MongoDB, Inc. (MDB) Earnings Call Transcript & Summary

Real delight to be able to host MongoDB. I think it's been 3 years now that we've had the pleasure of hosting you guys. CEO, Dev Ittycheria; Michael Gordon, CFO and I think your COO as well, right? Yes, I got to keep up with the promotion that happened quite a while ago, if I'm not mistaken.

Yes. Thank you.

And I'm joined by my colleague, Matt Martino here. Welcome to Goldman Sachs Communacopia + Technology Conference 2024. I think you heard the stats, 2,900-plus registrants, up from last year, a number of companies up from last year, 24,000 meeting requests. Insane. I will not tell you what the fill ratio is. It was not that high. So a lot of demand to see companies like you.

Thank you for joining us. Dev, you've been the CEO for now 10 years. I know it's hard to look out 5 years, but what are your aspirations for the company in 5 years? Where do you want MongoDB to be? When you come back to Communacopia + Technology 2029, what is this company going to look like?

Yes. So one, thank you for having us. It's great to be here. I'll start with saying kind of just reflecting our 10-year journey and what did we try to accomplish. When I joined, we were doing roughly $35-ish million in revenue, and we had about 350 employees, and our job at the time was to prove that MongoDB could be trusted for mission-critical workloads, because we were viewed as a cool -- fancy cool new toy, but everyone said, I'm not sure where and when I can really use MongoDB. Obviously, we knocked that off. Then we launched the cloud service, and there was a lot of skepticism, this was in 2016. Could we actually compete with the hyperscalers, and we were actually trying to partner and compete to saying like, why wouldn't Amazon eat you for lunch? Obviously, our growth since 2016 has shown that we actually built a pretty substantial business over that period of time. And then the third thing that we try to do is address a broader set of use cases. And so that was all about our platform and enabling customers to run more use cases on MongoDB versus just being an OLTP engine, and we launched Search, Vector Search, time series and a bunch of other capabilities that people use us for today. So that's where we are today. So over the next 5 years, we're going after a very large market that the TAM is enormous, and there's really kind of 3 priorities for us. One is moving more up market because we think that's where the best returns are, and we're seeing that in terms of the returns that was from the different channels. And also, I believe that AI will be more seen at the high end of the market. Unlike other platform shifts, you'll see the AI workloads emerge at the high end of the market versus the low end of the market. The second thing is that we think that there's a unique opportunity given with the sophistication of these cogeneration tools to really reduce the tax and the cost and the time and the risk of migrating these legacy applications. Customers are saddled with thousands and thousands of legacy apps. The technical depth is very high. The cost to run the management and apps is very high. There's end-of-life issues. And frankly, people also want to AI enable these applications. And so there's a confluence of factors that are getting customers to really be much more receptive to now modernizing these applications, and we're seeing a lot of interest on that front. And the third step would be for us to really become a core ingredient of the future AI tech stack. And we think that architecturally, we are well positioned to do that for inference workloads, and that's an important distinction, because one for -- for AI workloads, there's even more requirement to be able to query and manage complex rich data structures, which are designed to do -- you need a lot of flexibility and agility in your schema as data is always changing. And you need the performance and scale of a natively distributed system, which is what MongoDB is. And so from that point of view, that's kind of how we think of ourselves going forward.

Not that you're going to give a $10 million long-term aspirational thing?

No.

And Michael is looking like, no, that's not going to happen.

We tend to speak quietly and carry a big stick.

I like that. I like that. I forgot to mention, but Dev, I met you in 2006 or 2007 when we worked on your first company IPO...

That's correct.

Back then, I distinctly remember how you explained provisioning technology and how -- clarity of thought and clarity being able to explain the strategic vision has always been a very strong thing with you. Michael, shifting over to you. We're looking at -- we're already starting to look at calendar '25. When you talk to customers, what's on their mind as far as IT priorities, budgets?

Yes. So a few different thoughts. Everyone is focused on AI and today's point, trying to figure out how to get value out of AI, right? But there's also a lot of care and feeding that's sort of the normal course business has to take. And so one of the beauties of being a relatively small share player in a very large market is we're not particularly dependent on the underlying IT trends and spend environment. We've been able to successfully win kind of new business regardless of the macroeconomic conditions, regardless of IT budgets. And so I don't see any reason why that should change. The image that I often use for people as you think about the ocean, there may be a lot of froth at the top, there may be choppiness in terms of the waters. It may be high tide, it may be low tide, but that doesn't matter if you're at the bottom, and you're a 2% share player in a market measured in the many tens of billions of dollars. And so I do think that people are focused on getting value. I do think people are focused on driving efficiency. I do think incrementally, people are looking to turn the opportunity into impact within the organization. And so it creates a lot of exciting opportunities for us. But again, despite our relative size, the underlying IT spending environment isn't a hugely important factor for our ability to succeed.

Yes. It looks like the AI thing is real. Your customers are talking about it, you're talking about, we're talking about it. Just making sure. We lived through the dotcom buzz, I mean, it was not that good. So there's a natural skepticism, is this thing really going to work? And it takes a lot to...

It will take time, but...

Yes. Yes. So on that 2% share, Dev, you've got 50% logo share, which is incredible, the Fortune 500 including Goldman, we use your product, but only 2% share of the database spend. What could you do to unlock greater share? And in particular, what -- if it's a logical part of that question, I think the strategic accounts program is an important lever to be able to unlock that growth. Where are you today in that program?

Yes. So it's important for investors to understand, unlike other businesses like say a Salesforce or a Workday, that tends to be a top-down decision where you make that decision and everyone standardizes on either your Salesforce automation platform or your HRIS platform. It's not like marketing will use one HRIS platform and engineering will use another, that makes no sense. And our business is very different. It's -- the unit of competition is the workload. Like if are you building a new app or you're considering replatforming the existing app, the development team and the architects need to think through, okay, what is the new tech stack I'm going to use to build that application. So we're always trying -- even though we're -- a customer like Goldman may be a customer, we still have to go win that next new workload or that next new use case that they're planning to deploy. And so we win this business workload by workload. So in some ways, we don't get the big bang of like everyone's standardizing, you have some large multimillion dollar deal upfront. Our new business, the workloads start small. But as you acquire more and more of them and they start growing, that's when the growth really starts kicking in. And so that's kind of the buying behavior dynamics of our business. In terms of like what we've called out as our strategic accounts, what we have seen is that there's been accounts where they've grown over time, they've started to -- we negotiated the services agreement, the cloud services agreement. There's strong champions on the count, we said, why don't we deploy more resources and see what kind of returns we get, and the returns have been very, very strong. And so we are now trying to get -- repeat that same motion in more and more accounts. And another financial services firm, not Goldman, literally, we did this with them about 18 months ago, and they were probably the low 7-digit figure spend. And now they're close to $20 million a year spending with us in 18 months. So that kind of -- not saying that every account grows that way, but those are the dynamics of these accounts where once you penetrate and bring a lot of resources to bear, you can grow your share of those workloads and grow your business in those accounts very, very quickly. And so that's -- we're trying to do that now, not just in North America, we're doing that in Europe and also other parts of the world where once we see an account at the tipping point, we'll bring a lot of resources there, and that has a disproportionate return on our investment. So that's an important thread for us going forward. And what we're seeing actually, app modernization using AI is actually causing more and more accounts to be kind of put in that bucket because when we see the estates, the legacy estates that these accounts have and their real high interest in kind of working with us, we said, wow, these accounts could end up being the strategic accounts in the next 6, 12, 18 months.

I think Dev's tipping point concept is really important for investors to understand, because a lot of times we do meetings and investors say, you seem to be having a lot of success in the strategic accounts, why don't you just have more of them, right? But they're not successful because we call them strategic accounts, right? They're successful because the necessary conditions are a precedent. We've learned and experimented and integrated to get more intelligent and better from a capital allocation standpoint about that. And so to Dev's point, the goal is trying to get more and more of those accounts kind of ready or ripe or however you want to think about it so they are closer to that tipping point, so that when we do make that incremental investment, we can get appropriately rewarded for that. But just simply calling an account a strategic account doesn't actually suddenly make it more productive. And so I think that's something that investors sometimes don't understand what happens underneath the surface.

I wonder if there's some way you can feed this into a GenAI prompt and say, these are the characteristics that made this a strategic account. So how can we harness the nonstrategic accounts to become...

So we actually have a process where we think we can simulate those accounts that aspire to be in the strategic account, that we say, okay, because we obviously are forecasting where these accounts will be over the next, say, 12, 18, 24 months, so we can start seeing the signs of this account A, this geography could be a potential strategic account. But here's the science and here's the progress you want to make for us to deploy even more resources.

I hadn't planned on asking this, but maybe it's a natural question. Do you deploy GenAI inside MongoDB to get efficient at certain things?

Yes, we do. Like many, we've been eager to test and experiment. We've had a fair amount of success internally, a lot with sort of making people who are customer-facing more context aware. So case summarization, sentiment, things like that. We recently rolled out a next iteration of some of our kind of customer-facing knowledge base with generative answers in that. And that further helped them kind of self-serve and not have to open up underlying support tickets in a way that's actually been super helpful and obviously leads to both higher customer satisfaction, but also greater efficiency in terms of how we serve those customers. So absolutely.

Okay. So Dev, maybe bringing the discussion to more recent trends, the hyperscalers have put up solid numbers in recent quarters. And it appears to some extent, their database offerings are benefiting from the push towards generative AI. So we'd love to get your thoughts on whether this comes at the expense of MongoDB. And if not, why?

Yes. So when you -- it's important again for investors to understand, there's 2 dynamics in our business: there's the macro dynamic, which you see in the consumption of the existing workloads for Atlas, because that's now 71% of our revenue; and then there's new business. And our new business -- except for Q1, where we admittedly had some operational issues that we've since resolved, our new business has always been quite strong. So I think it's a function of, one, a big market, very compelling value proposition and the fact that deals start small, so you're not dealing with a lot of resistance of spending, like make a 7-figure expenditure. So our new business has been quite strong. So our win rates against the hyperscalers have been very high. And candidly, the hyperscalers see that, and they also partner with us. Our relationships with Amazon, Azure and Google have never been better. And we're all adults. We know that there are times when we partner and times we compete. And frankly, all 3 of them also joined our accelerator program, our MAAP program to really help the customers get comfortable with reference to architectures and proof points and integrations to get started quickly with AI. So the short answer is, no, we're not seeing any impact from the hyperscalers.

And Michael, Dev made reference to the strength of the new business. And I think that's been a really palpable part of the MongoDB story through kind of a weaker macro environment. So if we just focus on the expansion piece of the growth algorithm, like what's happening underneath the hood from a consumption perspective, whether it's the digital natives or the enterprise opportunity? Could you speak to that?

Sure. The couple of times that we've tried to call out the macroeconomic environment, what we're really describing is the underlying database activity. So things like reads, writes and transactions at the database layer. One of the benefits of being a general-purpose database is we've got use cases across industries, across geographies. And what we see, we call out the macroeconomic effect as we did in the Q1 call is less growth, slower growth in the underlying read and rights at the database layer. And we saw that in a broad-based way. So to your question, it wasn't some particular impact of geography, of digital natives, of anything. It was really quite broad-based and across the board. And we saw, again, in Q2, despite the results, we were pleased with the results. We think about the consumption growth as week-over-week growth in consumption. That consumption is the actual -- what we get paid, but it very closely mirrors the underlying usage. And so we saw as -- we saw that grow slightly slower on a year-over-year basis sort of pointing to this slower macro environment that we talked about in Q1. That said, we did slightly beat our own sort of forecast for Atlas in Q2. But as we described, it was really kind of within a reasonable range of alternatives. So there's nothing that would suggest the macro is getting materially better or materially worse kind of quarter-to-quarter.

Very helpful. And maybe just kind of shifting the discussion a little bit to kind of EA versus Atlas, right, it seems like EA continues to perform incredibly well. You talked about really solid pipeline into the back half of the year. So if you could just talk about kind of those customers that continue to kind of commit to Enterprise Advanced versus kind of the Atlas opportunity for Dev or Michael.

Yes. Maybe I'll start and you can -- so Enterprise advances, just again for people new to the company, it's essentially a self-managed product. You get our software and you provision, configure, manage MongoDB yourself. And our belief was that once people got more and more comfortable with Atlas, there'd be less and less people who wanted to do that themselves. Because part of the argument of Atlas is there's lots of undifferentiated labor in terms of all that management that you have to do, self-managing the deployment of MongoDB. What surprised us is the staying power of EA. And what's become clear is that people want -- especially at the high end of the market, want a hybrid world. There are some workloads that run on-prem, some workloads are run in the cloud, some works that may move to the cloud, bring back or vice versa, start on-prem and then move to the cloud. And so given that, we have changed our posture and now we're investing more in EA, first, through our community product, which is a free version, where we're now rolling out -- we'll soon roll out Vector Search and Search because it also helps developers who are starting with new product to just get started faster on those newer capabilities. And then that will get folded into probably a year later into EA. So you can see us invest more in EA. EA today is still more of an upsell to existing customers than driving net new customers, but that may change as -- over the next couple of years as we invest more in EA. But again, the long and short of it is we've been pleasantly surprised of the staying power of EA. And that was something 5 years ago, I would not have predicted.

Model transitions are always very hard to pin down, unless you rock the product that you want to sell less of. In maintenance mode, no more enhancements and sunset mode, that's not something that you would want because it's -- as a product, that's a really, really good product. Shifting to rate cuts, we just hosted Jan Hatzius, our Chief Economist. He talked about 25 bps, 50 bps maybe less probable. But how important is that as you look at your business and you get through multiple cycles? Could this be a tailwind or not? What are your thoughts?

So a few different reactions. I think to the earlier comment, if you think about our business and the ways in which macro economy affects our business, there are 2 possible ways: one is in terms of winning new business; and then the other is the underlying consumption growth in existing workloads. To Dev's earlier comments, we've been very effective despite different macroeconomic environments, winning new business. And so I don't think it's particularly a necessary condition or a particular tailwind there. We have a very large market, we have a great product, we have a small share, we have an excellent team. So that all adds up to, kind of quarter in, quarter out, a pretty steady execution. In terms of the consumption growth of existing workloads, certainly a better macroeconomic environment should help with the underlying reads and rights. As much as we've tried, I can't walk you through the 5 macroeconomic factors and their respective coefficients in terms of the -- what that's going to predictively mean in terms of underlying Atlas use. I think the last thing that I'd say, and I am not an economist, but is -- I think that part of the reason why the Fed is talking about cutting is to avoid more negative environment. And so I guess the absence of a negative would be a positive, but it's not clearly more positive than where we are today, it's just sort of trying to prevent a future mess.

In fact, Jan said that a year out, he expects rates to be -- a year plus out, 325 basis points, which is a big reversal from where we have been so if he's right, that's hopefully more optimistic. Dev, you've been through multiple tech cycles. Is GenAI overhyped or real? My voice is louder this time. Did we crank up the volume? Thank you.

It's the AI overlord. What I would say is -- so that's Kasthuri's way of saying, Dev, you're old. So you've seen some of these platforms before...

So am I...

And so I do believe that AI is not a question of if, but when. And actually, I've had this conversation in the last meeting. I said, I view the world we're in today, circa 1996, maybe 1997. Netscape was launched a couple of years earlier. People all get excited about the web, but the web was still very basic static web pages, and it wasn't that interesting. People were starting to build businesses on the Internet, maybe Amazon and a few others, eBay, but you weren't seeing this plethora of kind of companies exploding on the market. And I think in some ways, we're kind of at the same stage in the AI era. I think it's that old saying, people always tend to overestimate the impact of a new platform or technology in the short term but underestimate it in the long term. I think one of the things, when I look at the use cases, I typically see 3 patterns of use cases with customers today. One is chat bots. We haven't even talked about that. I think what you're going to see is -- the next version of chat bots will be embedding real-time information, right? So if you are working with a chat bot, say, your financial institution is now using a chat bot to interact with, you need to be able to, as an institution, know the last transaction that person did. So you need to embed real-time data. So you need to have real-time awareness of what that customer is, if they filed a ticket, if they executed a transaction. They tried to put an order, and if something has gone awry, you need that real-time visibility. And I think you're going to see more and more chat bots get more and more sophisticated because you just can't use legacy data to make that work, or old stale data. The second thing I see is on the research and summarization. The context windows are getting bigger and bigger, so you get essentially more memory. So you can essentially push more information through these LLMs. And by definition, you can kind of work with larger and larger data sets. So I think at a point, you will -- the memory factors of LLMs will make them much more interesting. And frankly, that will also increase the switching cost to go between LLMs. Because once people understand who you are, what you've done and the history of your engagement with them, all of a sudden, you may not want to lose that history to go from, say, OpenAI to Anthropic to Llama, et cetera. And then the third area around automation and what we're seeing is that obviously, everyone's talking about agentic workflows and having agents and multi-hop agents do a lot of work. I think that's still going to come. But I see a lot of people looking to build much more sophisticated solutions around this area. And I think you're going to see, again, all that, that's going to be fits and starts. The key factor for all this is going to be the research breakthroughs for the LLMs. And I think a lot of people are saying, why aren't we seeing more apps in production? Why aren't we seeing more impact? Why aren't we seeing more returns from these investments? I think it's like an iterative process, where as the research teams at these firms get these breakthroughs, you could argue -- GPT-5 was rumored to be out in the spring of this year. And why is it still not out? A lot of people say it's because actually, black ball was late. And so part of the research breakthroughs is also reliant on the compute architecture. And so I think this is all iterative. As you kind of get -- go through those kind of iterations, you're going to start seeing the breakthroughs come, and that's when the opportunity arises. For MongoDB, a lot of people -- we don't get the "buzz factor" like a Databricks or Snowflake does, maybe more Databricks than in Snowflake because we are not in the world of training. Our -- where we come in is in the world of inference. So we will be a beneficiary as these as LLMs become better, they become lower latency, lower cost, better accuracy, less hallucinations and people get more and more comfortable, then they're going to deploy inference workloads. That's where we come in. Old app systems are not designed for inference, OLTP systems are. And we are -- we think we're well positioned for inference because of the complex rich data structures that we enable developers to manage in query. So we are looking forward to -- all this money being spent on LLMs to us is a good sign because people will need a return, the research breakthroughs will come, and ultimately, those applications will start getting deployed in production, and we think we have a chance to win more than our fair share.

What does that inference world look like, if you can imagine that?

Yes. So I think my belief is that if you're just building a thin wrapper on LLM, your days are numbered. And we've seen some of those companies fall apart already, right? I think where you can build an app that the LLM is an enabling technology, not the key technology, but enabling technology, you marry that with some unique or novel data set that you particularly have or you have insights on and you embed that with deep workflow into your business process. That's where I think you can build a meaningful AI application. I think it's not a perfect example, one example of that is Perplexity. You could argue Perplexity did not create their own LLM. They wrap their search functionality around both the LLMs and different data sets. And so -- and you can pick obviously different data sets based on the search that you want. And again, it's not a perfect example, but it's one example of what a future AI workload or application will look like.

Sticking with that thread on real-time information and inferencing workloads, it sounds like MongoDB is making a big bet on kind of the RAG architecture as kind of the primary enabling technology for a lot of these applications. So can you maybe talk about the distinction between maybe RAG, fine-tuning and kind of how MongoDB's Atlas Vector Search and stream processing kind of fit into this?

Yes. So a lot of the -- there's this big debate between fine-tuning and RAG. And people say, if you can fine-tune a model, you don't really need to use RAG. I'm not -- given the way the models are being built and the -- the way the models are kind of -- I guess, the performance of those models are kind of all gravitating to kind of one level, I think the way to distinguish yourself is actually through RAG. I think fine-tuning time will go away and RAG will be the future. And now there's this thing called advanced RAG, where you do very, very sophisticated questioning, use multiple data sources, you do iterative questioning loops and so on and so forth. So I think RAG is going to get much more sophisticated. And we do believe that embedding -- using vector embedding and then using a vector search functionality to do -- to marry data with what data is sitting in LLM will become an important tool in your arsenal. Because data is where you'll actually define your business logic. Like in the old world of SaaS, software is where you define your business logic. Today, now data is the way you define your business logic, and that's why data becomes so, so important.

So why not become an apps company if generative AI were...

If we were at 80% share of the database market, maybe we'd contemplate doing that, but we're far from it. We're ones to -- I think it's also very dangerous for a company of our scale to try and be all things to all people. And so we really want to focus on our coordinating and where we think we're well differentiated.

Okay. And what are the things, Dev, you've done from a you -- you answered this in part, but from a product perspective, go-to-market perspective, what are the things that you've done to retool the company for success in GenAI ?

Yes, I would say if you've tracked us since we've gone public, which is since 2017, one of the things that I think we've been very constant on is that, we believe to build a great software franchise, you need to marry great product with a great go-to market. And then the magic happens when you bring those 2 together. And yes, we had a stumble in Q1, and we still feel terrible about it. But historically, we've executed quite well. And I would say, a lot of other companies who maybe focus too much on product and not go-to market or focus too much on go-to market and have a very same product struggle. And I think the real -- the bear case against MongoDB, it was like, how can you go compete with -- become a general purpose database? No one believed us. There's no way a SQL -- no SQL database could become a general-purpose database. Everyone said that's not going to happen. We proved that happened. Our cloud business, everyone said there's no way you can compete with the hyperscalers. We proved that. And I would say our cloud business is the gold standard in terms of how other companies kind of look at what they would try and do. And so I think that's all a function of like marrying a great product and innovating very aggressively on the product side as well as strong distribution output, our sales force against anyone day to day in terms of our ability to win.

Excellent. Do a quick pulse check and see if anybody is brave enough to -- if you've had your coffee. Yes, I do see 2 hands. Go ahead. You can just shout it out and I'll...

Maybe if you can dig a little bit on the GenAI workload as you see it in the next year, 5 years. [indiscernible]?

Yes. I mean -- I wish I could give you -- be so prescient as to tell you exactly how this all is going to shake out, and I don't think anyone really knows. But one thing I do believe strongly is that the inference market will be far larger than the training market. Because by definition, infer once -- or I'm sorry, you train once, maybe train once or twice, but you're inferring all the time, right? And the R are -- there's lots of I going in terms of investment in these AI models, a lot of people are saying, where is the R, right? SQL AI came out with that blog, a load of people who said, there's so much CapEx being spent. The R has to come through the inference workload. Someone has to get value out of all these trained models. Training is a means to an end, not an end unto itself. And so that's why I think we're well positioned. It's taking longer. I think the difference -- I think, if you had asked me this question 1.5 years ago, I would never have thought how expensive or how much capital is required to build these models, right? Facebook and the hyperscalers say they're going to spend billions and billions of dollars every year on training and building out these models. It just tells you that it's a very expensive proposition. But the good news for us is that when those models are trained, if people feel like they're comfortable really using them in production, we're the beneficiary of those train models.

So you're basically saying slow to learn, quick to conclude. There's another question, was there? You can just -- yes, if we can get the mic over.

You touched on this -- all of these investors are trying to figure out who are going to be the main players. You talked about Mongo being a core ingredient. You touched a little bit on the hyperscalers. You just mentioned Databricks and Snowflake. Of course, there's [indiscernible]. Can you explain to us why Mongo in the inference world will be a core ingredient of AI versus all these other players? What's the difference? What are the distinctions of what you do that will allow you to be one of the winners?

Yes. So what I would say is this is the same question asked of us 7 years ago about why are you going to win the NoSQL game, right? There's Couchbase data stacks, the hyperscalers had their own NoSQL variants, and we ended up winning the NoSQL game. One is that the document model is a superset model. So you can do things like key value lookups. You can do things like more traditional queries, you can do joins, you can do time series use cases, you can do graph use cases. So architecturally, we are designed to be a general purpose database. The second thing I would say is that when you think about an AI workload, an AI workload requires a need to work with very rich and complex data structures. We are designed to work with rich and complex and query rich and complex data structures. JSON is the example of a rich and complex data structure, and we're a JSON database. A lot of people are saying, well, we now support JSON. But like a lot of these relational tools, they have to do very complicated and convoluted things like off-road storage for a very large object, like an unstructured -- like a video or a graphic and then the performance overhead with those kind of off-road storage techniques become very, very high. And then we are, by definition, a natively distributed system, right? So relational databases are designed to be single-node systems. So our most basic configuration is a 3-node replica set. And we have customers who have deployed not just replica sets, but sharded clusters where you spread the data across multiple nodes because of data volumes are so high. And architecturally, we are designed to scale massively. And then the last thing I would say is the flexibility of our data model where you start with a data model, but then it changes over time, there's nothing better than MongoDB. I mean that's one of the strongest suites whereas like other data models become very brittle over time and then it becomes harder and harder to change and add new features and new capabilities. So for all those reasons -- and then the last one I'd say is developer mindshare. We are the world's most popular modern database, right? Now I would say -- long term, I would say there's going to be probably a relational standard and what I'd call a modern standard, though some people would still want to stay in relational. And -- but I think for those people who truly want to modernize, we have become the default choice.

I think there was a question here. [ Ishan ], you had a question, right? Yes.

[indiscernible].

Yes. So as you -- we have 50,000 customers, almost like 60%, 70%, I forget the stat now of the Fortune 500 are customers of ours today. So we get pretty good visibility into large accounts. I recently met with a large financial services customer, not Goldman, who has roughly about 45,000 developers. I asked them how many AI workloads did they have in production. They told me 15. Then I asked them, how many of those are client-facing workloads. They said 0. And the reason for that is they're terrified of the hallucinations that's -- because with a probabilistic system, you can ask the same question 3 times and you get 3 different answers, right? And so they really need to get comfortable that they can put enough guardrails around these probabilistic systems or AI models to feel comfortable to start exposing that to end customers who make financial decisions based on the recommendations or advice they're getting from these models. So I think as the research breakthroughs happen, as these models become more accurate, as it become more performant, as they have lower latency, et cetera, people slowly start becoming more and more comfortable with deploying these apps on production. But I think it's an iterative issue of like the research breakthrough is happening before people start deploying these en masse. The area that we're seeing the most interest in is app modernization. Because the biggest problem customers had is rewriting the app code to modernize the legacy, say, side base Oracle app to a more modern architecture. With these cogeneration tools, they can do 3 things: one, analyze existing code because by definition, a lot of these development teams for these old apps don't even exist. So you can understand what every line of say, 1 million lines of code does. Two, you can reverse engineer tests so you can then say, okay, with this input, I get this output. And then three, you can reproduce the code in a new, more modern language, and then use the same test to make sure the input matches the output. Because no one's going to cut over to a new system if they do not guarantee that the system works like it did before. So that's where we're seeing a lot of interest, because people say, like, finally, I have a chance to really modernize my legacy estate.

On that note, why don't we wrap it up? Thank you so much for your time...

Thank you.

Thank you for your attention.