Clips
News
More from Snowflake Summit 2025

Gerard Francis

Chief Product Officer: AI and Data

JPMorgan Chase & Co.

Transforming institutional investing: How JP Morgan’s Fusion and Snowflake are elevating data management

Robust data management is critical for institutional investors, including pension funds, insurers and mutual funds — as it underpins sound decision-making, effective risk management and regulatory compliance across the vast capital pools they manage.To meet those needs, JP Morgan Chase & Co. introduced Fusion — a cloud-native platform designed to streamline data management, analytics and reporting throughout the investment lifecycle, according to Gerard Francis (pictured, left), chief product officer of AI and data at JP Morgan.Snowflake’s Carl Perry and JP Morgan’s Gerard Francis talk with theCUBE about how Fusion is taking data management to the next level

play_circle_outline Fusion normalizes, harmonizes, and converts data into one semantic model.

play_circle_outline Snowflake enables fusion and harmonization of data for customers.

play_circle_outline OpenFlow allows unstructured data to be pushed into Snowflake for insights.

play_circle_outline Semantic view stored in Snowflake unlocks AI capabilities and BI vendors.

play_circle_outline Cortex AI SQL enables complex natural language queries for insights.

Info
Transcript

Carl Perry, Snowflake & Gerard Francis, JP Morgan Chase

Gerard Francis

Chief Product Officer: AI and Data JPMorgan Chase & Co.

In this Snowflake Summit 2025 segment, theCUBE’s Dave Vellante and John Furrier sit down with Gerard Francis, chief product officer, AI & data and head of Fusion at JPMorgan Chase, alongside Carl Perry, head of analytics at Snowflake. The discussion uncovers how the banking giant is redefining financial-services analytics while Snowflake pushes the boundaries of AI-powered data applications on the AI Data Cloud.

Francis explains how JPMorgan’s enterprise data mesh and Fusion platform turn messy, siloed information into AI-ready assets that drive meas... Read more

explore Keep Exploring

What was the intent behind JP Morgan's launch of Fusion in relation to data management for institutional investors? add

What were the three key components that were implemented in the data transformation process described in the text? add

What is OpenFlow and how is it a game changer for customers? add

What potential benefits does OpenFlow have for customers, particularly in relation to semantic views and AI capabilities within Snowflake? add

What is the most recent announcement made by Snowflake called Cortex AI SQL? add

bolt Powered by CUBE AI

Carl Perry, Snowflake & Gerard Francis, JP Morgan Chase

Rebecca Knight

>> Good afternoon everyone, and welcome back to theCube's live coverage of Snowflake Summit 2025 here at the Moscone Center. I'm your host, Rebecca Knight alongside my co-host, Co-Analyst, Co-Founder of theCube, Dave Vellante. I would like to welcome two new guests to the show. We have Carl Perry, Head of Analytics at Snowflake. Thank you so much for coming on, Carl. And Gerard Francis, Chief Product Officer, AI and Data at JP Morgan Chase. Thank you so much for coming on the show.

Gerard Francis

>> Nice to be here.

Rebecca Knight

>> So I want to start with you, Gerard, because a lot of investors say they want to bring AI into their workflows. They want to see ROI yesterday, but they don't always have the right data foundation to make that happen. I'm curious from your perspective, what is it that's missing and how has JP Morgan solved that problem?

Gerard Francis

>> Yeah, so I think that's a great question. I think when you look at the Gen AI models out the box and how you use them, there's value you can get for certain types of things, whether it's writing code, being creative, editing documents, and there is value there. But it's a certain type of value that's normally accretive to an individual. So it's hard to solve the bigger use cases with it. Once you start going after the bigger use cases, then it becomes really how you make the Gen AI personal to your own data, and that's where the challenge arises. I think when people begin this journey, they normally start off a little bit piecemeal. Everybody's got their data, they want to solve this specific use case, and they learn through doing. But where we think we need to be is to have a integrated data platform where all the data within the firm is visible, accessible, discoverable with the right governance and controls.

Gerard Francis

And that same platform plugs into the AI and the AI services and allows people to build applications really fast and then deploy them into production. And if we can solve for that, it's a virtuous circle, then that really makes the AI use cases go a lot faster. So we're in the early stages of that, but we're really excited by what the opportunity represents and the success that we are seeing.

Dave Vellante

>> Just a follow-up question on your title, what is Fusion? That's an intriguing concept, fusing the data, all this harmonizing, fusing the data. What is Fusion?

Gerard Francis

>> Yeah, that's a great question. In fact, the name itself is a giveaway because when you look at almost any organization, focused on institutional investors, they have so much pain with their data because their data comes from lots of providers, it's in different formats, it's in different models, different schematics, semantics, and it's a lot of hard work to bring it all together. And the truth is, everybody doesn't have all the experts to make all the data work. So when we as JP Morgan, as part of security services launched Fusion, the intent was, how can you take away the data hard work from the customer? So the idea is we onboard the data from the vendors, we normalize it, we harmonize it, we convert it all into one semantic model, and we push it back to the customer completely harmonized with analytics on top, so their entire data pane has gone away and they can jump right to value and extracting that from the data.

Dave Vellante

>> I love that. So last year, yeah, last year, George Gilbert and I wrote a post, and the title was Why Jamie Dimon is Sam Altman's biggest competitor. And the premise was that JPMC has all this data, I think we underestimate it, it's whatever, how many exabytes you have, and that's never going to get out into the public domain for LLMs to trade on or train on. Rather, you're going to fuse the data internally, harmonize it using partners like Snowflake, and then build your unique value add on top of it. So from Snowflake's standpoint, that is what your business is all about, keeping it protected, governed, easy to use, easy to harmonized, et cetera. And for JPM, that is your competitive advantage in the future. So I wonder if you could maybe discuss not only what's going on with JPM, but are you seeing that pattern across customers where they're saying, "This is our new value add vector?"

Carl Perry

>> Well, I think there's a couple of things that are really interesting happening right now. I remember when the iPhone came out, to use an analogy, and it was only a consumer device. Now everybody uses an Android or iPhone to do their job every single day. And Chat GPT has evolved to a similar thing. You could ask questions of anything that was on the web. Now there's an expectation by every employee in the world that they can ask questions of all the data they have, whether it's structured or unstructured. And this is a huge revolution in terms of how I can derive value out of my data and drive better business outcomes. So that's one of the big trends that's happening. Two, I think in particular is we want to be able to make sure that as you do work on data, you can trust it. Gerard talked about this, but you want to make sure that you have the right semantic view and model on your data, and ensure that you have feedback when you're asking these questions that you actually are getting the correct answers that you can make decisions on. And then I think the last reason, the last thing that we see oftentimes is a lot of customers still have data on-prem, that's just the reality of it. And we want to help enable them to move to Snowflake to gain value of all these incredible technologies. So we're doing a ton of investments in things like migrations with SnowConvert AI, to accelerate the movement to the cloud so people can get all the value of their data with all of these amazing technologies.

Dave Vellante

>> In the fullness of time.

Carl Perry

>> Well, hopefully we can accelerate that to actually be much sooner.

Dave Vellante

>> But you're right, the whole consumerization piece, that's always where it starts. That's how the economic flywheel gets going, and then it's like, "Oh, how are we going to govern it? How are we going to secure it? How are we going to trust it?" And that's the really hard part.

Carl Perry

>> Well, I think the way we're built from day one makes that really easy for our customers.

Rebecca Knight

>> Gerard, I want to go back to you about this Fusion, and I want you to break it down for our viewers about how it really helps clean up and connect that messy, siloed data and make it actually usable, and make it be AI usable?

Gerard Francis

>> Okay, so let me try and use some analogies. So when every data vendor, or every data provider provides data, their formats by default are different. You can imagine people send you a bunch of Excel spreadsheets. How the hell do you combine these things together? They're so different to look at. But yet they're all in the Excel format and yet I can't put them together. Imagine if every format is different to Excel. So now I've got a difference in formats, I can't put them together, and yet even if I put them together, it's going to be a mess of different sheets, none of which connect. So you go through a layered process. The first stage is, how do I make various things, whether it's a JSON, or an XML, or a CSV, or an old-style CSV, it doesn't matter. How can I first make the format look the same and pretend it's the same, to start off with that? Then I go through it and I look at all the column names. Everybody gave these things a different name. It makes no sense to me. How do I go off and re-correct all those names? Don't change it, but have something that makes sense assigned to all of them. So now I know at least what the columns mean when I'm looking at it. But now I've got an even bigger problem, because somebody's got companies and they have certain identifiers they use, and the other person has companies, but they use completely different identifiers. So how the hell do I know you're talking about Apple? Or you talking about Microsoft? Or I'm talking about Apple's fifth subsidiary that's hidden below and not something else? So now you've really got to go through every identifier whether it's a company, or a stock trading on a stock market, or the exchange it is listed on and you've got to remediate and make all these identifiers come together, clean it up, normalize it, and harmonize it. So now I've got my identifiers sorted out, I've got my columns with somewhat common names, and then I've got my data in the same format. But I still can't exactly write a query and it all will work, because I've probably got 100,000 columns and I've asked a query, it's going to be a completely ridiculous answer because the tables are so many. So I can't really operate on thousands of tables with hundreds of thousands of columns. No query really works. So then what we do is we create what's called a semantic model, which is a much more simplified model of the entire world, where you can take these thousands of tables and reduce them to just 10 tables. You can take these hundreds of thousands of columns and reduce them to a finite number, say 2,000. And now I have a chance of actually writing a query, and actually making it work because in the background we figure out the hard stuff, we join it all up together and connect it together. So we make it feel like a simple Excel spreadsheet where everything is beautifully organized and it just works.

Dave Vellante

>> So how did you harmonize that data? Was that humans? How much of that was automated? How hard was it or easy to build that semantic model? And how does that semantic model evolve and learn over time to keep up with the business changes?

Gerard Francis

>> Now you're asking for all the secrets. I'm going to get into that now. I'm kidding. I kidding.

Dave Vellante

>> It's your secret sauce. That is really cool stuff.

Gerard Francis

>> Yeah. So I think if you look at this historically, you have people who work on file by file and transform it file by file. So even though they could use good ETL tools, the process of doing that is highly manual. So what we invested in is really creating one technology stack that we can do data transformations at scale, and we can apply as we find common patterns, we can push the common pattern into the stack so we can take away repeat work that happens by different people. So we are constantly optimizing the process through technology to make that entire transformation cycle work better. So that's one thing we did. We built it ourselves. You could argue whether we should or we shouldn't. We think we've done the right thing in getting that scale. The second thing we had to do, was how do you make all these IDs consistent? And that's where we had to spend a lot of time. The good news as a bank, is we have a lot of information from lots of sources. So we have the knowledge on how these different things work together. We probably are the biggest purchaser of data in the world. So we understand identifiers, we understand nuances, differences, similarities. So we could build an engine that could account for the vagaries of everybody's different identifier systems. That's the second thing we did. And we built it, and we have great partnerships with our vendors, so we made sure we had the right licenses in place to understand the data, et cetera. And the third thing we did, again, because we are JP Morgan and we have a lot of expertise with every kind of private asset data, public asset data, et cetera, we hired a bunch of people who really understood all the complexities of data, and developed this semantic model that can account for every nuance out there. And you can only do that if you've lived through all the pain to go and build it. So it's a combination of these three things, this transformation engine, this magic thing we call the spine that brings things together and this semantic model that accounts for the vagaries. And we also had to build the semantic model in a different way so that you never get stuck in the modeling situation where every time people are trying to model stuff and then get stuck for years trying to get it right, we had to design a process to your question, that's easy to evolve, quick to change, and it's constantly working with the market.

Dave Vellante

>> So Snowflake was a data enabler in that?

Gerard Francis

>> Yes, so Snowflake is very important for us because ultimately when you have harmonized data, for the client it's still not good enough because they need to then go off and execute on it in order to get value. So Snowflake was the first partner we worked with in order to be the last SMILE delivery into the customer. So the customer's got Snowflake, we push the data fully modeled, fully harmonized, direct to Snowflake, and immediately the client gets value. They can write their queries, they can take advantage of Cortex AI, all of that right off the bat because the data's been harmonized to them.

Dave Vellante

>> So the point of our post was that consumer LLM foundation models, they can create synthetic data. Their data's not synthetic, and that's really the difference. But now of course, you're JPMC. Your goal, I presume, is to enable any company to be able to get to this level, which is again, probably going to take a while.

Gerard Francis

>> But this is where we are partitioning with Snowflake because as long as the client has Snowflake, we as JPMC have a service where we can do all this hard work, do all the normalization, transformation them and push it back into the client's Snowflake, so they're good to go.

Dave Vellante

>> Really? You're like an ISV of sorts.

Gerard Francis

>> Yeah. If you want to call us that.

Dave Vellante

>> I'm sure you-

Gerard Francis

>> .

Carl Perry

>> Actually, you guys have built this end-to-end experience for your customers that you deliver today using Cortex and Snowflake to them, which provides them the ability to interact with their data in ways they previously couldn't, right?

Gerard Francis

>> That's right. In fact, we did this experiment last year where we took our data, put it into Snowflake Cortex, and immediately people could ask these crazily difficult natural language questions and got absolutely precise answers.

Carl Perry

>> This was awesome to work with them because-

Dave Vellante

>> Do you monetize that service?

Gerard Francis

>> Yes, we do.

Dave Vellante

>> You'll sell it to other banks?

Gerard Francis

>> Institutional investors. So all the buy sides of the world is what we sell that to.

Dave Vellante

>> Okay, so a direct bank competitor you wouldn't sell it to, but a small regional bank.

Gerard Francis

>> financial institutions, we would, yeah.

Dave Vellante

>> Yeah. Interesting.

Rebecca Knight

>> Well, I'm just thinking about what it unlocks not only for the organization to have access to this beautiful data, but also the individual, in terms of how it really supercharges their career, and the kinds of things that they can do, and learn and the kinds of questions they can answer immediately.

Gerard Francis

>> Exactly what Carl said. Yeah.

Carl Perry

>> It was funny, because we started playing with the data that JPMC had, and we would ask questions, and like Gerard said, gibberish. Horrible queries came out that didn't actually make any sense. And so as we built a semantic model and refined it, it lit up. And the types of questions you could ask were mind blowing. It was like, what derivatives, do I still have open positions on this day in this quarter with these currencies? I don't know who can write that query, but it could take that and then generate the query for me, and give an answer and form a visual that actually answered it correctly.

Gerard Francis

>> That's right.

Carl Perry

>> It was pretty amazing.

Gerard Francis

>> Pretty cool. Yeah.

Rebecca Knight

>> So you've announced a bunch of new analytics innovations here at Summit. You've mentioned of course, NoConvert AI. Walk us through what is really new and what's a big deal for customers?

Carl Perry

>> Well, so I'll mention two things that are directly analytics-based, but I'll also talk about OpenFlow, which I think is a game changer for our customers. In particular, OpenFlow is built on Apache NiFi, an open source specification that enables us to actually have customers install OpenFlow and the control plane there into their on-premise, and then have that data pushed into Snowflake, or in their VPC in the cloud and push that data to Snowflake. And the reason this is so beneficial outside of structured data, like from Postgres or other databases, is it can pull all of this unstructured data, which is really, you bring the AI to the data, and as you pull that unstructured data into Snowflake, you can have amazing insights. So I think OpenFlow is a game changer for our customers. The two analytics things I'll mention. One is I think the semantic view work that we've been doing, you can now create a semantic view and have it stored inside Snowflake as a top level object. This unlocks pretty much all of the AI capabilities that you're going to use with Snowflake. It also helps our BI vendors who want to build experiences on top of that semantic model. And finally, with all the agents that people are going to be building, this is the critical foundation to have accurate correct answers come back. So that's the second thing. And the last thing that's the most magical thing that we are announcing is called Cortex AI SQL. We had a bunch of AI operators to say things like, "Tell me the customers who were satisfied this month, where this series of audio calls from customers had a satisfaction rating of four." And an analyst is now turned into an AI engineer to be able to intersect unstructured and structured data, and the insights they get back, it would've taken quarters to go build these types of capabilities. So those are the things I'm most excited that we announced at Snowflake today.

Dave Vellante

>> So that's OpenFlow. Cool. You don't have to buy Informatica for $8 billion. By the way. I love what Informatica and Salesforce are doing. Informatica's got a lot more maturity and customer base, but then the piece about text to SQL, that's an example of taking what people understand as things, and turning it into what databases understand as things.

Carl Perry

>> Well, I think in particular, I didn't mention this, but all of-

Dave Vellante

>> I'm sorry, strings,

Carl Perry

>> All of the capabilities we're building in AI is I think the foundational piece of building a more accurate correct set of data that comes back, is the semantic view. Whether it's in Cortex Analyst, whether it's in Snowflake intelligence, whether it's in all the AI agents everybody's going to build, this is the foundation for making sure you have the right data. So AI can actually give you amazing insights out of your data.

Dave Vellante

>> So you guys, JPMC, unbelievable, always ahead of the curve. So I just mentioned turning things into strings, but people talk about building a digital representation of the business, people, places, things and processes so that now you've got this basically a digital twin if you will, of your business. Is that a viable vision? Is that something that you think is pursuable, where agents are actually acting in a trustworthy manner? Is that something that we'll see in the next years?

Gerard Francis

>> So I think when we talk about the semantic model, it pretty much follows what you said, people, things, et cetera. That's really your-

Dave Vellante

>> And processes.

Gerard Francis

>> Model and processes, and that's where it gets more complicated, right? Yeah. So I really do think that agents, which effectively, if you think about traditional software development, you need to know all the knowledge, codify it all into rules, and that's how software works. With agents, it's software development. But now you're introducing non-deterministic points within that process, which means things happen that don't need to be encoded with very specific rules, in order to go and do something. That both unlocks creativity, because now you can tackle harder problems that you couldn't before, but it unlocks risks too, because you don't want it doing something it shouldn't be doing. So I think as we go down this journey, I suspect we'll start off with using agents, the technology is here, we are building it out, to solve for the simpler, less risky use cases. And as we build more confidence in them, get the governance in place, the controls in place, feel confident they can't do things that they shouldn't be doing, then we'll go after the more complex use cases to make them work.

Dave Vellante

>> So contact center, coding, would those be examples?

Gerard Francis

>> Yeah, coding is a good example where you can use agentic processes to work. I think in customer centers increasingly people will look at how they can in a safe way introduce more agentic processes to make sure the answers are quicker, better, faster. But agentic shouldn't be confused with human agents. I think sometimes people get very confused because the word is the same. Agentic effectively a introduces non-deterministic way of doing things. How people use it, what problems they try to solve is going to vary across the industry, and you'll have different consequences for that.

Dave Vellante

>> But my takeaway is it's going to take some time before you trust the agent actually affecting a transaction?

Gerard Francis

>> Oh, yes. Oh, yes. Absolutely. Absolutely.

Dave Vellante

>> Will that happen?

Gerard Francis

>> It's a little hard to... See the way I would think about it is software development. When software development began, could we think about an algorithm doing equity execution in a stock market? Probably not. Today the vast majority of trades and stock market happens through algorithms. I think we're really at the early stages of the journey where we're going to feel our way through to understand where it goes, while we balance the risks as we go along.

Rebecca Knight

>> Interesting times, interesting times. Thank you so much, Gerard and Carl.

Gerard Francis

>> Thank you.

Carl Perry

>> Thanks a lot.

Rebecca Knight

>> Really fascinating conversation.

Carl Perry

>> Thank you.

Rebecca Knight

>> I'm Rebecca Knight for Dave Vellante. Keep it right here on theCUBE for more from the Snowflake Summit, you're watching theCUBE, the leader in enterprise tech news and analysis.