NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hello, everyone. Thanks for taking the time today and welcome to today's webinar for Next-Gen Analytics in Capital Markets. We have a great set of topics to present in today's session. And with us today, we have Ashok Reddy, CEO of KX. Everybody knows KX in this market where KX is positioned uniquely as a AI database, who can handle both time series as well as other unstructured data, key to GenAI and other areas. So we are going to hear from Ashok. And we also have Satish, Director of Quantitative Engineering at ICE Data Services, who works on prototyping the latest and the greatest. And we will hear from Citi, in these slides on rates and market making in the interest book. And of course, you have me, facilitating all these models and solutions for our clients on key front office, middle office and back office use cases. So let us take a brief overview of what we are covering. And Ashok, please feel free to take us through the key use cases and items that you are seeing in the industry.

Thank you, Prabhu. Great to be here. I think for today's presentation, we'll focus on how companies are trying to become AI first and take advantage of AI, first to create differentiation. There's a lot of hype around generative AI and use cases. But end of the day, it all comes down to how can everyone create an edge or differentiation for their company. The second aspect of it is, you'd still have to manage risk and compliance. If AI is just a black box that you cannot explain, for those of you dealing with audit and compliance, it's a big problem. And also, you can use AI to manage risk. The third aspect is, how do you do this with the best efficiency and price to performance, which makes sense. In order to get to these 3 objectives, one of the things what we find is, adopting an AI factory approach that will allow you to use data, both external as well as your own internal data, as every company is a data business. How do you make that AI ready and how do you discover the use cases, develop, build, research and be able to then incorporate them in all your applications and use cases, both external applications such as whether you're doing trade research or execution or risk management and almost every business process within your companies. But all this will be powered by data, for example, coming from ICE in this case and we'll talk about how people use an AI factory approach and machine learning to develop use cases for both differentiation as well as risk management and we'll wrap up with who we can help all of you to take advantage of the AI factory here from NVIDIA and KX to help discover your own differentiated use cases. With that, let me just give a quick introduction on KX for those of you who are not familiar with us. KX powers every corner of capital markets. We started in 1993, same year as NVIDIA focused on high-performance data processing and analytics. We not only do time series but factor, we are into streaming and we are #1 in terms of a multimodal vector database that's called DB-Engines as well as we are used as a benchmark for testing stack, speed tests. KX, at a high level, it's not just a database, it so happens, we actually bring in a high-performance, analytical engine where both data processing and analyzing happens at the same place where you don't have to move the data around. We bring algorithms and models to data at scale. So it has a combination of speed and performance while you also are able to do things such as complex analytics and now AI to bring the best models with the highest accuracy, with the highest efficiency, at the same time, you are able to meet the compliance and audit needs. We're also available, whether it's on-prem or any hyperscale cloud, or on the edge. And we, our focus is really just like Nvidia, making things which are impossible, possible, which means which typically are used for mission-critical, business-critical applications. And in that context, you'll see that pretty much the world's biggest and most leading companies, whether it's investment banks and capital markets or hedge funds or market data providers, everyone pretty much uses KX as a industry standard. With that, I think our focus here is really around the use cases. And use cases are important in the sense, it helps you to look at what outcomes you're trying to achieve for your company. It's not just about using AI to just improve your productivity, everybody can summarize data and get some things from generative AI. But really, it comes down to the 3 things. One is, how can you help drive more revenue, better alpha or edge for your companies? How do I implement the best, efficient algorithms using the least amount of compute and storage and energy for getting the prized performance while also reducing risk? In order to really deliver this, all 3 front, back and middle office have to come together and each one has to work together end-to-end to create differentiation and -- at the higher speed and efficiency and lowering your risk. Some of the challenges in order to really get to these use cases. These are some of the things that we have find working with many of you, is these 4 areas. One, you want to do these things with speed and accuracy, with time awareness and not having hallucinations, if you are using generative AI. With the data, it's really intersection of the high-speed velocity of data to volume and variety, both structured, unstructured at scale. You're talking about high-frequency data, timestamp to nanoseconds, to petabytes of data you may have to process and find analytics and that feeds into machine learning and that feeds into how you generate data in the first place and then generative AI can be used to become much more accurate. At the end of the day, also, you have to manage end-to- end and you need to have governance and there need to be explainability and you need to have models which are -- you can understand. And the last piece in all of this is, also you need to have the best price to performance because without TCO and ROI, AI is not something which you want to invest in because it just becomes yet another tool where we are investing in technology for the technology sake. And all these things will lead to, again, the type of things what customers are trying to do, which is differentiation and how you can unlock through AI. AI is not just generative AI because recently, everybody is exposed to ChatGPT and others and they think AI is just what generative AI is. But really, it's really about uncovering all the things all of you have been doing, machine learning for years and now you're applying natural language processing and unstructured data but really it comes down to these 3 main areas of use cases, trade research with ideation and trade execution and risk management. I think at KX, working with our partner here, NVIDIA, we bring 3 things. One is unlocking your alpha, your edge, how do you look at all aspects of your data, not just limiting to one aspects of it. Is it relevant? Is it representative of all the things and then it's structured, unstructured together at scale, so that you can truly beat your competition or truly beat the market. If only 5 stocks are giving you alpha this year and what about the other 95%. So how do you find that edge and everything what you can do with unstructured data, combining that is structured, optimizing trades where you can find patterns, where humans are good in finding patterns but when you have a high intersection of these, the 3, volumes and velocity and variety, how do you find patterns? At the speed of thought in real time. And then, of course, you want to consider all data, you have multiple technologies for 1 thing for unstructured data, structured data, real-time, historical data. You have a whole bunch of stack but how do you get a full stack and where you can actually do search not just by keywords but also semantic search, a fuzzy search and time-based search. And you also want to bring all of it together in a way you can build an AI factory, which enables you to take your data, build models and be able to build these use cases, which differentiate you.

Ashok, I know you have been the leader in time series database, right. That's an area which has been your key strength. In the industry, did all these use cases with back testing in the quantitative world. And now with the qualitative information, are you able to get to the same areas with KX, adding in the new world of qualitative world combined with the quantitative world?

Absolutely. I think one of the things what we found is, when we look at machine learning and traditional AI, people only look to structured but then now with generative AI, people are looking at just qualitative information. Yes, it's helpful in terms of summarizing documents and finding things and similarity things. But bringing that together is actually very powerful. In fact, that's one of the things. It sets up here, as one of the example, I want to talk about is, in capital markets, when people do trade research and execution, typically you do first look for signals in structured data. And this is an example where we actually have structured data of a company, aerospace manufacturing company. It's a stock and their price, how it performs over a period of time. But looking at the unstructured data in this case, especially around negative quality problems, control problems they've had. And you can see that the 2 things, when you look at them together, you're able to find signals, way before it happens. And it's not just a correlation, it's a causation, cause and effect. And that's part of the problem and many times in AI is that when you just look at structured, unstructured, you can correlate things but it doesn't always have a cause and effect. And I will actually show -- we'll show a demo based on this data and how we can obviously partner with NVIDIA using the AI enterprise with NeMo and NIM to apply this quad search, where we bring these 4 different types of searches and rerank them to bring that signal with the combination of these 2. And here is the demo where we actually bring -- combine the hybrid search, where we take both dense and sparse vectors, both looking at not just keyword-based search. In this case they -- if you look for the company ticker symbol and they'll come in and look for the -- all the documents related in this case. And you'll find everything related to quality and they're able to load all the transcripts and everything ever said about this and be able to do all of it using Python and be able to find, in this case, supply chain constraints. And that would bring everything related to that in all these documents and then we are able to look at the causation. And I'll get to some of the numbers in terms of what the number of documents and how much -- how quickly this is doing. But it's something which we find that by combining this, you'll see the patterns and it's combining the unstructured, structured data to be able to overlay together, to be able to then you can actually start looking at time series-based pattern recognition, which I'll get to in the next part of it. But this is something which is something very timely because you're not just looking at transcripts separately and then looking at the market data separately but you're actually linking the price and how the market reacts. It enhances the user experience and it helps you make better decisions and it provides you the differentiation you need because you are getting that information nobody else has. Does that answer your question, Prabhu, in terms of how the unstructured data can help in the context of your -- not just time series but there is a time element to unstructured data also.

Yes, Ashok. I mean, it does, I mean, I clearly saw that and you are able to combine the unstructured area to the structured data, right? You're able to connect that information back to the tabular data, so you're showing a world where generative AI doesn't sit alone but it actually sits along with brownfield projects, your current market data and you're able to combine these.

Absolutely. I think it's -- that's -- one of the key aspect is, one of the things what happens with unstructured data is, again, many of these things, documents have time components. If you have the same documents available in different time stamps, how do you actually make sure that when you're training your models and especially on the generative LLMs, they don't always have the awareness of time and so the unstructured data, being able to combine that in the same time period and being able to find patterns and causation is one of the key aspects of this.

I wanted to understand that more, Ashok, [indiscernible] we know that LLMs are not time aware, right. Is this something, the concept of applies to the newer age LLMs as well?

Yes. I think if you actually look at any of the things about, anything that comes with time, typically, when you convert text, it's based on whatever the time periods, those things are but there's no statefullness. The state is missing. So when you learn, the models are learning, it's machine learning, by definition, is learning based on experience over time but that experience changes with time and how do you actually get those incorporated into neural networks. So that's a big challenge for the industry. There are new models which are coming up. And in fact, some other things -- Mamba and others are looking at statefullness. But that's where we can solve the problem by because we look at time as a key component, whether it's structured, unstructured. And we -- this is one of the things what -- this particular thing, also what I have -- we have in this chart here is using time to kind of look at patterns and trends over a period of time. So what this does is to -- if you have a pattern, just like we just went through the airline manufacturing company, you can actually say what that particular pattern that it happened -- what has happened during the time and you can use that pattern to take your structured data, be able to find it semantically, which very few people thought about it. This is one of the use cases. Many of our capital markets customers are using our product KDB.AI, where, not just trying to apply this for unstructured data but you're actually taking some structured data, making them semantically searchable and then you are looking for patterns, which then can be used to train your models to even be more accurate. And so that's a very important thing. And obviously, we have done this with using the NVIDIA software to give that accelerated computing as well as accelerated AI with now with some of the newer container technology and microservices such as NIM which enables us to incorporate multiple models. So this here shows, as we went through both structured, unstructured, the example I was sharing, here we actually went through, that demo was using more than 2 billion trades and quotes, 60 million news articles and the type of -- a number of analyst reports, all processed in seconds and if you look at the type of analytics, whether it's a trade, we were able to do all of this on disk as well as we were able to ingest in just in 6 seconds and 6.5 gigabytes of -- worth of disk. So many of the vector databases and others are not able to handle anything which is not in memory, you'll have to buy a lot of memory. If you have large amounts of data and documents, you're not able to handle a disk. We are able to do it both. And then other point is that in terms of everything we talked about going back to price to performance, it's something where everything you do here is aligning with that. And then it comes down to the ROI and the TCO. So that's some other metrics. I won't go into the details but you'll see some of those elements. The other aspect of it here, which we are very excited about working with NVIDIA who not only was known for GPUs but now they have Grace Hopper, which has Superchip, which has both a CPU and a GPU with shared memory as well as a AI stack on top of it, which I think is one of the game changer where we were able to, for example, use, in this case, 2 million documents in order to create vector embeddings. If you do it with any other technology, we found it took at least 40 hours. And here, we are able to cut it down without any optimization to 80 minutes, that is 30x faster. Think about, if you -- before you get to actually doing all the search and everything else, you need to ingest data, create vector embeddings, it costs time and money and what we have done here with very limited amount of memory and the type of things on the right side. And it also gives you a choice of models you can use. It's not just any one model, so you can mix and match. But that's the power of what NVIDIA Grace Hopper is. It gives you both CPU, GPU and removes the bottleneck today where GPUs are idle, many of the -- most of the time because the CPUs cannot keep up on the memory transfer of the data is a bottleneck. So that's really the AI factory here, which is really bringing together all aspects in terms of the different type of data sets, being able to discover, experiment and research models and improve accuracy and others. At the end of the day, you're building world-class differentiated AI analytics applications, search applications, bots and recommendation apps and do that with the type of numbers what we have seen with KX and NVIDIA, you can do search results so that when they're compared to any other competition, we're outperforming by 3 to 5x. Our efficiency of CPU, if you take any of the top 5 competitors, you'll see that we are far superior. And of course, we are the most efficient in terms of energy consumption and power because AI, they say it's going to need many more power plants but in our case, NVIDIA and KX are focused on making sure that what we deliver, it also is very efficient and sustainable. And I think, again, we talked about how we are creating vectors embeddings, 30x faster for unstructured data. But the good news about structured data is you don't need to create vector embeddings because for time series data and others, we have the world's fastest way of doing that and you can do semantic search without having to create vector embeddings and you can save all this energy, power and models. So now I think we're going to move on to the next-gen analytics in capital markets focused on risk management. Peter Decrem is joining us from Citi, a customer of KX for many years. And Peter, welcome. I want to just start with -- can you provide a brief overview of the role of market making in rates trading? As well as how critical is accurate price and risk management forecasting including the challenges, the methods and how you're evolving that?

Sure. So my name is Peter Decrem. I work in rates trading at Citi and I have oscillated between technology, machine learning and trading for quite a while. In terms of what we do as an institution in our market-making capabilities, we actually transform risk. And so we're taking liquidity, we identify liquidity in the different markets. We transform it and make it available to our customers. And in that process, we obviously identify any of the risks associated with those transformations and we're using a variety of different methods to do so. We -- and you see that in the box -- I believe in the middle, in the methods, we have the traditional methods but we are exploring together with folks some of the newer methods that allow us to scale in even higher than mentioned, than what we've done. If you look back -- on the history, I think when I started, we had about 7 sectors in the treasury market, that has obviously been reduced. We have integrated those together with the linear products and even in the process as have many institutions of integrating those over multiple asset classes. So these techniques, which you folks bring to the table and the streaming data together allow us to make more liquid markets and identification and the management of the risk associated with it. I thought it made sense to talk a little bit about the data, which we have a simple view on the next slide. And that is the kind of the markets as you see. And there's different levels of information. If you look on what is called Level 1, it's just the best bid which is, in this case, 99-25 plus, and there's 816 units that people want to buy at that price. The best offer is [ 99-256 ] and there's 790 units available. And so that is what is called Level 1. There's buyers, there are sellers and if somebody wants to buy at [ 256 ], they can do so. And the answer is they can buy 790 units at that. So that's the simplest way. The level 2 is what you actually see on this slide and that was the multiple levels. If you ask -- if you go back for 1 second on the previous slide, then you see actually if you want to buy not 790 but 1,500, well, then the price goes up, you can buy some at [ 256 ] and some at 26. So Level 2 is the multiple levels and price points and sizes that go associated with it. Level 3, exactly, as you see over here is actually the identification of the individual orders, obviously, on public data, you do not have identification of the firm or the trader who does so but you will see the individual orders. The size of the orders is kind of the point of this slide it'll be even for treasuries and linear products, you're talking about the hundreds of millions of those orders per day, which stream to those systems, which actually take advantage of the infrastructure you folks bring to the table and what is really nice about that they're going to a shared memory between Python and the KX streaming data. And some of it then makes its way on to the GPU for different models. And that has been going on for the last, I would say, 5 years, of which for example, on the GPU, you will find auto encoders, you will find some gradient descent methods but basically looking at the patterns of what is the true liquidity in the market. And as I mentioned to you, we are -- and the rest of the financial industry is actually scaling larger to provide more liquidity at more competitive levels and identify the risk with it. And so if you look at the next slide, you actually will find that one of the things that is driving it is obviously intuitive that if there's more buyers than sellers, then there's a higher probability that the market will move up in the reverse way. So the order book and balance just gives you an indication of the amount of buying, a pressure that is in a book versus the amount of selling pressure. And here's 2 securities. And I'm just showing you 1 example, which has been around for decades and you will actually even find in the news media and it's the correlation between futures and the underlying cash. This is actually an older slide but you see there's a slightly negative correlation between this order book and balance, between a cash security and a futures security and it is of the order of minus 0.08. And that means if there's more buying in the cash, there is less buying pressure in the futures and the reverse. But it gets more interesting. If you look at the next slide, what you actually find is there's actually a confounder. And the confounder is the price level. And -- well, actually, if you split it up and condition upon the price level, which is the basis, which is the difference between those 2 securities, then you'll find that these correlations go massively, obviously, in opposite side but they will become significantly more. And this is just 1 point of a pattern, which is sitting in these orders and these order books where you can actually apply machine learning methods. And so this gives you some idea. And if you then look at the next slide, this is not just -- it's a simple thing. It's actually in the example that I gave you, as you can see, the amount of buying versus selling is significantly, if you look at the last column, then you see that the percentage varies between 15% and 86% depending on where the price relationship is. So a large portion of this order book is actually determined where that cash point is and vice versa. This is 1 example of a simple pattern, been around for decades, very well known. But it shows you, yes, this is -- these things are not random. There are patterns in there. And if you want to understand liquidity and manage the liquidity that you provide, you have to understand where are these relationships and how do you manage. And that's where the machine learning comes in. The data analysis comes in. And it's in a dynamic environment. I think one of my focuses has always been dynamic systems. And one of the simplest example of that is iceberg orders. If I want to sell 1,000 units, I can slice it up. I can slice it up in time space. I can slice it up based on the volume that trades in a variety of different ways. If you have a static system, then you will just take the average over a period being 1 week or 3 months. But obviously, if you have a more dynamic system, you ought to be able to differentiate between the amount of these iceberg order in that order book right now at these price points versus where it has been on average. And so I think the benchmark is significantly higher and that is why streaming data together with some of the newer techniques are of tremendous interest to us. And we hope to be able to scale significantly larger than some of the things that we've done in the past. And in the past, we have dealt with 10 to 100 of different of these order books that we integrate but we hope that these new techniques on streaming data will allow us to actually integrate more liquidity and provide more liquidity in deeper markets at more efficient price points together with the risk mitigation that comes with it. And in here, you actually see some of the underlying dynamics. It's, again, to the same point I showed to you on the price level on an order book. This is a chart of if you look at the sequence of these orders and then we go maybe a little bit into detail, the symbols, for example, in the middle, you see L1A, that means Level 1 and Level 1 is the highest bid, lowest offer, if you recall, on the charts that I have shown you. And then thereafter is the security identification. This is an older slide but we were dealing with probably about per security at 10 levels on the bid side, 10 levels on the ask side. So that is a total there and the number of securities are in the, I believe, 40 or so. And point being and the decimal points is the number of orders that they follow. And so if you would have expected, you would have expected very small numbers across all the securities, that is not the case. These patters are in there. They obviously are conditioned on the price level but this is the kind of high level of the higher amount of orders that on a daily day streams to these systems and that we actually are monitoring. And that gives you kind of a blunt, things are not random. There are patterns in there. Hence, you ought to apply techniques to get better intuition and an idea about what the current liquidity is in those markets.

Peter, on that same front, I mean, are you looking at more advanced techniques that can be applied to the data in terms of evolution and what we have in the markets.

Yes. This is actually 2019. So in 2019, we were doing graph nets. And to be specific, I believe this was 9 securities. So 9 of these order books, where I showed you the first chart, you take 9 of them. There are notes in the net and the edges are the 2 charts that I showed you before, is the sequence of orders or the sequence of trades, or both. And then you can actually, over those 9 securities, you can put a graph net on top of it, put attention on it and we did a very low dimensional attention. And then you can actually identify and understand where are we in that liquidity provisioning [indiscernible]. So graph nets 2019, small amount of attention, multichannel and I think the next chart actually gives you an idea about what we actually did over here. And so here's 6 or 5 different models that we will show and we will look at the accuracy and also the standard deviation. How much does the results vary and how accurate is it? And we started on model 1 with a simple security and quite a bit of folks look at single orders and then look at single order books and they look at 1 single security and try to understand what is happening in there. I think you will find in the next chart, as we get more complex and integrate more of those, you actually have a higher accuracy. So the blue line in here that starts somewhere around 57% and goes up into the 70s is the accuracy over here with respect to a measure. And as we get more complex that is moving from the left to the right, you see that the accuracy is going up, not only that, you also see that standard deviation of the error goes down. So these methods that integrate information are significantly better than the single order book. And quite frankly, that is probably one of the most fascinating topics right now. Can one integrate dynamic systems together in a much larger scales, which have become available from you folks and from the methods out there. And so that is one of the things that, to me, is fascinating. Can you instead of -- actually apply it -- these methods to time series, dynamic system and streaming data? And let's -- to be very clear, streaming data is another couple of complexities and challenges on streaming data, applying gradients, it is not obvious what data is going to come through. You haven't seen it yet. And so how do you manage that and extract information that makes it more relevant to any static system or an average system, yet at the same time, dealing with gradients and making it controllable. And that is where some of these methods meet.

So Peter, thanks for that fantastic overview of things, what you have made progress. But I guess one last question I had was, how do you look at the challenges, what do you have in terms of scaling these systems for dynamic systems, especially if you start looking at opportunities for applying AI, what are those problems, which were impossible now, which you can solve hopefully with what, with KX and NVIDIA are providing you access to the AI factory and the lab.

Well, the thing is, actually, if you back up and say, where did we come from? I think about 15 years ago, is when I started on the [indiscernible]. And if you look what work we were doing at that point in time, the answer was, we were doing [indiscernible] paralyzable workflows, for example, [ torrents ] take 1 million [ Monte Carlo ] pass, value the security, take the expectation. And so what we had at that point in time was actually 2 orders of magnitude faster than a single trade at -- on the CPU. And so that was 1 part. I think on the risk management, we were doing linear products, linear products through matrix operations and a number of us had figured out that you can actually scale this and this is the way forward. But also, if you see where we were standing, I think we were running around 40% of the occupancy. The memory models in GPUs is significantly different than what we have on CPUs. And so how to manage that and get the maximum performance. And that's kind of the world, where we were 15 years ago. About 5 years ago, we started to make available streaming data, thanks to you folks and streaming data and the integration into dynamic systems. Relatively simple one, some of them on CPU and on GPU. And now what we're looking forward to is actually going significantly larger scale and with that comes both infrastructure being the power footprint that goes associated with these things. The risk control systems [Audio Gap] opportunity of actually scaling in significantly higher dimensions. If you think about some of the things that are around, when you're talking about vocabularies that's between 100,000 and 200,000, we're talking about context that are 1 million and above. As I mentioned before, we are looking at hundreds of millions of those orders on a daily basis that go to those systems, also it's lower dimensional space. But if you start talking about securities, exotic options, then you are talking about larger dimensional scales and that is the opportunity which is there. So both for risk management and for analyzing liquidity provisions. We hope to be able to scale like these models have the last 2 years and actually apply them to time series data.

Fantastic. Thank you, Peter.

You're welcome. Looking forward to share some results with you folks, too.

Thank you, Peter.

Excellent. That will be great.

Thank you.

I would like to move on to talk about AI factory and AI-ready data. I'm very pleased that ICE will be joining for this portion of the presentation. I think one of the things as we kicked off this presentation, we talked about how do you make sure that if you want to make money and if you want to create differentiation, it starts with the data and is it AI ready. In this case, I think part of this is, do we have the right relevant data, or we have data which represents all aspects of the things you are going to use to train your models. And is it clean, the quality and whether it can be trusted. So there's so many aspects where the data preparation aspects be, especially when you are processing large data combined with large language model-based data, I think that's very important. And we actually then help select any type of model and analytics to now real-time analytics. And now as Peter talked about, using machine learning models, which will then inform what we do, combining that with generative AI to bring hybrid RAG, where you take your own data, be able to fine-tune the model to make it relevant and more accurate and more precise, at the same time, not have hallucinations. And that will lead to the right outcomes around whether you're doing ideation or execution or managing risk. With that, I think it will be useful to hear Satish from ICE, a leading provider of data around how they help companies in the context of a AI enterprise and AI factory.

Thank you, Ashok. Thank you, Prabhu. Thank you for having me on this webinar. You hit on exactly all the right spots. The need for efficient price discovery, the increased data transparency needs and the demands of the modern trading technology flow, they just keep growing. And this is where ICE lives. Data is our core business. This is what we do. Exchanges, the exchanges that we run are natural producers and disseminators of data, right? Our combined experience and presence in the markets, clearing houses and the data business means that we are uniquely positioned to be your data partners. We operate collaboratively. We leverage all of our intelligence across the different businesses that we have and we focus on innovation. The end-to-end solutions that we produce are fit for purpose and robust and production ready. Towards that, let me just talk about some of the numbers that we have in our business. We have about 600 streams of data across 300-plus trading venues, right. We provide a consolidated tech history for more than 10 years. And this is central, if you're back testing any algorithms or if you're looking at trying out new cutting-edge technology, this is where you start. You get the right data set together and then you started rating on it. Of course, no data set is going to be complete unless you marry it with the right reference data. That's another place where we shine. We have 75 million -- more than 75 million assets that we address, we have all of their underlying fundamental information and corporate actions and business entity details like parent-child relationships and the sector classifications. And on top of all of that, we have our proprietary vault services, we have our credit spreads and pricing. When you marry all of these offerings together, it becomes a very compelling and end-to-end offering from a data perspective and that enables you to go ahead, take confidence on your data source and then build a top of that with all of your analytics.

So what are the advanced type of workloads that you are seeing at your clients?