OpenAI CFO Sarah Friar: IPO, AI Rivalries, New Device, and Spending $100B+ on Compute

00:00:00Open AI's CFO, Sarah Friar.>> We're going to get right to it. You havejust completed what I regard [music] asthe most successful fundraising round inhistory.>> We're going to raise [music] actuallynorth of a hundred and twenty billiondollars. We think AI is the biggest erathat we've seen [music] today. We'rejust starting to understand what it'sgoing to mean for global productivityand with that, you [music] know,hopefully more affluence, better livesfor everyone. Luck is whatever thepreparation meets [music] opportunity,but you got to grab it.Long time listener, first time caller.Quite exciting.To get to hang out with all the broshere. Hello.>> [laughter]>> We weren't sure how to start this off,but I thought the best thing was toallow our erstwhile crypto czar to maybe>> erst>> say a few comments and>> I saw an article today. I think it mighthave been in the Wall Street Journalthat the perception is that there's anadvantage to IPOing earlier if you're anAI company. So now we know SpaceX isgoing and then the question is when'swhen are Open AI and Anthropic going togo? And I'm curious, how do you thinkabout that? Do you think there is alittle bit of a race on or, you know,you haven't made a decision about thatyet?>> Like in the end, an IPO, I say this

00:01:15tothe team all the time, it's a milestone.It is not a destination. Do not run yourcompany as if that's some sort ofdestination. It's just another way tofund raise. We just did, you heard me onon the the the sizzle reel, raise ahundred and twenty-two billion dollarsin March and that was to give ourselvesmaximum flexibility. I feel like my jobas a CFO is create optionalityfor this, not just this company, butjust this era that we're living in.>> was thatthat point in fundraising, is that thebiggest private or public up until theSpaceX IPO?>> It is.>> Right.>> It is by orders of magnitude. I thinkthe largest IPO to date was the SaudiAramco, which was about $30 billion.So, it is actually incredible thatyou're going to have potentially threeIPOs at a scale that will be bigger eventhan 2001,2000 that that time frame. There was alot that went on in the market, too.But, the market has grown. And by theway, the other thing going on in themarket is like if you look at buybacks,M&A, and so on, it's actually a lot ofcapital keeps being returned back toshareholders in cash. So, there is a lotof money sitting on the sidelines. But,in the spirit of like the question,David, I think in the end

00:02:30you want toyou'll be measured, right? It's the Inthe end, the market is a weighingmachine, not a popularity machine. Noone remembers who won first, Google orYahoo, Lyft or Uber. And I say that notbecause well, I want to be first orsecond, but I just think it's you know,the the press loves a bit of drama, butin the end we're going to have to buildbig, sustainable, durable companies. Andfundraising will be a key component ofdoing exactly that.>> Sarah, breaking news.>> Oh my god, so many people coming at me.Hi, Jason. [laughter]>> I know, it is it is it is it is hardbalancing four interviewers at the sametime.>> This is my world, by the way. So, I'mgood with this. Jason.>> and Anthropic justuh confidentially filed their S-1. So,does that mean you're third place interms of the filing?>> It does not mean anything yet because[laughter] you have to run now thegauntlet of the SEC, and who knows howlong that takes for anyone.>> Yeah, is it Is there though abenefit to them going forward? And Ithink unpacking the rivalry withAnthropic is on everybody's minds. So,justI guess you can't talk too much aboutIPOs, so I'll just pivot to Anthropicwas far behind, and now they've reallyum I think everybody would agree in theindustry now blown past OpenAI in termsof

00:03:45developers and corporations, and itseems revenue.So, didhow did that happen at OpenAI when youhad such a tremendous lead? How didAnthropic blow past you guys?>> So, let's talk a little bit about ourstrategy. Our strategy is different,right? So, we are building the AI layer,the infrastructure, and it's reallyimportant that there's a singlefoundation, but then with manyinterfaces out into the world. So,ChatGPT is one to the consumer. Over 900million people use ChatGPT weekly, andit's become the noun and the verb. It'show most people experience AI for thefirst time. Kind of fun fact, oureconomic research team just showed methe fastest growing continents now areAfrica, probably not totally surprisingsince it started a small base. Fastestgrowing languages areAzerbaijani andwhat Kazakhstani? What is It's Kazakh.>> Kazakh.>> Um which is kind of incredible to talkabout where it's going. So, multipleinterfaces, ChatGPT, of course there'sCodex, um just hit 5 million over theweekend. We're really proud of thatcoming from almost zero in January.>> users.>> Go Codex. Um help me prepare for thislittle special up here, too.

00:05:00Um there's of course Frontier, ourenterprise offering, and everythingevery other way that we can get outthere to reach businesses of all sizes.That is a very different strategy. Wethink that because it's served up on onemodel, there's a compounding element ofadvantage that comes from that. Moreusers, more data, more ability topersonalize. ChatGPT acts as a frontdoor. As we as models get bigger,there's more efficiency that shouldlower the overall cost to give you atoken in the world. That should compoundto higher gross margins, ultimately moreways to pay for compute, and then accessto compute is one of the really bigcompetitive advantages at the moment.So, you know, we have to all run our ownraces, but we all have to recognizewe're part of an ecosystem that alsoneeds to bring people alongcollectively.>> Did you spread a little bit too then toomany projects? People are talking aboutthis new gadget, Sora, and and thenmaybe not enough focus on enterprise. Isthat a fair assessment of if there was amistake in the last year, that was it?>> No, I I think that the world loves to goto binaryisms. Like, are you a consumercompany, Sarah? Are you an enterprisecompany? The reality is we're very muchboth. We're not one or the other. Rightnow, our revenue's getting prettybalanced, about 50/50.

00:06:16We are incrediblyfocused on the enterprise. Like, I spendso much of my time with I mean, justeven in the last week, I could tell youI've been to see Thermo Fisher inBoston. I was with a bunch of banks inNew York. I was on the phone withTravelers on Friday. I spent thismorning on the phone with a techcompany. Doesn't matter the vertical,people are really moving on AI rightnow. Our new head of revenue, DeniseDresser, in seat since December, she isa force of nature. And so, I think theenterprise, broadly speaking, is reallyfiring on all cylinders. But, we don'twant to leave the consumer behind.Remember, our mission at OpenAI is AGIfor the benefit of humanity, not for thebenefit of humanity who can pay, or forthe benefit of humanity who live in anenterprise, but very broad-based.Um it's why we offer so much free,because we want people to get a taste.Once they get a taste of intelligence,the ability to come up with commitmentcurve is incredible. Our free users doabout seven turn seven questions a day.Our first paid tier do double that,about 15. Our our real paid tier, theplus, 20 bucks, hopefully you're all onit or higher, about 3x. And pro, aboutum 11xover a free

00:07:31user. So, remember when yougot your flip phone, and you're like,yeah, I don't know what it does, makesome calls. Now, that same phone, thinkof all the things it does for you.That's the path we're on withintelligence right now. Sorry, Sam.>> You said something very influential, Ithink it was about 18 months ago, for alot of us in the industry where youframed a very simple economic trade-off,which was gigawatts to cash. And I thinkyou said 1 gigawatt is roughlyequivalent to about 10 billion dollars ayear of revenue to OpenAI.So, comment number one was this 1gigawatt equals 10 billion dollars ayear of revenue for you. But, it's notjust you cuz you can probablyextrapolate that toAnthropic and other folks, Gemini.But, then you were really at theforefront of getting access to power anddata centers and powered land. It seemeda little crazy, but now it looks like,"Hold on, there's a huge deficit ofsupply." Can you just unpack all of thatand explainboth the spectrum of where we are andthen those specific economics and ifthat's changed?So, first of all, yes, compute is a veryscarce resource at the moment. We whatwe see in our business, we're going upthat kind of vertical wall of demandright now, and there's just not enoughtokens available. So, I'm very gratefulthat I got to work alongside Greg

00:08:46andSam. I think we really pressed hand onthis. And last year, we were definitelytaking some, you know, arrows in theback about, "Why are they out therebuying all this computes?" And I think,thank God we did because in '26, westill won't have enough compute. Um,where are we on the compute continuum?There's kind of choke points everywhere,and and I think they will continue tomove back and forth. I mean, you alltalk about this and know this.as well as anyone here, whether it'senergy, first and foremost, um, land,power, how we get regulatory, um,environments such that we can buildquickly. Um, when you get into the racksand chips themselves, clearly, do wehave enough, um, in that supply chain?Memory spike is is on at the moment.Access to great talent. Um, do we haveenough people coming through oureducation system? I really worry aboutthis right now. I'm a trustee atStanford and you know, I see just thatyou know, we need to keep the focus oneducation and science.And then trust. I mean, I actually putthat as part of the supply chain.Sam right now is in Saline, Michigan.He's going to be cutting the ribbon inabout 2 hours. So you are getting asneak preview, but they told me it wasokay to say it in the room.That will be, you know, sticking shovelsin the ground on a 1 gigawatt datacenter,

00:10:01which is part of our Oraclecomplex.Really important there on the trust sidethat we don't leave communities behind.I spent 7 years of my life working atNextdoor doing the hard work of what itmeans to be local. And you cannot tellpeople from top down what they need cuzthey will tell you thank you, but nothank you. I will tell you what I need.And so in a data center like that, we'reactually spending a lot of time in thecommunity saying, "Number one, we're notgoing to raise your electricity bills.We're going to pay for ourinfrastructure and our power. It willnot be the ratepayer that has to pay."Number two, we're going to bring jobs.2,500 union jobs.Good jobs. Like electricians, HVAC, andso on.We are going to pay our taxes. A billiondollars in taxes just for that datacenter into Michigan. And on top ofthat, we're going to invest 45 milliondollars going into education for Codexcredits to do what you all talked aboutthis weekend is like anyone who's notlike coming in facile to their new job.I have teenagers using Codex. It wouldbe like I would never hire a financeperson didn't know how to use Excel andI pretty much probably wouldn't hire afinance person today that doesn't knowhow to use a tool like Codex. So that,you know, so when I think aboutinvestment,

00:11:16we're having to invest aheadof demand. That means we need to both beable to find all of the compute and allthe pieces and then pay for it. So thatgoes back to your capital question onIPO. And then on the other side on theeconomics, look, the economics docontinue to get better. They're gettingbetter on multiple fronts. I think weare doing a better job of actuallyshowing true value to our customers. AndI think you get beyond kind of a costplus type pricing into something thatfeels more akin to the value beingcreated. Now, scarcity of tokens helpscuz it's causing a bit of a compressionin time.>> about that in just like without specificnames, where you know the landscapeexists today in terms of all the powerthat's available and all the demand thatexists across everybody?>> Yeah.>> What's going to happen over the nextyear just at the current course andspeed of what is available? Of the datacenters that's available, of the tokensthat's available, of the infrastructurethat's available for everybody because Iyou know I told this story last week butyou know I'll use Anthropic and one ofthe frustrating things is at some pointit just says you know 10:30 it's likeall right Chamath see you at 2:30.>> Yeah.>> And that's not a viable experience.>> Right.Um>> And in fairness to ChatGPT actually I'venever had that with>> Yeah, we we're quite generous with our

00:12:31tokens and again on purpose we're tryingto drive access so people understand cuzif you're on that free tier not actuallygetting the latest model but we'retrying to put it in your hands so youget a sense for it by the way becauseyou know if you're a kid um doinghomework like I think about when I grewup and the Encyclopedia Britannica'sshowed up at the front door in NorthernIreland in a tiny little community inthe middle of the troubles it was likethe clouds parted we want to make surethat people get that feeling by the way.But the landscape right now in 26 if youwant to buy more compute good luck toyou. Like tell me cuz I don't know whereelse to find it. I mean as you know Iwas going to say Elon ironically endedup being the one person that had toomuch compute in a way. Um but good jobon like figuring out how to sell thatoff.Um in 27 it's pretty limited as wellfrankly. Now there's a couple of thingsshifting around. When we talk aboutcompute, there's training that mostlystill all happens here in the UnitedStates. For USG reasons, for making surethat a national asset in effect ishappening on US soil. For inference, wewant that to be global. And I thinkparticularly in an agentic world, youwant much

00:13:47more kind of real-time. Evenfor things like Sora and video, which bythe way, yeah, we have you know, we hadto make a really tough choice cuz wedidn't have enough compute.>> And it uses a lot.>> Right now, yeah, video does. But videois not over. Like in particular, whenyou start to think about where AI istaking us into more multimodality. So,remember, we've all been taught by thelast generation of technology to talkwith our thumbs. It's a disease. Youwalk around, everyone's looking down,they don't look up anymore. Teenagerssit on my sofa at night and talk to eachother with their thumbs. I'm like, "Whoare you talking to?" And then my sonwill be like, "Him." I'm like, "Okay."Talk. Multimodality is here. Umhopefully, I think you all talked aboutit this weekend. You're talking to yourtool. I talk to Codex every day. And sothat is changing rapidly, but that isgoing to need much more kind ofreal-time compute cuz it's an oddexperience. If I was talking to Chamath>> Johnny Ive this puck and ear pieces. So,maybe tell us a little bit about thatproject. You've admitted it now.>> If I If I tell you it's an ear piece,Johnny will come and steal my teenageson. I might give it to him. Give him tohim.Uh but we arewe're changing into a consumer substratethat I cannot tell you what it is,

00:15:02butby the end of this year, we will unveilit. Early next year, you'll be able tobuy it. I have seen it. I've tried it. Iam a hand talker. Right now, I'm sittingon my hands.>> Did you have a Did you have a paradigmshift? When Yeah, when you used it, wasit like having an iPhone for the firsttime?>> It's verywhat Johnny and team are really good atisbringing humanity to devices and I don'treally 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 verylovable.>> Really?>> And I can't really explain what thatemotion is cuz so much>> Intimate in some way in terms of>> technology is>> Not taking your phone out and it's it'sseamless is what I've heard from peoplewho played with it.>> is very um can be very mechanistic, butwe all know great design just makeseverything fade away, right? It's whatum at the time, you know, the simple ishard.>> Yeah.>> Um but I I think this is a drop set.>> This story just going back to theearlier question, so putting onthe CFO hat, help us understand thecapital allocation model that you use.Cuz a lot of businesses over the lastdecade, two decades that have kind ofbeen these outsized returners have foundsome unique way to deploy capital at ahigher ROC

00:16:17than anyone else and then youend up plowing all your capital intothat higher ROC bucket. What is that foryou guys and how do you think about thatled that portfolio approach to havingmore of these kind of big returner shotsand is there an engine where that getsbetter over time?>> There has to be becausein the end, the durable, high valuecompanies created in this era, I don'tthink they're not going to be magical.They're going to look like the greatcompanies of prior eras. They're goingto create customer value. Starts withthe customer um and really helps thecustomer do something different, better,more revenue, more efficiency, right?Thermo Fisher wants to be able to get umpatient screening done faster so theyget FDA approval faster. That's reallyimportant. Like if you have a form ofcancer where you have weeks to live, thedifference between a breakthrough infour weeks and two weeks can literallybe life or death.They also have, I'm going to misquotethis, but something like 30,000, 38,000people in the field selling thoseamazing like if you walk into any lab inthe country, you'll just see ThermoFisher plastered all over every device.Those people want to be more efficientgoing to work. Like the the fastesttakeoff of Codex within Open AI rightnow

00:17:33is actually in our go-to-marketteam. Our devs are there, but like ifyou look at the pace of growth kind ofmonth over month, it's all in GTM. So,they want more productivity out of theirGTM team. And of course, um they'redoing things in areas like finance,which I get really excited about. Butso, customer value first. From that, nowyou need to get to a great gross margin.So, how do you get to a great grossmargin? You're looking at like the costof revenue. The main input is compute.The good news on compute is that thereis a massive deflationary curve on cost.Right from chat GPT uh five to fivefour, I think the deprecation cost wassomething like 97%. It's like a kind ofan amazing curve. Actually, I'm slightlyfrom four to five four, it was 97%. Butthat happened in like two years. It'skind of wowing, right?>> That's it.>> Um even our newest model, if you look atfive five that we just released, we'retrying to now translate that back to thecustomer. So, we actually raised priceson five five two X. But if you look atwhat the cost of the customer is,they're probably still getting a breakof about 20 to 30% cost reduction pertoken because just much more efficientper token. So, there's a lot to do inthat envelope.>> Yep.>> And and part of making an a capitalallocation decision is having

00:18:48toif you make it on today's cost profile,you actually might misprice theoutcomes. You have to lean in a littleon the cost profile. And then as wethink about like the builds, yeah, youare having to make like really my focustoday on compute is what's the compute Ican buy for 28 onwards. Like thatMichigan data center in Saline, I don'tthink we will be getting compute out ofit until probably end of 27 or early 28.So, that's where you're starting to makeyour bets. And in fact, where I feelmost short of compute right now isstarting to look at 30, 31, 32. So,you're having to create a businessmodel. Um now, the good news is eachyear goes by, we get more confidence inthe build. We're seeing it massivelyoutperform. And so, that's giving usmore and more confidence. And the marketis coming towards us much more.>> All right. So, how are you making thecompute need forecast multiple yearsout, accounting for all of thearchitectural and model advancementsthat are happening, where quality valueor utility per unit of power is goingup? And help us understand how you kindof estimate that given that there's alot of technology development going

00:20:03onthat has a high kind of variance to it.>> Yeah, yeah. So, we we do have to makemultipleassumptions both on the compute itself.So, we assume right now that compute itactually on a per gigawatt is gettingmore expensive cuz power is getting moreexpensive, memory is getting moreexpensive, and so on. However, the theintelligence that we get on the otherside out because of the deprecation onthe chip side is is more than making upfor that. So, in a terms of a per unitsold to a customer, it should actuallyget a lot less expensive for the>> improvement in that.>> Yeah, exactly.>> Yeah, exactly. That's just the chipitself. We don't want to overestimate onthe model side cuz sometimes like 55 isan incredibly good model on theefficiency side, but if you look atsomething like 54, the prior model, itwas a really large pre-trained model. Itwas very expensive. It was actually hardto serve. And sometimes we want to dothat really big pre-trained moment. Andthen we take multiple model turns to beable to kind of drive down on the costside. I mean, in the in the near term,like in 26 and 27, I clearly build amodel that's bottom up. So, I know whatmy products are, I have a sense of whatthe pricing will be.You know, P times you know, consumer Ptimes Q. How many

00:21:18wild do I think Ihave? I can see what the shape of theline is. How many of them willsubscribe? Advertising coming in is alsostill related to how many weeklyactives, how many dailies, how manymessages and so on. So, you can you cando actually a pretty good model job in26 and 27. That said, the shape of theline keeps taking us by surprise to theupside. When you get into the outeryears, you're actually looking more atthe compute you've bought and almostjust doing an algorithm the other waythat's saying this amount of computeshould equate somewhat this amount ofrevenue. I don't know for certainexactly where it will all come from.Like a year ago, I built a model forinvestors that showed agentic revenue.And the story was we're going to havethis thing, we're going to be in theagentic era, we're going to hand it to adeveloper with natural language. They'regoing to be able to build and we thinkthey will pay upwards of maybe $2,000 amonth for it, which is kind of laughablein hindsight. But nobody believed. Theywere like, I don't even know what she'stalking about. There's no way that willhappen and $2,000 a month? Remember whenpeople were losing their minds overChatGPT Pro being at $200? Like, oh mygod, no one will ever pay for that.>> Right.>> Yeah.>> So, why $122 billion? Does it take youto 2031,

00:22:332032? Like, how do you get thecalculus on the capital needs as you dothat modeling?>> Right. And you'll maybe get morespecific. So, the estimates I've seen isthat to stand up 1 gigawatt of AIcompute costs about $50in capital. Land, power, shell, chips,everything. All in around 50 billion.Do you have to front all of that moneywhen you create a new data center? Orhow much of it do you do? How much of itcan you get debt for? Does a hundredbillion raise only get you two gigawattsor does it get you five? Like, what doesit get you?>> It's It's a great question. So, if youlook at our compute strategy,um and it's crazy how fast the world haschanged. So, just two years ago, we wereliterally one. We had one CSP we workedwith, Microsoft Azure.Um we we sat on one chip, Nvidia. We hadone product, ChatGPT. One price point,$20 a month. So, I often use a Rubik'sCube as kind of my metaphor. So, we werelike one cube in the bottom. Today, ifyou look at our strategy, it's been togo, first of all, multi multiple CSPs.Because what CSPs do for us, in effect,is they shift CapEx into OpEx. So, youpay as you get the revenue, so as you'reactually utilizing the data centers. So,in effect, we

00:23:48are riding somewhat ontheir ability to build and have CapExand um financing. So, today, we sit ontop of every CSP, Oracle, um CoreWeave,um Microsoft, GCP, AWS, and a bunch ofsmall neoscalers. On the chip side,we've also um gone for a program ofbeing multi-chip. Um cuz we want to makesure you're always on the frontier. Ithink if you're only on one chip,there's just inherently a moment whereyou can't be on the frontier becausethere's some leapfrogging that happens.So, today, Nvidia remains our absolutepriority partner. They have the frontierchip. Our next big trading run in thefall will be done on Vera Rubin's. We'rereally excited about that. And now we'replotting kind of the Simon series that'scoming. But, we also now have chips inthe pipeline from AMD. Um Cerebras isalready online. It's been an incrediblelow-latency chip, great for devs, forexample, that want real-time coding. Andthere's our own chip that we're workingon with Broadcom. And then beyond that,there's other ways we've diversified.So, now think about that Rubik's Cube.It's become much more multi-dimensional,and it allows us to effectively utilizeinvestment-grade CSPs in order to beable to go fast and push it back to bemore opex, not capex. Now, we arestarting to

00:25:03shift gears into more of abuilt-to-suit type environment. Weannounced a data center we're buildingwith SoftBank Energy um down in Texas.That's the beginning of something that'sbeyond a CSP. There's a little bit morecapex required there. And then finally,I think as the world progresses,remember we've done all that just in 2years. The reason I like a Rubik's Cubeis again, please chat GPT this, but Ithink a Rubik's Cube has something likea quintillion different um forms it cancome up with. And so it just gives us alot of optionality. So remember what Isaid, my job is maximum optionality. Andin a moment where I'm not yet aninvestment-grade type of entity where Ican go get lower cost debt financing,being able to work with partners to dothat is really important.>> think that in 5 years from now the stackis just merged together? What do I mean?In traditional or historical markets,you'd have Nvidia sell the chips, butthat's all they do. And then you'd have,you know, Microsoft just run a cloud.That's all they would do. And then youwould have a consumer app. That's allyou would do. But now we see everybodydoing everything. You know, you guyshave silicon that you're spinning. Youhave models that you make.You may or may not eventually decidethat you need to be some form of a neocloud yourself. If you look at Nvidia,they have incredible

00:26:19silicon, but theyalso have their own open-source models.They're increasingly becoming anoff-taker. Google is a cloud companyfirst, but they also have a chip. Nowthey have models. So it's allmerging.Is if that continues to happen, doesthat makethe competitive landscape simpler oreasier?>> I mean, I think where everyone is tryingto make sure they reside is the layerthat is closest to the customer whereusually you take the the largest portionof the profits of the ecosystem, right?No one wants to find themselves>> They're searching for profit pools.away. Absolutely. And so, that's whytoday, when I think about our position,it comes back to where I started, why wewant to be that AI intelligence layer,is because a year ago people talkedabout the commoditization of the LLMs.Um and frankly, it's gone the oppositebecause as you start building an agenticlayer, and we all started use this wordharness, but the harness is what bringsthe context, the memory, right? It Ihave in my Codex, I have a wholeginormous memory file where it knowsthat I'm me. It knows I'm the CFO ofOpenAI. It knows how I like to writethings, well, how I like to say things.

00:27:34It knows what I'm interested in. Itactually also knows that I'm a momteenagers. I mean, it just carries allthis memory.And that makes the model more powerfulfor me. Now, think about what happenswhen that memory and that context isbrought into an actual enterpriseenvironment. So, now it's not just evenabout the data that resides there, but Ialways think about the the intuition oflike back when I worked on Wall Street,right? There was all the data in theworld that told you what a stock shoulddo post an earnings call.But, give me one 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?Like all the numbers say it did this,did this, did this." And he's like,"Yeah, no, but I know this fund is underpressure, and they need to sell downtheir book, and that is going to killthis stock for the next week." Right?That is the intuitionof an enterprise. Like it's the bestexample I always think of cuz I came outof a financing world, but there's thisintuition in every walk of life. Andthat's where I think the models are nowgetting very connected to the memory andcontext and intuition of your company.And that's what gets CEOs and C-suitereally excited cuz they're like, "Okay,now I really see how this is going toadd value to drive my revenue line,

00:28:49mytop line, but also, you know, I canthink about it as an efficiency play aswell." And so, back to what you'reasking, I think what people want to makesure is they stay as close to that valueas possible.>> And be flexible enough to pivot as youcan see we have to wrap, yeah.>> But>> Sorry, Jason.>> It's quite all right. Umit's been wonderful and and you've beenso great with the details. One finaldetail question, rapid fire. Three greatgreatest consumer businesses of ourlifetime, iPhone, Metaadvertising network, and Google'sadvertising network. Two of those threeare ad-based, and and even Apple has asprinkling of that ads.Haven't heard you talk about ads much.People tell me they're seeing some adsin the experiment in the free version.What is your commitment to the adversion? You guys got a littleuh trolled by Anthropic during the SuperBowl. Oh, you're going to have ads. Butis ads the solution to making this freefor the world?>> Yeah. So, first of all, on the ad front,you know, we we want to stick by ourprinciples. We want to make sure thatyou know you're always getting the bestresult based on the model, not bysomething that was sponsored. So, thathas to hold true. And I think the secondthing is that we'll always provide afree a tier, sorry, an ad-free tier forpeople that just don't want ads. But>> If they pay.>> If you took

00:30:04If you took what I Fiji saysthis really well. If, you know, Googleand Meta had a baby, it would beChatGPT. Cuz what you have in GoogleSearch, and by the way, we know we haveat least 11% of the search market. It'sa lot more because actually, when you doa Google search and the page refreshes,that counts as one. In ChatGPT, when youdo a whole conversation where you mightask 50 questions, that also only countsas one. So, in reality, we have a muchhigher portion. Very high intent. Thatis great for advertisers because I'meffectively telling you what I'm doing,right? I want really cool shoes to siton the stage. I'm telling you what Iwant to go buy. In Meta's case, right,they use this like people like you sortof intent, so they have the demographic.We have more than that cuz we havememory, right? I just told you it knowswho I am. So, imagine putting memory andcontext next to intent, you should havea very potent ad platform, which givesyou an ability to offer up massiveaccess to the world writ large becausenow you can pay for it. And I think backto a question you asked Freeberg, likeif you look at them the revenue pertoken right now. If I was optimizingonly for today, I would give every tokento the API.>> Right.>> Every token to the API. Order

00:31:20ofmagnitude more than to the consumer.However, I told you we're playing ourown game. We have a strategy where webelieve there's an AI infrastructurelayer utility like electricity. And in afuture state, you'll want to be able toserve 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. [music]

00:31:52>> [music][music]