OpenAI CFO Sarah Friar：IPO、AI竞争、新设备，以及在算力上投入1000亿美元以上

We're going to get right to it. You have just completed what I regard as the most successful fundraising round in history.

We're going to raise actually north of a hundred and twenty billion dollars. We think AI is the biggest era that we've seen today. We're just starting to understand what it's going to mean for global productivity and with that, you know, hopefully more affluence, better lives for everyone. Luck is whatever the preparation meets opportunity, but you got to grab it. Long time listener, first time caller. Quite exciting. To get to hang out with all the bros here. Hello.

We weren't sure how to start this off, but I thought the best thing was to allow our erstwhile crypto czar to say a few comments.

I saw an article today—I think it might have been in The Wall Street Journal—that the perception is there's an advantage to going public earlier if you're an AI company. Now we know SpaceX is going public, and the question is when OpenAI and Anthropic are going to go. I'm curious: How do you think about that? Do you think there's a little bit of a race on, or you haven't made a decision about that yet?

In the end, an IPO—I say this to the team all the time—is a milestone. It is not a destination. Do not run your company as if that's some sort of destination. It's just another way to fundraise. We just did—you heard me on the sizzle reel—raise $122 billion in March, and that was to give ourselves maximum flexibility.

I feel like my job as a CFO is to create optionality for not just this company, but this era that we're living in.

Was that fundraising round the biggest private or public one up until the SpaceX IPO?

It is.

Right.

It is by orders of magnitude. I think the largest IPO to date was Saudi Aramco, which was about $30 billion. So it is actually incredible that you're going to have potentially 3 IPOs at a scale that will be bigger even than 2000 and 2001, that time frame. There was a lot that went on in the market, too, but the market has grown.

By the way, the other thing going on in the market is that if you look at buybacks, M&A, and so on, a lot of capital keeps being returned to shareholders in cash. So there is a lot of money sitting on the sidelines.

But in the spirit of the question, David, I think in the end you'll be measured, right? In the end, the market is a weighing machine, not a popularity machine. No one remembers who won first, Google or Yahoo, Lyft or Uber. I say that not because I want to be first or second, but I just think the press loves a bit of drama. In the end, we're going to have to build big, sustainable, durable companies, and fundraising will be a key component of doing exactly that.

Sarah, breaking news.

Oh, my God, so many people coming at me. Hi, Jason.

I know. It is hard balancing 4 interviewers at the same time.

This is my world, by the way, so I'm good with this. Jason.

Anthropic just confidentially filed its S-1. Does that mean you're in third place in terms of the filing?

It does not mean anything yet, because you have to run the gauntlet of the SEC, and who knows how long that takes for anyone.

Is there, though, a benefit to them going forward? I think unpacking the rivalry with Anthropic is on everybody's minds. I guess you can't talk too much about IPOs, so I'll just pivot to Anthropic. They were far behind, and now they've really—I think everybody would agree in the industry—blown past OpenAI in terms of developers and corporations, and it seems, revenue-wise, as well. How did that happen at OpenAI when you had such a tremendous lead? How did Anthropic blow past you guys?

Let's talk a little bit about our strategy. Our strategy is different, right? We are building the AI layer, the infrastructure, and it's really important that there's a single foundation, but then with many interfaces out into the world.

ChatGPT is one for the consumer. Over 900 million people use ChatGPT weekly, and it's become the noun and the verb. It's how most people experience AI for the first time. A fun fact: Our economic research team just showed me that the fastest-growing continent now is Africa, probably not totally surprising since it started from a small base. The fastest-growing languages are Azerbaijani and Kazakh.

Kazakh.

Which is kind of incredible to talk about where it's going. So, multiple interfaces: ChatGPT, of course, Codex, which just hit 5 million over the weekend. We're really proud of that, coming from almost 0 in January.

Users.

Go, Codex. Help me prepare for this little special up here, too. There's, of course, Frontier, our enterprise offering, and every other way that we can get out there to reach businesses of all sizes.

That is a very different strategy. We think that because it's served up on 1 model, there's a compounding element of advantage that comes from that: More users, more data, more ability to personalize. ChatGPT acts as a front door. As models get bigger, there's more efficiency that should lower the overall cost to give you a token in the world. That should compound to higher gross margins, ultimately more ways to pay for compute, and access to compute is one of the really big competitive advantages at the moment.

We all have to run our own races, but we also have to recognize we're part of an ecosystem that needs to bring people along collectively.

Did you spread yourself a little bit too thin, then, with too many projects? People are talking about this new gadget, Sora, and maybe there wasn't enough focus on enterprise. Is that a fair assessment? If there was a mistake in the last year, was that it?

No, I think the world loves to go to binaryisms. Are you a consumer company, Sarah? Are you an enterprise company? The reality is we're very much both. We're not one or the other.

Right now, our revenue is getting pretty balanced, about 50/50. We are incredibly focused on the enterprise. I spend so much of my time with companies. Just even in the last week, I've been to see Thermo Fisher in Boston. I was with a bunch of banks in New York. I was on the phone with Travelers on Friday. I spent this morning on the phone with a tech company. It doesn't matter the vertical; people are really moving on AI right now.

Our new head of revenue, Denise Dresser, has been in seat since December. She is a force of nature. So I think the enterprise, broadly speaking, is really firing on all cylinders.

But we don't want to leave the consumer behind. Remember, our mission at OpenAI is AGI for the benefit of humanity—not for the benefit of humanity who can pay, or for the benefit of humanity who live in an enterprise, but very broad-based. It's why we offer so much for free, because we want people to get a taste. Once they get a taste of intelligence, the ability to move up the commitment curve is incredible.

Our free users do about 7 questions a day. Our first paid tier does double that, about 15. Our real paid tier, the Plus at $20—hopefully you're all on it or higher—does about 3x. And Pro is about 11x over a free user.

Remember when you got your flip phone and you're like, “Yeah, I don't know what it does—makes some calls”? Now, that same phone—think of all the things it does for you. That's the path we're on with intelligence right now. Sorry, Sam.

You said something very influential, I think it was about 18 months ago, for a lot of us in the industry, where you framed a very simple economic trade-off, which was gigawatts to cash. I think you said 1 gigawatt is roughly equivalent to about $10 billion a year of revenue to OpenAI.

So, comment number 1 was this: 1 gigawatt equals $10 billion a year of revenue for you. But it's not just you, because you can probably extrapolate that to Anthropic and other folks, Gemini. Then you were really at the forefront of getting access to power, data centers, and powered land. It seemed a little crazy, but now it looks like, “Hold on, there's a huge deficit of supply.”

Can you unpack all of that and explain both the spectrum of where we are and those specific economics, and whether that's changed?

First of all, yes, compute is a very scarce resource at the moment. What we see in our business is that we're going up that kind of vertical wall of demand right now, and there's just not enough tokens available. I'm very grateful that I got to work alongside Greg and Sam. I think we really pressed hard on this.

Last year, we were definitely taking some arrows in the back about, “Why are they out there buying all this compute?” I think, thank God we did, because in 2026 we still won't have enough compute.

Where are we on the compute continuum? There are chokepoints everywhere, and I think they will continue to move back and forth. You all talk about this and know this as well as anyone here: whether it's energy, first and foremost; land; power; or how we get regulatory environments such that we can build quickly.

When you get into the racks and chips themselves, clearly, do we have enough in that supply chain? The memory spike is on at the moment. Access to great talent. Do we have enough people coming through our education system? I really worry about this right now.

I'm a trustee at Stanford, and I see that we need to keep the focus on education and science. And then trust. I actually put that as part of the supply chain.

Sam right now is in Saline, Michigan. He’s going to be cutting the ribbon in about 2 hours. So you’re getting a sneak preview, but they told me it was okay to say it in the room. That will be us sticking shovels in the ground on a 1-gigawatt data center, which is part of our Oracle complex.

It’s really important there, on the trust side, that we don’t leave communities behind. I spent 7 years of my life working at Nextdoor doing the hard work of what it means to be local. You cannot tell people from the top down what they need, because they will tell you, “Thank you, but no thank you. I will tell you what I need.”

In a data center like that, we’re actually spending a lot of time in the community saying, “Number one, we’re not going to raise your electricity bills. We’re going to pay for our infrastructure and our power. It will not be the ratepayer that has to pay. Number two, we’re going to bring jobs: 2,500 union jobs. Good jobs, like electricians, HVAC, and so on. We are going to pay our taxes—a billion dollars in taxes just for that data center into Michigan. And on top of that, we’re going to invest $45 million in education for Codex credits, to do what you all talked about this weekend: anyone who’s not coming in facile to their new job.”

I have teenagers using Codex. It would be like—I would never hire a finance person who didn’t know how to use Excel, and I probably wouldn’t hire a finance person today who doesn’t know how to use a tool like Codex. When I think about investment, we’re having to invest ahead of demand. That means we need to both be able to find all of the compute and all the pieces and then pay for it. So that goes back to your capital question on IPO.

On the other side, on the economics, look, the economics do continue to get better. They’re getting better on multiple fronts. I think we are doing a better job of actually showing true value to our customers. I think you get beyond a cost-plus type of pricing into something that feels more akin to the value being created. Now, scarcity of tokens helps, because it’s causing a bit of a compression in time.

Talk about that, just without specific names: where does the landscape exist today in terms of all the power that’s available and all the demand that exists across everybody?

Yeah.

What’s going to happen over the next year, just at the current course and speed of what is available? Of the data centers that are available, of the tokens that are available, of the infrastructure that is available for everybody? I told this story last week, but I’ll use Anthropic as an example. One of the frustrating things is, at some point, it just says, “10:30.” It’s like, “All right, Chamath, see you at 2:30.”

Yeah.

And that’s not a viable experience.

Right.

And in fairness to ChatGPT, actually, I’ve never had that with—

Yeah, we’re quite generous with our tokens, and again, on purpose. We’re trying to drive access so people understand. If you’re on that free tier, you’re not actually getting the latest model, but we’re trying to put it in your hands so you get a sense for it.

By the way, if you’re a kid doing homework, I think about when I grew up and the Encyclopaedia Britannica showed up at the front door in Northern Ireland, in a tiny little community in the middle of the Troubles. It was like the clouds parted. We want to make sure that people get that feeling.

The landscape right now, in 2026, if you want to buy more compute, good luck to you. Tell me, because I don’t know where else to find it. I mean, as you know, Elon, ironically, ended up being the one person who had too much compute, in a way. But good job figuring out how to sell that off. In 2027, it’s pretty limited as well, frankly. Now, there are a couple of things shifting around.

When we talk about compute, there’s training, which mostly still all happens here in the United States—for U.S. government reasons, and to make sure that a national asset, in effect, is happening on U.S. soil. For inference, we want that to be global. I think, particularly in an agentic world, you want much more real-time interaction.

Even for things like Sora and video, which, by the way, we had to make a really tough choice on because we didn’t have enough compute—

And it uses a lot.

Right now, yeah, video does. But video is not over. In particular, when you start to think about where AI is taking us into more multimodality, remember, we’ve all been taught by the last generation of technology to talk with our thumbs. It’s a disease. You walk around, everyone’s looking down; they don’t look up anymore. Teenagers sit on my sofa at night and talk to each other with their thumbs. I’m like, “Who are you talking to?” And then my son will be like, “Him.” I’m like, “Okay. Talk.”

Multimodality is here. Hopefully—I think you all talked about it this weekend—you’re talking to your tool. I talk to Codex every day. That is changing rapidly, but it’s going to need much more real-time compute, because it’s an odd experience. If I was talking to Chamath—

Jony Ive, this puck and earpieces. So maybe tell us a little bit about that project. You’ve admitted it now.

If I tell you it’s an earpiece, Jony will come and steal my teenage son. I might give him to him. But we are changing into a consumer substrate. I cannot tell you what it is, but by the end of this year, we will unveil it. Early next year, you’ll be able to buy it.

I have seen it. I’ve tried it. I’m a hand talker. Right now, I’m sitting on my hands.

Did you have a paradigm shift when you used it? Was it like having an iPhone for the first time?

What Jony and the team are really good at is bringing humanity to devices, and I don’t really know how to explain that well, but when you see it, you feel it.

It feels natural in some way?

It feels very natural, but it feels very lovable.

Really?

And I can’t really explain what that emotion is, because it’s so much—

Intimate in some way, in terms of—

Technology is—

Not taking your phone out, and it’s seamless, is what I’ve heard from people who’ve played with it.

Technology can be very mechanistic, but we all know great design just makes everything fade away, right? At the time, you know, the simple is hard.

Yeah.

I think this is a drop set.

This story, just going back to the earlier question: putting on the CFO hat, help us understand the capital allocation model that you use. A lot of businesses over the last decade, 2 decades, that have been these outsized returners have found some unique way to deploy capital at a higher ROC than anyone else, and then you end up plowing all your capital into that higher-ROC bucket.

What is that for you guys, and how do you think about that portfolio approach to having more of these big-returner shots? Is there an engine where that gets better over time?

There has to be, because in the end, the durable, high-value companies created in this era, I don’t think they’re going to be magical. They’re going to look like the great companies of prior eras. They’re going to create customer value. It starts with the customer and really helps the customer do something different, better, more revenue, more efficiency, right?

Thermo Fisher wants to be able to get patient screening done faster so they get FDA approval faster. That’s really important. If you have a form of cancer where you have weeks to live, the difference between a breakthrough in 4 weeks and 2 weeks can literally be life or death.

They also have—I’m going to misquote this—but something like 38,000 people in the field selling those amazing devices. If you walk into any lab in the country, you’ll just see Thermo Fisher plastered all over every device. Those people want to be more efficient going to work. The fastest takeoff of Codex within OpenAI right now is actually in our go-to-market team. Our developers are there, but if you look at the pace of growth month over month, it’s all in GTM.

They want more productivity out of their GTM team. And, of course, they’re doing things in areas like finance, which I get really excited about. So, customer value first. From that, now you need to get to a great gross margin.

How do you get to a great gross margin? You’re looking at the cost of revenue. The main input is compute. The good news on compute is that there is a massive deflationary curve on cost. From GPT-4 to GPT-5.4, I think the reduction in cost was something like 97%. It’s kind of an amazing curve. That happened in 2 years. It’s kind of wowing, right?

That’s it.

Even our newest model, if you look at GPT-5.5 that we just released, we’re trying to now translate that back to the customer. So we actually raised prices on GPT-5.5 2×. But if you look at what the cost to the customer is, they’re probably still getting a break of about 20% to 30% in cost per token, because it’s just much more efficient per token. There’s a lot to do in that envelope.

Yep.

Part of making a capital allocation decision is having to—if you make it on today’s cost profile, you actually might misprice the outcomes. You have to lean in a little on the cost profile.

And then, as we think about the builds, you are having to make—my focus today on compute is, what’s the compute I can buy for 2028 onward? That Michigan data center in Saline, I don’t think we will be getting compute out of it until probably the end of 2027 or early 2028. So that’s where you’re starting to make your bets.

And in fact, where I feel most short of compute right now is starting to look at 2030, 2031, and 2032. So, you're having to create a business model. The good news is that each year goes by, we get more confidence in the build. We're seeing it massively outperform, and so that's giving us more and more confidence. The market is coming toward us much more.

All right. So, how are you making the compute-need forecast multiple years out, accounting for all of the architectural and model advancements that are happening, where quality, value, or utility per unit of power is going up? Help us understand how you estimate that, given that there's a lot of technology development going on that has a high degree of variance to it.

Yeah, yeah. We do have to make multiple assumptions, both on the compute itself. We assume right now that compute, actually on a per-gigawatt basis, is getting more expensive because power is getting more expensive, memory is getting more expensive, and so on. However, the intelligence that we get on the other side because of the depreciation on the chip side is more than making up for that. In terms of a per-unit cost sold to a customer, it should actually get a lot less expensive for the—

Improvement in that.

Yeah, exactly.

Yeah, exactly. That's just the chip itself.

We don't want to overestimate on the model side, because sometimes GPT-5.5 is an incredibly good model on the efficiency side, but if you look at something like GPT-5.4, the prior model, it was a really large pretrained model. It was very expensive. It was actually hard to serve. And sometimes we want to do that really big pretrained model, and then we take multiple model turns to be able to drive down the cost.

In the near term, in 2026 and 2027, I clearly build a model that's bottom-up. I know what my products are, and I have a sense of what the pricing will be: P times Q. How many WAUs do I think I have? I can see what the shape of the line is. How many of them will subscribe? Advertising coming in is also still related to how many weekly actives, how many dailies, how many messages, and so on. So, you can do a pretty good modeling job in 2026 and 2027.

That said, the shape of the line keeps taking us by surprise to the upside. When you get into the outer years, you're actually looking more at the compute you've bought and almost just doing an algorithm the other way that's saying this amount of compute should equate to somewhat this amount of revenue. I don't know for certain exactly where it will all come from.

A year ago, I built a model for investors that showed agentic revenue. The story was, we're going to have this thing, we're going to be in the agentic era, and we're going to hand it to a developer with natural language. They're going to be able to build, and we think they will pay upwards of maybe $2,000 a month for it, which is kind of laughable in hindsight. But nobody believed it. They were like, "I don't even know what she's talking about. There's no way that will happen—and $2,000 a month?" Remember when people were losing their minds over ChatGPT Pro being at $200? "Oh my God, no one will ever pay for that."

Right.

Yeah.

So, why $122 billion? Does it take you to 2031 or 2032? How do you get the calculus on the capital needs as you do that modeling?

And you'll maybe get more specific. The estimates I've seen are that to stand up 1 gigawatt of AI compute costs about $50 billion in capital: land, power, shell, chips, everything—all in, around $50 billion. Do you have to front all of that money when you create a new data center? Or how much of it do you do? How much of it can you get debt for? Does a $100 billion raise only get you 2 gigawatts, or does it get you 5? What does it get you?

It's a great question. If you look at our compute strategy, it's crazy how fast the world has changed. Just 2 years ago, we were literally at 1. We had 1 CSP we worked with, Microsoft Azure. We sat on 1 chip, NVIDIA. We had 1 product, ChatGPT; 1 price point, $20 a month.

I often use a Rubik's Cube as my metaphor. We were 1 cube at the bottom. Today, if you look at our strategy, it's been to go, first of all, to multiple CSPs. What CSPs do for us, in effect, is shift CapEx into OpEx. You pay as you get the revenue, as you're actually utilizing the data centers. So, in effect, we are riding somewhat on their ability to build and have CapEx and financing.

Today, we sit on top of every CSP: Oracle, CoreWeave, Microsoft, GCP, AWS, and a bunch of small neoclouds. On the chip side, we've also gone for a program of being multichip, because we want to make sure you're always on the frontier. I think if you're only on 1 chip, there's inherently a moment where you can't be on the frontier because some leapfrogging happens.

Today, NVIDIA remains our absolute priority partner. They have the frontier chip. Our next big training run in the fall will be done on Vera Rubin. We're really excited about that. And now we're plotting the Simon series that's coming.

We're also getting chips in the pipeline from AMD. Cerebras is already online. It's been an incredible low-latency chip, great for developers, for example, who want real-time coding. And there's our own chip that we're working on with Broadcom.

Beyond that, there are other ways we've diversified. Think about that Rubik's Cube. It's become much more multidimensional, and it allows us to effectively utilize investment-grade CSPs in order to be able to go fast and push it back to be more OpEx, not CapEx.

Now, we are starting to shift gears into more of a built-to-suit type of environment. We announced a data center we're building with SoftBank Energy down in Texas. That's the beginning of something that's beyond a CSP. There's a little bit more CapEx required there.

Finally, I think as the world progresses—remember, we've done all that just in 2 years—the reason I like a Rubik's Cube is, again, please ChatGPT this, but I think a Rubik's Cube has something like a quintillion different forms it can come up with. It just gives us a lot of optionality. Remember what I said: my job is maximum optionality.

In a moment where I'm not yet an investment-grade type of entity where I can go get lower-cost debt financing, being able to work with partners to do that is really important.

Do you think that 5 years from now the stack is just merged together? What do I mean? In traditional or historical markets, you'd have NVIDIA sell the chips, but that's all they do. Then you'd have Microsoft just run a cloud. That's all they would do. And then you would have a consumer app. That's all you would do.

But now we see everybody doing everything. You guys have silicon that you're spinning. You have models that you make. You may or may not eventually decide that you need to be some form of a neocloud yourself. If you look at NVIDIA, they have incredible silicon, but they also have their own open-source models. They're increasingly becoming an offtaker. Google is a cloud company first, but they also have a chip. Now they have models. So, it's all merging.

If that continues to happen, does that make the competitive landscape simpler or easier?

I mean, I think where everyone is trying to make sure they reside is the layer that's closest to the customer, where usually you take the largest portion of the profits of the ecosystem, right? No one wants to find themselves—

They're searching for profit pools.

Away. Absolutely. And so, that's why today, when I think about our position, it comes back to where I started: why we want to be that AI intelligence layer. A year ago, people talked about the commoditization of the LLMs, and frankly, it's gone the opposite way because as you start building an agentic layer—and we all started using this word "harness"—the harness is what brings the context, the memory, right?

In my Codex, I have a whole ginormous memory file where it knows that I'm me. It knows I'm the CFO of OpenAI. It knows how I like to write things—well, how I like to say things. It knows what I'm interested in. It also knows that I'm a mom of teenagers. It carries all this memory. And that makes the model more powerful for me.

Now, think about what happens when that memory and that context are brought into an actual enterprise environment. So now it's not just even about the data that resides there, but I always think about the intuition of back when I worked on Wall Street, right? There was all the data in the world that told you what a stock should do after an earnings call.

But give me 1 second. Then you called your trader. And the trader would be like, "Yeah, that stock's not going up, Sarah." Now I'm like, "What are you talking about? All the numbers say it did this, did this, did this." And he's like, "Yeah, no, but I know this fund is under pressure, and they need to sell down their book, and that is going to kill this stock for the next week."

That is the intuition of an enterprise. It's the best example I always think of because I came out of a financing world, but there's this intuition in every walk of life. And that's where I think the models are now getting very connected to the memory, context, and intuition of your company.

And that's what gets CEOs and C-suite really excited, because they're like, "Okay, now I really see how this is going to add value to drive my revenue line, my top line, but also, I can think about it as an efficiency play as well."

And so, back to what you're asking, I think what people want to make sure is they stay as close to that value as possible.

And be flexible enough to pivot. As you can see, we have to wrap.

Sorry, Jason.

It’s quite all right. It’s been wonderful, and you’ve been so great with the details. One final detail question, rapid-fire: The 3 greatest consumer businesses of our lifetime are the iPhone, Meta’s advertising network, and Google’s advertising network. 2 of those 3 are ad-based, and even Apple has a sprinkling of ads.

I haven’t heard you talk about ads much. People tell me they’re seeing some ads in the experiment in the free version. What is your commitment to the ad version? You guys got a little trolled by Anthropic during the Super Bowl: “Oh, you’re going to have ads.” But are ads the solution to making this free for the world?

Yeah. So, first of all, on the ad front, we want to stick by our principles. We want to make sure that you’re always getting the best result based on the model, not by something that was sponsored. That has to hold true. And I think the second thing is that we’ll always provide an ad-free tier for people who just don’t want ads.

If they pay.

I think Ilya says this really well: If Google and Meta had a baby, it would be ChatGPT. What you have in Google Search—and, by the way, we know we have at least 11% of the search market—it’s a lot more because, actually, when you do a Google search and the page refreshes, that counts as 1.

In ChatGPT, when you have a whole conversation where you might ask 50 questions, that also only counts as 1. So, in reality, we have a much higher portion. It’s very high intent. That is great for advertisers because I’m effectively telling you what I’m doing, right? I want really cool shoes to sit on the stage. I’m telling you what I want to go buy.

In Meta’s case, they use this “people like you” sort of intent, so they have the demographic. We have more than that because we have memory, right? I just told you it knows who I am. So, imagine putting memory and context next to intent. You should have a very potent ad platform, which gives you the ability to offer up massive access to the world writ large because now you can pay for it.

And I think, going back to a question you asked Friedberg, if you look at the revenue per token right now, if I was optimizing only for today, I would give every token to the API.

Right.

Every token to the API. An order of magnitude more than to the consumer. However, I told you we’re playing our own game. We have a strategy where we believe there’s an AI infrastructure-layer utility, like electricity.

And in a future state, you’ll want to be able to serve the world writ large: consumers, small businesses, large enterprises, governments. That’s our strategy.

Ladies and gentlemen, the CFO of OpenAI, Sarah Friar.

Well done.

Fabulous.