Subscribe
Share
Share
Embed
Welcome to The Payment Expert weekly podcast, brought to you by SBC Media. Each week we analyse the news driving the global payments industry forward; the innovation, the infrastructure, and everything that has to happen to make it all possible.
Louis (00:00.77)
Hello and welcome back to the Payment Expert podcast, your source for the latest news, insights and analysis on the payments industry. I'm Lewis Tompsett, news editor at Payment Expert and joining me today, I'm delighted to have Bahao Patanjia, head of global industry business development for payments at NVIDIA. Bahao, thank you very much for joining me on today's show.
Pahal Patangia (00:25.404)
Thanks for having me, Louis. It's a pleasure to be here.
Louis (00:28.194)
Great stuff. Let's get into it then. Let's talk about obviously Stripe and Mastercard. They've launched new payment foundation models. They were both built with NVIDIA. What do those models sort of actually look like and what can they deliver that wasn't perhaps possible before?
Pahal Patangia (00:47.164)
Yeah, no, I mean that that's that's one of the hottest topics that are happening in the payments world right now. And if you if you if you take a step back and think about financial services in general, the industry has been heavily laden with tabular style data, relational style databases all this while. And so far to to make sense out of data and to get in s to predict to get
predictions for downstream tasks out of that data is where you would have required machine learning algorithms. And that was the state of the art. The industry was doing fantastic with it. And and and only only very recently with all the developments that have been shaping in the world of generative AI and large language models, etc., there's this emergence of transformer architecture, which is like a deep learning
I would say paradigm that is very well suited to be learning patterns from sequences. anything that has like a sequential nature to it would be could be encoded through the transformer architecture. And then you you're applying that to the text domain, you were applying that to the videos and the images domain and and it turns out thanks to you the work that is being done by the academy and industry
that you could apply that that transformer architecture and the ability to learn sequences and patterns, say over time, apply to tabular data as well. And and and what what happens at the at the at the at the culmination of both the both the spectrums, that is tabular data combined with transformers, is that you are able to build what I call contextual representations or what somebody would technically call embeddings.
that come out of it. Now these embeddings would compose everything about a customer's behavior when it comes to when it comes to what the customer likes versus you know what kind of games or sports did the customer play back in the past versus today versus you know what the customer will do next. So when you think about that holistic view of a customer, you know, think
Pahal Patangia (03:09.659)
think three sixty degree, every enterprise, every payments company, every banking institution is kind of racing towards getting that full view. Because at the end of the day, if you know your customer better, you're able to sell better, you are able to, you know, get the right f facing, right catered products in front of the customer. And and and that's that's the whole pursuit and that's where the whole industry is heading, you know, to to understand the to understand
anything and everything they can about the customer and the means to do that is these payments foundation models, tabular transformer architectures, etc. And and and that's that's where you see, you know, recent announcement announcements and the work which we have been doing with the likes of Revolute, with the likes of MasterCard, all of them you know presented at our GTC conference, which happened earlier in the year in March.
And since then we have seen great momentum in the industry coming in from the likes of Plat, coming in coming in from the likes of other payment customers who who have like kind of doubled down on this horizon. So so so that's that's that's the holy grail. And then you know, there are like great components of NVIDIA's platform being involved as a part of it. if you think about it as NVIDIA as a full stack computing platform company.
you know people think of us as like the hardware play but you know what what's lost in that narrative is that you know that hardware is increasingly s supported by the software that goes beyond it. And these software is libraries which helps you know helps you train these models better and effectively and similarly such libraries are being used in the development of payment foundation models.
Louis (05:05.045)
Yeah, absolutely. I mean, it used to be for years, didn't it, that whoever held the most transaction data, they were the winners. But as data sharing has improved, and I suppose the understanding of the value of data sharing has increased, I suppose the advantage has moved forward to who can best interrogate that data, who can best understand that data, who can derive the most value and generate the most value out of that data.
Pahal Patangia (05:32.828)
Hundred percent. And and if you and rightfully so, because if you think about it, just a typical transaction funnel, there are like different players who would touch, you know, certain kinds of data and for a certain transaction the transaction will flow through different entities, I would say. be it the acquirers, be it be it the card networks, be the issuers, etcetera. And and and and what happens in the processes that you know, like everybody has some sort of data at scale.
You know, and they combine that with their own proprietary information which they have about that particular transaction or entity. so so so the differentiation at times to your point, even you know, within the in the world of open banking, etc., is not at the Rails level. The Rails were already developed and and data was flowing through them and transactions were flowing through them and business was going great. But again, in in in this world where you have to
grab a greater share of pie within a newer set of customers or continue to make your current products current I would say current to continue continue to be on the loyalty spectrum with your current customers is where you have to get from a system of record to a system of intelligence. And and and the means and and and that is why that is why payments foundation models are super important.
Because they they are they are the bridge between that system of record to system of intelligence. How do you build that semantic layer of understanding? How do you build that behavioral encoding of customers at the end of the day so that you are not just sitting at the data layer, which, you know, by itself means nothing, but rather you sits on you you sit on the insights layer, you sit on the actionable layer. So data layer is very reactive, I would say.
the the insights layer or the intelligence layer is is where, you know, the action happens because that then you could take it to downstream use cases, be it be it you know, your fraud or your dispute predictions or your cross selling, etcetera. All of them come into play.
Louis (07:40.779)
Yeah, totally. mean, I suppose perhaps a different kind of intelligence, but one that's coming to the fore at the minute is is agentic commerce and it's much discussed too, moving into sort of live production and pilots on the way. But I suppose the infrastructure behind that that you've spoken about so well was almost built for humans, all the logins, the onboarding, the authorization flows, et cetera, et cetera, was built with humans in mind, not for agents.
And I suppose today there are computing protocols, different card networks engaged in the space, all trying to solve agentic commerce at once. As of yet, there's no perhaps standardization. that, do you think, one of the problems that has to be solved first? your perspective, what do you think needs to be the things ironed out before, obviously, agentic commerce moves into full-scale implementation across the commercial world?
Pahal Patangia (08:39.311)
Yeah, yeah. If we think from if you think from our perspective and zoom out a little bit, you know, there's and then we think about what what do we do as a company which which we are best at. And and and and I and I think from first principles we have a full stack computing platform company that will provide you the infrastructure tools and the libraries to do X. X being, you know, the realm of agent e commerce here. Now who would help
build those layers of X are the different ecosystem players in the payments community. And that could be the networks, that could be the startups, you know, companies coming with protocols, everything who participates in that, in in the vicinity of things, right? So so as we are seeing the convergence, as we are seeing the convergence of different layers and where value accrues, I think I think the biggest the biggest near term thing where we are
where we are continuing to build you know, continuing to build a position of value is in the fact that if you think about agents and agents taking actions on your behalf, not let's not even talk about shopping for that matter, but any action on your behalf, an agent should know what do you want. And the only means for agent to know what do you want is to know you better.
The means to know you better is again payments foundation models. What happens is that w the embeddings and the contextual representations which I've been I've been I've been talking about that come out of payments foundation models have the right context about you. They have the right I would say caricature built in terms of you know, one customer, one persona. And then and then when agents have to take action on your behalf.
And you could put shopping in this context, yeah, they would they would actually grab that contextual representation and learn what exactly would you like and then shape their action accordingly. This shaping coming in from po foundation models, the payments foundation models, and converting into action on the agent e commerce side is a phenomenon which isn't discussed yet in the industry.
Pahal Patangia (11:04.301)
And that will continuously become important as as agent e commerce b goes at scale. Because because we are talking about autonomy, which means that to for autonomy to succeed, the level of personalization needs to be at the highest level. And that will only come through these payments foundation models. So in effect, payments foundation models and agent e commerce
Louis (11:24.171)
Hmm.
Pahal Patangia (11:31.899)
are like a flywheel to one another. Famous foundation model will feed into agent e commerce and the feedback loop as a result of that action would again go back into improving these models better and it goes like a three sixty degree loop.
Louis (11:48.449)
Yeah, very interesting stuff, pal. So I suppose the autonomy comes from that higher level of personalisation that you allude to. And by having that higher level of personalisation that the agent can offer through those payment foundation models, you, suppose, get that. I mean, the question I have for you off the back of that really is, I suppose when an AI agent initiates a transaction and it goes maybe slightly wrong,
Perhaps it doesn't have the level of personalisation, as you would put it, that is required. I suppose the question there is about liability.
I suppose regulators are pointing out, you know, who has the liability for if it goes wrong? Does it fall on the consumer? Does it fall on whoever deployed that agent that the consumer has used? How does payment infrastructure sort of need to be built to make that assignable in practice? Do you think it's just about getting those levels of personalization as high as possible so that doesn't happen?
Pahal Patangia (12:50.895)
Yeah, no, I I think I think that answer unpacks into a few facets, I would say, you know, and and first is like you know, begins on the personalization front to be at the lowest level. I I would give one example coming in from you know, you know like first let's let's talk about like being being rogue at the most basic level would be that, you know, you are
you have given a certain task and then it is performing totally orthogonal orthogonally to that task, right? And and the and and the and the means to do that, the means for that realignment lies in the lies in the piece of personalization. Like this is this is the kind of work which we're doing with the likes of Paypal, for example. so PayPal has been building the Neagic Commerce platform where they are where they are rolling out these capabilities to the ecosystem of nineteen million
SMBs. Now to do that at scale with the with the latency with the latency and cost effectiveness they need is is where they have resorted to leveraging open source models like what we produce called Nvidia NemoTron and and and and primarily sort solving the search part of the agent e commerce piece. Now
Now now this has been solved over the years with recommender systems and, you know, all the all the work that consumer internet marketplaces have been doing. and we have reaped the benefits of it. But but but but the bigger picture here is now that we have better algorithms that are able to align better, that are able to personalize better, that are able to search better, you know, in terms of multimodalities like
you know, your text, images, videos combine and get you that particular SKU into your chat or whatever interface you are in, thanks to thanks to thanks to, you know, them being fine tuned better. So the combination of open source models, fine tuned with your open fine tuned with your data, is, you know, getting you to a level which is you know, as good as the best of the best models there. at the same time
Pahal Patangia (15:08.325)
the same time being cost effective and latency sensitive as well. And and and and and when when you when you do that you are hitting that personalization check mark if you will to put it very simply. Now now the other pieces we as an industry need to solve and and and the whole idea of KYA becomes super important.
of course the industry we as an industry have solved KYC and KYB. And if you think about it, if you break that, if you break that, the underlying layers are always have always based on have been rooted on number one data, number two, what algorithms you apply on that. So data plus deep learning, you know, data plus machine learning has been has been the crux of you know how we have solved KYB and KYC. And
And and our belief is that a similar process will happen in the KYA world as well, in terms of in in terms of you know how different data points, etcetera, would be leveraged to find out whether this agent is anomalous or not. And that would require a certain level of modeling, that would require a certain level of in the machine learning language anomaly detection process to be developed.
And and you know and that will you know, that will be like a fundamental base ground for it anyone to anyone to go and test test out agents in the in the wild. And and and and the ecosystem has been doing a lot of good work. There there have been few frameworks here and there the other day. I was talking to folks at Experian who were talking about their KYA framework. and there are like really good startups.
the ecosystem who who have been contributing to the space. So I'm really looking forward to what's coming next. But again, all of them would require data and deep learning and that's where Nvidia would continue to play a role.
Louis (17:17.331)
Yeah, absolutely. suppose the KYA element is a big talking point and once that's kind of achieved or reached, then I suppose those agentic models will start to sort of really flourish in the market today. I want to pick up a bit on what you mentioned about those open source models because they are sort of reshaping, I suppose, the competitive dynamic that we see in financial services. I think Nvidia's own research points to that. Do you think that accelerates the...
the adoption of it or does it sort of create risks that perhaps government teams aren't ready for? know historically financial institutions like to conserve their data, the value of open source model, and I suppose the cost effectiveness of open source models is a benefit that many are starting to realize.
Pahal Patangia (18:04.015)
Yeah, absolutely. you know, just to just to be very clear, we we are agnostic of what our end customers and partners use. If they are if they want to leverage the best and the best of the closed source foundation models, definitely we would encourage them to do so because you know, at the end of the day, those foundation model companies are great partners for us and we would want to
you know, help them propagate into the broader ecosystem. And at the same time, you know, there are there are cases when institutions, financial institutions in particular, don't want to expose their sensitive data or and want to w want to keep them closer to the chest. And and and and and and then how do they, you know, leverage all the goodness that is coming out of the technology changes?
and and and and the answer to that is open source models. That you know, open source models today, broadly speaking are like six months behind, I would say, to the broader f foundation model race. And and and and if you think from that lens, what's required for open source models to be as good is is the is the is the is the is the idea of fine tuning. That when you take open source models
you combine it with your proprietary data. What comes out of it is the is is is is the product which is totally catered to you. It is proprietary to you and it composes all your intelligence into that model or whatever offering you are ultimately making it into. And that level of autonomy
is at times desired by financial institutions. And more and more so is is where we are seeing institutions you know gravitating towards this path. I just mentioned the P PL example to you. Around around March or April earlier this year, Wall Street Journal did an article on how Capital One, B NY, different different institutions are are
Pahal Patangia (20:27.869)
resorting to the use of open source models and largely because they see the value and you know bringing their own data and combining those proprietary aspects to ultimately make something which is equally or almost as powerful as closed source models. And it helps solve the use case. And then and then the sec second thing s other thing to think about it is the world tomorrow would be obviously hybrid
Of course you are thinking about an agent tick first world tomorrow. And if you think about if you think about the agent tick world, or if you think about the pipelines that go in a typical agent workflow, the you you you what is an agent? It is actually like a combination of your large language models plus a harness. So the so the large language models as we know, they are like we all know them.
What a harness? A harness is like an orchestrator. And and between these two there is a is a is a ecosystem of different I would not say ecosystem, but rather like a amalgamation of different processes that go around it. So there's tool calling and there's some calculations being done versus there's some search and retrieval, etc. And then ultimately your reasoning through to, you know, either assign
to a certain downstream model or you are doing a certain task. And and in the process, what you will see is in an agent tech system, you would require the best and the best models for your reasoning capability and this could be like closed source models for example. But at the same time you would require very small, fine tuned niche models for performing a certain task where you don't need that big of a model. And and
And that is where that is where the whole concept of like if you if you think about what are the recent trends that are happening in the in the in the industry, particularly from an AI side of things and large language model side of things, there there are two terms which are like popping up. first is the idea of tokenomics and second is model routing. And and these basically sit on that.
Pahal Patangia (22:52.673)
spectrum which I'm talking about where you know they they will become increasingly important in the world of agents as we go after.
Louis (23:02.229)
I'd love to talk to you all day about it, but we are nearly out of time, but some fascinating insights. I'll leave you with one last question before we go, which is if you had to pin, I suppose, a single structural barrier standing today between sort of a genuinely autonomous financial services, something to get the agents off the ground and all the things you've spoken about and the tokenomic as well, if there is one single structural barrier today, what do you think it is?
Pahal Patangia (23:32.89)
Yeah, I mean I would I would actually not answer in like just one one word, but rather I I think I think it's a it's a because it's a recipe and and you know you don't miss the ingredients in the recipe. so so so I I I think, you know, like data combined with models, combined with infrastructure, combined with talent. I I think all four pieces need to come together. And and and what we see is
Louis (23:40.426)
Okay.
It's true.
Pahal Patangia (24:01.939)
leading institutions who have been at the top of the game, think the large banks, the large payment fintechs, large payment institutions, all of them have have have been acing on these four spectrums and and and the ones who who who who have been who have not been you know, who have been behind are actually are are actually not the greatest at each e either of these spectrums, I would say.
So so that is where that is where it that is where the value lies. That is where the competitive edge comes in. And then it reflects into the downstream use cases as I mentioned. You know, you if you are cruise focused then you are able to personalize better and target better and ultimately those customers will be coming onto your platform and staying being sticky. at the same time if you want to reduce costs, then again,
all those pieces thanks to whatever we are seeing in the world of agents and generative AI, there are enough means to reduce costs as well. So like you could optimize for both levers, you know, your profitability or cost reduction. and at the at the end of the day it all comes down to those four facets, data, models, infrastructure and talent.
Louis (25:25.232)
Absolutely. Bahau, thank you very much for joining us today. Unfortunately for those listening and watching, that is all we have time for. If you're not already subscribed to the Payment Expert podcast, make sure to subscribe wherever you do get your podcasts with more insight and analysis coming over the weeks and months ahead. And for the latest news as it happens, head over to paymentexpert.com. We'll see you all next time.