You're ruling out a bubble bursting.
Did you know it was going to be trillions when you first got into this space?
No. That was probably a bit of a surprise. You know, when Sam and I have talked about the various, call it the pantheon of AI people, Mustafa is not on his list of the people who he most worries about.
The thing that I think pissed off Elon with me and then started like Elon's own relation with Epstein.
Today's guest is Reid Hoffman, entrepreneur, investor, best-selling author and co-founder of LinkedIn. Reid is known for shaping how the world connects and for backing some of the most transformative companies of our time. He is highly influential in political circles and most recently was named in Epstein emails. What they should do is literally take every single piece of intel that they have about Epstein and release it.
In this episode, we're getting into all of it. This is the Newcomer Podcast.
Reid Hoffman. Great to have you on the Newcomer Podcast. Thanks for coming back.
Always a pleasure. I look forward to it. And, you know, done this before and I'm certain we're going to do it again, but it's a great time to be doing it.
Busier audience this time though. I think it was just you and Eric last time. Now you've got a full set of inquirers in front of you.
We've got the whole crew.
We have the full Inquisition panel here. Now bring out the comfy chairs. I think it'd be hard through Zoom to bring out the comfy chairs, but you know.
Yeah. Not too comfy. Yeah. I mean, look, Reed, you've been early on on the whole AI boom. Maybe earlier than most, you know, investors.
So early. So early you didn't get equity, right? I mean, almost too early.
You're doing okay. But look, it's been an interesting, I would say, couple of weeks, at least on the public markets front. Like you saw a runup in all of these AI stocks over the course of 2025. Like we're entering the end of the year now and all of a sudden it seems like the market is blipping and has lost faith in at least a lot of the kind of tech fueled public stocks. I mean if you had to just give our macro look at like why there's a little bit of cynicism or skepticism about the technology, I mean what's your read on the space?
So I don't really, I'm much more of a private market investor, long-term multi-year, than public market investors. Public markets have a whole bunch of different randomness in it, because it is anything from, oh, the war in Ukraine is going badly, stock trades down, etc. All this kind of stuff. And there's reasons for it now. That being said, I'd say the tech companies over the last two years, over the last 5 years, over the last 10 years have been, when you get down to these indexes, part of the massive growth overall in it. And so even if there's a modest near-term correction, if you're looking at it over a year or two, it's still pretty far up there. Now that being said, part of what people are trying to track is at what level do you get to the different forms of business and industry and economic transformations that are anticipated by AI. People like to have a little bit of a, this is a classic, you guys are all expert in this. Week one is, oh my god, AI is going to be everything. Week two, AI is an overblown disaster and is a bubble waiting to happen. Week three, AI is going to be everything. And so you have a cycle that kind of goes through that. We may be in the, oh, look, it's overblown cycle. But I, you know, eh.
Ilia Sutskever, I think, is that how you say it? Ilia just said these companies might make a lot of revenue but he wasn't sure that they make a lot of profit, which as an investor has got to be sort of worrying. I think his point was there's a lot of opportunity but then the model companies are clustering around each other. Obviously we just saw great work from Gemini. What's your read on the state of the models and how much they're sort of clustering together and if that's changing your view on how they are?
So, a couple things. There's definitely, one of the things that was really a happy month for kind of AI being American intelligence and having Silicon Valley leadership is we got GPT-5.1, we got Gemini 3, and we got Opus 4.5, all of which are massively great world-leading models in this stuff. And you say, well, does that mean that you won't be able to command margins on any of these things because of competitive pressures? And I think there's a couple of different things. One, it may very well be that margins don't come from a uniqueness of a model, but from rather a kind of where you live within a product service ecosystem, whether it's the services of how you're delivering particular applications or whether or not you're better at some things versus others because, one of the things that's certainly true is each of the models are better at some things than others. And it's one of the things that I kind of track because when I have a serious thing that I'm doing, I tend to put it into five plus models and then compare the answers and it leads to interesting things. So like I'm going to have to try this with Opus 4.5, but like Claude tends to be better at fiction writing than any of the other ones, not just the code stuff, but these kind of things. So I think that's great. Now that being said, the margin comment, there's two things I think about margins that even if you said, hey, look, these trained models will be run in a hyper-competitive way, so there'll be relatively little room for margin. One, obviously if you tie that into an interesting business where you have an application, I'll give the two Microsoft cases. One is enterprise software and you're in enterprise that will obviously have margin positive margin capabilities. And then the other one is for all hyperscalers, part of the reason why the hyperscalers can invest in this is if you said, well, what comes out of this is the software is just running lots of cloud compute, well you can still sell cloud compute at your margins of your cloud compute. And so that's another place where there will be margin. That's one of the reasons I think you have to be careful when you're a startup trying to do frontier models is you better address this question because you don't have a fallback of the cloud compute. So the first one had better be your answer.
Right? Yeah. If you're a Microsoft or Google, you're like, well, I'm safe. I have my cloud services to provide anyway.
Yeah. Reed, I was curious though into this question about margins. Where are you seeing applications that have maybe better margins? Because we've talked a lot about coding assistance, obviously Anthropic's Claude Code is really high performing and Cursor obviously is on everyone's mind, but its margins aren't exactly the best. So as you play mostly in the private investing space, what are you looking for when backing a startup in this space to make sure it's defensible?
So what you're looking for, collecting margins early is usually a sign of either a very dominant business or a sign that you're not focused on growing enough. So actually initial margins don't tend to matter that much. What tends to matter is what your theory of the game is for margins when you're in the midscale to scaling. Like for example, in the very early days, I led the Series A in Airbnb. Some of the investors asked, what's the margin structure? And you're like, it's a software business, right? We don't care about the margin structure right now. We care about growing this business because it's irrelevant unless it really, really grows. And then after that, we can see what are the different things we tune to margin. So you do want margins. And the kind of things that give you margins are obviously one of the things that I have a multi-decade career in is things with network effects. But it can also be enterprise integrations. It can also be scale effects. It can also be by blitz scaling, being the first to scale that you then get a general market dominance that people tend to use you and prefer you. All of these things can lead to that, and that's what I think you're playing for. But you're playing for future margins more than current margins. And if you sometimes current margins are the best predictor, but most often it's the theory of how you grow to the future margins.
Can you make the case to me that especially at the foundational, like frontier lab companies, that these margins will get better, that they are going to follow the path of a traditional software company? And the reason I'm skeptical, or at least I'm presenting the devil's advocate side, is that inferencing and specifically these reasoning models are highly resource intensive and per query are extremely expensive to run. And I guess you can hope that there'll be some breakthrough in inferencing technology, but do you foresee this going down the same exact path that every software company that's existed before this will go down, that it'll get much more margin appreciative, or this is just always going to be a lower margin business to potentially negative margin?
So I can make both cases. I do think it'll ultimately be, like all good smart people, it's like all the universes in my head. Well, but I like because you're charting now. If you actually ask me to guess, I do think it'll be margin positive. And the real question is, is it lightly margin positive, like the fallback of hyperscaler compute, selling cloud, that kind of stuff, or majorly margin competitive, like classic software businesses at scale. The thesis because you've asked me to make the software case is to say actually what will matter, the reason why it's important to have multi-gigawatt centers that are training the largest possible models is that when you train those large possible models, they will have a bunch of software derivative products. So even though running them will be expensive, you can distill much smaller models and agents will be compositions of smaller models. Those compositions of smaller models will be a substantially better product than anything that isn't distilled from the largest possible model. They can be trained to specific kinds of high margin businesses, which can include coding, but also legal and medical and a bunch of other things. And the fact of having to get to the really large scale allows you to create the computationally efficient but very high quality small scale in terms of how you're operating, and those can operate in software margins. That would be one of several possible theses which give you software margins in the future.
And that's essentially the case that OpenAI and Sam are making for why they need to be raising hundreds of billions, trillions of dollars to build these multi-gigawatt data centers. It's essentially an R&D expense that will eventually allow them to create these smaller topic specific models that won't be money pits. You buy that.
Yeah. Exactly. And it's also I do, did you know it was going to be trillions when you first got into this space that this was going to be the scale that you were operating at?
No, that was probably a bit of a surprise. But part of what I love about Sam and the OpenAI team is it's literally the, shoot for Mars. No, we're going for Alpha Centauri.
You're like, no, it's like we're throwing all of our chips in repetitively to a complete bet.
Yeah. Not a moonshot, a galaxy shot basically.
That's a good line. Are you, has Greylock invested in Anthropic? Like are you exposed to OpenAI? What's your exposure?
I am, both earlier when it was the 501(c)(3) I donated money, but when its first commercial and when it set up its LP fund which converted equity, I led that round.
Sorry, oh man, okay. Was that when Kla invested or before?
Yeah, co, look, I love Vinod and he's very careful. He says first venture fund because it was me for my foundation, but I actually led that round.
And did you have more equity than him?
Do you feel like, are you happy with the for-profit conversion? Like I mean that round sort of got crammed down a little. Obviously it's a great return, but it could have been the best venture return ever in the history of the world, right? If the cap table had played out differently.
Look, it's obviously a strange path. When I led the round, it was a 501(c)(3) with some commercial technology that would come out of it. It's literally one of the strangest, probably the strangest investment I've ever made, because you're like, commercial technology coming out of a 501(c)(3), what's that going to be worth? But it was like, look, it's a potentially world-changing technology and they're playing a really smart coherent bet. And what's more, it's both for the way it affects societies and industries, but also the fact that it's an economic investment, but it's also for humanity. And part of the for-profit conversion that I think was really good is, and there's a whole bunch of noise about this in press, both criticism and positive and all the rest, but it really comes down to the organization has kept a true north of how do we create AI for the benefit of humanity, and that requires us getting partnerships, massive amounts of capital, and all the rest. And so the mechanism of doing that is a 501(c)(3), we now have to go to a public benefit corp with governance from the 501(c)(3), but it's still the same true north in terms of what they're trying to accomplish. And I think that I've never seen them vary from that.
And so you're happy with it?
Yeah. Yeah. Yeah. Delighted on both as an investor and as a human.
You're happy as an investor? How much did you try to protect that tranche? I guess like one of the things that's very good about the Silicon Valley process of this kind of Series A, Series B, Series C is you understand the whole format when you're going from, well, we had this LP commercial agreement that's now converting into equity. And by the way, did require billions of dollars from Microsoft in order to get there. It's like, well, what's the right fair thing in retrospect is always kind of hard because when you're doing it from the forward back, it's like, well, you should get a really good return, but how much of a great return becomes a, and I think all the parties balanced out the negotiation well.
Your relationship with OpenAI over the years has always really fascinated me because you're obviously this early investor. You're on the board of the company. You stepped down from the board around the time that you made the investment in Inflection.
That was about seven or eight months later. Yeah.
But you in your statement around that sort of said, I'm kind of increasingly conflicted because of my investments, so it wouldn't make sense for me to stay on the board. That was essentially your rationale for stepping down.
Yeah, it was more that I was getting this flood of startups from Greylock saying, please give me privileged access and so forth, and I was like, being a board member of a 501(c)(3), I can't even ask for that.
Right. Right. Right. So because it was a 501(c)(3) that I was on the board of.
Right. Right. Right. But if you look at the quirk of history, that decision some people argue to step down ends up opening up the vulnerability to Sam when the other board members vote to oust him. And my understanding, and I wrote in a couple of stories in the Journal, was that Sam prior to even his being ousted was pretty pissed off at you for investing in Inflection, which was headed up by Mustafa Suleyman, who Sam doesn't really like. And then down the road you end up leaving Sam vulnerable to this coup, this short-term coup. I guess like what's your take on that trajectory of history? And then also, what's the relationship between you and Sam these days?
Great reporter question. Look, I've asked you about this, so I've written in so many stories.
Yeah. Well, as far as I know, Sam and I continue to have dinner together, have very productive conversations, etc. As far as I know, my relationship with Sam is strong and good.
Was there tension though for a time?
Well, I think what I did learn was that Sam is one of those ultra-competitive people. And no shit, a little bit of how I got surprised by it is like, I'm doing OpenAI at that time which was a 501(c)(3) on the board, and I'm then doing the separate commercial thing, and I kind of don't think commercial things and 501(c)(3)s are in conflict. That literally didn't occur to me as an issue. Because commercial conflict is something I track a lot, but 501(c)(3)s are structurally and positionally not supposed to be competing with 501(c)(3)s. And I was like, well, but we're going to be competing for talent and competing for deals. And it's like, oh, okay, I see some of this issue, although I don't think there's room for multiple players here. Okay. So it did create some friendship conversations that get into conflict because in conflict you either get further away or closer apart. I think Sam and I got a better understanding of each other and higher trust and how to collaborate. But through a little bit of a, what are you doing investing in Inflection? Starting as a basis. So yeah, he was angry, but you're still friendly. Sounds like it's a good friendship. They can get over these things.
But look, this is historical stuff, so I don't want to spend too much time on it, but I have the opportunity here. Was it because of them as a competitor, or you think it was specific animus towards Mustafa that he was mad about your investment?
Well, as far as I know, it was competitor. Again, when Sam and I have talked about the various, call it the pantheon of AI people, Mustafa is not on his list of the people who he most worries about.
Careful.
I guess the next question is who? And I was like, ah, I don't think I could say that.
That is a list. That's the real reporter thing is like I got to get that list now. That's interesting.
Very succinct question. I don't want to go down a whole rabbit hole. Do you have an AGI year that you think right now? Like if you were to say my best guess for AGI it's this year. What's your answer?
So my typical quip about AGI is it's the AI we haven't invented yet. So therefore by definition it would always be in the future. If what you mean by, and obviously people mean different things by different things, you say AI is capable of doing more than half of the tasks that human beings currently do and could do at better than 50% competency or 70% competency, my guess is you're at least 10 years out.
Right. And it might be substantially further. I think the people who do this tend to go, well, aren't all the tests like coding? And it's like, no, there's a lot of great things in coding and we're all going to have coding co-pilots and assistants in what we're doing and that'll be a huge amplifier for all, but there's a huge amount of tasks. For example, most people don't track the level of metacognition that we do in the current uses of our AI agents and don't really focus on what that kind of context awareness, goal setting, pruning and editing and a bunch of other stuff that goes into this. And when you look at it, you go, okay, of course we're making progress, GPT-3, GPT-4, GPT-5, but there's still massive room for metacognition. Here's a simple way of doing it. One of the things that Microsoft has been doing for years now is having long longitudinal agents interacting with each other to see what it is, and they end up in lacunas all the time. As a simplistic example, it's like, Eric, thank you, no, Reid, thank you, no, Eric, thank you, Eric, Eric, thank you. And they do that in all kinds of weird work processes and so forth. You don't have to set them in a strange zone to get into that area. And metacognition is kind of the awareness. Here's a more, call it last year one, which is you'd ask one of these things a prime number question and it would get it wrong, and you go, hey, you got it wrong, it's, oh yeah, sorry, I got it wrong, and give you another answer. You go, well, you got that one wrong, it's, oh, I got it wrong, and give you another answer. It'll keep doing that for the 20th answer. If you ask a human being the third time I got a prime number wrong, I go, oh, sorry, I clearly don't have this problem, like I'm getting this wrong. I have the meta awareness of what's going on. These artifacts still lack that meta awareness, that metacognition.
That one of the most infuriating things about the models is their unwillingness to say, I don't know. They're programmed to give an answer every time. And so it's like, yeah, there's so many things. You ask me a question, I'd be like, I don't even know. I don't know.
By that philosophy, would you land more in the sort of Andrej Karpathy timeline of AGI? Like, we have several years to go, but that doesn't mean they're not useful now.
A thousand percent. Like if you said, is it AGI that makes them useful? No, they're incredibly useful today. One of the things that I tell everybody is if you're not finding uses of it, new uses every month that really help you, you're not engaging enough, right? Full stop. And I don't mean sonnets for your kid's birthday or that kind of stuff. That's all fun. But like things that matter to you in your life or in your work. If you're not discovering that, that's because you're not actually trying hard enough.
It's an interesting topic for us. A couple of months ago, we had an episode of the show where we had a real AI skeptic on and we've sort of been tagged as being kind of AI optimists or shills at times for the industry. But I'm curious about the capabilities of the models now and the valuations of these companies. For OpenAI to truly be worth $500 billion or Anthropic reportedly in the $350 billion range, do the models need to be much better than they currently are or are the capabilities where they're currently at enough to justify this level of market valuation?
Well, I tend to not overly try to say like I'm an expert at it should be valuation X versus Y. It's one of the reasons why I like the long compounding of private investing because it's kind of like, okay, whatever this is, if I'm right, it'll go into something really great, right? Even if I might have gotten lucky and underpaid by half or been foolish and overpaid by 2x, whatever. Now that being said, I think we are massively underusing the capabilities of these systems today. One of the reasons why I push people on this is that most people, even AI evangelists, don't actually use AI nearly as well as they could.
I know I have vibe coding projects I need to spend more time on it. I mean we use it a lot like copy editing. Yeah, we have a human proof reader but we also run everything through that and reporting ideas. But yeah, you're making me feel guilty. I'm like I should use it.
Yeah. But let me give you an example. So you guys are running a multi-party podcast. One of the things I realized is AI gives you ability to translate the language. So one of the things that we did with Arian Foster and I do this Possible podcast and we have this Reid Riffs section of it which is just Ari and my voice. We already had a Reid AI voice, we made an Ari AI voice, and now we're releasing it in different languages. We started with French, we've already released it. Like that's there now with AI. And so it's just simple, you develop the process by which you run it through the compute and now you have your podcast available in other languages. And while of course we have the fortune that a lot of the world speaks English due to the combination of the British and American empires, there's a ton of the world that does not. And what's more, even the world that does, can you get the depth and sophistication of it? There's a reason why you like the translations. And that's just the beginning of all kinds of things you can do. And that's part of like, wait a minute, we have the world's best translators. If you did it with humans, it'd be a huge amount of work. But we could set it up as an industrial process with AI and we can be in every language before too long, right? And this is stuff that it can already do.
I mean, maybe because I know we kind of need to transition to other stuff. And maybe this could be like the last section, a question about AI, but if you look at the spending that's going into these things right now. And Sam Altman has sort of created this incredible cycle where he will announce and OpenAI will announce these deals with these chip manufacturers or cloud computing companies and stocks shoot up. There's a huge amount of debt that gets involved in building the infrastructure needed to make this technology work. I mean, does this feel sustainable to you? Do you like this current set of AI being so intertwined into the economy that it's a huge percentage of GDP? This massive economy fueling process that has so many loadbearing beams built into it. What's your sense on the fact that AI has become this thing now that's so expensive and so economically motivated?
Well, so I'll translate this a little bit into what I think about, which is kind of a framing I have, which is AI bubble versus AI correction. And I'm not even sure we're going to have a correction, but I'm pretty sure we don't have an AI bubble because the bubble thing tends to be that this investment cycle is ahead of the economy and there is no there there on the other side of it. So the simplistic way of making this argument is Nvidia invests in companies, companies buy the chips to do training, and then Nvidia invests in other companies to buy more chips to do training, and that's the cycle of this. The thing is actually we're going to get to where intelligence and compute has the massive scale and always on 24/7 of electricity, and that is actually useful. This is the undercapabilities things. And it may be some retooling, and it may be that we paid over like too much premium prices on timing and other kinds of things for this, and you get some correction on pricing, market caps, economic flows. And so some people may have said, hey, I paid $60 billion for the data center and it's really worth $30 billion, so it's a $30 billion loss. But by the way, if you started running that data center now, you can actually run it operationally profitably. For example, the use of these AI models in inference is already profitable. It's the training costs that cause the issues. And so that's the reason why I say, look, the bubble is a nice language, kind of semi-historical but incorrect. I think what you might get is a correction because you might go, oh my god, the actual real economic flourishing is at 5 years and that's too far from this really massive escalation and it gets really, really expensive and there's bumpiness in between, with an overturn of capital to that, and it gets you pricing corrections and some companies go out of business and some things go bankrupt and get bought out of bankruptcy or assets get bought for dimes on the dollar. But I think that doesn't create the kind of reset that bubble collapses do. I think it's just correction, and that's what I think the worst possible outcome is. I'm not even sure that that is an outcome.
You're ruling out a bubble bursting.
Yes. I like to hear it. It's a relief for everybody who has their 401k tied up. But obviously, there could be real corrections. And that's what I'm saying. It's like smooth sailing guaranteed from November 26 on.
We got him. He said it. No, he promised. He promised.
I think one of the reasons we initiated, oh, we need to have you back on the podcast, was a couple weeks ago you came to the defense of Anthropic. Anthropic had sort of said they're worried about safety. They're like, oh, we're a corporation, but we care about doing good. We want to make sure this AI thing doesn't get out of our control. And I feel like David Sacks, who obviously now represents the Trump administration on AI, has basically said it's purely cynical. He seems to see it as a totally cynical sort of government capture strategy instead of a sincerely held belief that Anthropic seems to think we should be cautious. It's a big question to you. The newest thing is the Trump administration trying to say no state bills. I'm curious as you're at once the AI bull and you share this sort of democratic sense possibility that tech should still be governed some. What do you make of an effort to stop state bills on AI?
Well, so two things. One is, I think Sacks is showing his own true colors because if it was him in the Anthropic thing, that's why he would be doing it. It would always be insincere. That's literally what I tweeted something. He doesn't know a sincere belief when he sees it. Yeah. So I think anyone who has met Dario understands he is nothing but transparent, maybe too much so for his own good health and certainly in the current administration probably doesn't help him to be that transparent. Yes. And then on the states things, I thought Gavin Newsom did a very good job of saying the earlier SB 1047 was not the right bill.
The 53 was good because it was like, look, let's actually opt for some transparency, some measurement, some accountability. That's what allows us to learn, have society in the processes we're doing is not too burdensome either for large companies or small, doesn't prevent channels of invention and risk-taking, doesn't slow things down in global competition. Because, you know, I think focused on competition with China, which obviously the current administration should be more focused on, they seem to want to sell a lot of chips to China.
