AI’s hidden 20-year monopoly

0:00Reid, it is delightful to be here with you in person. We always try to do this, and being here in person is more fun for us and probably more fun for the people watching, but we try to make it always fun. Absolutely. So no surprise, we're going to talk about AI, we're going to talk about profitability, we're going to end with some politics. But let's start off, for those listeners who don't know, you in the last year or so founded a company called Manus AI with Siddhartha Mukherjee, who is the Pulitzer Prize winning writer of Emperor of All Maladies. and the pitch when it launched was ambitious.

0:31Take drug discovery from a decade-long process to just a few years. Can you give our listeners an update on where you are right now with Manus AI? So, you know, AI compression was actually two years. Right, but yes. So look, one of the things that we were just reviewing and you were in the meeting, we were reviewing kind of like where we've been and one of the things that was an interesting reflection is, look, one can argue there's a theme in Silicon Valley that the language AI chatbot is basically done with OpenAI as a leader, followed by Gemini and Claude and so forth. The coding is basically done with Claude, Copilot, and OpenAI Codex, and maybe Google and Microsoft will get in the mix.

1:16They're going to work to do that. But that's done. But then actually, in fact, now it's the time for medicine. And what people don't realize is that medicines, like they're all like, actually I'll use this pun because it's fun. They're googly-eyed about the anthropic revenue thing. And medicines are massively larger TAMs. Like when you discover an amazing medicine. And what we're doing at Manus is creating a drug discovery factory for monopolies because drugs are essentially legal 20-year-long monopolies. and it was like, yep, and we've actually already kind of discovered novel chemistry. We've done a whole bunch, like in our building of these models from scratch and having taken the time to do this, Shana Ujwal Singh has been the absolute perfect co-founder, CTO in doing this.

2:08And it's like, yep, we're on target and we're going places and we know things that other people don't know. And so that's awesome. And obviously, eventually we'll get to more public statements. I mean, there's obviously a version of public statements, But it's like, no, no, no, we want to have it more locked down and more like as opposed to pre-announcing, you announce what you've done versus you announce what you're going to do. But it was a really exciting couple of meetings here that we did, you know, basically kind of two days of modest meetings. And matter of fact, in the conference room, literally to the right here, here at Greylock, was where Sid and I came up with this.

2:44Because I had flown out to New York to have dinner with him saying I have this idea that I think AI is native for medicines and drug discovery and can increase the volume of different medicines, the accuracy of targeting, the de-risking of them, et cetera, et cetera. And this is why AI, custom-built technology, can do this. And Sid explained to me in Thin Slicing how drug discovery worked because Sid has brought multiple drugs to market. This is one of the things that people don't realize. He's actually one of those people who has actually brought cancer therapies to market. Which is insane. Which is like, look, here is how it works from a target and idea and like the work process.

3:21And here's the things and here's here are the loops and cycles that go into it. I was like, OK, I can do something about that today. We can build something about that today that we just got to have to go with straight, you know, you know, kind of classic process. And then that's what formed the strategic plan for the company. That's awesome. I mean, I think everyone is so excited about the possibility for AI to help cure cancers, cure many diseases. But when you think about the business side, some markets are winner take all or winner take most. When you think of drug discovery and these new sort of AI plus bio companies, is this a winner take all market?

3:56Or there are going to be many companies who are going to be able to come through with blockbuster drugs? Great question. Not surprising. The short answer is many. The TAM is so large. I mean, think about the number of different providers of GOP-1 drugs. There's a number of different providers. All of them are in the tens of billions of revenue on these with kind of awesome margin characteristics and all the rest. And by the way, they're all sold out in terms of how they're operating. So it's very possible to have a monopoly on your drug and have other drugs, even in the same space, in it to be really, really lucrative.

4:34But the question is, what other things are there to be discovered? And currently, the current process is you have some room of PhDs who are like, we're banging our head against the wall. And we got, Eureka, we got one idea. Let's see if we can make that work. And with AI, it's like, oh, we have 100 ideas. Totally. And we're working through it. And we're actually like using these super smart PhDs to accelerate, you know, prioritize, et cetera. Like one of the things we joke about in the company is we have SIDRL, right? You know, SID reinforcement learning for those people who don't understand what that is because they're like, no, no, no.

5:09In order to be well, like to identify a good target, to understand if you got something. Part of what we've done is we've replicated other work that took like a decade in hours because we went, oh, OK, try to solve this problem because we know there's a solution here. Can we come up with other solutions? Can we identify the solution in our outputs or the solution that has come up that is already done by these genius researchers? And the answer is, yep, that answer came out in the mix, too. And now we're looking at the other ones. It just seems like in this in particular, like the generative nature of AI is exactly what we need.

5:44We need so many ideas for what are the potential targets, what are the potential drugs, then we obviously have to test them in a wet lab and make sure, but the more possible high quality targets we have, the better. Yes. Let's move directly to AI. So back in the early to mid 2010s, you used to make an argument that new 100 million user social platforms would keep getting created. You were saying, again, this isn't a winner-take-all market. There's going to be many different types. Perhaps one of the reasons you started LinkedIn, you said people are going to have their social network for friends.

6:14They're going to have their social network from a business perspective. They're going to be, they have different identities. And so with each part of their identity, they might have a different social network. And these could be 100 million people. They could be a billion people. Does that same logic apply to AI? When we think about the AI agents that are going to be helping us, are we going to have one to help us in fitness, one to help us cook, one to help us to child rear, one to help us get us medicine? Or are we going to have that one general purpose agent who does everything? What do you think the personal agent landscape is going to look like?

6:49Is it going to follow the same properties as social networks? So one mod on the social network question side, but that also gives a lens into the eye, is that there are multiple winner-take-most networks. It's not there isn't one. It's that there's multiple of them. One really does social. One does video. One does da-da-da-da-da-da. It's kind of ways of doing this. And there's room for more. By the way, there's still room for more social networks as it is. but you have to be technology native to where you're going. You have to have a distribution plan. You have to, you know, like all of them try to be, you know, broad based on what they're doing.

7:29And so, you know, is there room for your particular one? And there's a whole bunch of different things. And that's the social network thing. So I think there's still a lot of zone for doing stuff there. And so those comments still plus one. And I think the parallel to AI is exactly right, which is there are some people who think, no, no, there's just only going to be one super intelligence model. And that's going to be the one for everything, because that's one that is just better. But there's a bunch of different limitations on that. One is, as you get to the superintelligence model, it's actually very expensive to run, right?

7:59And yes, you'll figure out how to bring down the compute cost. But if you're like, oh, look, the superintelligence model, it costs the equivalent of$10 ,000 an hour to run. Well, you're going to focus it on like, you know, help me cure cancer or help me with fusion or, you know, like other kinds of things. You're not going to say, help me with my email responses or, you know, like, hey, if I have this rash, you know, on my baby, is this something that is a normal thing, something I should be concerned about, something I should do something about, et cetera, et cetera. Now, part of what will happen from the super intelligence is I think there'll be a bunch of different, you know, kind of essentially distilled models and derived models.

8:40This is part of the thing that we're seeing is like people like legitimately amazed by some of the things we're seeing in China. But by the way, a bunch of them are efficiency and distillation and all the rest. Not all of them, but a lot of them have that characteristic. And so there's a bunch of things. But it's not just that there will be a bunch of different models and distilled from top models. And by the way, currently it looks like a similar kind of prediction of there will at least be three to five frontier models. And maybe that's going to seven to ten. And we'll see what happens in China.

9:12And we'll see what happens with various European efforts and a bunch of other things. But as ways of kind of doing this, there's also different ways of getting distilled models and partnerships. But then there's also different data sources. There's different user affordances. Like one of the things that people underappreciate Codex as a coding model. And it's actually, in fact, got some unique strengths that, generally speaking, people should pay attention to. But right now, when you go around Silicon Valley and you look at social media and everything, it looks like it's only clawed code all the time.

9:44And Cloud Code is an amazing product. There's a reason why there's all that heat. It's great. But it's trained on less compute than Codex. And so it's a lesson to all the other kind of lesson, all the other ability to develop specific AI models. It isn't just you might have unique data. But if you're actually going after a particular product surface, If you're doing something that you're learning a lot of what that affordance and the interaction works like, that will also allow you to create differential products. So there's multiple zones, not just smaller models, not just compute efficiency, not just whether it's distillation or new data or other things and all this.

10:29There's also the product surfaces and the interaction, what you learn from that and how you get there. Then, of course, you get to what's the zones that a person might have in the interaction of their thing. of their thing. And the thing is, is yeah, the tendency is I don't want to have to try to remember and go to lots of different AI models. So it's like, I'm going to have a hundred. Right. You want one front door. Yeah. Most people don't do that. That's part of the reason why, you know, like search becomes synonymous with Google, et cetera. Cause it's like, nah, like a general front door. I will want really, I'm using one of those.

11:04Right. And I think there's still contention of that. I think chat GBT is still the leader in that Gemini obviously trying to come in. Claude, not as much. But then it's like, okay, I have a general front door, but I have a front door for my enterprise. Well, that might be co-pilot. I have a front door for this other thing that's kind of much more specific for me and that's really important to me. I'm a parent. I may have something that's specific to that. And that's where the social network parallel, the opening of your question actually, in fact, really applies because it's like, well, actually, in fact, If you can get to a large enough, coherently valuable and growing thing where people go, this really matters to me.

11:43I would add another door for this. Then there's a zone there. It's willing to be an app on my home screen. This is what I'm willing to give up real estate for. And by the way, we do remember, like people are a little harder in remembering like, okay, going to different URLs and so forth. That's the reason why there tended to be a kind of winner takes most in search, plus AdWords is the back end and paying for distribution in order to do that. But we actually, most of us have multiple people in our lives. It's like, well, what would I call Aria for? What would I call David Z for? What would I call, et cetera, et cetera.

12:22And so our brains more naturally work in that. So I actually think the narrowing focus to I just want one front door, even though you tend to get a trust, I think it'll be broader here than the other circumstances. Because it's like, oh, once I discover the no, no, no. Like, look, chat GBD is great, but I really get much better legal from this or I really get much better medical from this. Fine. I'll remember to add it in because it's like, great. I have the thing to go to on that. And so I think we'll have a broader lens because Aristotle said we're citizens of the polis, which actually doesn't per se mean we're political animals.

12:59It means we're social animals. We remember a bunch of different people and kind of like have this conversation here, have this conversation here. Agents will naturally slot into that. So I think there will be a broader range of this. And it's one of the reasons I said there's a huge amount of startup opportunity here. It's not like the small and frontier models who are extremely expensive to build. The small and frontier models are going to run away with it. I think there's a huge startup zone here. We're just at the beginning. And so we're talking about these big companies, you know, whether it's you're talking to ChatGPT or Claude or Gemini.

13:34But should some of these agents be public goods? Like you often talk about that everyone should have a tutor in their pocket or everyone should have a medical assistant. But you'd also assume like right now everyone is afforded a public defender. Should everyone have a legal agent that is provided by the government? Should everyone have a tax helper who helps them file their taxes every year? Should some of the government goods that people are afforded now, they also be afforded an agent in that space? And should that be a public good owned by the government? How do you see that playing out? So I'll speak a little bit more bluntly than I've spoken about this before, because as you know, for, you know, maybe coming on eight plus years now, I've been going around to various governments and saying, hey, you should do this.

14:22And the reason I've been doing it mostly there and not publicly is because it's like, look, people tend to have an idea they can own and they can go do something. And it's like, I want it to happen. I don't want credit for it. Whatever. I want it to happen. And when you say it, you mean a medical assistant, a medical assistant, a legal assistant, a tutor, et cetera. It's like, this is something you could do that you could be a champion for in providing services for your people, whether your people are, you know, like Californians, Americans, Brits, French, Italians, etc. And I've literally been singing this song sheet for eight plus years.

14:56And of course, we see nothing, right? And it's the incompetence of being able to figure out how to deploy tech strategy, what kinds of things to do, not being able to imagine the possibilities in the future, you know, uncertainty, listening to too much press, negative chatter on AI and other kinds of things and not being willing to be bold and so on. So look, the short answer is every government that actually, in fact, really cares about their people should arrange, this doesn't even build themselves, right? This is a part of the thing. Does it build themselves? Of course they can't do that. They don't have the competence to do this.

15:28For a medical assistant, for a legal assistant, right? Not just public defender, but actually like most people, you know, sign rental contracts. Absolutely. Like, okay, what is that? Like most people, when they're engaging with medical services, it's not just the thing. It's like, well, what is this thing that I'm agreeing to with my data and all the rest? And how should I navigate that? What are my rights in a number of different circumstances? What are the opportunities for me in a number of different circumstances? Tutor, right? And obviously, sure, it could be kids and all the rest. But also like, oh, I would like to learn new jobs and new skills.

16:03And how do I do all that? And that's just the beginning on all the things. And what's more, like governments, it's not hard to do. And the reason it's not hard to do, please like get out of your own way and make this happen, is if you go to any number of these providers, and it's not just the Microsofts and the Googles and so forth, but it's also like Anthropic and OpenAI. And by the way, just to substantiate earlier comments I was making where people said, well, are you worried about antitrust? It's like, look, I said, look, if we were five to seven hyperscalers going to three, I'd be worried.

16:40We're five to seven heading to 10 to 15. And now, like, I was right. Right. You have NVIDIA. You have OpenAI. You have Anthropic. You have whatever is going to happen with SpaceX and that random. I mean, 10 years ago, people might have laughed at you if you had said there was going to be these, you know, multi-billion dollar companies. Because it's like, no, what I see is these more of these coming and they're competing with each other on all these things. And that competition is precisely what creates the speed of good. So how do you create these public goods? Well, you negotiate with them saying, OK, like if we if you provide this, we help you in the following ways.

17:17Like, for example, a simple one for a medical assistant is we will give you the following parameters of safe harbor for providing medical advice. If you do that, we want something that runs for free in every phone that's connected on our domestic networks. Right. So and then we can go to the carriers and say, you must provide this app on every single phone free of charge. for doing that because then all of our people have this. It feels like the future we're going to have to have these public-private partnerships that is delivering for our citizens. The government cannot build this technology, but perhaps the companies aren't going to be delivering the technology in the way we want unless they have these government partnerships.

18:05If instead what you say is we're beholden to the trial lawyers association for lots and lots of suit, we're going to decrement these capabilities and not provide them. We're going to be much more conservative. Right. Well, of course. Well, why should we? Because what you want to do is provide, for example, medical advice, legal advice. You want to provide it for free. For free. Right. Totally. And just like, OK, so like how do you make the economics of that work? Well, you can't be the, oh, well, we're going to make you massively legally liable for this. No, absolutely. And so one of the companies you mentioned was Anthropic.

18:40And it feels like Anthropic is everywhere in the news these days. But a new reason they're in the news is because they're actually, this is their first ever operating profit this quarter. So they're projecting around$10.9 billion in revenue for the June quarter, up roughly 130 % from the previous quarter. And that's$560 million in operating income. And so, so many people talk about these companies and say they are spending, spending, spending. They're spending on compute. They're spending on inference. And they're never going to make a profit. Perhaps Anthropic is proving them wrong. Certainly Anthropic is proving them wrong, but keep going.

19:18Okay, okay. For Anthropic, the growth here is driven by coding tools, enterprise. It's not coming from their consumer chatbot. It's not coming from everyday Americans talking to Claude every day. Does that tell you something about sort of the durability of the AI revenue that's going to come? Does Anthropic, will they look fundamentally different because their revenue is coming from a B2B situation as opposed to a company that's chasing hundreds of millions, billions of consumers to interact with their platform? So this is standard pattern in technology, which people say, well, there's no revenue here.

19:55I remember this from the earliest days of the Internet. Absolutely. Et cetera, right? And then later getting to Google AdWords, which is the best business model. They're thus far invented in software technology for doing this. And so they start with no revenue and they say, well, then it's just that. and like for example, when the internet one-oh winter boom happened, it was like, you know, bust happened, you know, the internet winner, like, oh, like the only winner is in consumer. And now we're all going to go to enterprise software and clean tech. And that's part of what allowed, you know, a web two revolution to happen and, and all the rest of the thing.

20:28And then like, oh, no, one's ever going to make any money with Facebook, you know, blah, blah, blah. And so it's, it's, it's just, it's a standard pattern to go. It's, it's nothing. Then it's only this. What I guarantee is, yes, there's durable revenue in enterprise. Yes, there's durable revenue in code. Those will continue. And there will be other major revenue streams. Right. That was part of the modest gesture earlier. Like, look, all of the stuff with like if we can get the drug factory going, which seems very early promising and awesome, like that kind of thing. Then there's a huge amount of revenue there.

21:01There's going to be lots of different revenue things, including new models. And by the way, of course there's going to be revenue and consumer chatbots. Like I can think of three easy ways in order which to do that. But do those companies look different? Are you fundamentally a different company if you're going for enterprise? Well, almost for sure, right? Again, when you have the platform shift tsunami, lots of things change. So stating everything with 100 % probability is always a little cautious. But like, for example, there's a lot of different rules between enterprise and consumer. And they tend to fork for particularly good reasons.

21:39And companies that tend to be good at one tend to be not good at the other, even when they have some things that allow them to do that. And it's usually good strategy to go, especially as a startup, but even as scale companies, to go, we're going to really prioritize this. And the other thing is just an option that we're going to kind of try to play at. And so, yes, and I think they will look different. But I think that one of the things is not only are we kind of early days in this where people say, well, is all the enterprise AI stuff going to be Microsoft and Anthropic? And the answer is, I guarantee you there's going to be others.

22:16Right. Is all of the use of AI for consumer and search going to be open AI and Google? And I guarantee you there'll be others. Right. So it's just like now, which prediction? What do we do? That's part of the venture gig we do here at Greylock. We're sitting in the Greylock offices. That's one of the things we try to do and make smart bets on. And then, by the way, there will be others too. That's part of the reason why, I guess it's now three years ago, I first went, oh, drug discovery for this stuff. And we've got other things, as you know, that we're cooking on this stuff as well. So it's, I think that, and then the most interesting mind-boggling is, what are the different ways the business models come together?

23:02Because this is one of the things about AdWords. It didn't start in the computer revolution of the PC revolution of the 80s. And it took a while in the internet to get to it. And part of the reason why people were willing to hand search distribution to Google was because they went, oh, this is the wrong business model because the right business model is a media business model, not an AdWords business model. and it was that discovery of the AdWords business model, which, by the way, wasn't initially invented by Google. It was by, you know, kind of GoTo and Overture and so forth, but then massively improved and kind of put it in a loop with buying distribution and all the rest as part of how that played out.

23:45And by the way, you know, people said, well, no one will ever pay for the kind of ads you see on Meta. Well, we'll actually take a look at their revenue stream, et cetera. So we build new business models in this. And one of the things that I am always paying attention to in AI is what are the new business models? Because, yes, what is the new distribution? What are the new engagement patterns? What are the ones that co-opt existing ones? But also what are the ones that create new ones and create new business models? So we're going to end with some politics. This is I have a lot on this sheet of paper to give people context because there has been so much, dare I say, corruption happening in our government this week.

24:28So back in January, President Trump sued the IRS for$10 billion over a leak of his tax returns. And that was just settled, I believe, with Trump on both sides. And he wants to build a$1.7 billion fund to give essentially restitution for those who were prosecuted and then pardoned during January 6th. So he wants to take$1.7 billion and give it to people who were arrested, who are criminals, who fought cops, who did violence in our state capital, in our country's capital. But there's also other things. He got out of$100 million IRS fine. He secured tax immunity for himself, for his family. Like there are so many things happening this week.