Google has the driver’s seat in the AI race. Who will trip it?

Hi, welcome back to ZeroShot. I'm Rohan Dharmakumar. I am the lead host for today's episode. And before I get into today's episode, which by the way is in most senses, you know what? I'm holding myself back from saying it's a great episode because that's so fake, right? I mean, how can I say that it's a great episode? That you guys have to help me. And frankly, I'm just disappointed not enough of you are rating us. You know, more of you need to lead us. And this has turned into a trumpisk rant, right PGK? I've gone and told everyone in my family to rate us up ZeroShot. And we said, okay, yeah. All right. I also asked my sister to rate us. A lot of family pitching in. Make zero-shot great again. Thank you very much. So we have large families. And on that wonderful stereotype, I want to dive in into today's topic. And I want to start with this very famous quote. It's anyone who wants to recognize, even though you're not here on the call, but you cannot overtake 15 cars in sunny weather. But you can when it's raining, who said this? This is Atten Senna. That's right. The late Atten Senna. He said it. And of course, the context being when the weather is sunny, everyone is driving well. Everyone thinks they have a chance at winning the race. But when it starts to rain, tires start to skid. So people become more careful. They slow down. And it's at that point when things slow down. This is very poetic. I think the less poetic or somewhat poetic one is something that they say also in Game of Thrones where chaos is a ladder. Chaos is an opportunity. It's all the same. What is that saying this? There are only what how many stories in the world and everything else is just sort of like a remix of it. Like how many stories or something like that. But why did I say you cannot overtake 15 cars in sunny weather, but you can when it's raining. Today's episode is not about Atten Senna or F1. I'm going to pose it by saying that the largest players in the AI race are all dealing with their own versions of rain. For a minute imagine the AI race is a giant F1 race where these, you know, the biggest AI companies are driving in their souped up, you know, they are racing vehicles around the track again and again. You don't only think that comes. No, you know what? You know what? Absolutely. I don't. You know what's coming to my mind when I think souped up vehicles, F1 vehicles are not coming to my mind. Mad Max type of vehicles, which is more apt for the AI doom loop than F1, but let's for a minute assume that they're F or they're like formula one vehicles, which are the players. Let's actually start with open air and anthropic. They're both working towards their upcoming IPOs, which means this sort of slowing down. They're slowing down because both of them need to show more revenues and profits to investors, convince them to invest in their IPOs, even if of course we know that for a fact that even when their IPOs come, they'll still be probably lost making overall, but they need to show that we're accelerating and growing revenue rapidly. What that means is, you know, that other things suffer. If you take open AI, open AI is moving to inside, adds into chats. It's taking a cut, 4% cut out of shopping transactions from Shopify Merchant. In a sunny world, open AI wouldn't do these right now. Instead, it would probably want to improve the quality of its models, which no longer are best in class, or it might want to subsidize usage more and get enterprises using it more, attack say, Claude's customer basis. Anthropic tool has been aggressively cutting off access to third party tools at embed Claude. For instance, they cut off access within cursor ID, to extra AI, and extra AI people just woke up one day and said, "Damn it, we can't use cursor ID because Anthropic cut off access to us." Anthropic also did that with open code. I mean, to be fair, Claude's and Anthropic's message is simple. Pay us $20 to $200 a month to use our product, so they're also not focused on growth. Well, what about GROC, NX AI? Well, sure, they used AI to generate over 3 million sexualized and non-concentral images in 11 days. Sure, they figured out how to put up data centers in the US quickly by ignoring existing regulations and generating bar onsite through massive generators. And while it may be possible to stem the flow of advertisers by threatening to sue them, I don't think you can get organizations to embed AI inside them by threatening to sue them. So let's just put GROC and XDI on one side, meta. Sure, I mean, I'm absolutely sure. There is one area where I'm sure they're world class, which is the use of AI and LLMs to profile their users and to shove more ads into their feeds, right? But frankly, it's been a while since-- Or to improve recommendations. Which is all the same. Recommendations for what? Fair enough. For more content. More content for what? To show more ads. Right? I mean, they're transparent in that way, right? Everything for them is an ability or an avenue to show more ads to you because that's how they make money. And I'm sure nobody comes close to touching them with the exception of perhaps TikTok, right? But it's been fairly a while since anyone felt that meta's AI efforts are threatening anyone else, but their own neurosis and anxieties, right? But what? Nvidia. I know it's odd to be mentioning Nvidia here, but do remember that Google's TPU chips are now competing with Nvidia's chipsets in the market because Google after many years of allowing TPU access only through data center services is now finally ready to be selling TPUs directly to enterprise users, right? So even Nvidia is sort of on the defensive. Wait, tell us what are TPUs? I know it covered tens of processing units. Google's own custom chips, which I've written about in an earlier edition of the zero-shot column, which we will link to in the show notes. See, I teed you. Yes. Thank you. What about Microsoft? It's a company that's desperately trying to make co-pilot appear in the same sentences as Claude, OpenAI or Gemini, but it feels when when was the last time you saw a sentence that said AI models and it said Gemini, OpenAI, Claude, and co-pilot. But just a very small observation of something that happened last week. So as you know, Claude has been on a tear off late, Anthropic and Claude. People have been using it for all kinds of things, which you won't talk about a lot today. But I think just last week Claude released a functionality where you can now plug it into Excel. Microsoft Excel, I'm not calling it co-pilot, and use Claude to write sheets, edit sheets, create pivot tables, etc. Almost natively in Excel. Right? And somebody put a comment saying, how on earth did Claude do this before co-pilot? How do you do this? How is this even possible? But yeah, it was fun. Right? Well, the answer to that is Claude has nothing to lose by using AI to disrupt Excel. Microsoft is in this fine balance between we want Excel to be useful so that people continue to pay subscriptions for office, but at the same time, you also want it to be AI enabled. So always like companies that don't have any in-compancy are able to do things faster. I also get the sense that Microsoft is desperately trying to make a leap from what I call hardware to AI. I mean, much of its software, like, let's say, Windows or Office, etc. were sort of designed in the physical era where you bought PCs and installed them on them and then used it in paste license fees. And then of course, there was an internet era in the middle where which Google, which was Google's era. And then now we are in the AI era. I don't think Microsoft actually made that jump proper to internet era, right? Which is why Microsoft has to, in India, we call it leapfrog. We say that we went from plain old telephone systems, copper wires to just wireless, right? So Microsoft is in this unenviable position of having to leapfrog office into the AI era, which is co-pilot. I don't want to get too much into Microsoft right now because I want to be writing my zero-shot newsletter edition this Saturday is going to be about Microsoft. So please do read it and wait for my detail arguments in the newsletter itself. But you know what? Microsoft is also a great parable of what happens when you peak too early and announce your victory too soon. I'm going to quote something. At the end of the day, Google is the 800-pound gorilla. I hope that with our innovation, they will definitely want to come out and show that they can dance. And I want people to know that we made them dance. Who said this in roughly exactly 22 years back in February 2023? I wrote about this so I know the answer. I'll let Brady go if he wants to try. Brady who said that? The Microsoft's Satya and Nutella. That's right. Back in February 2023, Satya and Adela proclaimed victory and said we made Google dance. They say revenge is a dish best of gold. It's probably, I'm definitely sure Sundar Pichand knows about it because some Google CEO Sundar Pichand took 15 months to respond to that jive when in May last year he said on the Olin podcast. I think that maybe only one of them has invited me to a dance, not the others. Pichand was referring to a question on which which were the key players driving AI innovation. Open AI, Meta, XAI and Microsoft. Well, of course, he said only one of them has invited me to a dance. It's not a great comeback, but it will do. It's a brilliant comeback, right? Because I'm continuing with what where Google is right now. In January, Google became Apple's preferred AI model partner and will now pass Siri as a long-time Siri user. I'm just happy that something is happening there. But let me see, this is actually a pretty big win. It is absolutely. I'll read a bit from Mark Gurman, one of Apple's most, you know, the analyst that always scoops them and Apple has a terrible amount of anger towards Mark Gurman. He writes from Bloomberg. I'm quoting from his recent newsletter. We link to it in the show notes. At the time, Apple was already in discussions with both Anthropic and Open AI about supplying models to rebuild Siri and power parts of Apple intelligence. At the start, Google wasn't even considered a likely partner for this. The alphabet-owned company wasn't seen as the leading tech provider, and it was locked in a US government anti-trust lawsuit tied in part to its Safari Search deal with Apple. By August, negotiations with Anthropic had stalled. The startup was seeking several billion dollars annually over multiple years, terms that weren't favorable to Apple. And partnering with Open AI posted its own problems. The company was actively poaching Apple engineers and pursuing its own hardware ambitions and a former Apple designer, Johnny Ive, an obvious strategic conflict. That left Google. Apple revisited the company's Gemini software and found that the technology had improved significantly in just a few months. No shit, those words are mine. Moreover, Google was willing to agree to a financial structure Apple deemed reasonable. I'm guessing a lot many billion dollars were involved, but the flow or the direction might not have been the same. Again, those are my words, not Mark Gurman. Isn't this some form of a barter deal? Like, remember that Safari Search is being powered by Google. That's right. And Google pays Apple billions of dollars for that privilege. So I'm sure it's fair. Yeah. So now on this side, it is okay. You power our Siri, we power out this thing. There was also some fortuitous timing in September. Sorry, just a second. I have, wow, I've just been given a warning by my macOS that I've never seen this. Your system has run out of application memory and it's recending that I shut down Microsoft Word. I kid you not. Were you using cloud code over the weekend to build something? No, I have never. I'm using a Macbook Pro. I have never seen such, you know, but I like that the AI of the laptop is really good. It knows the most useless software to kill. Oh, no, actually, I disagree with that because I am using Word right now. Yes, I do write in Word. You mean co-pilot? Yes, soon to be co-pilot, but till such time that I'm a subscriber, it's still word to me. Yeah, I mean fortuitous timing in September, last year, a judge ruled at Apple and Google's roughly 20 billion per year. That's a deal that you were referring to. Google pays Apple 20 billion dollars a year for Google to be the default search on Safari. That would not need to be unbound. That made it less risky for Apple and Google to expand the partnership. By November, the two companies were finalizing their agreement. Google would supply Gemini models for Siri and future Apple intelligence features. And of course, they would run on Apple's private cloud servers, but down the road, they would run on Google's custom hardware, right? So this is a fantastic huge win. It's like Apple is not even insisting that they run it on their own silicon. It's just like, okay, fine, run it on your servers. Yeah. So how is Google doing? This isn't obviously just about the Apple deal. Gemini, its usage, active usage has gone from 450 million users globally to 650 million users from just July to October 2025. Gemini API calls for enterprise and business users doubled in just the five months to August 2025 to 85 billion as per the information just last week. Gmail now is stuffed to the gills with AI. I've mostly turned it off, but Gmail has a 25 to 40% market share globally depending on who you ask. What about Android? 70% market share approximately globally to 99% in India controlled by Google. YouTube, Google has exclusive access to YouTube, which is why if you ever try to use another AI tool to summarize a YouTube video, it says, sorry, I don't have access to that and which is why I don't book it. So brilliant. What's YouTube's market share? 97%. What about search? You search for everything and the first thing that comes up now is AI responses. Well, again, it's Google 90% market share. Add Apple's deal, which bring in the iOS and Mac ecosystems. Add TPUs, which means Google's own chipsets. And well, you have this situation where if you look at it in 2026, Google is among all the tech giants, whether you call it Mac 7, AI giants, etc. Google is the one whose stock is up the most, most others are actually down. Google today, I posit, is in this virtuous cycle of A, being able to control all aspects of AI from consumer services to enterprises to consumer hardware, to data centers to even its own chipsets, all running its own models. I don't think there's any other organization that has this sort of ability to do everything top to bottom, not even Apple. Be, it's in the sweet spot today where adoption and appreciation for its products services models is rising. See, investors are rewarding it for all of this, which means its stock is going up and D, its rivals mostly all are on the defensive. So I posit that Google, this is absolutely Google's year of rain where Google should and is you can clearly see putting its foot on the accelerator because it's able to kind of do all these things. It's able to say, you know what, we are not in a rush to monetize users because like nobody else can say that. It can essentially go everywhere from going after Nvidia to data centers to Android phones to web devices, everything while everyone else is being circumspect. So I do feel that this is going to be Google's year for better or worse, for it to make or lose. My question is to you, therefore, this wasn't just a random solid lock way is what could trip Google or most specifically who could trip Google. I'll kick off the discussion by actually just announcing one person who I definitely know has the ability to trip Google, who might that be? Any, any guesses? Brady, any guesses? There's only one person in the world who has the ability to trip Google. Well, obviously Donald Trump has the ability to trip Google, all right. So let's leave Donald Trump aside, right? Because, you know, he just has a changes, Wikipedia page to see your Google, and that's it. That's right. What could trip Google, who could trip Google? But of course, let me pull back because Brady, you were saying you actually disagreed with my, well, I give fantastically researched hypothesis on Google. I am disappointed you would do so. I thought you disagreed with me and I said it was a good deal, but I let Brady go. I just agree with both of you. Oh, okay. Wow. I like it. Brady, coming out. We'll go today morning and chose violence. Yes. Go on. Yeah, I'm going to be violent today. So I'm going to start with one idea. And then there's kind of an arc that covers Microsoft and bite dance. And hopefully we land somewhere. But this, this main idea is I don't think distribution guarantees AI dominance in the long run. Especially when models or I should say tokens become commoditized. And by that, I mean tokens going from a scarce resource to something that is abundant and low cost and easy for anyone to access. So, you know, Google is accelerating. Rohan, your very well researched introduction, you know, lays that out very nicely. And I do believe that any model that isn't used at scale is worthless. But I want to reference what Nadella said much more recently, not two years ago. So he was speaking in Davos around a week ago. And he was thinking about how how many tokens can be produced per watt of electricity. And that sounds a little boring, but, you know, give me a, give me a minute or two here. The big themes here are token factories and energy efficiency. And I think there's a lot of merit in that. So taking Nadella's idea as a starting point, the consequences that the real competition in the AI industry is infrastructure efficiency, not distribution. And then I think it's fair to say that a model is only continually used at scale when token prices are low enough. And of course, you know, I'm assuming that the models perform as advertised. So those Google's installed base advantage matter if cost discipline isn't quite there yet. I'm not sure. I lean no. And then let's take this a little bit further. If token costs do fall industry-wide, and I'll talk about how that can happen in a second, you know, does winning distribution just mean subsidizing more users at unit economics that are like everyone else's? I think the answer is yes there. So, you know, as Rohin said, every company is dealing with their own rainstorm. And Microsoft has that with co-pilot. We've talked about this. And even though that's the case, Microsoft is the only one that's talking about the actual nitty-gritty of the AI industry's economics. When Google is taking a victory lap for its distribution win. So, Nadella sounds a lot more like somebody who understands the big picture. And that big picture is not so much about model capabilities and more about operational excellence. Right. Of course, Nadella will say that. I'm going to let BGK come in because I do have strong ones. I'll just take this and maybe I'll make it a slightly imperfect analogy. What you're really saying is that energy efficiency matters and operational excellence matters and more than distribution. And as token costs are going to come down, you know what this really sounds like? This sounds a lot like Telecom. This sounds like basically saying maybe like 20 years back that look, you know, it doesn't matter what the distribution is at the end of the day. You're just going to, you know, per MB or per GB or this thing is going to go down and instead of putting it in terms of data centers in the old world, it was about fiber optic cables, etc. Right. So, you're saying that it's a very similar thing. I get that. But that does not the same as saying that distribution doesn't matter. In fact, I'm not saying that I don't think distribution is going to necessarily be the where the margin exists. But what is probably going to happen is that owning the distribution is going to be an immense advantage. It's like saying would Google or Microsoft be, let's take Google, would it be better off owning a Telecom network? Of course, it is. There is this company called Geo. In India, it owns a Telecom network and it is able to use that to power its application. It does not make money on the network side. It makes money by doing all the applications like OTT and payments and retail and everything else, which is what Google is going to do. So, or whoever wins the, you know, the token war in terms of creating that pricing thing is going to use that to leverage applications on the other side. I agree with you that model companies are not where the money is going to go. The company is going to be made on the application side. And guess who is an application company? Google. Let me come back to, let me, as they say, let me unpack Brady's arguments. Let's start with the very interesting one that he said, which is the token factory. And there are two things that I want to kind of compare. One is that the token factory. And I think you brought up the, you know, I'm going to do this thing that I do very well, which is being able to connect, like, you know, themes. You mentioned telecom networks. Yeah. Do you remember what Airtel was called when it became super successful India at scale? Yeah. You know what it was called? What factory? What was it called? When Airtel? Yes. Airtel used to be called the something factory. The. Or telecom companies in India were called the something factory. They were called the minutes factory. Okay. Correct. Yes. So, I want to bring in Brady's token factory example to the minutes factory example from telecom networks. And let's, let's start backward. Let's start with telecom networks, which is a great example, right? You know what's, what differentiates minutes from AI tokens? All telecom minutes are the same. If you're talking about voice, it's a commodity. Yeah. If I want to call you, I don't care, you know, what that like, you know, and as long as the call is not breaking and, you know, somebody else isn't like listening to my calls and stuff like that, your minute is as good as my minute as opposed to somebody else's minute, right? So minutes in telecom networks are a fungible product, which means that the more minutes you create, the better you can run because, you know, minutes are a commodity. Now, let's come back to AI and the point about token factory, right? Sure. At one level, you can say tokens are nothing but, you know, I mean, in some version of like data packets going back and forth. So obviously, there is some truth to the, you know, the claim that at the end of the day, if you can do a better job of that, you're more efficient. But what tokens to our tokens completely interchangeable, are they fungible? Does it not matter to you whether you're exchanging a token with co-pilot or whether you're exchanging a token with open AI or with cloud, etc. Of course, it matters. Why does it matter? Because the underlying models matter. Why does that matter? Because there is a huge variance. They're not the same. You cannot in a telecom network or in a minute's factory. You can say, you know what, I'm just going to shift from and let's let's talk about the voice era because that's where the minutes factory thing, right? Once you bring in apps and data services, it's sort of become different. You could say, I don't like this telecom network. I'm just going to go to another telecom network. What's the difference? It's all the same. I call you. I'm still able to talk. Is it the same though in the AI space? I mean, my entire preamble was really about who exactly is saying, I'm going to use co-pilot because Microsoft is doing a much better job of keeping its own cost low around the tokens of co-pilot, right? Sure, it helps Microsoft. I'm not saying it doesn't matter. It allows Microsoft, I mean, let's take their claims at face value and say Microsoft is able to figure out how to operate at the lowest possible token costs in the industry. What will that allow it to do with it? Right? Obviously, there's no point of claiming, you know what? I have the best margins if nobody's buying a service. So, I mean, if you're a Microsoft or any company, you'd say, because I'm able to operate at such an efficient level, I'm going to use that to go to my consumers, whether they be, you know, consumers or enterprises, and say, you can use my products cheaper than others. Correct. Right? But at the end of the day, that is the reality. You still have to go to them and convince them to use your product because your product is better. It's not just cheaper. Remember, we still haven't reached that stage of AI where there has been certain convergence across all models. No, I'm pretty saying that that is inevitable, which I tend to agree with. No, I'm not sure of that. I don't, I mean, I don't see anything right now to say, I mean, that's like saying singularities inevitable. So, or in the long run, we're all dead. What evidence do we have to say? Claude is going to Claude's performance or Gemini's performance or XAI/Grocks performance is going to ultimately be the same as, let's say co-pilot's performance. What is that path together? Unless, like I said, at the core model level, it might be, you know, like if you look at all benchmark metrics, et cetera, it might be these are all like on top of it, you Claude will build things like Claude code, it'll build things like Claude for financial services, et cetera, et cetera. But those are all applications that are built on top of models. And what happens? How do models evolve? How do models get better? I think we have already discussed on this podcast that there is going to be some kind of limiting this thing over there. No, no, no, it's, I mean, fundamental reason models get better is through better, through more usage. Right? Sure. A model which is getting, if you stack 10 models on a chart, and one model is getting used by quality consumers and businesses, small and large and enterprise who are using them to code, to generate images, to refine them, et cetera. And that model is able to learn from those and become better. Okay. If that model, that proprietary model is going to evolve to become a better model. I mean, right now, let's assume Citrus Paribus, we are not talking about cost. Let's assume all models have the ability to spend or sustain their spends on tokens. The models that get better are the ones that are able to learn and improve themselves better. That's correct. I mean, Brady, you see this right? So it's, it's not exactly the telecom analogy because there is a network effect part of this. Thus, I come back to distribution. And I'm saying that, yes, Brady is right. But, you know, there is this thing of, he's going to be right, but right when, because we're in the period right now, where, you know, large AI companies are trying desperately to get consumers and businesses hooked to their specific version of AI. They're trying to build loyalty. Correct. So all these articles that we're seeing about how cloud code is so great, why cloud is releasing, cloud work or Gemini is so great or Nano-Granana is so great is because behind the scenes, everyone is saying, you know what, my model is better. Can I tie you into a subscription? Can I turn you into a loyal user? Hopefully, and include open AI itself, right? So that at some point, you'll be like, you know what, I can't just switch over to another model because this model knows me so well. It, right? So that's the phase that we're in right now. Correct. So while this phase is going on, if Microsoft sits it out by saying, you know what, let everyone else focus on distribution and consumer adoption and improving models. We'll just make our token cost cheaper at the backend. It's essentially shutting itself out of whatever the next time is, whether it's 12 months, 24 months, I don't know what that time period is. By the time, and by the way, what's also going to happen is these other companies are continuing to grow, they will do their IPOs, they will continue to improve, right? Of course, they are going to reduce their own costs as well. Yeah, yeah. We're in agreement there. I mean, so that's like saying that Microsoft's only decides, oh, we're going to become a telecom company. It doesn't make sense. That's right. So Microsoft could absolutely say, you know what, our costs are the lowest, but to actually answer that, we have to ask us, honestly, like who exactly today, which enterprise user is saying, I chose this model over the other because they cost our low. Have you come across one instance where a company or a business is saying, you know why we evaluated these three models, but this model quoted us the lowest cost, so we picked them up. So again, it seems like this mindset to me reflects, I think, you know, I mean, I mean, the two ways to kind of analyze what Sathya Nadella is saying. One is, of course, he's picking the thing where he can say that I have an advantage, and also nobody else will bother to refute him on that, right? Do you really think right now, open AI is going to come into an argument saying, you know what, our model token costs are lower, or Claude, you know, they don't care. They're like, look, you can continue to claim this crown. We are going after users and subscriptions and loyalty, right? So I just don't know, right? There's thing of, you know, it's somewhat, I mean, it is accurate, maybe in two years or three years, once sort of everyone has shown in their respective user basis through subscriptions, etc. Then people will say, all right, who's operating more efficiently for every 100 million users, Google is operating at X efficiency because it has its own TPUs, Microsoft is operating at Y efficiency, but today we are not in that. Not in that era. We're in the, you need to win this consumer and adoption and usage race before you ever get down to the efficiency of token zero. Correct. Last point and I let Brady jump in. I mean, I'm just trying to be charitable here. Maybe another interpretation of what Nadela is saying is nothing to do with the models themselves, but more to do completely with Azure, where he's basically saying that look, we are going to build data centers and run Azure on it at that is going to be way more cheaper for anyone to deploy models. So imagine a future where Claude deploys its models and its software and everything else on Azure and it's hosted on Azure as opposed to say an AWS or a monel. Sure, that I can see because in some senses, I mean, a data center, a data center, right? Microsoft Azure is going to compete with Google Cloud is going to compete with AWS with Corv, with so many other, with the Adanis in India and for some reason, maybe Nadela is thinking that look, our costs are going to be, we are going to be much more efficient there than everybody else. But then it becomes more like a question is why? The question is why again, like, you know, we must interrogate, right? In today's world, like, you know, why will Azure be able to operate lower costs? What does it have a particular hardware advantage? I mean, sure, by the way, I think Microsoft did come out with a new bunch of chipsets, just I think 24 hours ago. So again, we're not getting into those technicalities. Is Google able to operate TPUs at a better efficiency than Microsoft's new advantage? Because it uses NVIDIA / Corv, etc and stuff like that. But again, I think those are, I mean, I don't have enough information to say that Microsoft has so much of a hardware advantage that it is able to operate much more efficiently. I have not seen any convincing data that says that Azure is able to do that, right? So to come back to essentially Google. Oh, sorry. I mean, I countered your point. Do you want to counter my counter? Maybe just like an addendum. Yeah. So, you know, when I was thinking about token economics and Microsoft, you know, Microsoft has Office/Copilot, which we've gone over. Teams GitHub Copilot, which is a different thing, and Windows, and it's jamming Copilot into all of these things, right? And, you know, all of these products, especially when they're used in the enterprise setting, create like natural demand that would drive that token economics efficiency. And, you know, I was immediately reminded of how bite dances, internal consumption forced that cost discipline. So it had a couple of products in particular that used a lot of tokens. One is Douyin, which is like TikTok for the China market. Another is like a content aggregification platform that's super popular. And of course, it's own content moderation. And in bite dances case, AI was kind of thought of as internal infrastructure cost from day one. It wasn't a product to monetize. So they managed to drop the token prices close to zero, actually, by leveraging just like their internal consumption at scale. And, you know, when I look at Google, it has YouTube, it has search, you know, Rohan went over the numbers earlier. But I don't think those are necessarily the right token consumers to drive that cost discipline. I haven't really thought this through completely. So if you think about search, right? Every time a Google search user hits search today, it is consuming and like tokens because AI searches everywhere. Look at Gmail. I mean, the Gmail AI interfaces and like, you know, the way it's consuming this all the time. Look at YouTube now, the amount of like, you know, AI features which are coming to YouTube. So I think Google has this ability because of all these hundreds of millions, but actually billions of users at scale across its devices. It has the ability to optimize it, both is generating an incredible amount of token usage every single day. But also in a way that it can control the experience. Not many others have that because I mean, if you're a, let's say you're Claude or open AI, you know, the interface of token consumption is either through an API or someone comes to you and says, here is my search query. Now tell me what? I think Google because it has all these embedding, right? So when you go to, like, I'll give a trivial example. Just the other day, I was discussing with a friend. What's, I mean, I cannot for the life of me watch long YouTube videos or even short YouTube videos if they're beyond five minutes, right? So the first time someone sends me a YouTube video, I just take it to notebook LM and I just use it to summarize. Now, I'm thinking, think about this, right? Only notebook LM and Gemini, et cetera, can summarize YouTube videos because when you Google has access to that. But think about it, do you really need to run a fresh query every time a video needs to be summarized, right? Because, you know, it's incredibly efficient. Once somebody summarizes that document, all you need to do is watch for any changes to that video and if there are no changes, virtually you're, it's almost like Google search serving you the results, which has been pre-cashed on its network at a server really close to you because it knows that enough people will ask this query so I can just send it to you real fast. So, you know, those kind of advantages are just, you know, I mean, now it's apparent how the Google web sort of comes together, right? Or what it looks like when Sundar Pichai dances. I almost want to make an AI video Sundar Pichai dancing. I'm sure it exists. You started by asking who can trip Google or what can trip other than Donald Trump. Other than Donald Trump. I mean, if you keep regulatory aspects aside, which anyway, they have sort of gotten over that because they were under a lot of regulatory risks some are right, which they seem to have gone past. Actually, that's interesting because the their regulatory savior is the same one who can use their regulatory threat, right? Globally, American companies, like they're like, let's say if EU wants to regulate Google, like, you know, this thing is like, Trump is going to say, no, you can't regulate. But Trump is also the same threat. So it's like in Hindi, there's a saying of my bop, right? Like, you know, so but yeah, that aside. Yeah, so this is not a case of somebody who can trip up Google, but I'm just saying that it's and it's something interesting and I'm not able to figure this out. So it might be a disruptive aspect or it might be very opportunist. Asteroid crash. I have no, no, no, I'll explain what it is. Okay. No, no, no, no, we don't have to go all the way there. We can go to this one company called OpenAI. You know the news and Rohin, you mentioned this sometime back about OpenAI launching ads, right? Have you been following the OpenAI ads thing for a while? We spoke about it. I think last week where we said that they said that all it is our last resort. And now it looks like they have to do it. So they are doing it. So I'll just say a couple of interesting things about this ad push that OpenAI is doing. And again, I used to do a little bit of mobile advertising in the past. So it's just very interesting and pop lexing for me. And that's a panel explain. So first, if you remember, first, the news came out that said that OpenAI is going to try to do shoppable ads. So they're going to basically put these links up. And if you click on it and you shop, they're going to take a cut out of it. Rohin, do you remember the percentage from it? Some four percent, five percent. Some cut they're taking from it. But the e-commerce transactions. They come as four percent, right? Now, I'll start at the first point, which is that this is a very strange place to begin because in general, shopping ads or trying to do ads that do shopping and commerce is like the holy grail. Like if you remember at one point in time, Facebook and even Twitter now X had this thing called the buy button where you see a pro, you see something, it comes as an ad on your feed and you can immediately buy it. It's a really complex and a weird thing to do because people don't really shop like that. They tend to essentially buy, you know, things like they want to go to the website. They want to experience a little. You don't just buy off and add that easily. Again, the first thing that I did when I joined in movie was work on shopping ads. So I can tell you that it's actually really, really hard to make it work. You need to know the merchants, you know what the merchants have. Their inventory is just really hard to make it work. So that's one. The second thing that it did was they launched this thing where they said that, oh, we're going to do ads and by the way, the information broke the story where OpenAI has gone to advertisers and said, oh, we're going to do ads, but you're going to be premium. And here's the weird part, okay? They're going to sell these ads and getting a little bit into the weeds here, but it's interesting. They're going to sell these ads on a CPM basis. Do either if you know what CPM basis is, but for me, a thousand, a million, right? So you basically sell what you're really saying is that for a thousand times that you show something to use us, isn't it cost per thousand? It's cost per thousand. For a thousand. Mill is the French one. Mill is the French. So it's called, yeah, exact. So it's called CPM, right? So you sell. In my word, mill means mill, short, a million. So you sell and you basically say that for every thousand people who see this, you pay us X, right? Usually like $1, $2, $5. Now here's a weird thing. Number one, OpenAI is charging $60 per CPM, which is like a ridiculous amount. First, second, I don't even understand why you're selling something by CPM. So as a very, very, very quick two-minute ad history, you start by selling something in CPM because, and the very initial stages when advertising started coming on the internet, nobody knew what people are going to do, what users are going to do. It was completely, like, uncharted territory. So you sell eyeballs. That's exactly what you do. So you sell eyeballs and say, well, we're going to sell eyeballs and that's it. We don't know. For every thousand eyeballs, we give you, we pay us X. We don't know whether they'll buy anything or not, but we'll give you eyeballs. Exactly. And so over time, nobody does this anymore. Nobody really sells our buys based on eyeballs anymore. People buy based on what is the action they're going to take? Are they going to download an app? Or you sell based on here is the audience that we will show it to based on these interests, based on these demographics, etc. Because things have gotten most sophisticated on the internet. And I am very perplexed why on earth is open air trying to sell this and go back to an era of 20 years ago. That's the first point. Now, you would think that, well, they may be they don't know any better, but that's also not very convincing because I went and looked up their head. So their CEO of, you know, applications is this person in Fiji Simo. She's a 10 year monetization person coming back from Meta, right. So clearly, she knows how to do this. The entire team knows how to do this. And in fact, the people who actually tried ads sometime back and screwed it up and had to walk back was this company called, they even know the AI company that tried ads and had to walk back. It's a favorite company of sealer shut. Open AI. No, the second favorite. Microsoft. Anthropic? No. Publicity by Fred's. Publicity. Okay. Publicity like everything. Walk back. Publicity should do this thing where they start walking back things. We are going to give free publicity pro to all the hotel users. We walk back. So they also did ads. They tried this for like one and a half years. They got this person who's I think an influencer marketing person who had a startup that got sold to one of the ad agencies. They tried for one and a half years and they said, okay, this is not working and they scaled it back. So and now last week, the CEO of Deep Pined of Google was asked at some event where, oh, I mean, Open AI is doing ads. What about you? And he said, we have absolutely no plans to do ads on Gemini. So this is one thing. I'm looking at this and I'm wondering what's really going on here? It's not like Open AI is doing ads the conventional way. They're doing it an unconventional way. Google is looking at this and saying, yeah, be my guest. Please go for it. We are not interested in this at all at this pointed type. And so does opening. I know something that Google does not, which sounds very strange and unlikely. Exactly. So it looks like it's an incentive thing where Open AI is just going to say because one of the advantages of selling based on CPM is that you can lock up huge brand deals. So imagine going to like all the ad agencies of the world and saying, I know you all have some part of your budget. You want to spend on AI and AI users because that's definitely going to happen. Just give us a portion of that budget and it's going to make it really simple. We're not going to go deeper into who clicked on it, what happened, etc. You can go and tell your clients that, oh, we put money on AI as well. Your ads are coming on AI and that's about it. And so that's really what the plan is. Okay. So Provenessing, maybe Open AI could trip up Google and you know what Open AI becomes after it runs out of money, right? Part of Microsoft. That honestly, because I was actually thinking about Microsoft's survival or long term strategy. That has similar spontaneously to, you know, has to be Sam Ottman's biggest goal to essentially become unacquirable by Microsoft. And I don't think anyone else can do that, right? It's only because of whatever structures and deals, Microsoft and agreements that Microsoft has with Open AI based from the early days. And at the same time for Sathya Nadella and Microsoft, that has to be the one that look. I mean, we brought this thing into the world and we gave it all the money and now it's grown up to be this giant. If we let it get away, you know, so right, that has to be the move to watch out for. So from Microsoft, I was like, you know, I mean, we but the incentives are like very strange, no? Because it's like walking a tightrope. You want to watch Open AI's tumble and struggle, but not so much that it becomes like after you acquire it, it's just going to be dead. So you want it to like really, really struggle to a point where its valuation crashes and nobody wants to touch it. And then Microsoft can swoop in and take it, creates a consumer play. Yeah, sure, I can see that. I mean, I don't know. I mean, why do you have to assume that it has to be after the IPO? Why can't it be before the IPO? Yeah, but opening AI is dealing with multiple issues, right? It has Elon Musk fighting it in the courts over its conversion from a not far profit to a far profit. Correct. It has all these other deals that it had signed with its earliest investors. Yeah. Biggest of whom is Microsoft to whom it had promised certain, you know, amount of returns till such time that they get to Asia and while Asia is not coming anytime soon. So Open AI has this like thing, you know, don't lose to Elon Musk on the case. Don't lose to Microsoft. Do the IPO. And therefore, so all these, I think the IPO is strategically important to Open AI to essentially eliminate many of these because once it's a public company, it's not going to be so easy to essentially take it down, even given its skill. All right. That was a wonderful episode. If I might say so myself on Google, tell us what you thought about it. I told you at the beginning of the year that we're going to be changing up zero short episodes based on who the lead host is. Sometimes we have three stories. Today we have only one story broadly and that was on Google. What did you think of it? Please write to us with your feedback. Of course, you can write to us as zero short at the ken.com. And of course, rate us. Not enough of you rate us. I may say it. With that three, have you rated us? Yeah, yeah. Okay, fine with I should check myself. I need to rate us. Thank you. And we'll see you back next week with another wonderful episode of Zero Shot. That was Zero Shot, the Ken's weekly podcast on the biggest developments in artificial intelligence. Our hosts and commentators are Praveen Gopal Krishnan, Rohan Dharmakumar, and me, Brady. Our sound engineer is Rajiv CN who makes everything sound spectacular. Don't miss our Zero Shot columns which are published every Saturday. We'll be back with more for this podcast next week. [BLANK_AUDIO]