Mobilewalla on the Data Behind Agentic AI

Hello and welcome to another SPAC Insider Podcast, where we bring an independent I in interviewing the targets of SPAC transactions and their SPAC partners. Identity AI is here and working in a number of industries, but the question of which model can win is increasingly defined by who has the data to back it up. I'm Nick Clayton and this week I speak with Aninja Data, CEO of MobileWalla, and follow Padma Kumar, CEO and chairman of SPACSphere Acquisition Corp. The two announce a $250 million combination in June. Aninja explains how MobileWalla's long tail of data forms the backbone of its own AI offerings and how it is innovated internally to give it an edge. Paula gets into why he sees MobileWalla as standing above the noise in the AI space and how SPACSphere concluded it was ready for the public markets. And so Aninja, MobileWalla was founded in 2012. I imagine so much has changed in that time in between. You have a quick version of the story of how MobileWalla has changed and how it's modified its approach and solving problems for clients with all of the technological change over that time. Yes, super interesting questions and takes me back in time. So to understand how MobileWalla started, you know, I'll take you back to, you know, my background is, I started my career as a practicing computer scientist as a sort of young faculty member at Georgia Tech in computer science and my sort of computer science work research work was in AI. When I started my career in the late 90s, early 2000s, AI was not sexy at all, right? In fact, in fact, you know, those of, you know, listeners who were on the same boat as me would empathize with me that in fact AI was really, you know, people would make fun of AI other other computer science groups and they would say that really nothing is ever going to come of AI, you're going to build toy systems forever. I mean, that situation has now changed. You know, so my, my work was always at the intersection of AI and data and to tell you how mobile wall has started, let me just, just, just give this sort of preamble that, you know, all AI is basically regardless of what flavor of AI, you know, these days we talk of generative AI. Generative AI is really a tiny, tiny, tiny part of AI, right? So all AI, the machine learning, neural networks, any, any flavor, all AI is basically a marriage between a technique and data, right? So you have a technique, you have an algorithm and you train it on some data and this data is really a manifestation of some period of history, right? Rainfall in Georgia or mango production in the Philippines or whatever, right? You take this data that, that manifests history over a period of time and you, you apply this technique or algorithm on it and what it does is that it finds patterns in the data that, that repeat. And then once it has found those, it basically looks for those patterns in ongoing stuff and it predicts like, hey, next year, mango production is going to be good or bad or next week, what, what how much rainfall is going to be there in that environment? So all AI, all prediction at a very abstract level is a marriage between algorithm and data, okay? So my work was focused on sort of the interrelationship between these two. So when you build a AI model, you build it by applying an algorithm on data, what are the relative impacts of data and algorithm on that model, which is more important, right? In the field of AI for 40 years, every practitioner, every researcher, chase the algorithm because it was believed that better algorithms are going to lead to better predictions. The reason that predictions were not good was, you know, it was thought that techniques that were being applied were not powerful, right? And you'll see that one of my thesis was that, perhaps that was not the case. Perhaps data is more important than than what we believed, right? So now that consensus in every major AI lab, whether you go to MIT or Stanford or Google lab or answer big lab is that it's really data is a dominant partner, right? The reason that a simple image recognition algorithm, you can feed it a picture of a cat and it'll tell you it's a cat is not because innately the algorithm is very superior, very simple algorithms are going to do this is because you've trained it on a trillion cat pictures, right? I mean, that's that's so a lot of my thesis was that, hey, data is really important to predictive science, right? So mobile wallow was founded on that basis, Nick, right? So my bet was simple, the bet mobile wallow was making was simple. So it was if better data beats better algorithms, let's build a data asset first, right? That build a solid solid solid data asset and build a kind of data that nobody can catch up to. Now, now, of course, data could mean a lot of things, all my work for a long time, we're looking at consumer data. So that's the kind of data we are looking at how people behave in different places. So we just said that, hey, let's let's build an asset that captures consumer behavior over a long time, right? And right when we were starting Nick, you know, this is like 2013, 2014, mobile apps were getting really, really, really, really powerful and mobile apps represented some of the most instrumented technology of all times, right? And there was a lot of data that was coming out of mobile apps. So basically, we were lucky in a sense that we said, let's go collect people behave and there were this highly instrumented entity that was kind of capturing how people behave. So we did that, right? And the thesis was again, let's build this very, very, very long behavioral longitudinal asset. And the second part is the thesis, right? It's not just build a data asset first, but build an asset that nobody can copy because you can copy an algorithm sort of on a weekend, right? But you cannot copy time. I mean, when you capture data over 10 years, a decade of longitudinal behavioral data, you know, the same mobile devices observed continuously across 40 plus countries is literally uncompressible, right? If a competitor started today with unlimited money in 10 years, there would still be 10 years behind because we would have collected 10 years more data. That's how mobile wall has started. And that's what we sort of started quietly building. Now, now, you know, the industry, of course, arrived at the same point. You know, there's the belief that data is the dominant partner in AI is, I mean, there's consensus on that now. But the exact answer to your question was that mobile wall has remained sort of a little science experiment to you over some period of time, right? Because we were collecting data without quite knowing all the cool stuff would be eventually going to be able to do with it, you know, but we had to generate revenue. I was lucky in the sense of having good investors because I already made them money before. But and they were tolerant of my science experiment, kind of technique, but you know, so initially mobile wall has started by selling insights from the data that we were collecting, right? So and it was pretty easy. So because we were getting consumer data over long periods of time, you know, we were, you know, we were figuring out where people visit and what kind of people visit, what locations and so on. And we were selling insights like this primarily into the advertising ecosystem as we were building the underlying asset. But eventually, what happened was that we got to a point where we had enough history to be able to start making predictions. So that was the first kind of, and I wouldn't really say it's a pivot because it was the underlying thing didn't change. We said that okay, now instead of selling data, we can actually sell features, meaning these are features is also data, but features is predictive data. For instance, we were selling data into the telecom companies, right? And telecom companies predict churn. And we figured out that something that is a very strong predictor of churn is something that we call carrier heterogeneity in households, meaning if you are an AT&T subscriber, and you lived with all AT&T subscribers, you are less likely to leave AT&T than you lived with non AT&T subscribers in your household as well, right? So we saw that we could make these features from the data that we're collecting because we had enough history. We started making these features. We started selling these features and then came sort of the third incarnation, which is where we are now. We said that, hey, instead of selling predictive features to people who are building AI systems, why don't we rise up the value chain and build AI systems ourselves, right? And of course, the recognition at that time was that the type of AI we all know, the gen AI we all know, which is the foundational LLMs, like a GPT-like quad, you know, they operate on publicly available data, right? So you cannot go and ask them questions, need proprietary data, which we had. So we said that, okay, let's now rise up the value stack and instead of selling the data that we are building, let's be actual predictive systems that can operate on that data. And that brings us to today and vertical AI. Yeah, and I want to get into some of the more of the details of that as well, but, you know, I want to get ballin here as well. And, you know, so ballin, you know, with SPAC sphere, it appears that your team, we're looking at a variety of industries with your initial target search. How did mobile walla and this opportunity around agentic AI specifically come up in your process? As you indicated, yes, we were looking at a variety of industries, though very clearly, either directly AI or AI adjacent. So we were looking at agentic AI systems. I had certainly seen a lot of opportunities in healthcare. We had seen a lot of opportunities in energy and infrastructure support. See of data centers, small modular reactors, you know, cooling systems, battery systems that do support these power architectures that are required for today's data centers. So in India and I were introduced in February of this year about a couple of weeks after after the IPL and why this deal kind of came together so quickly was that not only did it check all the boxes that we were looking for it was it was an agentic you know AI system it was a product that was launched it was revenue ready they had a team that was highly credible in both in India and his broader team with Jay and Laurie and and Mehear and Chandrera and and the rest of the team but what kind of made the entire deal come together so quickly was the fact that these the support from his investors were there that is investors were not passive in the process and they were actively involved in the process and continue to support the company not just through the despaque but post despaque so that was a significant difference from you know frankly every other spack process that I've been involved with and second generally what an India has is multi-customers so his customer base are top tier telecom industries in a customer's over here in the United States in places that are in Santa Far East and most significantly they had expressed interest potentially in being part of the process too so we had we had support from all of the interested parties that an India was bringing to the table which was fairly unique and give us give us a degree of comfort that this is a deal that's going to close and not only close but it's also going to trade well you know post despaque given the revenue opportunities so we have we have come to the table and close frankly much quicker than the than the average spack we are expecting comments back from the SEC this week on our first round you know and we are six months post IPO so if things go well we there is an outside chance that we will actually close the deal before the end of the year otherwise I don't go to early next year that is our hope right now great and I want to dive into the the deal of it as well but I did want to bring things back to the mobile wallet for a bit as well just you know as you're talking about there in India that about how really the data is the foundation of these AI companies that are now coming to the fore can you talk a little bit about how you've been able to form your own proprietary data set and how are you able to continue to scale that as you go along great question so if you look at the mobile wallet kind of product architecture right at a high level so the mobile wallet products look like this we have multiple products I'll give you the names of those but at the at the base of all our products is this thing that we call the mobile wallet data platform and on top of the platform we have products in verticals right right now we have three verticals in a telecom consumer lending and sort of consumer data solutions and in each of these verticals have multiple products but really the question you're asking has to do with the base all of these basically rests on on this base called the mobile wallet data platform what the mobile wallet data platform is you can think of the mobile wallet data platform as a machine that ingest data from ingest commodity data from a variety of sources you know like the mobile device ecosystem you know people like SDK providers the RTB ecosystem we ingest things like household data sets that are commonly available basically all commodity data that anybody that has money can buy and then we run it through sort of a proprietary computational process to produce the eventual artifacts that we sell right and right now the easiest way to understand and I'll give you both sides and the middle first talking about what we produce so you can think of the mobile wallet data platform the output of the data platform is a very large data lake and the data lake has thousands of attributes each of those attributes you can think of as a predictor of something right so so these attributes are generally consumed well it's now consumed by us in building our applications but for a long time and even now to some extent they're consumed by enterprises that are looking to do certain things right I mean telecom companies looking to predict charm emerging market lenders that are looking to to lend to to someone that has absolutely no credit footprint so they need to assess their risk and so on and so forth right so so and the and the left side of the platform is the data that we buy right which is then transformed into these features of this predictive data items the data that we buy is I told you is all commodity but what's interesting to mention is that we ingest the data in large quantities right so we get but 50 terabytes a day of data flowing into the mobile wallet ecosystem right and this data is then run through but but you know but but you know data that comes out of devices the SDK stuff the RTV stuff is very noisy it's very noisy it's full of fraud and so on so this middle layer the compute process does really two things one is well three things but but you know the third thing is actually the algorithms to produce the the eventual data items if you leave that aside it is two sort of key steps one is it defaults because there is so much bad stuff in it right you might get a location you might get a record that says the devices in this location but in very overall majority of the cases that location value is fraudulent right people just put it in you know so we we do a lot of sort of denoising of the data and there's a lot of proprietary kind of there's a lot of AI in that as well so that's one thing we do we do a lot of denoising the other thing we do nick that's basically competitive mode for us is I told you we are getting 50 terabytes of data day so you can imagine once you get this over you know now almost almost you know over 11 years the underlying raw data tonnage becomes very large right so if you were to multiply 50 terabytes a day by 365 days a year by 11 12 years you'll see that you exceed it's almost exabyte scale you exceed 500 terabytes right and you know we need to store it because at the end like I told you in my last answer is this history that allows us to do all the cool stuff that we eventually do find patterns but the economics of storage gets very complex right so if I just you know we use very cheap storage on commercial cloud even on that 500 terabytes of data is going to cost you many millions of dollars a month like six seven eight millions of dollars a month that just you know it's not practical for any startup or really any company to be able to pay that right so that's one of the base impediments to building what we build so what we did was a few years ago one of my PhD students basically designed a new class of compression techniques Nick right so we all know compression we use things like zip G zip to send files to send pictures compressed and a typical zip will give you you know two to one compression at best right so we are getting 10 to 1 to 25 to one compression now we are taking advantage of the type of data we are getting in the structure and and and and and semantics and all that but that gives you tremendous leverage right instead of whatever 536 petabytes of data we are compressing it down to 17 petabytes right and that instead of paying whatever six million dollars a month we are spending 150 grand a month on on Amazon that's a very key more even if you give somebody 10 years 20 years to collect data just keeping it around becomes economically very complex right so so that's what we do we we buy tons of commodity consumer data literally that anybody can buy but we buy that in large quantities we process it such that a we we remove noise from it or we I shouldn't say we remove noise for we reduce noise from it and and I think we do that reduction better than than than anybody else and then we can can bring it down to manageable size so the economics work to process it and then we run our proprietary techniques to produce the features that I spoke about and the way that we we keep it fresh is because we are continuously getting new data right and the other thing what I'll end by saying is that but there is also limitation of the type of data that we get right because the type of data that gets is is the type of data that we get we don't there are other types of data that that we do not get that we would like to so one of the one of the reasons for joining hands with bala and and going through this back process and not perhaps a classic fundraise which is what I've always done private front in in my career building companies is that we would like to acquire companies that are very unique data sets right and and we know many exist so one of the things that I'm looking forward to is bringing in other new kinds of data sets that that is not purchasable right because it's unique to these companies both augmenting and enriching the the semantics of the of the data that we have and semantics of the history that we have such that we can impart even more power to the top level vertical engines that we are building yeah and and so talking about you know kind of arriving at the deal what was sort of the main factor there that made you decide that now was the time for mobile wall to go public and and that the SPAC option was going to be the the right route for you for the first time I'm going to give you somewhat of a hand wave the answer okay I did not know Nick that I wanted to go public I'm not going public for the farthest from my mind right if you look at my background you know I've always raised money from pretty tier one venture capitalist and private equity funds I've sold a company to another large company and so on but at mobile wall I you know mobile wall is the first company where I've not raised a lot of money I mean it was consciously done to be able to do things at modest budgets okay and and it took us some time we are at a point Nick where I believe that we are at the inflection point where we can scale very fast and I truly believe that the next sort of you know I'm going to use a very cliche phrase the next trillion dollar opportunity in AI is no not in horizontal AI. I think, you know, the foundational LLMs are there, right? There is no play in building, and there are plenty of open-weight models that are catching up to them. You probably saw in video acquiring hugging face for 13 billion dollars a couple of days ago, right? Foundational LLMs in many ways. I mean, that's not the opportunity anymore. The opportunity is to be able to, you know, much like in the software business, Nick, if you look at how the software business developed the first company that got big for the foundational tech companies, right? Database systems, you know, messaging systems, you know, cybers, Oracle, TIPCO, MQ, these guys, right? But then, you know, companies realize that, hey, I can just buy Oracle licenses and TIPCO licenses and build a payroll system. You just doesn't work that way. So then came the next wave of companies that got much bigger than these guys, which were like SAP, like people saw Seabull, now Salesforce, that build basically these vertical enterprise stacks on top of these foundational tech companies that allowed companies to solve operational problems. I think AI is going to the same arc as well, right? Your answer pick and your open AI are the Oracle's and SyBases and TIPCOs of AI, right? They're setting up the foundational tech without which nothing is going to happen. But companies still need to use these to solve company problems. And I think we are at the Vanguard of, you know, so you can think of as like the SAP of AI or the people soft of AI. We are building enterprise stacks with proprietary data on top of these foundational airlines. And we are at a point, Nick, where I feel we have proven out a couple of key thesis, right? That proprietary data can do things that these horizontal engines cannot do. And you need specialized computer science in the stacks as well. And now is the time to scale. So how does one scale, right? So I knew I wanted to raise capital. So one of the things in AI, you know, in software systems is pretty much a technology bill. Unfortunately, in AI, it's largely a data bill. I mean, you need domain expertise, you need data. And the cash requirements to grow are a lot more. Eventually, what I ended up as, you know, going public offered us faster avenue to scale. I mean, I want to scale three, four X a year over the next couple of years. And you know, it's not possible to do that without inorganic means. I was convinced by one of our investors that this is a way we should look seriously. So we looked at it and, you know, and then quite honestly, SPAC was, I was talking to SPACs. I was looking at RTOs. For me, having come from sort of this tier one VCP ecosystem, the small cap, go public ecosystem was a little weird because, you know, the investors were, were offered different flavor and so on. Then I met Bala. To be quite honest with you, the reason I did SPAC sphere was, I mean, we were already doing something with another SPAC earlier. We walked away from that. To be honest, I thought Bala was one of the first real knowledgeable, really solid kind of SPAC CEOs I met and felt that there could be a fantastic partnership here. And there are things he, I mean, I knew how to build companies and he knew how to navigate the, at least a go public ecosystem. And I felt that that was sort of, again, using a cliche phrase, a marriage made in heaven, right? To grow fast, why go public, I told you. And why SPAC sphere, why SPAC is basically, I met Bala and fell in love with. That's why. - Right, and Bala, on your side of that, you and your team, you've done deals and kind of all the different corners of the space, really software, hardware. And now, you know, coming into AI, I'm interested in how you see the market. And it's also, you know, doing a deal right now is gotta be an interesting time, just because on the terms of the valuation question with these mega IPOs, all these other things kind of swirl around, it seems like the landscape is kind of constantly changing. But so how did you approach those questions as you were engaging in your site of the marriage? - Absolutely. So, I mean, among the biggest, among the biggest challenges that we, that we perhaps have, which we needed kind of grapple as a SPAC, was in fact a valuation. You know, I mean, to your point, some of these, some of the valuations that we've seen in, and the private world are just outlandish. I mean, just completely outlandish. The most egregious one that I saw was company that was generating $10 million in revenue that was valued at over a billion dollars. And then of course, you've got all the other, you know, major ones, you've got cursor and all the other folks were valued at, you know, 60, 60, 70 times, you know, ARR. And the, and the other issue that we, that we obviously have in the SPAC world is the fact that we can't reduce valuation too much for two reasons. Number one, it's driven by the SPAC size. You know, we have $172 million. In the SPAC has to be a reasonable size of the surviving entity. And secondarily, you know, suddenly given our SPAC subcreated, you know, post-D SPAC and we wanted to stay well away from the $50 million or so market cap that SPAC didn't in fact fall to, and then wanted to deal with the territory in the event that the stock does momentarily fall to some point. It's in addition to that, you know, given the new rules, we did not use any forecasting. All the valuation was done only on 2026 numbers. So we came up with an analysis that said that valuations in the private sector were around a little off about 20 to 25 to a higher for about 60 to 70. And India's, you know, 2023, 2025 numbers were about 14. And we set up a number at about, at about 250, which basically set up a valuation, a training valuation of about 18 to 20. And we have forecasted in the S4 that mobile wallows looking at an ARR between $16 and $20 million for 2026, which will put the valuation somewhere between 12 and 15. So we felt that both those numbers were, were eminently defensible and a significant discount to the numbers that we were seeing in the private sector. They were no really direct comes to what in India was doing in the public sector. So we kind of felt that felt that from that perspective, it was a defendable deal, it was an attractive deal. In fact, one of the, you know, one of the big pushbacks that I got from my voters to whether this was a deal that we should take was the fact that the revenue was on the light side. You know, we had initially set a target of a revenue, you know, between $30 and $70 million would have been an ideal target. But given all the other positives that we talked about in terms of investor support, in terms of market customers, in terms of customer engagement, we felt that that balanced out the issue of the revenue being lower than we would then we would like. Now, with respect to the other issue of the market segment, my previous facts were that are largely focused on the energy space because that's my core domain expertise. You know, if you take away my, my SPAC, and my public equity, you know, kind of expertise. You know, back on the day when I, when I used to be a technologist in an engineer, I came up through the energy world. I came up through, through generation, through storage. And certainly given, given the issues that all the data centers are kind of facing today, those are substantive issues. You know, especially transient conditions that have got to be supported as data centers switch from one source of energy to another. So we had looked at battery companies. We had looked at, you know, high-fathers of the capacity companies. We have looked at, as I said, healthcare, you know, we had seen multiple, multiple healthcare opportunities pre-dease back. I mean, pre-IPL. That was certainly a point of interest. So when an engineer came out, you know, frankly, out of the left field to us, it just clicked. You know, we had an L.O.I in place seven weeks after the IPL. So, you know, we think that the deal just fell into place and I think it's moved along very well in the months since then. You know, for people like me, the question is not, hey, is AI going to be a bubble or is AI going to work, right? I mean, I've spent my life on AI and I think AI is absolutely going to work, absolutely. The real question for me is where is the value of AI going to accrue next, right? And that's the question. And I firmly believe it is not going to accrue in horizontal foundational models anymore. I mean, that game is over. I truly think the reason AI is going to be big is because everyone and because I work in always in, you know, selling things to companies, big, small, I think that AI is going to get pervasive in companies. People are going to do operations using AI. But for that to happen, more stuff needs to happen. Repritory data, domain expertise has to be developed. The stack, because the current technology for doing things like retrieval, orchestration is not going to work in vertical AI. And I truly think that we are at the absolute frontier at the vanguard of that move. Right. And so in India, you know, now that you're moving ahead and becoming a public company is now very much on the horizon, what is the thing you're sort of most excited about being able to do as a public company to continue expanding mobile walls, offerings, and footprint? Yes. Here, Nick, I think very much like an engineer technology's founder, right? So the thing I'm most excited about the belief that drives me every day, which unfortunately is not immediately going to happen now, when enterprises use AI, which I'm going to call vertical AI, because telecom companies are going to ask telecom questions and retail companies are going to ask retail questions and so on, insurance companies, insurance questions. So the belief that drives me is that there are a couple of very nerdy reasons why you just cannot, Let's say you are a large insurance company. And you wanted to build insurance AI application, where you can ask things like, hey, when we wrote policies for these two people, they look exactly the same, why is this person turning out to be much riskier than others? You just cannot take the insurance company versus proprietary data. You just cannot take foundational, like a GPT or a quad or a hugging face or something, just train it on the proprietary data and build it. So there are some very nerdy reasons why that cannot happen. You need this other stack, other computational capabilities on top. And the belief that drives me is that we are one of the first to recognize that problem. It's a hard computer science problems. And we are the first to actually have Gen 1 solutions and hopefully we'll have very robust solutions the next year and a half. And then really the next, you know, in my mind, that is the trillion dollar play when you launch that stack where that anybody can take and build applications, right? So what I look forward most to Nick is to build that stack, right? Now that's, and the goal is that stack, we should be able to launch that stack and you know, early thoughts had to do open source, I'm not sure about that yet. And what this back and the go public should enable us to do is to put the pieces together to make that happen, right? One, it should enable us to hire the 15 incredible ML engineers that we need, where each of them going to cost probably half a million to $700,000 a year. It's going to allow us to put together enough compute capacity, either in the cloud or I don't know. I mean, we are thinking of sort of going away from commercial clouds now, but it's still expensive to be able to test out the thesis, you know, because basically the biggest thesis we have to test out is the stack works across domains, right? And in a bunch of domains, there are public data available, so we would like to be able to get the data and run tests, right, that's expensive, so we ought to do that. And all the while doing that, had the resources to be able to still grow the company through some organic sales, like, you know, sign up every major telecom company, the world, for instance, but largely in organic means, acquiring interesting data and teams, right? So that's what I look forward to, to build a stack, to build that fundamentally disruptive piece of technology, which I think is whether we do it or not, somebody's going to do it and have the resources to get there by growing the company during the time we are doing it and putting the resources together to make that happen.