Please welcome to the stage Mary Erdos, CEO of Asset and Wealth Management.
Welcome everybody. I'm Mary Erdos. I'm responsible for most of the investors across JP Morgan Chase, and on behalf of all of my partners, Jamie, and everyone in the room, thank you so much for joining us in what is likely to be the most consequential IPO of all times, and we are so thrilled that each and every one of you are here for this. We are really, really looking forward to this conversation. We have spent a lot of time figuring out how to pull this together. We have so many — I'm looking at the front row here — so many of the earliest investors within the fund, within the company. We have so many people that have flown in from around the world. We have people from Egypt. We have people from Turkey. We even have people from your hometown in Italy — so your family's hometown. So we're really excited about that. But most importantly, as you know, JP Morgan Chase has 5,500 branches across this country. We have 62 other countries that are dialed in, and of course countless other people that are streaming here. And so we couldn't be more excited about today.
I remember three years ago, almost to the day, sitting in Dr. Kissinger's office with Eric Schmidt, and Dr. Kissinger had just met with Dario, and he said to me in that office, he said, 'We are entering a new period of human consciousness, and it will be the equivalent of the Enlightenment.' And Eric Schmidt said after that, 'We will be talking about this period of time for thousands of years to come.' And that's why it's so exciting for all of us to be here today.
And before we get started, I just wanted to make a quick shout-out to a person that many of you know in the room. But Krishna, could you just stand up? CFO extraordinaire, the person who is helping us to get through things. I called Phipe and asked him for a quote on you. You want it? I'll give it to you. He's probably said it to you before. He said he's extraordinary at building trust. He simplifies complexity, but he is an excellent steward of capital. And then this morning, somebody who you were an alumnus of, John Gray, said to me, he is a world-class financial executive and leader, but he's never an even better human being. And that's probably the most meaningful thing of all.
Now, we are going to introduce the master, Dario, who really needs no introduction. A physicist by training, his thesis professor, when he won the Herz award, said, 'You were the smartest PhD ever at Princeton University.' He believes that you have been a leading human being across all of this. And when Ashton Kutcher talked about you in the Time Person of the Year, he said there are three things that drive this man to be the person that he is: one, his physics background, which gives him his analytical expertise; two, his neuroscience background, which allows him to help us to understand the intelligence layer; but most of all, he is kind. And that's actually the kind of person we need driving what we are doing here today. And so I believe we are about to see the leadership of Dario and his amazing management team lead us to the largest IPO in history. And for that, Jamie Dimon is going to lead us in conversation, introducing the two people that don't need their last names: Dario and Jamie.
Wow, Mary, thank you very much. Dario, Cammy who's here, his wife. So, welcome. Welcome everybody. You know, I'm looking at the faces. I haven't seen a bunch of you for a while. I wish I saw more of you a little bit. I hope you enjoy this building and upstairs and some of the food afterwards. And this company — and Mary, thank you very much. I mean, the effort you make for your clients and this company and our country in some cases is extraordinary. So thank you for all the JP Morgan Chase people in the room.
Dario, welcome. You started this company in the year 2020. But I just want to go back. I first saw AI because we hired Palantir to help us in 2012. And once we saw, we started our own AI department, and we now have, you know, thousands of people doing it, embedded by Anthropic, a thousand use cases. And I always remind people we are going to use AI to do a better job for our clients. That's it. And we as best as we can. It's still at the tip of the iceberg. So, I do have some questions. I want to make sure I ask the right ones from you, Dario.
You worked at OpenAI, but you were a pioneer in this field. And obviously now a lot of companies doing AI. So, how have you been able to differentiate from the other ones around the world?
So, first of all, just wanted to say thank you for inviting me here. Thank you everyone for coming. Thank you Mary for the intro. Yeah, you know, there's obviously, you know, a lot of people building AI technology. You know, I've worked at most of the major companies in the space. And as I've thought about my time working with various people and what has made projects succeed and what has made projects not succeed, the conclusion I've drawn after many years and after working at many places is that talent and culture are the most important things. That it is the talent and the culture that generate the technology, it's the technology that generates the products, and it's the products that generate the go-to-market and the sales and everything else that the company does. And the company's presence, how the public sees it, also comes from the talent and the intentions of the people who built the technology.
So when we set out to start Anthropic, when we left the place that we were before, the thing that, you know, from day one, myself and the other six co-founders said, you know, we're going to focus on talent and we're going to focus on culture. And right from the beginning, we set things up in that way. We have a culture interview that everyone goes through, and we've been very disciplined about turning down people who are deeply technically talented, but who don't fit our culture, right? We don't want people who are jerks. We don't want people who try to steal the show and don't work well with others. We don't want people who aren't oriented towards the public benefit. And it turns out if you stick to that consistently, it's painful in the short run, but you get a team where everyone works together and everyone is on the same side.
One of the things that most causes difficulty for companies is if you have an incredible team, but people aren't working together. They're working against each other. You have this team that works against that team that works against the third team, and they spend all of their energy fighting each other. Whereas at Anthropic, what we found is that everyone is able to work together when there's an important goal, whether it's training a model or achieving a safety objective or building the right product or making something happen in policy, the whole company comes together to make it happen. And we see it in the results.
Many more people come to Anthropic from other companies than leave Anthropic. We have the best retention rates in the industry. All seven co-founders of the company are still present. That's something that's not true of most of the other players. Of the original 100 employees to join, more than three quarters are still at Anthropic. It's relatively frequent that even those who leave Anthropic often come back. You know, one time we had, you know, two major product people leave, and they came back right within two weeks. You know, they said, 'Sorry, we didn't appreciate what we had.' So every — you know, I pay probably 40, 50% of my time focused on talent and culture.
And, you know, every two weeks I get up in front of the company. We have something called DVQs, Dario Vision Quest. It's kind of a goofy name, but the idea is that I speak honestly to everyone at the company. I tell them what's going on. I don't try to hide things. I don't speak in corpo-speak, and I encourage the other leaders to emulate that. And so it means that the leaders of the company can trust the employees, the employees of the company can trust the leaders of the company to be candid. Everyone feels that we're being honest with them, and that has this enormous value.
Thank you very much for that. So we're one of your largest customers. It's been transformational for us. We have 250,000 employees that use it many times a week in various different forms. By the way, how are the other Fortune 100 companies leveraging Claude?
Yeah, so, you know, we apply Claude across all kinds of industries. I think one of the most incredible things about this technology is its generality. We have what you might call a Moore's law for cognition or a Moore's law for intelligence. And so one of the amazing opportunities is that I've gotten to learn a lot about industries that I didn't know anything about, as well as apply Claude to industries that I know a lot about or that I used to be in, and see what incredible things can happen. So this ranges from, you know, certainly coding — that was the area that was the earliest and the fastest adoption within enterprises. You know, of course, I used to write a lot of code myself. So did most of my co-founders. And it's incredible to see what the models are able to do. And, you know, of course, every company within the Fortune 100 has an incredible need for coding. That's usually where the first adoption is.
We then go to things like knowledge work more broadly. So within financial services, you know, we see Claude used from the simplest paperwork to the most complicated strategies within areas like insurance, legal. Claude's great at drafting contracts, finding flaws in contracts. Life sciences is one I'm very excited about. We'll probably get to that more later, but everything from drafting clinical study reports to helping with the fundamental research, you know, predicting protein binding, helping make manufacturing more efficient, making fundamental biological discoveries. Similarly, we've seen things in, you know, things like e-commerce where Claude is more efficient. Some folks will deploy Claude as customer-facing agents. So there's really an incredible range, right? Industries like manufacturing that I didn't know anything about, I'm starting to learn about because I talk to the manufacturing CEOs, and they're increasingly starting to use Claude to make their business more efficient. It covers also areas like semiconductor manufacturing, physical construction. So really across the board.
You raised the question of agents, which of course is news today for everybody using agents. So if you need to be trusted in a company, you're regulated, what do you think people have to do before they actually point agents at their sensitive information?
Yeah. So, you know, I think the number one thing is actually safety. These agents are getting incredibly capable incredibly quickly. The difficulty is not whether an agent can do this incredibly difficult task in, you know, life sciences or, you know, writing software architecture. The pace at which they're progressing is incredible. The thing that actually limits us in our development of the technology, limits us in our internal deployment, and also limits customers is doing it safely. We don't want these agents to run wild, and that's why Anthropic has invested, I think, more than any other company in the space in building this technology safely, in making sure that we have the right safeguards that run from, you know, when we originally trained the model, making sure we get safety and alignment right. You know, we have a very large alignment team, we're very paranoid. Every few months we say we've put all this effort into safety and security and alignment, and we need to do another sprint to put even more effort into safety and security and alignment.
From the beginning, Anthropic has been a company that's been focused on enterprises, and we know that enterprises care about this more so than consumers do. We sort of deliberately chose enterprise as a focus because we understood that it was synergistic with our focus on safety, right? We care about safety for values reasons. We thought about that at the beginning when we started the company. Safety is what matters. We want to get this right. We want to build this technology in the right way. And then we said to ourselves, what is the focus? What is the business focus that will give us the right incentives to pursue safety? And that turned out to be enterprise, because enterprises care about trust. They care about long-term relationships. They care about preserving their reputation and making sure that models don't wreak havoc. And so it's been very helpful to us as an incentive, as a pro-social incentive, in aligning our own business incentives with what we ultimately want to accomplish in society, which is building this technology in a way that's good for society.
Are we going to be able to control the agent, like limit what they can do, to actually allow them in incentive systems, but know for a fact they can only do A, B, C, and D and not do other things?
Yeah. So, you know, one thing I would compare it to is, you know, it's a little bit like, you know, something like the evolution of cars that drive themselves, right? You know, this is an incredible technology that has incredible benefits. It's in the end going to be much safer than humans doing the same things, right? You know, humans driving cars are actually quite dangerous. AI systems driving cars can be 10 times or 100 times more safe, but you have to get them across that line, right? And you have to make sure you have the right safeguards, you build them in the right way. You're never going to have a 100% guarantee, right? You're never going to be able to say this self-driving car, no matter how many billions of people I deploy it to, it will never once crash. That's just — the world doesn't work like that. But what I think we will be able to get to is this is much, much safer, many times safer, 100, a thousand times safer than if you had humans operate the car. And I think that's what we're aiming for.
So we move to compute. You guys disclosed that you have 900 billion of forward commitments for compute and leases and data centers, etc. And I read somewhere that 80% of compute is doing new models and something like only 20% is inference, which most of us actually do. So how do you know you're going to get what you need and make it very cost-efficient down the road?
Yeah. So I would say, you know, compute is the deepest business dilemma that the company faces. So the issue that we have with compute is that because this field is growing so fast from a technological perspective, and we don't know how that is going to transfer to the creation of economic value on the business side, we had this huge uncertainty in 2024 and 2025. In both of those years, the company's revenue grew by roughly 10x. And every year we look at it and we say we don't know what's going to happen next year. There's a huge uncertainty in what's going to happen next year. And, you know, deepening the dilemma, every year the numbers are larger. At the same time, we have to buy compute a year or two years ahead of time. And the compute depends on how much the company can afford to pay for.
If you buy too much, of course, you can financially overextend yourself. If you buy too little, then you can't support inference for customers. And even more importantly, as Jamie is hinting at, you can't do the R&D, which is what produces the model six or 12 months from now, which is what determines your ability to generate revenue in the future. So it's a very difficult problem. Now, there are two ways we soften it and one strategy that we take for investing it. The two ways to soften it are: one, the better your gross margins are, the bigger the buffer between too much and too little, right? The gross margin is basically the difference between how much you pay for the compute and how much money you make per unit of compute. And so, you know, as you can see in the S-1, our gross margins are quite good, especially compared to other players in the space.
The second form of buffer is cash on the balance sheet. And we have an enormous — you know, as you've seen, we have an enormous amount of cash on the balance sheet and seeking to get more. But even with that, right, the commitments are very large. They're over many years. And of course, the revenue is growing quickly, but it's still a large amount. And so the way we think about that is we actually tend to buy on the low side. We buy relatively little compute for relatively low cost ahead of time, and then, you know, if things do, you know, not as well relative to expectations, we're well covered. If things go really well relative to expectations, then we have an opportunity to buy additional compute on the spot market, which so far has always been available and which we believe will continue to be available. That costs more, but we're buying it in worlds where we're doing particularly well. So you can think of it a little bit like we're buying an insurance policy. And of course, when you buy an insurance policy, you have to pay more in worlds where you do well, where something doesn't go wrong. So that's the general strategy we're taking, and we're very confident that with that strategy we can both invest in R&D when we need to and therefore, you know, capture the upside, and for periods of time where things don't go as well, we're well covered and we're well insulated.
So I think a lot of — if I made a chart of all the people in this room who are buying tokens today, it's probably like those numbers are going up like this. And so a lot of people have learned very quickly to take some of the expensive stuff. You don't need expensive models. You can use far cheaper models, cheaper tokens to do the calculations. That part of that's becoming a commodity. Where would that be wrong?