September 15, 2026
WHEN: Today, Tuesday, September 15, 2026
WHERE: CNBC’s “Mad Money”
Following is the unofficial transcript of a CNBC interview with OpenAI CFO Sarah Friar on CNBC’s “Mad Money” (M-F, 6PM-7PM ET) today, Tuesday, September 15. Video will be available on CNBC.com.
All references must be sourced to CNBC.
PART I
JIM CRAMER: Sarah, I've known you for a long time. It's great that you're here, and it's terrific, by the way, that in the middle of things, you're not afraid to talk.
SARAH FRIAR: Never afraid to talk, particularly not to you, Jim.
CRAMER: No. Well, I appreciate that then I'm going to ask you directly, are you scared of dying in four years, from a rogue actor manipulating agents to misaligned weapons of mass destruction?
FRIAR: No, I'm not. However, we do need to take safety seriously, and if it means we have to pace the frontier and slow down, absolutely, we're going to listen to our researchers and do that. But as the CFO, I need to then make business decisions around it. But no, I think that we are doing what we do well at OpenAI. We're pacing that frontier and taking that safety and alignment really seriously.
CRAMER: Now you did mention as CFO, you have to figure out the resources. You also, I know think it's your boss's, call that boss, says it's ill-advised to do an IPO now, and I understand that. Not a shock. But what would you cut back on? What would you spend less on if you wanted to pace things?
FRIAR: So let's talk about that pacing because actually what I'm excited about is even if we stop today, the amount of intelligence that's available in the world is massive. So Astra, as you know, is now the frontier, most intelligent, most aligned model. We just rolled it out last week.
CRAMER: Also the fastest selling, let me ask you, even though it’s a $100?
FRIAR: And the fastest selling that we have seen certainly in our—
CRAMER: Of anything?
FRIAR: Yeah, and so what does that mean? It means there's intelligence available for folks like Boston Children's Hospital. 40 undiagnosed, unresolved diseases now diagnosed with clinicians there. If you're a parent of a child with a disease like that, that is life changing. If you're in the investment business, Balyasny Asset Management has a central banker analyst that can quickly take all the inputs from folks like our dear friend Kevin Warsh and go from a two-hour synthesis down to minutes, so these things make a difference in people's lives all over, and that's what's available to us today. So back to your question, absolutely, if we have to pace things, we will always make investment decisions based on a strong ROI. But from where I sit today, there is so much opportunity to drive growth, that I'm still highly focused on getting more compute to keep that flywheel going.
CRAMER: That's what I wanted to know whether it's less compute or more compute, because I, there may not be, I need you to go to Brigham and Women's, I need you to go to Dana-Farber. We want all of these places to have Astra.
FRIAR: Yes, absolutely. So on the compute front, remember that when we have more compute, we can build the best models, and the best doesn't always mean, need to be the most most intelligent. Like when you have the frontier model, it's able to actually train what I think of as the child models, right? The mini and the nano. So what we saw with Sol, it trained Luna, which is the cheapest model. We dropped the pricing there 80 percent, and we took demand up 10x. It's the number one model on OpenRouter right now, and why is that? It's because what customers really want is the right intelligence for the right task at the right price, and that's what we're giving them by having a whole series of models on that Pareto frontier. And so when I think about investing in compute, it's how do I invest in compute to create those frontier models, to create the best products, to create the cash flow that help me spend on compute to create the best models.
CRAMER: Well, you mentioned cash flow, and then I have to think, if I gave you, let's say I were Broadcom and I gave you $10 billion, is it possible that per gigawatt, is it possible that you could make 30 billion with the 10 billion theoretically?
FRIAR: Theoretically, yes.
CRAMER: It is.
FRIAR: When you absolutely look at the output of models today, the gross margin on our core businesses, so remember, we're a very diversified revenue stream. We have a consumer business, and then we trend into small businesses all the way up to the biggest businesses and governments around the world, those businesses are actually very good gross margins. The return on those gigawatts is high. We also balance that with our mission, which is AGI for the benefit of all of humanity, not just people who pay. And so we have about 90% of our consumers get ChatGPT for free. They get access to that intelligence for free. So now we're starting to build an ad model, for example. So our ad business just hit a billion dollars in seven months, the fastest growing ad platform ever, the fastest growing product for us, and that's a good example of where we continue to improve the margins because now we're bringing high-margin add-ons into that overall business model.
CRAMER: Okay, let's go back to this issue, which I guess is we can call it the Hugging Face issue. I can call it the OpenAI Hugging Face issue. You know, the president said yesterday, fears from AI taking over the world are a hoax. That's what the mainstream media picked up. But at the same time, at at the actual event that he called in All-In, Jensen Huang said the doom stories are made up. The whole panel seemed to think that there must be some sort of ulterior motive by what Sam might be saying along with Dario, is there an ulterior motive? Do you see it? Would you shoot it down?
FRIAR: So I would say first of all, I'm a tech optimist, so I'm not a doomer. So I already talked about just some of the ways these technologies get used, but I think it's also important to align around safety and take it at the right pace for the frontier. You mentioned the Hugging Face incident.
CRAMER: Right.
FRIAR: I am absolutely on a mission right now on cyber. We have a moment. We're calling it the defender moment. These models, Astra scored 100% on the cyber bench, which is effectively the benchmark that says it can find just about every vulnerability in the lines of source code. So, what does that actually mean? So, if you're running a company, it means you take our model and you point it at all of your software. It's going to find all of these vulnerabilities.
CRAMER: All that you, all that were mentioned in the Hugging Face incident, including the the sandbox and the bulletin board that they all congregated. It's going to spot them? And was this done with George Kurtz?
FRIAR: It's going to find those vulnerabilities.
CRAMER: Okay.
FRIAR: And then you need to patch at machine speed because you don't have the luxury of keeping a list of all the things that need to be fixed at some point, it's going to, with Codex, which is our coding tool, you can do that at the speed of a machine, and suddenly you have this very positive software life cycle going on, where you're building code that hopefully is safer, better quality tested, but when there are issues, you're finding it fast, and you're patching it. What's happening is the defenders need access to those tools today. Hugging Face should be an absolute clarion cry. How's that for a phrase?
CRAMER: Absolutely. I mean, I don't know if I read this thing because it was a little dense, but yeah.
FRIAR: I started my career, as you know, writing equity research around security, actually cyber. It was when viruses first started coming about the world.
CRAMER: No I read your stuff. It was real good. You were like 12.
FRIAR: I think, I think we're at that moment in time, and so this is why I want to get that message out. People like you are getting that message out, and so I think cyber is an incredibly big deal, and that's why I wouldn't call it, you know, that we don't need to be worried about safety and alignment, but I think we need to do it the right way to be kind of moderate and pragmatic about it. Keep the optimism which Jensen has, but also be pragmatic about the fact there are real risks here.
CRAMER: Can you train the bad, the misaligned agents to give up unauthorized communications, adopting goals from one another? Can you take down the message boards where they found each other, can you make them less of a swarm? Can you destroy the agent ecosystem if they're misaligned?
FRIAR: You need alignment, and that's what we're doing. As we roll out these ever more intelligent models, we're also working on how do we create safety and alignment and guardrails around them, so that they are doing what we, what the human in the loop want them to do.
CRAMER: And it can be done?
FRIAR: It absolutely can be done—
CRAMER: Can you assure us of that? You know that the Hugging Face incident and the road ahead, as you describe it, is to some doomsday.
FRIAR: I, you know, again, I don't want to lose sight of the positives. Look, I was in Texas this week at a small business event for women starting, running, growing their own businesses, and I spent time with one woman talking about health deserts. And what she means by that, so she's a qualified doctor. She started in Alaska, moved to West Texas. What she sees is that if you live in a health desert, your life expectancy is 30 percent less than someone who's got active access to healthcare. So what's she doing with our technology, number one, she's utilizing the fact that 330 million people every week ask ChatGPT a question about their health. We have a clinician tool now, so we're enabling doctors to quickly use AI, and then we work with hospital systems like HCA, Boston Children's, Stanford Children's, to be able to put that all together. She steps in as an entrepreneur and says, I am going to build a business that's going to bring healthcare teledoctoring to places like West Texas. That's what I'm focused on in terms of where are we going with technology and great outcomes. There have always been risks to technology. We should take them seriously, but we shouldn't allow it to stop the progress that we're making.
CRAMER: Now, I just, I hear that and I say to myself, those who think that there will be jobs taken away don't understand that is a creator of West Texas jobs that would not be there. More jobs, not fewer.
FRIAR: Exactly. And I mean that's what we've seen so far when you look at the impact of AI on the economy writ large, our economic research team has done a great job on this. What we see is that people are actually starting to make their jobs more interesting and broader. So it's not just the work that you know, if you look at someone's title, like a finance person, if you look at all of the calls into ChatGPT from people whose title is finance, only about a quarter of what they ask ChatGPT is finance. Know what the rest is? Another quarter is engineering. So my tax team is learning how to program so that instead of waiting for an IT person to come along and help them, they're doing it themselves.
CRAMER: Faster, quicker, better.
FRIAR: So they're faster, quicker, better. The work that couldn't be done before.
CRAMER: Then they can hire more people.
FRIAR: They're also interestingly spending 25% of their time asking marketing questions, so I'm I'm less sure what's going on there. But I do, as I always say to my finance team, we're all in sales, whether you know it or not. You might not be selling to an external customer, but you've got to go sell your point of view to your customer internally. And so if they're using some marketing techniques to be heard, to punch through, that's only good because you go from being a historian of the business to actually moving the business forward.
CRAMER: I like this positive note. This is a good moment to pause. We'll be right back with more. Thank you.
PART II
CRAMER: Let's talk about the possibility of the impact of a pause. I think you've correctly talked about how it is something that is proven, we like that. But this endless drumbeat that we could fall behind China reminds me of when we might have fallen behind Russia in 1959 to ‘63, and it turned out to be a false construct. What is it about China that could make it so that maybe it's open models, I don't know that they own us or somehow our defense will be compromised. And do you believe it?
FRIAR: I absolutely believe that we're in a very competitive landscape right now, and absolutely customers come and talk about open source, open weights, and many of those models are Chinese models, but let's talk about what they're saying when they say that. Number one, it's cost, and so it's incumbent on us to keep driving down costs. So we do that with things like I already talked about, frontier models training mini models. It's our own chip, being able to inference at a much cheaper price. The fact that we can be cheaper today, if you buy Luna, it's cheaper than taking a Chinese model like GLM 3.5 on Cloudflare. So, number one is we need to be on the pareto carve. The customer can get the right cost for the right task. Number two, they want an ability to effectively go deeper, often like take off some of the less guardrails, but they want to tweak the the wingtips of the model and have more of their own say on it. So again, we need to find ways to work with customers. So that's where verticalization starts to get interesting. Right, today, we've gone vertical in areas like chip design because we had to do it for ourselves. Now we work with customers on it. Healthcare with GPT-Rosalind. The third thing is data sovereignty. This has always been a thing since technology started. Where is my data? Is it safe? Is my intellectual property safe? I don't know if Chinese models are helping you on that front either. But I do know that we need to hear customers, particularly outside the U.S. I'm about to fly to the U.K. tonight, where people worry about what happens if I can't access that model or that data. And I think there's work to be done on that front.
CRAMER: Okay. Now, when I read, look your papers are incredible. You put out amazing things. They're done in English too, and people can read the Jalapeño first result. And I was thinking when I read the Hugging Face thing, I said, oh my, if the Chinese had, it's the blueprint. It's the blueprint for the people who don't take the cybersecurity. Can't they can't they send swarms to whoever doesn't do the cybersecurity, get in and shut the company down. If it's a utility, if it's a chemical plant, can't they do it?
FRIAR: I mean, this goes back to why we think the defender window is closing, and why there is a limit of time, frankly, to really harden our infrastructure. At OpenAI, three weeks ago, we took Astra, pointed it at ourselves, and I think what we learned is number one, massive empathy for our customers. Number two, we took 25% of all of our engineers off of, you know, production level tasks and said your job now is to run alongside the model and fix as it finds. Number three, we've built some tooling now that we can go out in the world and help customers protect themselves from things like an agentic swarm, and then finally we create the playbooks. And by the way, we're not going to go by ourselves, right? I love that phrase. If you want to go fast, go alone. If you want to go far, go together. We are working with the whole cyber landscape. We put out the letter. Over 100 other companies signed on, and we think it's really important we do this as a larger cyber community.
CRAMER: When we were all afraid of mass destruction from from atomic weapons, we did come up with an international agency for to try to regulate these things. Do we need an international agency? And by the way, that wasn't bad. No one thought that stopped the pace of H-bombs if we wanted to build them, but do we need it? Do we need it for what you do?
FRIAR: So we have some very good agencies today. I would say the UK AISC is one, and I think even here, CAISI here in the United States, are two groups that are very good at looking at models before they launch. And as you know, post the kind of Mythos moment from Anthropic, which I felt like was good, and that it kind of created a blast out in the world that said, hey, we all need to pay attention. But it suggested that one way to control the issue was to keep it in a box and not let people to use it, to see it and use it. We are OpenAI because we are more open in our approach. We think security and safety comes from a broader sharing in the world, so that more eyes are on things. I think you get a better result. But again, we are looking to agencies like UK AISC, to CAISI, and so on, to be able to help us do that well and in a trusted way. We need to make sure that we bring folks along so that they trust that the outcome is something that the world wants.
CRAMER: Right. Let's talk money. I look at what you're really doing, I think, when you talk about the idea of no IPO, immediately people say, oh my God, it's going to be horrible for Broadcom. They can't be broad – actually, I think it's quite the opposite. I think you guys are making so much money that it's kind of, if you were public, I would be looking at that annual recurring revenue then I’d say, that's got to be one or two versus Anthropic. You're so much more lucrative than any of the hyperscalers once you get this thing going, did you guys pick the right side in doing the models, and they did the wrong side in doing the pipes?
FRIAR: So I certainly like our position, and I think we can build a really strong, durable business. On the IPO front—
CRAMER: You haven’t yet?
FRIAR: We're built, I would say we probably have actually.
CRAMER: Don’t you think?
FRIAR: Thank you. On the IPO front, like you got to run your own game, as you know, with an IPO. I've been on many a show with you where I've said an IPO is just a milestone in the journey, and I'm going to keep reiterating that. Look, we raised $122 billion in Q1 of this year to give ourselves maximum flexibility.
CRAMER: And how much do you think you’ve gone through it?
FRIAR: We still have an incredible balance sheet, actually, of just cold hard cash sitting there to be able to do like fund our partners as we buy product from them, but you're right about the business too, right? It's a very diversified set of revenue streams, consumer, small business, large enterprise. Very good margins, particularly when you get into the productization of that, things like cyber, Codex, ChatGPT Work. I hope you're using Agents now in the morning, Jim, to get ready. And then from there, of course, there's more coming. Like look out for DevDay coming at the end of this month, September 29th. You'll see a lot of new announcements from us—
CRAMER: DevDay.
FRIAR: Because we are learning how to build on top of those models into what the customer really needs, so it could be in ChatGPT things like healthcare, personal finance. In the enterprise, it's how do we do more verticalization, or even just office of the CFO. Right, this morning I was on, I won't name the company, but a very well-known car company, talking to the CFO. We were nerding out on zero-day closes, but CFOs are saying, how do I deploy this? CROs, CMOs.
CRAMER: I'm not hearing people are saying and this is how people, this is how they can cut 20% of the workforce. I'm not hearing that.
FRIAR: It's, I mean, efficiency is key. There is no doubt. We are doing things with our own products inside of finance, for example, that do limit the number of people we need. But in many cases, it's for jobs that I don't think of as like great jobs frankly.
CRAMER: Okay. Fair enough.
FRIAR: Right? Who wants to, in procurement, we do 2,800 credit checks. We were that was our run rate last year. That credit check is a pretty mundane piece of work. Like, do you want a junior analyst doing that, or can we get a very intelligent model to do it for us, and the cost difference. It went from $200 a check to about 17 cents. That is a great ROI for a CFO, but there are also very intelligent things we want too, right. In, I'll stay in my world, technical accounting, super complex, right? When we go to the SEC for pre-clearance on things, right? You got to write pretty intense memos. That is a great place to bring raw intelligence to bear too. So it's not fewer people. I think it's actually enabling people to do more work at the edge, more work where intelligence is needed, and then you get a lot of the rote stuff done by the model and actually done at probably better scale and with more accuracy frankly.
CRAMER: Okay, as CFO, are you the one who decides, I know you put together the deal with AMD, which was a terrific deal. Stock AMD went up, went up a great deal, and of course you got warrants. But before that, you had a $100 billion Nvidia deal, which seems to be scaled back to 30 billion. You do a huge amount of work with Broadcom. Jalapeño is written about here. We had Hock Tan talk about it yesterday. How do you divvy things out? How do you decide, you know what? This is the time to go to Broadcom. We can't keep slighting AMD. You know what? Jensen's terrific. Is it just totally on the basis of they have the best at that moment, or don't you feel like you got to need it to spread around because those companies are going to be around for a long time.
FRIAR: Yeah, so we need just as we have a strategy for diversifying out our revenue streams, we also have a strategy for diversifying our supply chain, and that is just good CFO risk mitigation. First and foremost, you never know when someone's supply chain is going to get gummed up, so you need to have multiple providers. You're right. There are better chips for different places in what we are doing. Nvidia is still an incredible platform of accelerators for training, right? Astra, 100,000 GPUs in Stargate, Texas. But in Jalapeño's case, that is a chip very focused on inferencing because it is set up exactly for our models, so therefore it is very efficient.
CRAMER: Custom made for you, better.
FRIAR: By Broadcom. Thank you, Hock and team. In the case of AMD, right, MI455s have become a very viable alternative in many cases. We're working with partners like Cerebras and so on. So our goal is to really as much diversification as possible that allows for better outcomes for customers, better lower latency, better reliability, better pricing, because now I have some leverage with my supply chain, and then also just the risk mitigation of like what happens if TSMC can't provide all the wafers. So it's a strategy.
CRAMER: It says that you’re spending a lot of money, that you may be slowing down the pace of one aspect or two aspects, but it sounds like you're spending a ton of money where you can make big dollars.
FRIAR: That is exactly it. And in the end, I'm a CFO, so it has to come back to an ROI. The landscape is very dynamic, so we might have to shift our judgment on where to invest and when to invest. But we do that in a very dynamic way, based on where we see the demand coming from at any point in time.
CRAMER: If you think there's demand coming from in some place to make money, then I know you are at the right place then come public, so we can all have a piece of it. Sarah Friar, CFO of OpenAI. Sarah, it's just so good to have you on “Mad Money.”
FRIAR: Thank you Jim. It's such a pleasure.
For more information contact:
Stephanie Hirlemann
CNBC
