So, the Wall Street Journal has reported that Anthropic, which you invested in, and OpenAI are in a race to go public this year at huge valuations. Meanwhile, JPMorgan is estimating 5.3 trillion in spending on AI infrastructure and data centers over the next 5 years. Are we in a bubble?
Oh, boy. Starting off with a big one. So, I look, I would say it's very nuanced. And I always say that bubbles are good because they actually mobilize talent and capital into important areas. So, let's not be afraid of bubbles. But, I will say it's sort of more nuanced because on one side you're seeing real application of AI and how it's transforming product labor into productivity. I think that's a huge economic opportunity. We can talk about the complexities of it. But, on the other side, there is a lot of sort of science project stuff being funded around is there a better model out there that's going to beat OpenAI and Anthropic and Gemini. And so, there's a ton of capital going into the venture capital industry around things that are unclear what will happen to them. But, then there's also a lot more where the bigger economic opportunities that is practical. So, it's not quite simple an answer there.
I don't know if you fully answered that question, but maybe we'll get more to it later in the interview. So, you first invested in Anthropic last year at I think a 61.5 billion-dollar valuation. Why did you wait so long to invest in an LLM and has your view of that business model evolved?
I mean, I'll tell you so, I've been investing for 25 years in venture and this has just been the hardest time because these models take a lot of capital. And so, how much capital would you put at risk before you know there's a practical application? So, when we invested at $60 billion valuation, the business was doing approximately about $880 million run rate on a revenue basis. And we modeled it to be okay, it will probably do two to three billion, which was last year, and now it's starting to look like there's a real business. We're still paying 18 months ahead. But we can start to see what these things are going to become in terms of real applications. There was a cloud code application emerging. You could see that this is not just the promise of AI, but rather in real good economics that you could get behind. And I think the market has just significantly changed. We did not invest in these things before that because we couldn't get our arms around it.
Really interesting. Are you worried about the argument that this was brought up last night that Yann LeCun has been making that the current big LLMs have reached a dead end and can't get smarter?
Look, Yann's been saying that for a while. I think that there are two things happening. One is scaling laws have continued to work, and that's actually the fundamental belief as to why there's so much infrastructure spend going in because with more capital, more compute, more energy, you could continue to scale these models. So, I do think that hasn't really changed. But intuitively, I do think language is a constraining view of the world. And you think about the fact that as human beings, we understood and appreciated this universe and sort of put it in language, and these models are based off of that. And Yann's thesis around, well, there's this world model which has observed the world in a different way than sort of being constrained by language, and could you create a next level of intelligence theoretically, that's possible. And I do think there are good teams working on that attempt, but that does not mean that the potential for language models has tapped out.
So, on a recent episode of WSJ's Take on the Week podcast, we discussed how the hyperscalers say that there will be only one winner in sort of this AI race. And is that do you think that's true? And what about with the LLMs? Do you think there will be only one winner? Is this a winner-take-all market?
See, I don't subscribe to that. So, if you think about the three models that have scaled, you've got Anthropic, OpenAI, you've got Gemini. You would say Anthropic, every one of them has a killer app that's actually driving the use of the models. Obviously, there is a general intelligence that they're building, but if you go understand what Anthropic's doing with Claude, it really is about how does software get written? And what is our relationship with software as we sort of digitize our businesses? It's a killer app. It's working really well. I mean, it's becoming more and more powerful. If you look at OpenAI, yeah, they talk about an enterprise business, but a lot of their traction is really ChatGPT. It's a ubiquitous application, changed all of our lives. We all sound a lot smarter in our cocktail conversations because of it. And I think there's a business there. And Gemini, if you look at Google, it's actually more in the consumer end as well. So, they really are becoming app companies that have models underneath. The horizontal opportunity of, you know, is there going to be one model that takes on is hard to believe. I really do think they'll specialize in different areas.
So many companies are talking about integrating AI into their operations right now, and everyone's investing a huge amount. But so far, it seems like, you know, the gains from that investment haven't quite showed up in a lot of different areas. Where have you seen it most successfully integrated, and why do you think that was success?
So, I think first thing, when we talk about enterprise diffusion of AI, you have to understand what are all the things that have to happen for this to really work. So, from our standpoint, you need your data infrastructure to be ready, which is a lot of the draft behind Databricks and companies like that. You need the models to be adapted to your language, your secret sauce, your understanding of the business. And then you have to think about this immense workforce transformation, because you're going to have human beings and AI agents all working together. Some humans are managing agents, some agents are managing humans. This is a fundamentally different organizational structure of a company and it's going to get run. What does culture really mean? All those things have to be rethought. So, when you look at all that and you say all this really started 3 years ago, this idea that this was going to diffuse so fast to me it was a bit of a fool's errand. So, it does take some time. Now, we fund a lot of companies and we build a lot of companies that are taking on this problem. And I look at their progress and the pace of progress, it's actually pretty stunning. And we will look in 5 years and say the labor productivity mix is fundamentally changing. And I'll give you an example of that. We launched a company called Hippocratic AI. This was about 2 and a half years ago. We got a lot of health systems to work with us. We trained the data model. We went after it. And that agentic nurse made over 10 million phone calls, fully automated, in hospitals, checking on patients, either the ones that had a procedure the day before, or the ones that are about to come in to make sure they're ready. That would have cost $90 an hour with a nurse and we don't have enough nurses. And that's now being done at, you know, a tenth the cost. And you can actually take care of so many more patients as a result. So, 10 million phone calls will become 10 billion phone calls just in terms of just scaling compute capacity and these organizations saying they're going to go deeper. So, you're starting to see examples like this where there truly is the potential of fast diffusion. And I've said this before, last year was the year diffusion truly started to take hold. We started to understand how it's really going to work and it will be on a relatively accelerated path.
It seems like the reward for successfully integrating AI is a lot of job loss. So, for the average person, how is AI actually benefiting them in their life?
So, I think first thing is this is where the trickiness of adoption of AI comes in. That there's a fourth element to how companies have to adopt, which is the courage of the leadership. Because if you bring AI in, there's a fundamental fear that what happens to my jobs. And what I advise, we do a lot of work with organizations on their AI transformation, I always say don't focus on trying to drive efficiency with AI, focus on your growth. And what that does is your people have more interesting work to do, they grow, they are excited about winning together as opposed to efficiency means there's going to be a job loss. So first thing is culturally, how do you really transition yourself to be an AI company?
