Netweb Technologies India Limited (NETWEB) Earnings Call Transcript & Summary

Ladies and gentlemen, good day, and welcome to the conference call of Netweb Technologies. [Operator Instructions] Please note that this conference is being recorded. I now hand the conference over to Mr. Hardik Rawat of IIFL Capital Services. Thank you, and over to you, sir.

Thanks, Rio. Good afternoon, everyone. On behalf of IIFL Capital, I welcome everyone to Netweb Technologies conference call. We have the pleasure of having with us the senior management team of Netweb Technologies, led by CMD, Mr. Sanjay Lodha; CFO, Mr. Ankit Kumar Singhal; Whole-Time Director, Mr. Navin Lodha; Chief Sales and Marketing Officer, Mr. Hirdey Vikram; and Head of Uirtus Advisors, the IR advisory firm to Netweb Technologies, Mr. Sanjeev Sancheti. Without further delay, I'd like to hand over the floor now to Mr. Sanjay Lodha, post which we'll have the Q&A session. Over to you, sir.

Thank you, Hardik. Good afternoon, everyone. Thank you for joining the call -- joining us today. I would like to take this opportunity to address the recent global developments surrounding the emergence of DeepSeek, the new AI large language model platform and provide clarity on its potential impact on Netweb's growth trajectory. As a leader in high-end computing with AI as an important vertical, we view the emergence of DeepSeek as a significant opportunity for our business growth. This perspective is grounded on the 3 key factors. Number one, market expansion. DeepSeek paves the way for inclusive AI adoption, expanding the market further by lowering the cost barriers associated with advanced technology, it enables a wider range of customers previously hesitant due to high adoption cost to access and utilize appropriate computing resources. This empowers them to leverage AI effectively to address their business challenges, driving greater demand for our solutions. As you know, basically, all our is GPU-neutral solutions. Netweb AI solutions portfolio is designed to provide multi-GPU public APU platform that cater to both inference and training architectures. Our offerings include hardware, middleware and utilities that seamlessly integrate with end-use application like DeepSeek. As such, the performance acceleration enabled by platforms like DeepSeek will drive better adoption of our solution. Government and enterprise adoption. The third and critically important factor is the adoption of platforms like DeepSeek aligns with the interest of the local governments and enterprise. This will accelerate the development of similar platforms within India. The Indian government's current AI policies explicitly emphasizes developing indigenous large language models and domain-specific AI models as a key pillar. This focus is aimed at harnessing global technology disruptions, which will further propel India's AI-driven efforts and investments. In fact, such disruptions will only hasten India's commitment to advancing its AI initiatives and allocating resources towards this transformative technology. These factors clearly demonstrate that evolving platforms and technologies will significantly boost the adoption of our AI solution, reaffirming ample growth potential. These developments align seamlessly with our product and solution design strategies. As a result, we do not view these advancements as a threat to our business, rather we see them as a catalyst for future growth and innovation. Netweb's fundamentals and growth momentum. Netweb's fundamentals remain robust, both technologically and financially. This is reflected in our consistent growth over the last few years. We have consistently achieved strong year-on-year growth. In financial year '24, our revenue from operations increased over 62% and in 9 months financial year '25, it grew by more than 60% on a significantly higher base. This demonstrates our ability to sustain robust growth momentum even as we scale. Strong order pipeline. Our order book stands at INR 3,603 million with L1 at INR 3,481 million and a pipeline of INR 38,149 million as on December 31, 2024. Our R&D-driven in-house design and manufacturing capabilities enable us to produce technology-agnostic world-class compute platform. With a strong business pipeline, continuous enhancement of our capabilities and ongoing product expansion, we are strategically positioned for sustained growth while preserving our leadership in the technology. Our road map for future growth is guided by the following key pillars; curated product lines. Our product are designed to be resilient against sudden disruptions with each line offering robust growth potential. Our business model is dependent on these 3 growth pillars like HPC, representing over 30% of our revenue, private cloud and SCI contributing over 30% of our revenue, AI systems accounting for more than 14% of our revenue. The growth driving technology with these product lines provide ample headroom for continued expansion and clear road map for coming years. Our products and solutions are strategically positioned in the market, becoming the preferred alternatives to many globally recognized technology platforms for customers' high-end data center and computational needs. Thanks to our continuous product innovation and competitive pricing, we are confident our offerings will drive significant disruptions in the technology space in the years ahead. Minimal CapEx requirements. Following the completion of our recent capability enhancement, our capital expenditure requirement to support near-term growth will be minimal. This will go a long way in ensuring that we continue to maintain strong return ratios. Netweb distinguishes itself as a leading technology OEM with a unique market position. Thank you.

Should we begin with question-and-answer session?

Yes.

[Operator Instructions] Question is from Aditya Moona from YES SECURITIES.

My first question was regarding DeepSeek. So, we've seen that DeepSeek is showing a rapid growth, signaling a shift in the AI industry dynamics, especially towards the lower training costs. How do we, as a company position ourselves in this environment to have a much more positive impact?

So, Aditya, thanks for the question. So, just to answer that, DeepSeek is ultimately, it's a user application in the form of an LLM platform. And such platforms are ultimately enabling the solutions which we provide by creating a possibility of accelerated computation. So as such, it is not leaving any negative impact over our business prospects, number one. Secondly, it is only helping us to accelerate the performance of our platforms because the platforms what we provide or solutions we provide, this comprises of hardware and the middleware and the utilities which sit over -- above that. And after that, only the user applications come. So, whatever acceleration we can see through DeepSeek that can be utilized by us while delivering performance through our solution. So that way, it has got a positive impact only on our solutions.

Okay. So, in terms of the -- just a follow-up on that. So, in terms of that, do we feel that we have to readjust our AI aspects towards the R&D aspects or the product portfolio?

No. Actually, as I said that the solutions what we design, the product which we manufacture, so that has got support for all the architecture based on all the different designs. It comprises of low-end GPUs, midrange GPUs and even high-end GPUs also. And the same goes with even APUs also. So from that perspective, if DeepSeek as an application utilizes low-end GPUs, let's say, for reference, then that can also be straightaway be used by our customers who wish to basically use DeepSeek or a similar platform in case of our solution. So, we don't need to make any change to our architecture. Our architecture very well accommodates not just DeepSeek, but even similar platforms as well.

Okay. So, no matter what the platform of this, basically, whether it's a middleware, hardware, your product will be adaptable to the -- whether it's DeepSeek model or not matter which model, AI model.

Yes. Our solutions are already designed as per those architectures as well.

Next question is from Chirag Khasgiwala from Neo Asset Management.

So, a couple of questions. One is, I mean, there is a lot of news flows going around that with the DeepSeek coming in, the demand for data centers will go down significantly. So, what's your view on that? If that happens, then what could be the impact on your demand? And secondly, you have a tie-up with NVIDIA. Now DeepSeek has shown that the AI LLMs can be built up without using NVIDIA chips. I don't know how correct is that statement. But assuming if that is correct. So, what happens to your tie-up with NVIDIA? Does it become irrelevant? How soon you'll be able to shift your servers and all from the technology provided by NVIDIA then to the technology which is being used by DeepSeek and all?

So basically, I'd like to -- unfortunately, the answer to both your questions is no, actually. Basically, first is that you say the data center demand will go down. I strongly oppose this actually that data center demand will not go down, data center demand will go up rather because basically, what is DeepSeek doing? DeepSeek is helping so as to basically train the models faster, okay? And basically, so more and more models will get trained. So basically, I personally feel is that, and I think that's the industry view also data center demand is nowhere going to go down. It's going to go up only. So that is to answer number one. The number 2 is that basically DeepSeek is doing without NVIDIA. I think you need to check it very clearly. DeepSeek is also using GPUs and it's not -- basically, primarily they are using NVIDIA GPUs, but basically, this can be done even without -- but it cannot be done without GPUs. So, DeepSeek, they have not developed some GPUs or something. It's primarily a LLM model which is more efficient and basically designed in such a way that it runs on a lower GPU power actually. But primarily, it cannot be done without GPUs. If you -- the entire model is based on NVIDIA GPUs only, which they have tried to do actually. So, this will not derive the demand away from NVIDIA. Rather, what happens is that this will make adoption of what we have been -- which I covered in my opening remarks also, like basically, this will make AI more affordable, okay? So basically, more and more applications, more and more users will like to basically train their models and take use of -- and make use of AI. So, it is democratizing the entire AI infrastructure. So primarily, this will drive demand in a big way. And you see in that direction only the India government yesterday also had a great announcement wherein basically they are -- the India government is trying to work on India's own LLMs actually. So all the countries, everybody will try to work on their own LLMs, which will be efficient and that the total cost of deploying these LLMs will go down. So basically, the more adoption will happen. And hence, we feel this is a great demand basically generator rather than demand killer.

The next question is from Manish Choraghe from KJMC Capital.

So recently, we just saw the announcement of Cabinet Minister, Mr. Ashwini Vaishnaw about the India's own AI mission for which they are -- they have allocated INR 10,000 crores of budget and they will be procuring more than 10,000 GPUs. And they have also empaneled a few domestic companies. So, Netweb is directly or indirectly is any -- is getting any benefit from them? Or do we see any opportunity there for the growth?

So see, just to share with you, the current RFP, which they had floated, there is basically an interim arrangement for the empanelment of cloud service providers to basically provide GPU cycles available through cloud. So that is basically -- you can say that they are trying to use the available resources by means of a cloud. So, that is quite a stopgap arrangement, which you can see. And the actual requirement or the actual infra, which Government of India is trying to build, that is a separate one. So, we are focused around that. And definitely, you can see in coming quarters the actions around that. So absolutely, the opportunity is definitely ahead us and we are focusing on that.

Next question is from [ CA Garvit Goyal ] from Nvest Analytics.

You mentioned like DeepSeek is using low-end GPUs, right? And they are also using NVIDIA chips only. So, I want to understand like what is the difference between low-end GPU and high-end GPU. Is it going to affect us any way in financial terms like in terms of margins or anything like that? That is one. And secondly, you mentioned like India is also working on their own LLM and like other countries will also do it. So, all these are going to procure the GPUs from NVIDIA only, which we are having the tie-up?

So basically, to answer your question, basically high-end GPUs and low-end GPUs, actually, what is happening is basically, DeepSeek has basically made the models more efficient, okay? So, they are still using NVIDIA GPUs. But basically, those models can be run on -- not on the top grade model of the GPUs, even on the -- basically the GPUs, which are #2 or #3 in the range also, even that they can also perform the task because you understand that necessity is mother of invention. The basically, because it's a Chinese company and they have denied GPUs, that's the reason they made their LLMs more efficient. So definitely, it's running on NVIDIA. They have not developed their own GPUs or something. It's primarily running on the GPUs, which have been manufactured only. So basically, the LLMs have become more efficient. So, if they can run on the high-end GPUs and even result faster, they can run on slightly lower GPUs also and still provide you the result. So that exactly is the meaning of that. And as regards basically, the business or margin is concerned, what we personally feel is that it has got a positive impact on us because please understand the technology becomes more affordable. So, more and more use basically of -- because we are not manufacturers of GPU. GPUs, we have to buy from NVIDIA, from different OEMs. Actually, we are vendor agnostic in that case. So, what we do is that basically the demand for systems, demand for GPU-based systems for such kind of basically LLMs, training and all will increase and this will further fuel the inferencing demand. So, I think this is basically a great India moment wherein government has also understood wherein basically such demand will really go higher.

For low-end GPUs and high-end GPUs, there is no difference in the pricing is what you're saying?

No, I never said that. There can be difference in pricing, but adoption will increase. Now basically, like if you're running a model on basically H100 and you are running a model on the lower-end GPU, definitely H100 will perform better and faster. But you may not need that kind of fastness. So, you may need lower number of GPUs maybe for the training. So that will reduce your cost or the lower number -- if you are using the lower GPUs also, basically, your total cost of the infrastructure will go down. But the infrastructure which is being used currently will multiply. Please understand that, that more and more applications will clearly try to take advantage of this and try to do it.

Understood. So right now, what we have -- like the arrangement that we have with NVIDIA is we are procuring the chips from them so that those chips will get -- the cost of those chips will get reduced and we are integrating those chips with our systems and then we are supplying to our end customers. Is that understanding correct?

I will not say that the cost of the chips will reduce because NVIDIA has not yet decided to reduce the cost of the chips at all, okay? But we were still supporting all the chips from NVIDIA and we'll keep on supporting all the chips for NVIDIA. So basically, previously, we were selling 100 chips, we'll be selling maybe 500 chips to explain you in very simple language.

Next question is from Lakshay Agarwal from GrowthSphere Ventures.

So, I just wanted to understand that due to the recent developments, the amount of compute required for training would reduce as we would not be requiring like high amounts of data to train the model. But likewise, inference compute would also increase. So -- but do you feel that due to these companies like our customers, they would be investing less on the infra as of now because they were investing more, but will slow down their decision-making? And accordingly, we could see some slower growth in our order book at least in the shorter to medium term.

See, basically, just we explained this a little a minute back also that basically the requirement is not going to get killed with such kind of LLM applications or any other application. Such platforms will only enable a set of users who basically want to focus on certain type of workloads. For example, DeepSeek is largely oriented around FP8 kind of workloads and all. So, they will focus on that. So, this will basically help the users to adopt AI in a bigger way. So, those who are basically sitting on the fence and they were not finding an option to basically utilize AI technology, AI as a technology, they will also start using it now. So, this is not killing the opportunity or taking away the opportunity. That is one. And second is that for those applications where the need of workloads do require higher-end GPUs and they require higher precision as well for the processing of data, they will have to still use high-end GPUs only. So I mean, both the cases are completely different. Whereas with the help of DeepSeek, et cetera, it will help users to adopt AI in an easier way. So, you will rather find that those who were not earlier using it, they will also start coming to that bracket wherein they'll start using AI. So, this is only going to increase the demand, nowhere basically taking away the opportunity.

And as you know, basically, our business is split into 3 major pillars of verticals. One is the HPC. That is around 30%, 35% of the business, 30%, 35% of the business is private cloud and HCI, around 14% to 15% business is AI systems. And we have very categorically clearly mentioned that this 15% will become maybe 20% in next 2 to -- maybe 1 or 2 years. So that's the kind of growth which we are seeing on that since we are a company which is growing at 30%, 35% CAGR year-on-year. So, all these segments are growing for us. So, we personally feel this will rather boost our growth and anyway not hamper our growth.

Correct. So, like I did understand that. Actually, I was asking from a short- to medium-term basis because I do understand that in the longer term, obviously, we'll be having more traction and it will just increase our demand. But okay. And secondly, I had missed on the order book size. If you could again repeat that, what is the current order book size? And how much is the -- how does the pipeline look like?

So basically, as I have been mentioning since we got listed for 7 quarters, that we -- the first thing is that our order book is INR 360 crores as on 31st December, okay? And basically, it was around the similar number last quarter also. So -- but basically, please understand that. The order book seems -- will look to you same, but it's completely building for us because the order book got 80% rebuilt because our order book gets invoiced within 8 to 12 weeks, okay? So, what happens is that if there is an order that will definitely get built in 8 to 12 weeks. So, the new orders will come in and they will be there. So, order book is really -- we are not -- because once the order comes, we consider it sold actually, really speaking. So for us, what matters is that for our forecasting and all, we use a funnel. And we have a funnel of around INR 3,800 crores, okay, which basically which will suffice us for around 1 year to 18 months. with a conversion ratio of 60%. So, we will be seeing growth in that. And we are very confident that basically that's the reason for the -- I presented the second quarter results and we have shown consistent growth actually. And whatever we have committed, we have over delivered and we'll continue to do that.

Next question is from Nitij Mangal from Jefferies.

One, can you talk a little bit about how the lead times on various chips, especially on the NVIDIA GPUs? And secondly, when you look at your overall hardware bill of material, is it possible to roughly split that across various companies? How much is NVIDIA, how much are other companies?

Yes. So basically, as regards lead time is concerned, basically, we don't get much -- because since we are an OEM partner with NVIDIA, they have our forecast. And even in the most crisis period also, we are getting good support from NVIDIA. And since we are the only basically a design and an OEM partner for them in this reason, we get adequate amount of support from them. So, we don't have any particularly as regards to lead time, we get it as per plan. We don't have much challenges in that. As regards basically split around because our major technology providers, I can tell you is Intel, NVIDIA, AMD, Seagate and Samsung. But basically, components come from various companies, various people. It's very difficult to basically distribute and segregate who are the basically in terms of what kind of revenue, what kind of we generate for each of them. But all our direct relationships and basically different since we manufacture, design and manufacture our boards and everything. So -- and we have 8,500 components on one of the motherhood itself. So basically, it's really a huge list, which is there.

Just one follow-up on the first one. At an industry level, are there reduction in lead times or the wait times are still pretty long?

Actually, for us, basically, I really personally feel is that wait time for what we were getting earlier, it's almost all similar for us. It is a little bit -- you can say it can be on a reducing cycle only. But as I mentioned to you, it has not impacted me when it was impacting the world also at that point of time for the crisis. That time also we had good support that was reflecting for revenues also. And I don't think delivery basically has been a challenge or will be a challenge for us as regards chips are concerned. Because please understand, we are not into the volume manufacturing. We are not manufacturing basically. We are not in the contract manufacturer or a volume manufacturing. We are basically high-end compute manufacturers. So basically, our demand is well planned. We have complete -- basically products which are listed, which we have designed, which they are to be manufactured. So, we have a very, very good purchase mechanism wherein the forecasting and everything is done accordingly. So, we don't really have to bother much about the lead times.

And just one more thing, if you have some assessment of this. So, when you're selling your systems or at an industry level in India, how much of the compute you think is really going into training activities versus inferencing activities?

I think currently, the major demand which is coming is from the training side only. And I think slowly, the inference side of the market is also catching up. And that's the reason we were clearly saying that the adoption of such LLMs will ultimately lead to inference as a market also become bigger with the -- in the coming time.

The next question is from Hardik Rawat from IIFL Capital Services.

Sanjay, firstly, the understanding that we have currently is that out of the total top line that we have today, roughly 15% coming from the AI enterprise workstations vertical. Our current exposure to LLM as a business should be less than 5%. Is that understanding correct?

Yes. Actually, really speaking, what you are saying maybe 4 to -- if you only LLM is a basically business would be around 4% to 5%, not more than that in any case because the GPUs are not only used for LLM, GPUs are being used for various kind of things, various applications, even HPC. Even AIRAWAT, if you see today, basically, the AIRAWAT is one of the fastest AI supercomputers in the country, they have a queue for more than 6 months actually for application users who are really waiting to get time on that. So -- and they are not running only LLMs actually. So, LLMs as you rightly said, 4% to 5% is a good number.

Got it. And since you are envisaging a better growth rate for the AI enterprise workstations vertical versus other verticals, which is leading to the share increase from 15% to 20%. What is the extent of LLM-led growth that we were factoring in when we were estimating this increase in share?

I think basically, it's very marginal. Maybe 7%, maybe 5% may become 7%, 7.5%. It's not really basically LLM related and plus again, it depends as the government is putting focus, I am seeing more growth now because basically, what -- after hearing yesterday, whatever Ashwini Vaishnaw has mentioned that basically India LLMs, they will get a lot of traction and all. So, I think basically, what we were envisaging was around, I think maybe 5% will become 7%, 7.5% to 8% at the most.

Got it, sir. One last question was with regards to the outlook, like you mentioned with the DeepSeek thing coming in, the CapEx cost for developing an LLM has reduced substantially considering if people are able to build more efficient models. Are you seeing anything on the ground in India as well, people starting to looking to deliver like setting aside the announcement that was made yesterday by the IT Minister. But apart from that, private enterprises getting into action, probably developing more efficient models using the limited resources at hand and some demand from there. Are you seeing that on ground?

See, discussion. See, currently, if you see even the policy also clearly creates or shows this as one of the pillar of the mission. So definitely, LLM development, not just only for PSUs, even for -- I mean, government enterprises, even for the private enterprises also, this is going to happen. But yes, it largely depends that who really want to basically take an advantage of it. It also depends on the workloads of each and every vertical of the market. So, it is not something that every organization is going to start straight away start working on LLMs. This is something will be largely be need dependent. So, I think that's how the adoption is going to happen. And as we already mentioned, that usage of AI is not only going to remain dependent on the LLM part. There are many other workloads also. For example, there are scientific codes also available. There are high-precision workloads also available wherein FP64 kind of workloads are there. So, we cannot just discount that set of the market also, which is there with the existing enterprise or the government enterprises also. So that's how it is going to progress. As we mentioned that adoption of LLMs will only be adding what we foresee is that making the market for us, for example, from 5% to 7%, that kind of adoption is what we are seeing.

Next question is from Akshay from C D Integrated Services.

My question is more on the fundamental side, like, for example, some x company is using some legacy products or old system. And now they have started using our system like our [ Tier 1 ] range of products, then what are the better yield companies can generate? And can you give some practical example like after using our system, whether it is because of the faster speed or large, huge workload or something like that? Can you give some practical example?

So practical example, just to share with you that overall, if you see the way we serve the solutions, it is basically a very good amalgamation of the hardware, the middleware and the utilities. And the advantage which we deliver is basically in terms of, on one hand, the performance acceleration. Second is the ease of usage of the entire solution. So, the customers who buy our solution, they basically look at us from these 2 perspectives. So practically, if you see -- you talk about any application, be it related to private cloud or related to supercomputing, the best way we serve the customers and that is for which they look up to us is the -- on one hand, the complete solution, end-to-end design with a seamless architecture coming from us. Second, the acceleration in the performance which we deliver by means of our architecture. So that goes with all the applications, whether you are looking at a complete private cloud kind of architecture or the scientific application to be used in a supercomputing environment.

Okay. So, how different are we from the other OEMs like Dell and HP. So they might also be using the NVIDIA chips and the high-end chips. So, how different -- are we more customizable or something like that what front?

So basically, our -- we are very focused. We don't target basically the normal run rate servers or those kind of things. We are not box pushers, please understand. These are box pushers actually. Their margins will be in single digit. Our margins are much better than them. So basically, because of value addition we bring on the table as regards the solution, the design is more focused and more optimized, okay? That is one. The second is that what Hirdey mentioned is we have our own software and application stacks, which we bundle along with our system, okay? In their case, they have to basically buy or get all those applications from third party, whereas basically these applications, we give it to our customers. So, I think these are the 2 major differences, which makes our solutions more efficient and more acceptable in the domains we target.

Okay, sir. And my second question is...

Akshay, I'm really sorry to interrupt, but maybe request you to rejoin the queue as there are several participants waiting their turn also. The next question is from Himani Shah from Alchemy Capital.

Just wanted to check with you on this, Ashwini Vaishnaw also mentioned it, and then there is some empanelment that the government has started to do for the 18,000 GPU units. Are we one of the empaneled ones? And if not, how would we be involved in this?

So this is basically -- it's an interim arrangement, which is basically for the purpose of empaneling the cloud service providers who are going to basically provide GPU compute cycles through the cloud. So first of all, that's the reason we are not directly participating, first of all, because this is quite a stop gap arrangement as we mentioned earlier also while answering the other question in the session that the infrastructure which Government of India is trying to build, that is not which is getting built in this case. This is only for the purpose of renting some compute cycles. So that's the reason we are not directly participating in this case.

We are not CSPs.

We are not CSPs actually. So, it is not basically meant for our solutions actually. But just to share with you, those who have participated, I mean, we are backing some of the -- those CSPs by providing them the solution at the back end.

Okay. So in terms of an order book outlook, should we say that the outlook for our order book is -- or order inflow has increased from here on?

Order flow, I did not get it actually because basically, because the order book outlook, we have indicated very clearly that our order book cycle is also basically 12 to 8 weeks, and we have already indicated our order book. So, growth momentum ma'am, I'll tell you is that basically the funnel, if you want -- if you're asking me, the funnel, we have not basically once the government RFP comes, then the India AI mission funnel will be added into our funnel. Till the time, we are not projecting that into the proposed funnel, which we are showing to you.

Next question is from Keshav from Niveshaay.

So, as you rightly mentioned that inference will grow in the future. So, if you have heard about Grok and Cerebras. So, they have developed their hardware around it. So, do you see that as a threat to NVIDIA and ultimately to your business? So, just wanted to know your views on that.

Cerebras and...

So basically, the compelling technologies will always keep on coming. And nobody stops me from getting with the compelling technologies. But basically, there are definitely, the compelling technologies for mass adoption, it will take time. If they become popular, we are vendor agnostic. We can collaborate with them. And I think since we have the end user base, any new technology will always like to collaborate with us.

So how much like configuration that you have to do in your systems to run those solutions?

See, we do the complete architecture designing at our end. So that includes the sourcing the chips from the back end. So, in case if there are other comparable technologies, which may get available, for example, alongside x86 architecture, we had also introduced ARM architecture as well. So that was a right move for -- in our case because we look that as a good option available for certain section of the market. Same way, so in case in future also, we find such similar technology, which may be a comparable one to NVIDIA and which can be adopted by masses in the market, we don't hesitate designing our architecture basis on that. So that would be a swift move in our case because we have got our in-house designing. So that helps us to move in that direction easily.

So, how much time do you think like it will take to design that architecture?

First you should ask how much time will they take to become comparable to NVIDIA? They are there for 1 year. How much time we are waiting for them to come.

They are there for 3 years actually.

They're there for 3 years, but still, you should ask them that how much time will they take to come to NVIDIA's level. Once they come, we are ready.

Next question is from Sandeep Shah from Equirus Securities.

Just to understand correctly, a reply to a previous question, as of now, LLM-related revenues are 4% to 5% of the top line and may increase to 7% to 8%. So, is it fair to say within the pipeline also, which we track, the LLM-related pipeline is not that big? Why I'm asking is clients may take slightly longer time to understand the new architecture and to award the deal or convert the pipeline into revenue or deal. So, can you share some thoughts regarding this?

Sandeep, thank you for your question. Actually, we don't basically split our pipeline into various segments here, okay? But basically, I'm telling you in the total business what we do, LLM is, as I mentioned to you, is around LLM based would be -- the total AI would be around 15% and basically AI, which is being used for LLM development would be around 5%. And that we feel that it can go up to 7% to 8%. So basically, it will be very difficult to say because in the pipeline, the conversion rate is 60%, which gets converted, what gets converted, it becomes a bit difficult. But you can take it in a similar range, actually, really speaking.

Okay. Okay. So in a summary, we don't foresee any major slowdown in our client decision-making because of this new architecture?

We are looking at increasing the -- basically the client decision-making. You see what did yesterday, Ashwini Vaishnaw has mentioned. If you go through that press interview, very clearly, the India is trying to -- he was talking 10,000 GPU, he's talking 18,000 GPUs. The government is very clear on spending more and more money. And basically, this is the India moment, which India wants to grab upon. So basically, the -- so that is only the government story. The enterprises will also like to do that. So, there are a lot of model development will keep on happening. So, I feel rather the recent development will be -- will really help so as to grow the market.

Okay. And last thing, the newspaper articles, media articles indicate that out of the INR 10,000 crore budget of Government of India on AI mission, roughly INR 4,500 crore budget is towards the CPU or the compute architecture. So in that scenario, are we tracking -- we will have a role playing in that INR 4,500 crores because that is may be controlled by the cloud service provider or our addressable market would be balance, which is INR 5,500 crores and for which you believe the RFPs may be floated by the mid of the calendar year?

So, as Hirdey mentioned earlier, Sandeep, this is basically, even the minister also clarified that basically since this is an interim arrangement so that the India cloud -- India AI story should not get stopped and it should get the momentum, the government wants to fund it. So basically, that's the reason these are -- this is the GPU renting, which is happening from CSPs actually. But government intends to spend the money in setting up its own infrastructure and we are definitely part of that discussion. We are definitely part of that.

Next question is from [ Vidyadhar Ginde ] from [ Sohum Asset Managers ].

Just wanted to get your thoughts on the fact that there is a view that the cost of training of DeepSeek's reasoning model at [ $6 million ] is actually the comparable number for OpenAI's reasoning model is [ $15 million ]. And the view, therefore, is that this model came roughly 1 year thereafter. So, the fall in the cost is roughly in line with expectation that the cost of compute halves every -- falls by 50% every year. Do you think that is true? In any case, do you think cost of compute will keep falling? And is that positive for your revenue growth?

So basically, to tell you very first thing is that, that number, we've also seen that number. But you know the country which has declared the numbers, how true the numbers are, we have to really understand that. I think slowly as the time passes on, we'll understand the number. The definitely is that number or -- but definitely, it's a lower number than since the models are efficient, the number will be lower. So that is there, one is that. So, one is that the GPUs or basically AI is not only for developing LLMs. So that is one. The second is that basically if the cost of compute reduces, what we have been telling from the very beginning is that this will democratize the entire AI development. So, more and more models will migrate since because you are only seeing the cost of the GPUs or the cost of the infrastructure. You're not seeing the cost -- the manpower cost. That's a major cost. Like basically, previously, so many man hours were needed to develop the LLMs, now less number of man hours needed actually on that. So that will reduce the manpower cost also. So more and more, basically, it will become more efficient. So more and more users and more -- since the models have become efficient, more and more use cases will get into the AI actually. So that will significantly boost the demand actually for the AI that is there. So, I personally feel this is all basically this will help to increase the demand. And I don't feel the reduction in cost will anyway impact basically the demand for systems or something of that nature if it happens.

Mr. Ginde, we request you to rejoin the queue. We'll take the next question from [ Rudresh Kalyani ] from [ Kalyani Private Business ].

Yes. The thing is, since DeepSeek is based out of China, are there any regulations from Indian government on using it?

Basically, the government has got regulations, but basically, please understand DeepSeek is an open architecture. You don't need to use only DeepSeek. DeepSeek like models can be used, okay? So basically, whether you use DeepSeek or DeepSeek type of models is almost similar. So, they have really showed how basically models can become efficient. So, you can -- these are all open source. Even the OpenAI is also all open source actually. But a lot of programs in OpenAI, basically Llama is open source actually. So, I don't think that has -- putting any restriction on DeepSeek will create any kind of impact.

But just to share with you, you can also see that none of the countries will basically try to build their sovereign AI infra by sourcing such kind of imported LLMs and all. Please understand. That is the reason if you see the way policy has also been defined, the one pillar is completely dedicated for development of such models. So, you will have to understand that AI sovereign infra is not just limited to the hardware part of it. It basically reaches to those LLMs also. So, it will become the -- it is actually the need of the hour. So, sourcing LLMs would not be the solution for large enterprises or the public enterprises.

Next question is from Ankur Jain who is an individual investor.

My question is for Sanjay. Regarding this AI mission, what I have understood is that 10 players have procured 18,000 GPUs. And they will provide an infrastructure for others to use because the small companies cannot create their own infrastructure. So, they will provide it on rent. And in last earnings call, you said 2 things. One is that you said that AIRAWAT is a proof of concept. So, one thing I want to know that what exactly you mean by that proof of concept. Second thing that you said is that you are providing government some advice on the design. So, probably it was related to this thing only, this AI mission only. So, the design and the advice you are providing them, so is it free of cost or you are some part of advisory or what? This is my first question. After that, I'll ask.

So basically, to answer your question, one is that AIRAWAT is basically what AIRAWAT is a proof of concept for the AI mission only. That what has happened is that the government basically wanted to -- because people -- the AI infrastructure is costly. And in India, the government from the initially only wanted to make basically AI infrastructure affordable. So basically, India -- the government created a large cluster called AIRAWAT, wherein basically such start-ups, Indian scientists, Indian and all those people, they got time on that so that they can run and train their models and see how to use it and how to understand it, which basically they are seeing a lot of demand on AIRAWAT. It was successful. So they understand that basically, this is the way forward, how the employment can be generated, how the Indian AI -- India can become the AI factory of the world, so that is. So, these are the 2 perspectives on that. As regards India AI mission is concerned, India AI mission, as we have categorically told you that at this point of time, basically, so as not to delay since they are offering basically whatever infrastructure was available with the CSPs, the cloud service providers in India, they are trying to provide that to the -- basically to the researchers, to the -- basically to the start-up and to different organizations so that they can start their work actually. In the meantime, government is setting up its own infrastructure so as to basically build a complete large AI cloud. So, as regards design and all, we are not charging any -- we never charge anything from the design. We have been involved -- we've been working with the government for 20-plus years on various kind of areas. We don't charge anything for design or anything. We co-work with them. We help them so as to basically make the design more and more indigenous and more state-of-art. So, I think that answers your question?

Yes. One more thing that now I understood whenever you said all the con calls that we provide the full stack of services -- so regarding this DeepSeek issue -- DeepSeek issues, I got that DeepSeek has actually worked on lower requirement of GPUs, but with a superior algorithm. So, I think our work is also a similar kind of thing that we use some GPUs of NVIDIA and we integrate our software with it. So, we can also provide some superior algorithms and help reduce the costing of the things. Am I right?

No, just to correct here that we have never said that we are into the space of providing end user applications. We do not provide so. We only provide the hardware, the middleware and the utilities like cloud stack or similar, which basically enable customers to run their [Audio Gap]. So first of all, we are not in that race to compete with DeepSeek I mean by introducing our own LLM. So that is not our forte or we don't wish to basically enter into that space.

We take the last question from Vidyadhar Ginde from Sohum Asset Managers.

So if I'm not wrong, your AI and enterprise workstations revenue, which was 7% of your -- it was 7% of your revenue in '23, and it's up to 15%. So, what were the main drivers of the strong growth? And how do you see going forward, what are the factors which could drive this growth as strongly, if not more? And what could be the risks to growth?

So basically, really speaking, as the growth as regards from 7% to 15%, we have been -- it has been gradual. And the adoption of AI, the more and more applications moving to AI, you see today is India's enterprise, whomever we are dealing in various kind of things, they all basically -- because you see the annual report of maybe 200 or maybe the top 500 companies, they have the AI world in it. So, everybody wants to experiment with AI. And this is not a new tendency. This has been happening for the last maybe a year or so. So basically, people are investing on AI. People want to test AI, try AI and try to incorporate more and more AI. So, I think that is an answer to you that basically how AI adoption is helping us. And plus even the large enterprises, all of them, they have all been trying to do it. The government is trying to do it. So basically, that is the reason that momentum is there. And we personally feel with the current situation also as the demand generation will increase, as we have mentioned that this 15% can become 20% within a year or 2 actually.

Okay. And what are the risks, if any, or you don't see any real risks to this?

Really speaking, I don't see any risk.

We'll take that as the last question. I would now like to hand the conference back to the management team for closing comments.

Thank you so much. Thanks for your questions. We are very confident of basically continuing the growth trajectory, which we have mentioned, which we have indicated that we keep on growing in the same way. And basically, that's it. Thank you so much. Thanks a lot and have a great weekend.

Thank you.

On behalf of IIFL Capital Services, that concludes the conference. Thank you for joining us, ladies and gentlemen. You may now disconnect your lines.