[0:00:40] that is responsible for one of history's most pivotal technological shifts. Please welcome OpenAI co-founder and president Greg Brockman and my colleague, the journal's Keech Hagee. Thanks so much for being here. Thank you. Great to be here. So, many of the Democrats who won last night support moratoriums on data centers. What does that mean for OpenAI and its ability to compute it needs to grow?
[0:01:19] Well, something that I believe is that we need to do a better job of really communicating to people, to local communities, about why data centers are something that can be beneficial to them. And I grew up outside of a town, Grand Forks, North Dakota, where we had a sugar beet factory. And by the way, the sugar beet factory smells incredibly bad. You don't want to live right next door to it. But it was really a staple of the town. And I think it was something that had a...
[0:01:46] Created a lot of jobs. It was kind of, in many ways, the centerpiece of the community. And I think that helping people understand not just that data centers, the investment, the community benefits, for example, in Georgia, that we're sponsoring free lunches for students, that all of that is beneficial, but even the output, right? Like the AI technology is also something that is important for them, for their lives. And I think that all of that...
[0:02:15] Communicating it and really substantively living that is something that we need to do a great job of. Were you surprised by how unpopular data centers became over the course of this year? Well, I don't think there's anyone who could say that they were not surprised by that. I think that it is definitely an issue that's very singular. No one's really seen something like it in American politics. And in some ways, though, again, I step back and really think about
[0:02:42] The part that we can play, like what difference can we make? And that a lot of our shift over the course of this year has really been to realize that from San Francisco trying to build infrastructure in Texas and Georgia and other places, you can't just tweet and say like this is going to be good. You need to be in those communities. And some of the most effective programs that we've run have been when our teams are there on the ground.
[0:03:11] And stay, right? Don't show up for a week, go back to San Francisco, but are there and really continue to show up. And so to me, I think that there is something that we have been learning. And again, I think we need to do even better, and we intend to do so. So it's been really scary out there in the world of cyber. And OpenAI's tools recently helped uncover a cyber vulnerability that basically underpins a vast amount of the internet. If companies now have to spend, though,
[0:03:40] Vastly more on cyber to defend against this threat that AI has unleashed. Haven't you just sort of traded enterprise productivity for an endless cyber arms race? How do you square that? Well, my view, again, to really zoom out, is that we as a society, as an industry, have never prioritized security the way that it should. Like, when was the last time that a CISO was over-resourced? And I think we all kind of know the internet wasn't built
[0:04:09] For security, right? That security was added on. You think about the history of trying to even add encryption to HTTPS, to the internet. That was a retrofit. It wasn't how the internet was designed. And so I think we now have 40-ish years of tech debt that has been building, building, building. And AI in some ways is almost like shining a light into the cracks that we have all created. And I think that what AI can do...
[0:04:35] is that we have this opportunity to put the best AI in the hands of defenders so that the defenders can get ahead and actually close those gaps and actually secure the internet in a way that has never been before. And that's something we call the defender's window, this advantage that the defenders have over the attackers. And so I do think that we need to build in a different way, in a more secure way, in a safe way.
[0:04:56] I think we can move from a world where just breaches are just a thing that happens to one where that's just unthinkable, right? That you have real security guarantees, formally verified code, and code that was written in the first place to be secure by models that are just far better than people are at producing that kind of guarantee. One of the really interesting responses I heard you suggest to this was that maybe there'd be some kind of an always-on entity that is scanning for vulnerabilities, and maybe even the government would
[0:05:25] So this is something that, first of all, it's a very controversial concept, but it's also we're not the first people to propose it. I think as you really think about a world that we're heading towards where there are going to be more vulnerabilities found and that every company, every organization needs to increase their rate of patching, and you can just see
[0:05:55] Many companies have a quarterly patch cadence. That is just not going to cut it anymore. People need to bring it down to a month, to a week, to a day, to an hour, to a minute. And I think that that change, it's possible. We have the technology and the tools to make it possible, but it really is going to require collective action. And so if you're in a position where people aren't patching, where people do have servers that can be attacked, then you have a question of, is there someone who can
[0:06:24] Help prevent that from turning into big botnets that are then causing damage to others. And I think that this question of, well, if you do have vulnerabilities, then the government could use the vulnerability to get in and apply the patch themselves. But again, it's controversial for a reason, right? It's a little bit like, well, this server possibly would have some other side effects, right? It's possible you'd actually take down the service. And so I think that this needs to have some really delicate thought. But it's not without precedent. For example,
[0:06:54] The FBI over the past five-ish years, I think, has three or four times gotten court orders to go and clear out malicious botnet software from private infrastructure. So hundreds of thousands of computers, the FBI goes in and actually clears out this malicious code. So it's the kind of thing that has happened. And I think it's really just saying, OK, we see what's coming. We see the shape of these vulnerabilities emerging. And how do we get ahead of that and have a structure that really keeps the internet safe, secure for all of us?
[0:07:21] And have you talked to anyone in the administration about it? I think we've started to kind of float this idea. And I'd say, again, this is not the only idea possibly on the table, but we definitely have been engaging very well with the government on cybersecurity, on the vulnerabilities. PHP, for example, we talked to not just the US government, we talked to many other governments around the world to give them advance notice so that they could prepare. And we've been starting to tell them it's not just this one vulnerability, but this whole
[0:07:51] Sort of mode that we're going to be in for the next six, 12, 18 months. And I think that we collectively as a society need to respond differently to it from really any period before. So since you returned from sabbatical at the start of last year, your portfolio has steadily increased from compute to product to go to market. And I'm just curious, when you got under the hood of each one of these things at the company, what did you find under there that needed to be fixed and how are you fixing them? Well, first of all,
[0:08:21] I think we have an incredible team at OpenAI, and working with them is just one of the best privileges of the job. One of the things that I've found, and maybe the core theme I've brought to each of these functions as we've started to bring them together, is co-design, is really thinking about the end-to-end. You can't just think of any of these functions in isolation, because what AI is, is changing. The research every year, we have a new model that obsoletes everything we've done before. And you can see this with...
[0:08:51] Just the advances consistently across different modalities, across individual functions, like software engineering, totally different today from what it was at this conference last year. And before then, it was barely touched by AI. That's what's really happened in this time. And so, of course, you can't just have an isolated product organization that doesn't talk to research. That makes no sense. And I think the same is true for every function. So to me, one of the sort of...
[0:09:18] The easiest and also most high leverage things I've been able to do is just provide this cross connectivity, this connective tissue and say, hey, on the marketing side, this is kind of what we're seeing coming on the research side. So we should think ahead three steps to this is where we're going to be. Let's start messaging ahead for that so people can understand it. And so I think that this just like really trying to have a through line to everything we're doing.
[0:09:42] Make it simpler and more streamlined in a single direction. That's the general theme that I think has really yielded a lot of the fruits that we've seen this year. So when did you personally realize that Anthropic was going to overtake you guys in revenue? And what did you think was happening there, like when it was happening? So I think, to me, this became fairly clear, like honestly, even at the very beginning of the year. And I think that if you look at what's really changed this year,
[0:10:12] Q1 was pretty rough. Q2 was pretty rough. It really looked, in many ways, like coding, we weren't even in the conversation. We had Codex, but no one was talking about it. And I think that we just kept our heads down and focused and executed. And Q3, entirely different. We had 70% quarterly growth in our annualized revenue. And if you just break it down by the enterprise side, doubling. That's like crazy growth for a single quarter.
[0:10:41] It wasn't an accident. It wasn't something we did in that quarter. It was all the investments that we'd been making up to that date. And so we got a frontier model out in the form of GPT-56 Sol. We got ChatGPT Work released as well. And there'd been all this sort of painful behind-the-scenes work that resulted in this incredible revenue inflection. So what do you know about selling to enterprise now that you didn't know a year ago? I think that selling to enterprise is, first of all,
[0:11:12] You hear so much from every single conversation. Maybe the single biggest surprise that I've had is how every single customer meeting that I have, every time that you're talking to almost anyone, regardless of if it's a CEO, developer, salesperson, it doesn't matter, that there's always some insight in there that I have this aha moment, right? And this like, wait, I've never really heard anyone say it this way before. And so I think that there's all this kind of
[0:11:40] It's like the elephant problem. Everyone's touching a slightly different part of it. And you realize that transforming an enterprise and really delivering value, everyone is doing their part. And I think that a lot of our job is to figure out how do we actually help you figure out what are the most important milestones that can be achieved with our technology? How do we show up for you consistently as a partner? How do we build trust and build these relationships and ensure that you as an enterprise know that we're here for the long haul and are a reliable partner?
[0:12:07] And so I think systematizing how we sell has been a very important part of what we've done. And if you think about where we were a year ago, we weren't even really in the enterprise game, right? We were really perceived as a consumer company. And a lot of the companies we'd start talking to would say, is OpenAI really committed to the enterprise? But it's so different today. And I think that that is due to this continuous investment and the way that we show up to our customers every day.
[0:12:35] What is your enterprise market share now? Any news on that front? Well, so within the company, we were probably, I don't know, like I think our enterprise revenue was about 30% versus consumer at 70%. And now we're probably somewhere around, like we're majority enterprise. And it's not that consumers shrank, right? It's been growing, but it's the enterprise segment has really taken off. And so I don't know exactly how to quantify it in the broader market, but you can see the way that we're winning deals. So for example,
[0:13:04] Actually, one that's hot off the press, has not been previously announced, is that we just papered a partnership with JPMorgan Chase, and that they're going to be using Codex, they're going to be using our cyber models to secure themselves, deploying our models for enterprise transformation knowledge work across the company. And that that, for me, is such an iconic institution in such an important domain, a highly regulated...
[0:13:33] Enterprise that trust is like so key to what they do and that they chose us and that we're here to be a great partner. Anthropic is going public before you and my colleague reported that OpenAI has raised less than half of what it had targeted in its latest funding round. So will you be ready to go public next year? I think so.
[0:13:58] Yes. I mean, look, the way I put it is that we've been very clear this year that we want to go public when we feel is the right time for us, and I feel extremely good about the business. I feel like we've made so much progress this year, and I've honestly never felt more confident in what we're doing. So one interesting little nugget in Anthropic's filing was about infrastructure efficiency. So they reported that their revenue per dollar of compute expense jumped
[0:14:27] 5.2 times from 40 cents in early 2025 to over $2 in Q2 of 2026. So as a person leading OpenAI's compute and scaling engines, how does OpenAI stack up on that kind of metric? Well, look, I would look at, I mean, to me, just at a qualitative sense. Like if you look at what's really going on in these models is that we have been able to make these incredible efficiency gains very repeatedly year after year. For example, we dropped the price of Luna
[0:14:57] 80%, like an 80% price cut this year. And I think that our historical average has been that for a fixed level of intelligence, we probably do a 10x or even sometimes almost 100x price cut year over year on it. And I think that you can look at that from two ways. One is that you can just keep some of that margin. And the other is you can just say that these models, you can get such incredible efficiencies. And so...
[0:15:20] My view of where we're headed is a world where the frontier models, and you can just see this even for looking at open router stats and these other publicly available ones, is all of the dollars basically go to the frontier models. And then you have these lower margin but extremely high token count models in the form of things like Luna. And I think that the margins, basically people are almost price insensitive.
[0:15:47] Sometimes to an irrational degree on the frontier models. People are just like, give me the best thing. And I think it also shows you how much value these are delivering for customers, certainly in many industries, that if you give them the smartest possible model, that they can make incredible revenues out of that. And so you almost can tune your margins almost arbitrarily there. And so to me, I think that...
[0:16:07] The opportunity and kind of the engine and the pace of progress, all of that is kind of coming together in a way that I think that you'll just continue to see these numbers skyrocket. So we have seen OpenAI gain ground on Anthropic really because of some of these price cuts, right? So it sounds like we can expect more price cuts. Is that right?
[0:16:28] Look, I think that the way, like if you just look at the ethos of what we do, like we believe this is technology that can lift up everyone. We want to try to put this into everyone's hands, and we want to empower, transform every enterprise. And you can see that in our technology, you can see that in the products, and you can see it in our pricing. And what we've found is that, for example, when we did that 80% price cut on Luna, I think that the volume grew like more than 10 times or something crazy.
[0:16:56] And so you actually do the math. It actually ends up being pretty favorable. And so I think that we are in a world where competition is resulting in the customer really winning, which is really fantastic in my books. And I think that also what people can use the technology for is shifting. For example, we used, I'll use the term agent fleets, so big fleets of agents to solve the Navier-Stokes problem.
[0:17:23] About a million dollars worth of compute, pointed out one problem. Solved this problem that was out of reach, or at least had been unsolved to date by some of the smartest minds in humanity. Take that same capability, apply it to cybersecurity. Today we announced that we'd found, also for about a million dollars worth of compute, a thousand vulnerabilities in Linux. Produce patches for them. This is the kind of thing where it's like, typically Linux would get like 20 of these a year, something like that. A thousand all at once with this compute.
[0:17:50] And you're like, what else? What other problems can you point a million dollars worth of compute and deliver incredible results that no one has been able to do before? And so that, to me, is the new business, that really trying to not just sort of speed up a little bit of workflows, but really help people solve fundamentally new problems that would have been unsolvable otherwise. So you've been getting a lot of buzz with these always-on products like DOTS. People like it, but it seems very expensive to have some...
[0:18:19] What are the economics of something like that? Well, so I think this, first of all, is something that all of the players in the industry are kind of encountering now. My philosophy is that first, we should focus on the peak experience, right? Even if it's so expensive, just like the most transformative, you have the highest compute model that's always working for you with all of the different modalities turned on, all of those things.
[0:18:48] If it's expensive and inaccessible to most people, it's still worth building. But you don't stop there. Then the next step is you say, OK, we'll make a more affordable version that's more accessible that you can scale it to others. But even that first one, by the way, I think there's incredible PMF. I think there's incredible demand that we're seeing for this AI always on, always working for you. And it just does all these surprising things. Even, for example, yesterday, my dot flagged me a phishing email.
[0:19:17] It was actually a very well-done phishing email. I never would have realized what it was. And my dot just had been reading through my email and was like, this looks so sketch because of this mismatch in the domain. And I was like, that's right. And that's what you want. You want an AI that's there looking out for you, even in ways you never would have anticipated. And that means that you actually have real returns to smarter models that are able to do more.
[0:19:43] What are you seeing on advertising? And is OpenAI eating any nibbles of Google's lunch in that arena at all? So, yeah, our ads business is also something that's been hockey-sticking. I think we announced maybe, I forget if it was maybe June, July, somewhere around there, reaching a billion in ARR, which is, if you think about it, from $0 at the beginning of the year to a billion dollars in annualized revenue on that.
[0:20:12] Time period. It's unprecedented. I don't think there's been any other business that has done that with ads. And I think that, to me, if you look at the value that I think can be delivered through ChatGBT, it's really incredible. And I think that ads is just one of the ways in which we're able to monetize and make sure that we can sustain delivering all of this compute, all of this value to the world. Do you think open weights models are dangerous?
[0:20:43] I really believe in democratizing this technology. I really believe that access to AI is something that's very, very important. And if you look at how we've approached releases, we think very hard about we have a new level of capability, and we have things like on cyber, things like bio, we have different guardrails, we have monitoring, we have all of these things. And I think it's actually been a very sort of, I think, positive world.
[0:21:12] Where the frontier models have sort of a lead on other models. And that means that we kind of can try out some of these capabilities in a more restricted form or in a form where at least we can think about what the implications will be and have a little bit of time. But I think that democratizing this technology in some way, whether it's open weights or others, I think that's really important. I think avoiding a full concentration of power is something that's kind of core to the DNA in how we think about what we're trying to do at OpenAI.
[0:21:40] So you would be against any ban on open weights? Look, I think that for me, it's kind of the wrong question in some ways. I think that the question is really about how do we make sure that we bring AI technology to fulfill its potential, to really be something that lets humans live better lives, empowers us, solves all these problems in biology, drug discovery, all those things, and also protects us from our phishing emails.
[0:22:08] All of that is what has to happen. And there's going to be diversity of different actors, different ways of deploying it. And so I think we just need to look at kind of the whole ecosystem that's evolving and to make sure that we figure out what does it mean to do responsible development? What does it mean to do responsible deployment? And I think that sometimes that the mechanisms are going to vary a lot depending on the need. And I think even, for example, thinking about deploying behind a firewall, right?
[0:22:35] Industries highly regulated. Their data cannot leave the premises. How can we ensure that AI technology can be delivering benefits there? And so I think all these questions are really worth thinking about. So you and your wife pledged $50 million to Leading the Future. It's a pro-AI super PAC. And gave the first tranche of that $25 million. And then you decided not to give the second tranche. Why? So I really believe, and my wife really believes,
[0:23:06] Safety and AI innovation, they go hand in hand. They're inseparable. They're so core to what this technology needs in order to advance in the way that I think will benefit everyone. And rightly or wrongly, LTF didn't end up being perceived that way. It ended up being perceived against that. And I think that...
[0:23:29] There actually are a lot of very positive impacts that have come from LTF. And I think that even over recent months, I think that a lot of the, there's been some very nuanced messaging that is put out around thoughts around regulation and things like that, that I think, again, have been just like very different from the perception. And do you regret giving the first 25 million? I don't think of it quite that way. I think for me, I'm really looking towards the future and just really trying to think about how to really promote this point of like,
[0:23:58] Safety, innovation, they're inextricably linked, and we need both in order to really deliver. So speaking of safety, we had this sort of moment a month or two ago where all of a sudden the world realized that a lot of people who are AI safety researchers believe that AI has a greater than 10% chance of killing all of us in a decade. What do you believe about that? The way that I think about...
[0:24:27] AI, AI progress. What does AI represent? Is that this can be the most empowering technology, but you have to also think about the risks. You also have to think about what are the ways that it could go wrong in order to mitigate those and to deliver on those benefits. And so that's how I look at it, is that I think it's hard to multiply together numbers and come to very certain conclusions, but I think that the way that I've thought about OpenAI from the very beginning, like part of the reason, like I wasn't in AI before OpenAI.
[0:24:55] But I just felt like this technology, it was happening. It's something that's almost the logical conclusion of technology broadly, of compute in particular. And I wanted to see if I could help it play out in a slightly, even slightly more positive way with me than without me. And so I think that really thinking about alignment, safety, security, and capability all as one package and encountering and mitigating and deploying and learning, like all of that to me is the way that we get to the best possible future.
[0:25:23] So you and your wife have both said that AI should ultimately help animals, not just people. What is the single most compelling way you think super intelligent AI will change life for species besides us on Earth? Well, look, I think that you look at something like biology. And that's not something that's unique to just humans. I think that we all suffer the question of even just thinking about
[0:25:50] Death becoming a choice. All this is possible. That having cures for diseases. This is something that's become also a trope in this field. But I think that you're going to see real delivery and real significant strides and impacts on biology, health, and curing of diseases over the next 12 months, 18 months. I think it's going to take time to really deliver on the full impacts just by the nature of this field. But I think that...
[0:26:19] I think it's all possible and doable, and it's exactly the kind of thing we have these models that are solving these unsolved problems. Clearly, we need to apply it to biology and life sciences. So looking five years out, how do you expect the AI landscape to have changed? So five years in this field is quite an eternity. You think about five years ago, right, so 2021, and picturing the world of today, quite hard to kind of sort of put it into...
[0:26:48] Into perspective and detail. But I do think that we should be in a future, in a world where AI is broadly available to everyone. That it is something where people, for education, for their day-to-day life, for you want to start a small business, that the idea of the question of do you use AI or not AI is almost like do you have electricity or not electricity, right? It's like a core.
[0:27:16] And by the way, I also think that we spend so much time today typing on our computers, getting our carpal tunnel, the whole thing. None of it is very natural. It's something that we weren't doing 50 years ago. I think all of that will have changed. I think we will spend so much more time doing this, just like actual people interacting with each other. And I think that people will be much more free to do what they want with their lives and their time. And by the way, I think that people will
[0:27:45] Have lots to strive for. I think that that is going to become, like you kind of see it with the software engineers, where they're actually working harder than ever, even though they're far more productive than they were before. And you just feel that it's like there's so much opportunity. There's so much that you can do. And I think that people won't want to miss that. Well, on that note, thank you so much for being here. Thank you for having me. All right. This is exciting for our final segment.