with expert systems or help all the people who were kind of like not that expert but with the help of AI they can like develop that expertise borrow that expertise from AI and do a faster better more accurate job. So all through and through. Now

16:56 home loans another area. So you look at a whole bunch of documents and figure out whether to pass or reject a home loan. Now that is again very very inconsistent. Two people very similar similar profile, similar finances, similar family situation like uh one guy could get his home loan passed, another guy could get his home loan rejected. Now ultimately what are you doing? You're reviewing a bunch of documents. So visually the documents can be inspected, they can be OCRD, you can look at like what the uh application has been filled up on. You can look at hundreds of weak signals and come up with an answer. Now these were implicit in the minds of people who were doing that. Hundreds of weak signals that they are processing. Now rulebased systems

17:43 cannot capture all the weak signals that were being processed over there. But uh human mind was able to do that job much better. Now with rule-based systems whatever these human minds uh have answered in the past 50 years by looking at the data from past 50 years, 10 years, 20 years, 30 years, the AI systems can learn with that proprietary data and come up with rapidly changing rules in the form of weights and biases inside the AI models that can tell whether to pass or reject that home loan. Now home loan is an application of AI. When people talk about AI today, they are talking about like building the models, training the models. So what

18:29 happens is that like uh typically uh what media is talking about and where you see lack of impact in the real world systems is the gap that needs to be filled up. So that gap currently a lot less people are filling up than we assume. So there in lies the need for kind of like start doing the AI today so that that that gap can be filled. So which means that AI is not done and dusted today. AI has shown promise in terms of like what we have seen with uh like open AI like being able to do stuff that it was not able to do. people being able to chat with AI much better than they were chatting with the old world chat bots where you asked

19:14 for like okay change my reservation on flight from one day to another and like chatbot would get it wrong. So we most of the businesses are still at like that early stage of using the AI where the AI is not really working for them because on at one end of the spectrum the data scientists are working on theoretical problems uh from a very theoretical standpoint. Yeah, we can handle millions of uh weak correlative uh uh data points and kind of like generate outcomes but to a lot of real world situations we are still not applying AI. Now BFSI thankfully is one of the sectors which recognizes this need and is working towards that. So

20:02 over next couple of months, next couple of years, I think a lot of these stories uh kind of like will come out that where people figured out the use cases of deploying the AI. Yeah. So we spoke about like basically like uh uh there is like so whose loan to pass, how much credit to give, what's the uh like thesis on repayability. So almost all the uh being able to like answer back questions because people don't tend to call to ask questions, they tend to email. Now it used to generate like a lot of like trails and trails of emails like 30 40 50 email trails before someone would

20:48 actually get the correct answer because ultimately everyone was looking for that man Friday who could actually answer the question on email. If it was initially visible from rulebased systems people would have done that. So sentiment analysis which means that while the customer is running a journey with you as a brand is the customer at at what parts of the journey is the customer satisfied at what parts of the journey is customer dissatisfied. Every brand whether it is a BFSI brand or any other brand which has interaction with people is ultimately looking at like creating satisfaction throughout the journey with the brand. So sentiment analysis is very important and then again sending the people to the right people within your organization where there is

21:35 dissatisfaction. So then generically like BFSI sector amongst all the sectors is very concerned about security. So like at a IT infrastructure level the BFSI really cares about like okay security of your data the data not crossing the boundaries exfiltration of data and uh any kind of like uh uh detection of like inside and outside threats and of course like when we talk about CX when we talk about brand customer journey one of the outcomes of that is to kind of like customize is the sales activities uh as per the uh audience. The medium gets customized as per the audience and dynamically not in

22:21 a 3 month or 6 month cycle by the time which brand is customizing the messaging to a particular audience. So again same thing. So again on the IT part uh so not just BFSI but like uh like essentially like learning programs through an organization whether you are sitting close to the headquarters you know more about the brand journey over there compared to sitting in a branch where things don't really work. We've seen that all seen that in hotels like same branded hotel in Mumbai would be great as and as soon as you land up in like a tier 2 tier three city you don't feel the same experience of the brand over there. So ditto for BFSI sector like if

23:06 you are going to a big city the managers know about like okay how to do foreign transactions how to receive money from outside how to do X how to do Y when you go to a branch outside in a tier 2 tier three city people don't know uh like how to do X so very big problem in BFSI sector today is that like only at certain places like where certain teams are able to help you same man Friday problem so uh Again AI can help a lot with that. So again we have spoken about like any kind of conversations uh that customers are having like a lot of those c conversations the first 60 seconds first 120 seconds like uh

23:52 usually where there is some part transactional help required that like you received my money yes or no what's the UTR number of the NEFT that I did etc etc. So all of those can be done in like first 60 seconds without involving any humans. So next level is where you start training specialized experts in AI to do uh multiple things where people were working 24 into 7 on various areas which are not like front end related which are not your uh uh where you're not looking at like the uh uh satisfaction scores. So NPS so these are the non-NPS areas. So where people are working at a back end to do certain activities. So

24:38 until now what we have spoken about is like uh how to do high NPS activities. So all of AI compared to all of the software today the rule-driven software and AIdriven software the basic difference over there is that like uh the state-of-the-art and open source is such that all you need is underlying infrastructure you can swap and swap out models uh and there are like the compatibility infrastructure layers in between that allows you to have queries and responses to the new models running on the same data uh without making a lot of changes to the software. So all of that is helped by having like

25:26 the right kind of AI infrastructure in place, generic infrastructure in place which is performant which is a no compromise infrastructure and which ensures the software layer to do all the rule-based activities very easily and for getting access to the AI based activities which are increasingly uh solving for newer problems without having to write additional code. Now example of that would be that uh when you deal with current systems uh you find an anomaly, customer finds an anomaly, customer complaints, then 10 more customers complain, then thousand more customers complain and you have an audience of 10 million. Sometimes uh like when those thousand people are making more noise, you would see those

26:14 uh changes being made at a high priority. High priority means that in 3 months the problem gets solved in code gets deployed in production and those thousand people stop complaining. in AI uh as soon as those thousand people complain like somebody can look at the data because of which like those thousand complaints came in uh do a human reinforced feedback to the AI system and the problem can be solved in a matter of hours or days as opposed to taking 3 months in a rulebased system where the rule engine had to be modified tested and deployed in production. So over here the whole cycle of changes to how to deal with anomalies can be compressed and that compression is

27:00 possible because of the size of the compute infrastructure that is now available to you. So rulebased systems required you to run things from a relational database on a CPU that required a lot of humanled code changes 3 month cycle. AI systems are working on billions of rules which are generated uh using open-source software and running on top of infrastructure which is uh high-speed infrastructure in terms of bandwidth uh high bandwidth infrastructure GPUs with that many thousands and then large scale supercomputers with like large number of cores. So that allows you to make those changes by telling the expert systems to make those changes for you rather than

27:46 you making those changes by feeding it curated data. So now that completely changes the way you looked at IT systems in the past. So the infrastructure for BFSI completely changes. So earlier you were looking at like okay code reviews and code changes. Now you're looking at like data scientists making those changes for you uh without like a lot of code level oversight but you need to build like now newer systems for oversight of this AI. So when we talk about ethical AI you say that okay like the garbage in garbage out still applies. So which means that like it is very important to ensure that like what you are feeding to the AI

28:31 systems on infrastructure and where you are again getting the right kind of output and then output gets tested and then you kind of like uh uh throw it at the world. So that's that's that's where AI is today. Now this is all kind of like an ongoing work. Uh have we reached this stage? No. Because the first step is to deploy the AI infrastructure. Second step is to start feeding it data. Third step is to see in non-production how is that helping you developing the confidence on those systems. Fourth step is to roll it out to the customers. So I think like BFSI is still about like uh at least like a year away from rolling out and deploying these systems. But

29:17 once these systems are deployed like it's going to be rapid progress. AI is kind of like all or nothing. At 99% accuracy, your false negatives and false positives should could be so expensive that uh it doesn't make sense to deploy that 99% accurate system. So imagine like a 1% failure of KYC systems to recognize who you are like uh uh for that means like from 10 million customers like that 1% is a huge number of people who get impacted like every single day. So you can't deploy a 99% system. So which means that your training cost goes up by 10x before that 99% increases to 99.9%.

30:05 and you actually are able to uh deploy it at 100x the cost. So 99% accurate system will cost you one10th or 100th of 99.9% accurate system or 99.99% accurate system. So today in AI mostly we are happy and looking at like yeah 99% this works my humans also do 99% good job. So let AI do 99% good job. But there would be areas where 99.9% is required. 99.99% is required. I think like that journey requires like more investment into infrastructure. So today we are at a stage of seeing that investment in infrastructure and that investment in

30:51 infrastructure there would be some validation some proving that yeah this works for us. Uh the promise of AI in terms of like automatic generation of rule sets is going to work for us. We have started seeing the decline of uh uh complaints in the customer experience. We are starting seeing sentiment improvement in the customer journey and then that will grow another 10x because what you are achieving at 99% to go from 99 to 99.9 you have to grow the infrastructure 10x. So okay cool. Um so broadly coming back to how we help so we built a a IML platform to do

31:38 this entire journey uh very very quickly for any organization. So essentially when I said that there is a world where data scientists are living there is a world where humans are living to marry the two you need a platform. So tier is that platform where you are marrying the journey of a data scientist with real world feedback coming in from deployment of an application in production and allowing that data scientist to look at the anomalies that got detected in production and then refeed that to do the training runs again and again and again and then again simply go and do deployment of that in the uh uh production uh for inference. So that

32:24 entire life cycle and journey I've skipped over like lot of technical parts uh because I was told that primarily this is the CFA audience. So I didn't want to go deep into technical. Uh there is enough AI out there in terms of like OpenAI or Google Gemini to kind of like try out the uh MS office like capabilities or co-pilot like capabilities out there. But over here what we are talking about is to build B2B enterprise systems. So that was would not have made a very good demo in terms of like uh so BFSI coding jobs are one of the most boring jobs. So which is why I went into infrastructure. So uh so that would not have made a very great demo. So with this I would uh kind of like uh open uh

33:10 the discussion with satic um and uh then we'll proceed with uh I guess like audience questions. Uh thank you everyone. Uh so so Tarun uh there has been a recent chatter about deepseek like a lot of impact on it on uh as to AI. So uh what do you think will be the impact of deepseek on the industry as a whole? So overall it's a positive impact. So yes we keep talking about like uh building more and more infrastructure. So the idea is that like more infrastructure is required because more number of people are adopting. See initially everyone today is adopting AI in very small parts. If there is a large organization

33:57 they would have identified like maybe a couple of use cases to kind of like work on AI. Now what deepseek does is brings down the cost of doing AI not just for those five use cases which means that if there was one organization considering five use cases now it would be the same organization considering like maybe in the same budget we can do 10 use cases and outside the organization where people were doing zero use cases would also start looking at doing another five use cases. uh very good example of that is that like uh I think IBM famously said that there is probably a compute market of grand total of five supercomputers in the whole world and that was because those supercomputers costed like tens of millions of dollars or hundreds of millions of dollars. So

34:44 now as many computers came in and then PCs came in the compute exploded. So originally in India I think like we were like at 10 million mobiles the mobile data was expensive the mobile calling the roaming everything was expensive as cost comes down the adoption increases so ultimately that is to the benefit of the industry on a whole where somebody advances the state-of-the-art in open source and uh the promise of AI actually starts getting delivered to more people. So in fact there was this chatter that there is an incubator in the US where in a particular cohort there were a lot of companies with the coming of reasoning models from deepseek a lot of things that they were working on now they

35:29 started working without any extra effort. All they had to do was swap out their old open source models with the newer open-source model and whatever they had set out to do now started working for them. So their time to market actually got cut down. So now that being said like uh there is a lot of like misconceptions also about deepseek. So deepseek is not like super replicatable uh outside of very specialized teams. Yes the cost was less but you are really talking about like cost of one model training run. Second uh you are not talking about like uh people like you and me building those models with the current state-of-the-art. you are literally talking about people who know how to code assembly language computers.

36:15 So the language called C was a massive advance and C is considered as very complicated. Most people write code today in a a higher level language like a Python or uh similar. So like uh what they have done is they have not used the libraries provided for making programming programmers life easy. They have rebuilt those on their own. So which is a very specialized job which takes a lot of time and requires very very specialized people who really understand what's happening on the wire. So it's not that like uh the cost is actually that less. Uh lot of that cost has been covered in terms of like the people who were doing that job like literally if you you and I were to try to hire those people the cost would be

37:00 very very high. uh so so yes it's definitely from a technology standpoint a reasoning model is a massive advance some of the techniques they have mentioned uh how to do multi-attention with a mixture of experts compared to the old beginning of transformation uh transformer models where we said that attention is all you need so that has they have definitely advanced the AI science AI data science so but whether it has like any negative impact no So it is expected in the open source world that technology would change very very rapidly. What was there one year back would not be there. Now in AI world like I said that like since uh you don't have to go through the whole cycle of

37:46 rewriting the code every quarter you can actually do that cycling uh once every week once every two weeks uh because of the way development on AI systems developing new models or training new models or finetuning fine-tuning new models works so the cycle time for new state-of-the-art and AI is literally in weeks so the rate of change is is going to be very very rapid over here. So it is to be expected. So nothing unexpected about like a deepseek like model. So now that one has come in I think it's going to be a dozen more very very quickly. So so so given that it is originating from China uh what are your views on the privacy concerns the public? I think the privacy concerns are very very similar. So if

38:33 your business depends on uh your middle management's intelligence. So I think like in the management theories like 20 years back 30 years back people were talking about uh where a middle management in a company leaves and causes organizational amnesia. So uh lot of those theories were there. Now AI can prevent that organizational amnesia. Now on the other hand when we use something proprietary which is not self-hosted. So examples of non-self-hosted would be where your email is being read by AI systems not under your control where your chat bots are running on top of a SASbased AI system not running on your on-remise infrastructure. So when I say onremise what I mean is that where

39:20 infrastructure is fully certifiably completely under your control. So which means that any sort of multi-tenented systems are a no no. So it doesn't matter whether your multi-tenented AI systems provider is an Indian company in India is an American company based out of US and sending your data to US data centers or is a Chinese company sending your data to China. Uh from your and my perspective they are all equivalent. So when you build B2B AI systems where the core of your organization's knowledge which is what your organization practically is which is basically the knowledge stored in the brains of your people which you are slowly trying to make the AI learn if you send it out

40:07 you're losing the source of your competitive advantage. So which means that there is going to be a massive movement. So I'll go out on a limb and say that like look uh SAS is uh great SAS was great in the rulebased systems but now you need singletoned SAS that you're running on your own compute infrastructure. So when I say your own compute infrastructure uh it means that it should not be shared with anyone your data should not leave your organization boundary howsoever you define. So two ways to do it even on hyperscalers you can use the APIdriven parts of their infrastructure where your data uh is going and getting shared with their systems howsoever much they certify and do everything your data is crossing your organization's boundary. Second part is

40:53 even in a hyperscaler you have the possibility of using the raw compute where your data is not crossing the organizational boundaries. So in AI it is very very important that we don't all look alike uh in terms of like okay I have a writing style in which I answer my emails. Now uh when somebody complains that like look this looks like chat GPT generated. So as a uh executive in a company I don't want to be the one who is leaking out that information out to chat GPT on how people in a company are thinking. So which means that you need to have like a personal domain AI which is provided by the organization to you where whatever you are writing is not crossing the organizational boundaries.

41:40 So which means that eventually we will need to have a very close look at all our supply chains. So the IT supply chains especially are super vulnerable. Uh when your OS boots up, you are on a foreign provider who is deciding whether it is safe and secure to boot up your machine. when you write your email, it's another foreign provider who has the ability to look at everything that is passing through their systems using AI. And the purported reference over there is that look, this AI is not doing anything. It is just trying to do anti-PAM. It is just trying to help you autocomplete the sentences. Uh but the answer is no. Like like something is being learned by these systems which is going to help your competitors at some point of time.

42:26 Whether that point of time is three months down the line, six months down the line, six years down the line, I can't say. But ultimately AI has to lead to people bringing back their data at a place where they feel confident and secure about where their data is. So whether that is fully onremise or whether it is in the form of like raw compute where like uh expert systems from outside your organization's boundaries in multi-tenented environments are not looking at like how to learn from your data and give advantage to your competitors. So there is a very interesting question from the audience. You mentioned about AI agents in various fields. to what extent do you think they can take over from humans and if possible how long

43:13 will you will it take to implement? Okay, sure. So I think like at one point of time there were a lot of uh humans doing some very dumb jobs then IT systems came and we said that like look why do you need to do this dumb job why don't you do something more creative so did it lead to decline in number of jobs the answer is no so as the technology advances we all get to live like kings so because the overall productivity of any economy increases and what was possible only for the elite or for the kings of the old now becomes possible to do for the entire populace and there is profit in doing so which is what capitalism is all about like how to make life easier for uh like people uh with

44:00 the smallest amount of money. So ultimately this is what AI is going to do to economies where like certain jobs will disappear they will be replaced with other jobs. Now it may cause temporary uh problems in terms of one generation to another because ultimately at some point of time the humans stop to learn. They get set in their ways. So obviously people who were running the PCOS they are not getting their jobs back as PCO operators right so but there are other things to do for their next generation of kids. So I think that change will always cause some pain but if you avoid that pain today that pain is going to come back to you

44:45 multiffold. So because the world is not stopping for one country who decides that okay no this is not what we are going to do. So we are not going to let computers be implemented or we are not going to let AI be implemented or we do not accept job losses due to AI. So short-term pain we will have to accept uh with a view that like ultimately it is going to benefit everyone and the overall economy in a family it is probably going to benefit the younger generation the most. So like that that change is inevitable and uh do we see any policy from the government side which can help in uh like you know advancing AI adoption across the country? Sure. I think like the starting point of everything is education. So uh now education of course

45:33 like it's a it's a topic close to every middle-aged person's hearts because they are seeing their kids educated right now. So broadly if you look at like how India educates its kids uh the elite spend about hundred billion dollars a year to send their kids abroad where uh they learn from best of the best universities abroad. So I think like AI has the potential to change all that. So and that is the starting point of like reskilling people not just the people who are in jobs but also the kids who are going to learn. So I think like uh uh that's where that's that's the biggest imperative for a uh any government to work on today to how to make sure that like we do not get left

46:19 behind. Given that most of us are from the wealth management or the asset management field, what do you think are the top three uh use cases uh into those specific I think it's the easiest when you talk to wealth managers they ingest a lot of data they are always talking they are always finding new information they are always trying to make sense of new information that is what the AI models the reasoning models do so they are your best buddies so I think like uh uh practically being able to ingest a lot of information. Being able to make sense of a lot of information. Uh being able to search for information without getting bogged down by reams and reams of massive amount of digital text that gets thrown at you when you use your

47:05 traditional search engines. I think that's the major change that is going to be there for wealth managers. And then of course there is a personalization aspect where you say that okay uh I am not just giving you humanly possible answers yes and no variety answers uh saying positive or negative on every new development but I can probably put a learning model where I am feeding it the correct curated data and it can answer most of my customers like uh directly where it says that is this news or development positive for X sector yes or no. So like you asked the question about deepseek probably there could be a model that you could give to your customers where it tells you whether deepseek is good for AI infrastructure good for AI application development guys and how so

47:52 it can give a both a quantitative and qualitative information with the reasoning it has put behind it. So you can validate the reasoning as a human to say that okay I agree with this reason or I don't agree with this reason. So but it makes like life easier for everyone who is working in any field. So Tarun over the last 15 years you have had a very inspiring journey right from scaling up P2E networks to raising capital from marquee VCs like Bloom ventures to do doing one of the first theme IPOs in this sector to then having a strategic with Larsson and Tubro. So what would be the uh advice you would give to the young students out here as to building on a business as most of them would like to like you know go on that entrepreneurship path. Sure. Sure.

48:40 Sure. So uh uh I think like uh entrepreneurs are like a breed who see things that other people do not see. So and second entrepreneurs like obviously uh work in cycles. So now with new technology in place that cycle time for doing anything is coming down drastically. What used to take like so many years because of lack of capital availability. Now India continues to suffer from massive lack of capital availability. But what the newer tools and technology gives you today is the ability to do far more things that you could do with smaller much smaller amount of capital than in the past. So again like anytime in the future anytime in the present is a good time to become

49:26 an entrepreneur where you are seeing something that other people are not seeing where you are seeing value. So like that's the point at which like you become an entrepreneur and say that okay I'm not able to convince the existing players this is the right way to do things. I really believe in this uh this idea has value. Go ahead and do that. Prove the naysayers wrong. Nobody can tell you what is the right idea, what is the wrong idea. You yourself can go into the market and validate that idea. The market never lies. So if an idea works, the market will tell you. If the idea doesn't work, the market will tell you. And uh the tooling for doing things has become cheaper and cheaper. So make use of that tooling. So we used to say that okay like uh

50:11 uh sending rockets to space is the simple part. Designing chips is the more complicated part. So we're not doing something simple. We're not sending rockets to space. We are designing chips. Now if you look at the cycle time for designing chips today. So where people have built those processors over multi-deade efforts. Now we see newer chip companies coming in and designing those chips so rapidly. So because the tooling has drastically improved, the cost of doing that those things has drastically reduced. So which means that if ship designing has become like a half a decade cycle today compared to a multi-deade cycle it used to be I'm talking about fabless again like uh fab will also eventually improve in the

50:57 cycle time uh people will catch up on the building the fab technology as well but I'm talking about the software layer in which the chips are actually designed so that cycle time has drastically come down today like there are newer CPUs being designed there are newer LPUs being designed today and the cycle time has come down to less than 3 4 5 years today and I think it will decline further. So this is the most complicated thing that you can build today. Uh applied materials research uh where you have the ability to simulate the material properties based on uh the molecular formula or the folding of the molecules like you can do that in like uh on uh GPU backed systems today much faster. So again the cycle time has drastically come down. The newer

51:43 materials would get invented much faster. So especially in the plastic space. So similarly biomolelecular research like uh people working with like synthetically modified organisms like fungus or bacteria to do certain things maybe even producing some biochemicals. Uh so cycle time every everywhere is coming down. So uh so that's that's the key like reducing the cycle time on whatever you are doing is the key. So uh reduce cycle time means that like any new entrepreneur has the ability to build upon what has already been built. So you are no longer building things from scratch. what you need to do is like uh use the expert

52:28 systems to understand the existing supply chains, existing gaps, uh how to do stuff is uh much easier to understand today than it was 10 years back. So like you could sit in a uh skunk works kind of a place and say that okay I want to design a new geyser. I want to design a new refrigerator. I want to design a new washing machine or maybe a altogether new appliance for the consumer market which has not existed in the past which is based on the unique circumstances of India. So you could do that. So like the possibilities are endless. So if you want to be an entrepreneur like it's always a good time. So since we are running short on time, one last question from the audience. Uh what would be the

53:15 cost comparison on capeex versus OPEX of implementing artificial intelligence in small versus large companies and what is the trend you are seeing whether companies are going mostly on the OPEX model or the capeex model. I think uh what we are talking about is that like if you have the ability to make capex today in AI, it gives you a definite advantage in terms of building up your AI systems fast with uh expensively running open source. For example, a typical LLM with like a trillion parameters. Uh cheaper to deploy, cheaper to use, uh cheaper to make use of in your applications, uh compared to SLMs which require specialized workforce, uh more expensive to build

54:01 but cheaper to run. So I think that's the trade-off. Now it's a question of capital availability. Where capital is constrained, you will end up building SLMs, uh small language models. where capital is unconstrained uh you would go for fastest cycle time to uh get to solve the problems that you're trying to solve. So which means that like your capex would be higher. Sorry about that. So so that's that's that's the difference between a larger organization today and a smaller organization today. Although uh we are sitting with finance people over here the goal is financialization of the economy. The goal is to bring capital where capital is in short supply. So ultimately like uh if capital is a commodity, everyone

54:47 would want to do things fast. If it is worth doing, it is worth doing fast. So thank you Taran for all the knowledge you have shared. Uh I think I can request the team to take up. Uh can I have uh Mr. Shame Choki uh felicitate the speaker and the moderator please. Thanks. Thanks Dun. Thanks so much for this uh amazing session.