Good morning, everyone. Welcome to the conference. I'm Pinjalim Bora, a software analyst at J.P. Morgan. I cover mid-cap. I'm delighted to have here with me, Tom Siebel, who's the Chairman and CEO of C3 AI. Tom, welcome to the conference.
Thank you.
Let's start with a brief introduction, maybe about yourself and maybe a little bit about C3 for the audience who might not be familiar with the story.
Okay, I'm a computer scientist. Did my graduate work in relational database theory. Went to work for a little start-up with about 20 people called Oracle Corporation. That turned out to be a pretty good idea. I was ultimately one of the guys who ran that business, and we were involved in the, in the commercialization of relational software, relational database software, and then built a large stack of enterprise applications. About a decade later, I spun out and thought about... I thought about the application of information technology and communication technology to sales, marketing, customer service. That was a market that was largely unserved by information technology at that time. We founded a company called Siebel Systems in 1993, July of 1993. That also turned out to be a pretty good idea.
So six years later, we were doing about $2 billion in revenue and had, I think, 4,500 customers in 29 countries, and that, I believe, is the fastest-growing enterprise application software company in history.
Mm.
In 2006, that company was acquired by my colleague, Larry Ellison, and became the basis of their CRM stack. So I invented the CRM market as you know it today. This year, I think, a $120 billion business. And then we thought about what was next, and we believed that next was going to be about elastic Cloud computing, big data, Internet of Things, and predictive analytics, what we call enterprise AI. So we began work in 2009. In the next decade, we spent about, well, really, $2 billion building a software stack. And the purpose of that software stack was to provide an integrated platform that provided all the services necessary and sufficient to design, develop, provision, and operate very large-scale enterprise AI applications.
We used that stack to build, today, 90 enterprise AI applications that address the markets of manufacturing, agriculture, pharma, oil and gas, utilities, defense, intelligence, and then in the last couple of years, we've expanded that very significantly. Those 90 applications include, I believe, 30 applications that address some of these opportunities associated with generative AI. So I've been talking about enterprise AI now for some time. Understand, we started this effort, you know, when AWS was this big, before the Azure Cloud, before the Google Cloud, before the GPU, and, you know, now we find ourselves in, you know, 2024, and the rest of the world has discovered enterprise AI. So the world has kinda come our way, and it looks like the addressable market opportunity...
So I've been talking about this for more than a decade, when others were speculating about whether or not there would be an enterprise AI market. I can assure you, there is, and there will be, and this is not ephemeral. So as we address this market opportunity in, you know, 2024, 2025, 2026, 2027, what have you, we're looking at what is going to be certainly greater than a $1 trillion addressable market, maybe a $2 trillion addressable market. There is one company in the world with 90 applications that address this market opportunity. There is one company in the world. And, so the game we're playing is to see if we can establish and maintain a market leadership position globally in enterprise AI.
Now, that sounds like, you know, a little bit ambitious, but in fairness, we said this at Oracle in 1983, and we said it at Siebel in 1993. In CRM, we did have 85% market share in sales, marketing, and customer service when that company was transitioned to Oracle in January, I think, of 2006. And, you know, we might fail. We might have to end up being number two or number three. And so, you know, that, so that's the opportunity that's before us.
Yeah, that's a great overview. You obviously have been doing it for a while. But I want to ask you the inflection that we are seeing on the infrastructure layer right now, right? I was looking at the numbers. I think NVIDIA added something like $30 billion in revenue in the last four quarters. Obviously, revenue up over 100% year-over-year. Huge inflection on the infrastructure layer. But we're not seeing kind of a similar inflection on the software stack at this point. We have seen the hyperscalers accelerate slightly, but still seems like there's a big gap. So I want to ask you how do you see this play out, right? The massive build-out on infrastructure, how does it impact software of the stack?
Well, there's only one reason for that silicon and for that infrastructure, and that's to run applications. Okay, so obviously, somebody thinks there is an application opportunity. If we look at that stack, at the bottom, we have silicon. Above that, we have infrastructure. Above that, we have foundation models, and above that, we have applications. Now, the way that these markets work, you know, when we came out, when the PC came out, let's think, as we think about the IBM PC, think the 1980s. You know, and you had that PC on your desk, that Compaq PC or that IBM PC. Okay, the whole value stack was in the silicon and the infrastructure, and you had, you know, a $200 VisiCalc application on it, okay?
Today, that PC that's on your desk costs you or your company about $200 a year in cost. The infrastructure costs you about $200 a year, okay, and, you know, by the time you depreciate the cost. Okay, and the applications that you have running on it, Bloomberg, whatever it may be, that's, you know, costing you $10,000 a year. If we're looking at this guy that came out, you'll recall, in 2007. When this came out, you know, 100% of the value was in silicon and infrastructure. Today, where's the value? It's in the applications that run on it. Same thing will happen, okay, in enterprise AI. In the long run, the silicon gets commoditized.
I'm not saying that NVIDIA isn't a great company, 'cause NVIDIA is and will be a great company, but silicon does get commoditized. Hard stop. See Intel for details. And, the infrastructure gets commoditized, and 70% of the value will be in the apps. Now, the idea you have companies like, you have enterprise application software companies in AI growing at 15%-20% compound annual growth rates. Now, then I cut it nearly where I know you guys grow a lot faster than that, and you're like but that is the top 10% of the software universe. So I'm not certain that, you know, of your software universe that you cover at your company, when you get north of 20%, that's the top decile of growth rates. So I think that, we are seeing accelerating growth.
Yeah. So basically a matter of time that it permeates into the software layer, and, you're seeing it already happen.
Come on, it's a $1 trillion, $2 trillion addressable market. More than 50% of that, of that stack is going to be applications.
Yep.
That just happens in PC, it happened in a minicomputer, it happens on the phone. I mean, this is the way the world works.
Yep, got it. Let's talk about C3 AI, the enterprise AI platform. Maybe talk about... You have, you have some big brands that use C3 AI, see real economic benefit. I think you've quoted over $1 billion of economic benefit. I think Shell, one of the companies-
Shell, $2 billion a year.
Two billion.
Yeah.
Maybe at a core level, help us understand the main value that C3 AI platform offers to customers. Why is it so difficult to scale enterprise AI? What does C3 AI provide in that journey to scale?
We make it simple to scale enterprise AI. It is difficult to scale enterprise AI because people, for decades, have been trying to build these things out of piece parts. You have, you know, you have IT people that some guys consider themselves. They get a little bit confused, and they think that they're software engineers. Like, man, they know how to install, you know, single sign-on, or they could maybe, with enough—if they spend Accenture—you know, spend enough money with Accenture, they can figure out how to get SAP installed. Okay, that, that does not make a software engineer, okay? But every now and then, they get confused, and they think they are.
So they take all these piece parts from used to be the big rage was the Apache open-source stack, and then, you know, all the piece parts that are offered from the infrastructure providers and other things, and try to cobble them together into something that works. Nobody ever succeeds at that. Now but people don't want to buy tools to build applications. People don't want to buy tools to build CRM applications. They want to buy CRM applications. Okay, I proved that. People don't want to buy tools to build relational database applications. They want to buy relational databases. We proved that at Oracle. They don't want to buy tools to build ERP applications. They want to buy ERP applications. See SAP for details. Same thing in enterprise applications when we get into enterprise AI, what are the big...
Predictive maintenance, stochastic optimization of supply chain, demand forecasting, fraud detection, whatever it may be. Companies don't want to buy tools to build these things. They want to buy applications that they can install and optimize their supply chain. So we make it easy. So we bring our customers live in, you know, whether it's a Dow or whether it's a Cargill or whether it's a Tyson, we bring these applications live. Complex enterprise applications that we bring live in order of six months at the enterprise level, and then they pay consumption pricing afterwards. So it... That's the value of what we do, is we make it easy, so you don't have to spend, you know, years and hundreds of millions of dollars trying to build it yourself.
So, you're saying you're basically stitching the different parts of enterprise AI scale, right? That you have the data prep side, you have the data engineering side, you have the feature development, you have the model training, and then you have the observability side. If you look at the entire stack, and people can use different tools on their own, but C3 is providing one platform that's-
Yes, we provided a platform that you can think of as an orchestration layer. It's $2 billion worth of software engineering. It's quite a piece of work. Now, all of the various components are interchangeable, whether we're dealing with persistence, whether we're dealing with nFactor authentication, whether we're dealing with, you know, whatever generative AI tool you might want to use. You might want to use Anthropic, you might want to use Llama, you might want to use ChatGPT-4, and you just plug it into that box. But, you know, out of the box, it comes with a generative AI tool. Out of the box, it comes with, you know, relational database persistence. Under the box, it comes with a key value store.
It supports various, you know, supervised learning, unsupervised learning, deep learning, but all of those components can be interchanged with open-source capability or proprietary capability, but it works.
Yep, understood. The C3 AI, you said, 30 gen AI applications already. But from the basic level, maybe help us understand how does the C3 AI part differentiates versus some of the other stuff in the market. Is that also, again, trying to ease-
Okay, for those of you who are interested in generative AI, and I've been looking, and that would be everybody in the room, okay? And these large language... I've been looking for a good book on this subject for a year. I finally found one. It's called, you know, What is ChatGPT, okay, and What Is It Doing? by a guy named Stephen Wolfram. You can get it for, like, $19 on... This guy, Stephen Wolfram, if you know him, this guy is the real deal, right? Very highly respected member of the academy, physicist from Caltech, WolframAlpha, Mathematica. This guy is the genuine article, okay? Okay, and in 91 pages, you read these 91 pages, it'll take you 45 minutes, you'll know no more about these language models than anybody in your organization. Okay, and it...
I'll give you. If you don't like it, send me an email. I'll give you a money-back guarantee. Okay, I'm literally printing 5,000 copies of this just to give it away, okay, so people understand. Now, these large language model, how big, how big is this generative AI market? Holy moly, guys, this is huge! Okay, but there are problems associated with these things. What are the problems to the extent that you've used them? Okay, and if you're not using them today, how many people use generative AI every day? ... Guys, you gotta use it, okay? You gotta use it. Download a copy of Microsoft Edge, okay? It has ChatGPT-4 Turbo with it. You gotta use it, okay? And start using it, okay? And because if you don't, your competitor will, okay?
They're gonna be doing better work than you. It is, what this thing does is amazing. But what's the problem with all these learning models, with every one of them, okay? Number one, the answers are stochastic. Every time you ask an answer, you get a different question, okay? They don't enforce, you know, other problems are, you know, we can't. When the answer comes up, you can't tell where it came from, okay? They're not. Even with retrieval-augmented generation, with these RAG architectures, that'll work for unstructured data, but that won't work for structured data. Say, it came out of ERP system or a CRM system or any sort of structured data. You can't figure out where the answer came from. None of the enterprise access controls are enforced.
Hey, Bank of America, JPMorgan Chase, CIA, name the company, you have access controls, and, you know, the CEO doesn't get to the guy who works on the production line doesn't get to say the same thing as the CEO. Well, enterprise access controls are not enforced. We have enormous problems with the cybersecurity risks that are introduced by these large language models, which is why at companies like JPMorgan Chase, they don't let you use ChatGPT, okay? Because you know, the cybersecurity issues are daunting, and we have these data exfiltration problems that you've read about with Samsung and others. This hallucination problem is just wonderful. If it doesn't know the answer, it just makes up something, it makes up a story that, I mean, these are just wonderful.
By the way, if you read this book, you'll understand where these hallucinations come from. There's enormous IP liability associated with these, okay, with these language models, because they're trained on information over which somebody else owns the copyright, be it the Weather Company or Bloomberg. And by the way, they want their piece of the action, okay? Finally, they're LLM-specific, and anybody who bets on any LLM, say, in May of 2024, has lost their mind. I mean, with all of the innovation that's going on in this space, with these guys leapfrogging each other, I mean, there's just no telling where these go, and you want to be able to change LLMs, you know, maybe week over week as these guys out-innovate each other.
And by the way, there's some assumption that one of Google or OpenAI or, or Anthropic is gonna win this. Why would you think that? I mean, OpenAI could be out of business tomorrow. And was anybody, like, read the newspaper last Thanksgiving? It almost went over, in one weekend, it almost went out of business. These, any one of these could be gone. So when you combine the large language model with the $2 billion worth of platform. Oh, the other problem is they're generally unimodal. Unimodal means you can put in text, or they say they're, they're multimodal. Maybe you can put text and images. Guys, text and images doesn't cut it. You need text, you need images, you need, you know, you, you need telemetry, okay? You need all structured data, unstructured data, so it's gotta be omnimodal.
Now, remember the $2 billion worth of work that we did in building the enterprise platform? I mean, we've nailed identity, we've nailed access control, we've nailed cyber. Okay, we've nailed omnimodal, so we basically solve all those hurdles. That dog I described doesn't hunt, okay? Whether it's coming from Anthropic, whether it's coming from Gemini, that dog does not hunt in any reasonable enterprise. But when you combine it with the C3 AI platform, we've solved all those problems. So it's secure, it doesn't exfiltrate, it gives you the same answer every time. You can figure out where the answer came from. You don't have IP liability problems. So we're installed today at CIA, NSA, NRO, National Reconnaissance Office, the United States Air Force, United States Navy, United States Marine Corps.
We're installed in some of the most secure installations on the planet Earth because we've solved those problems, and that's what distinguishes our applications from others in the space.
Just to be clear, the CIA, NSA, NRO, all of these are using GenAI to do-
I did not say NSA. I said CIA.
Okay. These are using GenAI applications to-
As I announced on the last conference call.
Okay.
No news there.
Okay, got it. So what has been the feedback so far from these customers on the GenAI use cases, the GenAI applications that you have?
Oh, I mean, it's, I mean, it's just the productivity increases that people get from these are just amazing. Let's take a law firm, for example. There's one law firm that is that all of you know, that takes a lot of companies public, that you guys are following. And what we've done there is we've taken the... Think about, it's a small language model. Think about an enterprise language model, okay, and what we're... ELM, okay? And what we've done is we've loaded, for this law firm, the corpus of SEC.gov, EDGAR, okay, into a language model. So this is every S-1, every 10-K, every 10-Q, and that's what we've used to train the language model.
So when they want to take the next company public, when you guys are, you know, all these enterprise LLM companies that are being financed at, you know, $1, $2, $3, $4, $5, $6, $7 billion valuations today, that you guys will be taking public next year, okay, when, when, when you, you need to write the S-1, you put in the name, you put in the financials, you put in the first few risk factors, hit the carriage return, and, you know, 45 minutes later, you have 200 pages of the first draft of the S-1 written. That's two weeks' worth of work done, like, right now. I mean, what's the productivity increase? It's huge. So those are examples.
Yeah, wow, that seems, like, not that great for lawyers.
No, no, I don't think it replaces lawyers. It makes them more productive. I mean, they're just not doing this drudgery. I mean, why... you know, how hard is it to develop the proper list of risk practices for that company when you can look at the universe of risk factors that have ever been registered at the, you know, to the SEC and select those that apply? I mean, it... So I think it doesn't replace any jobs, but it makes these people certainly more productive.
Yeah. Let's switch gears and talk about macro. What is your sense? I mean, I'm still surprised that we're talking about macro, by the way, but it started in like June of 2022, but here we are. What is your sense of the spending environment at this point when you're talking to people, you know, can you characterize kind of the demand environment or the business confidence?
In Enterprise AI?
Enterprise AI overall from your customers.
So I can't talk about Silicon, I can't talk about anything other than enterprise AI, 'cause that's all that I do today. But as it relates to enterprise, I mean, at JPMorgan Chase, there is no budget for AI. It's whatever Jamie wants to spend. Okay, you think he's constrained by a budget? You think your people out here are constrained by budgets in AI? You think they're constrained at Bank of America by budget? You think they're constrained- I mean, nobody's constrained by budget as it relates to AI. These are CEO-level decisions, and he or she makes the budget, and it happens. So you're looking at a... These people at TSMC and NVIDIA are not building all this infrastructure for the sake of, you know, having big, you know, warehouses full of silicon. Okay? These are- these things are running learning models.
These learning models are supporting enterprise AI applications. So I think that, you know, this market is developing at a very rapid pace.
Well, but specifically with C3 AI and the demand environment around, around C3 AI, how do you characterize that?
Well, I think it's pretty good. I mean, what was the subscription growth rate year-over-year last quarter? 23%?
Twenty-three, yeah.
Twenty-three percent?
Twenty-three.
23%? I mean, that's not too bad. I mean, 23%, that means you're just at that rate, up from... So we changed from a subscription-based, subscription-based pricing to consumption-based pricing over a period of about 10 quarters. Our growth rate went from order of 50% year-over-year, down to negative, and in the last four quarters, it's gone to, you know, from zero to 11 to 16 to 18, and so we're seeing accelerating growth, and, you know, once you- so, what's not to like?
Yeah, no, I agree. The sequential growths are good, but the 23% obviously has an easier comp from last year, right? So you're comping the easier numbers, and you're accelerating growth. But the sequential numbers are also improving, which kind of is your point, that your consumption trend, the transition is working right now. And we shall see how it goes through fiscal 2025, since we can't talk about it at this point. But I was just trying to ask you about the normal kind of activity levels from your customers around C3.
Well, I think that last quarter we announced what? Let's see. Subscription revenue was up 23%, total revenue up 18%, customer count was up 80% year-over-year. We have $750 million cash in the bank. Things looked pretty good.
Yeah. Okay.
Could be worse. I know that, you know, right now, you know, for those of... You know, everybody's pounding on the table that this year, you know, right now, everybody's got to be profitable. You need to be profitable right now. Hey, we have a structurally profitable business. What does that mean by structurally profitable? Okay, our gross margin is greater than our cost of selling stuff, okay? So I could throw this thing into profitability tomorrow. Okay, would that be in the best interest of the shareholders? No freaking way! I mean, what do you mean? I went out and raised $1 billion in the market in December of 2020, as I recall. I think we have rough numbers, three-quarters of that left, okay?
And the idea, the purpose that we did it was to invest in market share, okay, to invest in brand, and guys, that's what we're doing. Now, I know that, you know, we have our mood swings in capital markets, where all of a sudden, everybody's got to be profitable right now. But I mean, let's put this into perspective. Anybody remember how many years it took for Amazon to be consistently profitable? Anybody remember? That would be 29 years. Okay, how'd that work out for the investors at Amazon? What's the market cap at? Well, roughly $2 trillion. Worked out pretty well. How many years did it take for Marc Benioff to get Salesforce profitable? I was an idiot at Siebel. I mean, we ran a consistently cash positive, profitable business from the day we shipped a product. How stupid can you get?
Okay, what did Marc, how long did it take Marc to generate a profit? Anybody know? Okay, 25 years. 25 years. Not quarters, 25 years. Okay, how'd that work out for him? He invested in market share, invested in brand. How did that work out for his investors?
Pretty-
What's the market cap of Salesforce?
$200 billion.
About $250 billion. It worked out pretty well, right? What's his market share in enterprise applications? Pretty damn good. Okay, how many years did it take Apple to generate a profit? That would be consistently. That would be a quarter of a century. So I know that right now, we're all thinking in weeks. Guys, this is a $1-$2 trillion addressable market. This is the largest addressable market, okay, in really, arguably, that we've seen in the history of software, and I've been there since the beginning, okay, of enterprise software. I have been. And right now, we are investing in this market, okay? And honestly, we're probably not investing enough.
Got it. Let's, let's talk about the transition itself. I mean, you had a lot of assumptions when you started, right? 70% conversion rates on pilots, the number of deals that reps would be closing, all that stuff. It seems like you're faring better from at least last quarter, what you were saying, some of those assumptions-
I think I said, I have, I've been reporting, okay, for some quarters. We're exactly on track with what we said. Okay, the number of pilots is growing, the conversion rates is about what we predicted, and the consumption is almost identical to what we predicted before the fact. It's actually uncanny. But when a pilot converts, we'll see that, you know, in the first quarter that they go live, they'll generate - they'll consume on the order of 400,000 CPU hours. In the tenth quarter, they'll consume about 1.5 million CPU hours. And so it kinda. And then it just kinda grows over that entire period of time. So it's kind of exactly as predicted.
What is driving that pilot conversion of 70%? We talked to industry contacts, and they basically say every AI project, 75% of them fail, and 25% goes into production.
No, no, no, no, no, no. 99% of them fail.
99%?
Yeah, no, not 20-something, 70% of them fail.
Okay.
99% of them fail, okay? What is driving the conversion? When we're doing, let's say, going into Dow, we're going into Cargill, we're going into wherever it might be, ExxonMobil, okay? And, you know, our value proposition is, we will bring the application live, okay, in six months. With the little turnkey enterprise application live in six months. So that pilot is $500,000, and, I mean, I assure you, we're six months later, that application is live. And so what's not to like? Application's live, they convert. And so they're not still trying to piece it together. The other 99% that are failing, people are trying to cobble it together from DynamoDB or SageMaker or the Apache open source stupid Hadoop stack or one of these toolkits.
Yeah, that seems like a perfect segue to the competitive landscape. You have a lot of applications prepackaged that you have created, probably more than the others, but others do have some applications, I think-
Who has... Anybody have 90?
Numbers, I'm not sure of, but-
No
... Amazon has some, I think predictive or some, I remember a few-
Really?
... probably not at that scale that you have.
I don't think so, but that's okay.
Okay. But how do you-
Amazon, a great company. I sell with them every day. Okay, they're a huge partner of mine, AWS. I mean, Matt, you know, Andy Jassy, Matt Garman, we sell with these guys every day, and our stack is entirely complementary to them. They're not, they're not competing with us. Same, I sell with Google Cloud, I sell with, with Microsoft, so I don't think they have competitive applications.
So you're not competing with SageMaker or Google Vertex AI or-
Oh, no, no, no. No, we are actually. I would say.
Okay.
I would say not for the application.
Right.
Okay, they're building-
Okay
... we're building... So I would say our competition isn't SageMaker. Our competition is the CIO, and he or she thinks that he can use SageMaker and another hundred tool components to try to build an application that will do customer churn or stochastic optimization of supply chain at Volkswagen, okay? And so that's the competitor.
Yeah.
But when... Understand, when whoever it is, whether it's Koch Industries or whether it's Dow, is using the C3 AI applications with the AWS Cloud, it's making use of SageMaker, it's making use of S2, it's making use of all the componentry that they provide. So they're entirely complementary.
Yeah, I see. Let me stop there and see if the audience has any questions.
Hi.
Hi, I was wondering... First of all, thank you. My name's Kathryn from Durable Capital, and I was wondering, going back to where you're seeing just no-brainer productivity increases, you talked about creating documents, it could be S-1s or contracts. So I was wondering if you could just go into that, like the top three areas that are just no-brainer applications. And then secondly, all of these companies that you mentioned, they've been around for like 100 years. They're very, very complex. And just wondering, in those cases, like, how much of what you do for them is very, very standardized, and then how much of it is very particular to their supply chain? Thanks.
The largest application of enterprise AI today is predictive maintenance. Okay, hard stop. Okay, whether we're dealing with predictive maintenance for aircraft, whether we're dealing with predictive maintenance for machines in production lines, for production facilities for pharmaceuticals, and when this production line goes down and you're making interferon or something like this, I mean, the costs are unbelievable. Let's take the United States Air Force. United States Air Force has 5,000 aircraft. On any given day, 50% of those aircraft will deploy, and 50% of those aircraft will not deploy, mostly due to unscheduled maintenance.
Pilot goes in at dark, oh, dark thirty, pushes the button, some red light comes on and says, "The system's broken, you're not flying." So in the case of the Air Force, we have fused all the underlying data sets from 22 of their weapon systems, F-15, F-16, F-18, F-35, KC-135, what have you, into a unified federated image, including the telemetry. A B-1 bomber, guys, has 42,000 sensors. Each airframe has 42,000 sensors generating signals at 8 hertz cycles. This is a lot of data. So we aggregate these data, we process them at the rate they arrive. This happens to be running in a air-gapped, secure AWS GovCloud. So now we're, AWS is our partner here, very much. And we're using all those AWS resources that you talked about.
So this is what we call a JWICS environment, which is as secure as it gets. It processes the data, and what we found that we can do is we can identify system failure before it happens. Auxiliary power unit, a flap actuator, igniter in the afterburner. And we can identify it 50 or 100 flight hours before it happens, then dispatch the personnel and the materiel to converge with the air, with the airframe in Stuttgart or Bahrain or wherever, fix it at night, and the aircraft doesn't fail. Net, net, okay, the savings to the Air Force, this is one of the largest enterprise AI applications in deployment on Earth. It is the only AI system of record that I'm aware of in the United States Department of Defense.
The benefit to the United States Air Force, 25% increase in aircraft availability at any given day. Hey, guys, at the scale of the United States Air Force, that's a big deal, 25%. It saves them money, but that's almost irrelevant. Okay, and that's an organization. United States Air Force has been around for a while, not hundreds of years, but you know, since shortly after World War II. It's been around for a little while. Same thing when you get into large oil companies, predictive maintenance on you know, offshore oil rigs, oil production facilities. You know, the failure in an offshore oil rig, that's not simply a cost, that's a catastrophe. You know, a CEO gets to write his resume if he doesn't get to go to jail.
So, you know, if we can identify these failures before they happen, the economic benefits are huge. And how long's Shell been around? Since roughly World War I. Okay, and so we take existing companies, and we make them, in the case of oil companies, they can deliver safer, cleaner, more reliable energy at lower cost. In the case of the Air Force, you know, stochastic optimization of supply chain, this is a problem that goes back to, you know, the Phoenicians, okay? That's how old this problem is. So these are the classic use cases, demand forecasting, customer churn, and fraud. Yes, sir.
Thanks. Hi, Bob Kovaci, Barclays. Maybe a silly question, but what's kind of the rate of false positives on that? Like, how do you, how do you find out if that was actually gonna happen or not?
How do you find out if what is gonna happen?
If the failure you're predicting was going to happen or not?
Well, you tune these machine learning models for precision and recall. So I would say in the, you know, I'm not sure what it is in the Air Force application, but I would say it's probably about, you know, 80% precision and sixty to seventy percent recall. So that would mean that twenty-- there's 20% false positives in there. And but these, these algorithms are self-learning, and so they do improve over time. But you can, if you can get 80% recall, eighty percent, precision and north of 50% recall in an application like that, you are generating value for the company.
Yeah.
And it's in $100 million-$1 billion in the kind of companies that we tend to do business with. Other questions? So what's the play? You know, here's the thesis: large addressable market, first-mover advantage. Come on, we got started before anybody even talked about enterprise. Nobody talked about enterprise AI until when? November of 2022, and now that's all you can talk about, okay? And, so, you know, all of a sudden, that, you know, so that's. There's no CEO today, no government leader, no military leader who's not thinking about, you know, AI today and how they're gonna use it. And, so we did a lot of spade work, and, we're in a pretty good position to take advantage of this market opportunity, and that's what we intend to do.
Great. With that, let's call it a day. Thank you so much.
Thank you.