Transcript & Summary: Who's winning (& losing) the AI race?

[00:00] We just got our most detailed look at AI adoption yet. 70 pages from Sensor Tower showing which AI companies are winning and which are losing and what people are spending their time and money on divided by gender, geography, and more. In addition, the finances of both OpenAI and Anthropic have recently leaked, giving us a look at what running an AI company actually looks like. I've linked to the full reports in the description, so you can read them yourself as well. But this means we can now take a look at [00:24] what the AI industry actually looks like. Is this a bubble? Will it burst? Let's take a look. [00:34] This video is sponsored by Nebula. I think by now many people have heard that ChatBT has been the fastest app to reach 1 billion users. And even on phones, this has been remarkably quick, right? So this is just mobile users on iOS and the Google Play Store. And here they've reached this in 3 years, which is significantly faster than any other app in history. I think this is pretty wellnown. But while Chad GPT was absolutely dominant in the early days of AI and they've also continued to grow, right? Like they're well over a billion [01:03] users right now. They do have some real competition now. There's Google Gemini that stands out especially which now has well over 600 million users. That is a lot. Obviously, they gained this in part because they pushed Gemini into existing Google products like Android as well as all the Google services on the internet. But it was also done partially by Gemini just catching up with chat GPT and capabilities over time. But the second big standout here is very obviously Anthropics Claw. They now have well over [01:30] 200 million active users and they've achieved this without having the early lead that chat GPT had and also without having the massive breath and catalog of products that Google had that they could fall back on for promotion. So that means they've achieved this mostly by just having a product that people actually wanted. Now, interestingly, Deepseek, which received a lot of hype when it launched, yeah, it popped a little bit when it launched, but then it also faded away almost as quickly as it grew. Uh, Perplexity also kind of faded [01:57] into the background and you can see that uh, Grock, uh, Meta AI, Microsoft Copilot, they're essentially all in the also ran category. Not really something that we have to think about. In other words, right now this is basically a three- horse race. There's Chat JBT, there's Gemini, and there's Claude. Now, I have four things that I'd like to add to this graph. First is that Siri has just launched and it has the same advantage that Google's Gemini has, which is that it's built into the operating system and the services that [02:23] people actually use. So I wonder if it will start showing up in charts like this very soon. Second, there was a lot of talk about how open- source models would win by being cheaper and more broadly available in general. And at least until now, that doesn't seem to be the case, right? Deepseek has kind of come and gone while the actual winners of the AI race until now are the companies that have built the most capable models and also made them available to the users very easily. Third, I noticed that while the chart [02:48] says that this is worldwide usage actually, if you read the fine print, this only includes 25 markets. Now, most of those markets are the ones that you'd expect, you know, US, Canada, India, Japan, etc. But it doesn't actually include China. And I'll get back to that later in this video. And fourth, the chart says that it measures unique users across mobile apps, mobile web, and desktop web. Now, while that actually captures most people, I was thinking about how desktop applications are explicitly not mentioned here. And I [03:18] would think that Copilot in particular would be mostly used on the desktop built into Microsoft's desktop applications uh or the Copilot app itself, which is a desktop app. So, I'm guessing maybe that gets under reportported here. All right. Now let's move on to the next chart which shows the exact same data but as market share percentages. We can see that ChatGpt has been losing market share since the beginning and that it now sits at below 50% for the first time ever. This is 46.4% market share for Chat GPT. The runner up [03:47] is obviously Gemini with 27.7% and then we have Claude at over 10%. All right, so now we know more or less how many people use each of these assistants and whether they're growing or not. And in this light, I find their valuations extremely interesting. Enthropic and OpenAI are both expected to IPO soon, and both have been rumored to raise money at valuations of $1 trillion. If the two companies will really be priced equally, then I think this means that investors think that Anthropic's cla will keep taking share away from OpenAI [04:16] aggressively and also that Enthropic is probably better at actually making money from its existing users, which we'll see later on. Next, Google's stock has been doing great, which aligns with Gemini's market share going through the roof. While Metas and Microsoft's stock have both been doing poorly, which also makes sense because their AI bets have mostly had limited success. The only AI story that doesn't really seem to be matching these numbers at all on the stock market is Grock. Like, their market share is [04:40] minuscule. They're at 3.3% which is basically the same as Microsoft Copilot or Meta's AI, which are not exactly considered great successes. And yet, investors don't really seem to care. SpaceX, the parent company of Gro, has just IPOed, and at the time of writing, they're one of the 10 most valuable companies in the world. Basically worth as much as Amazon or Microsoft. And if you read the IPO filings, SpaceX itself said something like 90% of their total addressable market is in AI, not in rockets and Starlink. So the company's [05:09] leadership says that investors should value them as an AI company, not a rocket company. Now, if SpaceX really is an AI company in a trench code, then having a market share of only 3.3% for their Grock assistant, combined with the fact that their AI business is actually losing money every quarter is rather confusing, but okay. Okay. Next, let's take a look at how much time people are actually spending using AI. So, first looking at mobile apps on iOS and the Google Play Store, both time spent and [05:38] total session count are growing rapidly. So now we're at the point where people are spending 36 billion hours in half a year using AI apps and they're also doing almost a trillion sessions in half a year using generative AI apps. Those are some crazy high numbers. Now on the next page we can see the same numbers but for AI websites. This is both on mobile and on the desktop. So the first one was uh mobile apps. Now we're doing websites and we can see a couple of interesting trends as well. First, uh [06:09] pay attention that we're looking at quarters here, not half years. Uh so the numbers are not exactly comparable. And we also see that there's a little bit of a flatlining here, but overall there's growth. And the numbers again are incredible. So we have something like almost 70 billion visits to AI websites and also more than 20 billion hours spent on AI websites on mobile and desktop. Those are once again incredible numbers. So, if there's a case to be made for AI being the next big thing, I think these numbers are definitely a big [06:39] part of it. It's clear that people are using AI all day, every day, at least many people are. In my comment section and also in other parts of the internet, I often read of people who are saying that actually nobody wants to use AI or if this is all a scam or that we're all just tired of AI being jammed into everything. And I'm certainly sympathetic to this line of reasoning. I think there's too much of that everywhere. But it's clear looking at these numbers that a lot of people do want to use AI and they want to use it a [07:04] lot. Okay. And another part of this data that will make AI companies very happy and people like me, tech YouTubers very unhappy is charts showing how AI is actually changing shopping habits. Starting with this one, we can see Amazon conversions based on whether the user used Rufus or not. So this one shows people not using Rufus. This is the average and this is people using Rufus. So very clearly people who did not use Rufus converted at significantly lower rates closer to 20% versus people who did use the AI assistant Rufus those [07:36] are converting at 40% or above. Meanwhile, if I switch to this chart, we can see how much time people actually spent on Amazon depending on whether or not they used Rufus. So people who did not use Rufus, the AI assistant, they spent something like 14.6 6 minutes on the website versus people who did use the AI assistant, those spent more like 40. That's more than a doubling of the time that the user spent on the platform. So the way I read this is that people who just want a simple purchase, they go to Amazon, they do for example a [08:05] search and then they quickly leave the platform without using AI. But people who have a complex question and they actually have to do research, they do use Rufus and then when they do, they also end up making a purchase decision way more often than those who did not. Meanwhile, if we look at shoppers across the United States, not just on Amazon, but also in other web stores, then we find that computers and consumer electronics were the category that were the most impacted by AI assistance. I'm guessing that this is a highly complex [08:33] category, people need to do a lot of research when they decide to buy a pair of earbuds or a new computer or something like that. And whereas in the past they used to watch reviews of reviewers or they read in-depth tech guides on the internet, now they just ask an AI assistant and that AI assistant summarizes everything that we do for them without giving us any money. Great. Not exactly an uplifting thought for someone like me, but uh let's move on before I have some existential thoughts. And specifically, let's move [09:01] on to the part of this video that will give AI companies some existential thoughts. Because really, I can see two potential problems for them. slowing growth and profitability. Starting with growth, while the AI apps are still growing overall, there actually some first declines that we're seeing. Asia here has actually declined for the first time ever. North America is essentially flat. Africa has been flat for a while. The Middle East has actually declined, etc. Keep in mind that this is only for [09:27] iOS and Google Play Store installs for mobile apps. And also that it only includes iOS apps for China. So, it's not a complete picture, but yeah, that's actually kind of worrying. Now, importantly, there's still over a billion app downloads every quarter, but it does look like we've started to see the first signs of some kind of saturation. And second, let's talk about revenues. And here, the numbers that we have from Sensor Tower are not like the most amazing thing in the world. So, first of all, this is only showing inapp [09:54] purchase revenues. So, those are only the revenues that companies made from people buying AI subscriptions, for example, through the Google Play Store and the iOS App Store. And there too, it only includes the iOS app store in China, not the Play Store because they don't have the Google Play Store. Plus, then also keep in mind that the bars that you're looking here are for halfyear figures. And this last one is actually estimated, but still these are the numbers that we have and we can already see a couple of things. First of [10:19] all, for half a year, inapp purchases made about $4 billion in this chart. Now, these are actually growing relatively rapidly, but $4 billion is just not all that much. You will not be surprised to hear that in terms of revenue, North America is leading the pack at least on these app stores. Then is Europe in number two and Asia is in number three. Of course, if China was in this chart beyond just iOS users, I think Asia would be in a higher place, but this is what we have. Now, in terms of companies, we see a very interesting [10:51] breakdown. The average monthly revenue per user, at least in the United States, shows that Claude is the clear winner. They're now making almost $3 per user from each user compared to all the others which are significantly lower. And when we're looking at iOS users only in the United States, then we can see that Claude has a share of paid users of 13%. 13% of iOS users in the United States actually pay for Claude on their phone. That is a crazy high number and it kind of explains at least in part why [11:23] investors are actually enthusiastic about Anthropic despite it having fewer users. And by the way, I'm especially surprised by cloth here because I thought that their users will be paying on desktops primarily with things like cloth code, but apparently they're strong even on mobile. But now onto the bad news. While inapp purchases have increased to over $4 billion in half a year, at least I guess they are now. Uh that's just a very small drop in the bucket. We're only talking about mobile users here. So there's more revenue that [11:51] we'll talk about later. But this kind of revenue does not justify any of the valuations that we've seen so far. It doesn't even justify the investments that we've seen into data centers at all. We're looking at something like 700800 billion just in data center new buildout this year alone. $4 billion is small potatoes compared to that. And sure, the revenues are increasing, but then so is the amount of money that we're actually spending on data centers, right? Just the buildout growth this year alone is apparently going to be [12:18] well over 50% once again. In other words, these are moving targets and the profitability of these companies is far from guaranteed. And to see what this looks like for an actual company, let's take a look at the leaked financials of OpenAI. The data that is shown in this chart was obtained ahead of their IPO filing by at Citron and then it was put into chart form by RS Technica and I'll link to both of those in the description. As we've discussed, OpenAI still owns roughly 50% of the market and they're also I guess medium profitable. [12:45] That would be my guess. Like some companies will be doing better than them, others will be doing worse than them. But I think OpenAI's financials should be a a pretty good baseline for understanding how the industry is doing overall. Anyway, you can see that the revenues actually grew pretty fast from 2024 to 2025. They went from $3.7 billion to $13 billion in a year, which is great news for them. But expenses also grew very fast. So, they're actually still extremely unprofitable. Now, keep in mind that these are the [13:09] numbers that OpenAI has prepared specifically for the IPO. Like, they've done all the accounting tricks. They've reduced all the numbers as much as they can to look as good as possible for the IPO already. Every company does that and yet there's a pretty sizable gap still. Anyway, the question will be if they can close this gap and for this there are two things to consider. All the AI companies are in the process of jacking up their prices massively right now. They're charging more for tokens and [13:33] subscription. They're placing more ads, etc. Now, I think they're doing this specifically right now because they're trying to juice their numbers for their IPOs. But if they can keep going with this without alienating the users that they want to keep, then that could lead theoretically to profitability. But then on the other hand, we have to take a look at their costs. Research and development is basically the cost of training new AI models. We're in the process of all the AI companies being in a mad scramble to win the AI race to get [14:01] ahead of the competition. And so they're spending a lot of money getting ahead of the others. research and development costs alone are more in OpenAI than what they have in revenue. So that's pretty crazy. Just new model training alone is more than the money that they're actually making. Uh theoretically in the future if we get to a point where there's maturity in the market, a couple of competitors have died out etc. Then theoretically these companies wouldn't have to keep training AI models at breakneck speeds and so this cost could [14:29] go down as a percentage of their revenue. Theoretically will it? Who knows? Now this next line is called cost of revenue which is essentially the cost of operating the data centers. It's essentially the inference cost so to speak. Will this go down over time? Maybe. Well, on the one hand chips keep getting better, faster, more efficient, but on the other models keep getting more complex and more resource inensive. So which one will win? Your guess is as good as mine. And the third line item is something called sales and marketing on [15:00] which OpenAI spent a mind-boggling $5.7 billion in a year. That is crazy for an internet service and especially for a company that is deeply unprofitable. Now I understand that uh OpenAI is in a race like there's the AI race going on. This is the one that they have to win and so maybe in this case spending a lot of money marketing your services could make sense. But uh I've never even seen really an open AI ad to speak of. But apparently it's a thing apparently especially a thing in the US. That's [15:31] what this paper says. Uh but I expect that this part at least will go down eventually as the competition decreases a little bit. Now to say whether OpenAI will ever reach profitability would require a crystal ball. I don't have one. I can only say that right now they're deeply unprofitable and it's also theoretically possible and not even particularly unlikely that investors might get cold feet before OpenAI reaches profitability. But before we move on, we should also take a look at Anthropic's leaked finances which have [16:01] been pretty widely reported on. The Wall Street Journal recently reported that quote mindblowing growth is about to propel Anthropic into its first profitable quarter. They say that quote, "The startup expects a 130% revenue surge to $10.9 billion in the June quarter and its first operating profit, defying skeptics of the AI boom." Now, that is a lot. It will mean that Anthropic is making essentially as much money this quarter as OpenAI did the entire year last year. I think it is part of why investors are apparently [16:31] excited about the company. But there are some caveats. First, these are leaked numbers from a single quarter that hasn't even finished yet. And we don't know how Enthropic does any of their accounting either. Enthropic is actively in the process of building hype for their IPO. And I've seen plenty of companies juice their numbers before the IPO just to make sure that they go out with a bang. And so, it's really hard to say anything about the longevity of the business based on these. Indeed, Aditron [16:58] points out that Anthropic signed a huge compute deal with SpaceX, for example, that temporarily gave them reduced fees for the exact months before the IPO to make them look profitable, for example. And who knows what other tricks they might be doing as well. Another trick that companies like Anthropic typically do before their IPO is that they start charging their users a lot more money all of a sudden. This shows crazy growth numbers, everything going through the roof. Of course, later on, the actual [17:23] users will get pissed off and then they'll start to leave the platform. they'll churn or they want to get a refund or they'll switch to a cheaper plan or something like that, but that will only show up a year or two or whatever later down the line and so by then you will have invested and that's it. Now maybe anthropic and open AAI will be huge profitable businesses going forward which is certainly possible but it's really hard to say that right now. I've shown you all the numbers that we have at the current moment and so let's [17:47] move on to wrap up. We have two more really interesting things to go through the different usage between people based on demographics and geography. Starting with demographics, we can see that the classic AI agents such as chat GPT for example are more heavily used by men rather than women. The biggest difference in ratios is for Grock which is very heavily used by men. 81% male to female ratio. But then if you look at the right side of this graph, we can see that polybuzz but then especially character AI and talk are [18:19] disproportionately used by women. Not only that, but the age distribution also shows that these apps have particularly young user bases. In other words, the primary target audience for these three apps are essentially women that are aged between 18 and 24 years old. Now, all three of these are apps that are marketed as AI companions, which I find particularly interesting. I think they were originally marketed as essentially AI girlfriends for horny dudes, but apparently the actual target audience [18:47] that they ended up with is as AI boyfriends for horny girls. I guess I keep hearing a lot about the male loneliness epidemic. But apparently based on these charts, at least the female loneliness epidemic is a thing that is happening as well. And maybe that's worth exploring for somebody more knowledgeable on this topic than me. But then another super interesting chart is this one. It groups people into so-called selected consumer personas. So these are stereotypical things like for example there are crypto traders, there [19:14] are home cooks, there are uh console gamers, there are shopaholics etc. And then it shows us which AI assistant they're most likely to use based on this persona. You can pause the screen to pick out any information that you want from it, but the highest match by far of any consumer group and any AI assistant is crypto traders using Grock. Crypto traders are almost four and a half times as likely to use Grock as any other AI assistant. That is a crazy high match. Now, I have to say being a crypto trader [19:47] and hanging out on X where Grock is embedded in the product like intuitively makes sense to me, but I also think it's kind of a partial explanation for the SpaceX IPO as well. There's a large group of people who personally likes investing their money, likes using Grock or at least they use it a lot because they're on the platform and presumably also like Elon Musk because again they're on the platform. And so when Elon Musk launches a new company on the stock market, it kind of intuitively makes sense that many of these people [20:14] would buy the stock almost regardless of what the underlying financials are. Like I found that many Apple users, for example, who use a lot of Apple products in their lives somehow assume that the rest of the world runs on Apple products as well. It's a similar case for other operating systems as well. Like think of the Linux user for example who thinks that it's essentially uh uh the underpinning of everything and actually secretly all the people want to use Linux or something like that. And I think it's there's a similar case for [20:37] Grock as well presumably just the people the the thing that the people use themselves are presumed to be more important than they actually are. But okay, let's move on to the geographic split. On the left you can see which AI assistants people are using the most right now. Whereas on the right you can see which AI assistants are growing the fastest. So the left is the current leader. The right is essentially the biggest challenger. Clearly Chad GBT is the dominant model in most of the world right now except for in China where DBA [21:10] is leading. That is a Chinese model that was developed by Bite Dance that is the company that is uh building Tik Tok for example. And then in Russia interestingly enough the leading model appears to be deepseek but it's kind of wild. Deepseek is more popular in Russia than it is in China. And I'm guessing that that is based on political conditions. Anyway, on the right, Gemini seems to be the clear challenger both in North America and in Europe. They're growing very fast. And if we add the fact that the new Siri will at least in [21:36] part be built on Google's technology as well, we can see that Gemini is indeed becoming quite a substantial challenger. But then when you look at this, there's also Purple in Argentina and even the UK for example, plus a couple of other smaller markets. And this is Dola. I haven't even heard of Dola before. Meanwhile, if you look at the rankings, it's pretty crazy to see that Dola is in the number three spot worldwide. Whereas in quite a few markets like Mexico, for example, but also Brazil and Argentina, [22:05] it is at the top or very near the top. Now, again, I've never heard of Dola myself before, but our first sign of what it might be is that its icon looks suspiciously similar to DBA. And indeed, Dola is basically an international version of the DBA agent. It is very successful apparently especially in South America but then also growing rapidly in markets like the UK. So that seems to be a a pretty not very well reported success of the Chinese AI industry. I've never used the Do app myself before but if you have let me [22:34] know what you think about it and if you're from South America let us know why it is so popular over there. I'd be very interested why Bite Dance specifically found its success in South America especially. Now China has a fascinating and unique approach to many things including AI. And if you want to understand how the country work, you have to start with understanding how its politics work on a deeper level. And if you're interested in doing that, I think a new series called A Grand Theory of Xi Jinping is a fantastic place to start. [23:00] It digs deep into everything from Xi Jinping's rocky childhood to the state of the country that he inherited when he took over. And the series also lays out in great detail the vision that he has for the country that he's building right now. It's one of the smartest and most nuanced takes that I've seen on China's leader. and it's both entertaining enough that you'll actually enjoy watching it while it also manages to avoid being dumped down so you don't actually lose any of the necessary details. A grand theory of Xi Jinping is [23:24] made by Polyatter and is now available on Nebula along with many other fantastic Nebula exclusive series. These include The Anatomy of Everything where Real Engineering makes fantastic visualizations into everyday electronics such as phone chargers or the original iPod to show how they actually work and to see what challenges their designers and engineers had to overcome. But of course, I upload a ton of exclusive content to the platform as well. These include loads of dedicated videos and also a Q&A segment at the end of every [23:50] Friday checkout episode, which has become so popular that now a few thousand people watch it every week. Nebula has no ads and instead it is completely funded by subscriptions from viewers like you. And this means two things. First, it means that we get to create more and higher quality videos for you. Not for some advertiser, not for the algorithm, but you, the people who actually support us. And second, it also means an ad-free and distraction-free experience for you. Nebula is also selective about only [24:16] having highquality videos on the platform. So we explicitly don't support publishing AI slop either. And with my link go.nebula.tv/techalar, which is also in the description. The price of Nebula is now also down 50% to just $30 a year. Not a month, but a full year for $30, which is an extremely good price. Also, if you absolutely hate ongoing subscriptions, we even have a lifetime deal, so you can just pay once and never again. So, be sure to use my link in the description or the QR code to get that 50% discount. Happy watching [24:46] on Nebula, and I'll see you over there. Bye.