The AI Story Wall Street Doesn't Want to Hear

There's a simple way to talk about whether or not AI is a bubble. It's our old friend's supply and demand, supply, meaning in this case, basically data centers, and demand, meaning the money the companies and individuals spend on AI models that live in those data centers. The supply story is, I think, what most people talk about when they talk about the AI bubble. You start with the fact that the biggest tech companies in the world, Amazon, Microsoft, Google, Meta, are spending so much on AI chips and data centers that the cash from their own businesses no longer covers the bill. So now they started borrowing. And I do think that should make lots of people nervous since debt is that so-be, sudsy thing that has always allowed previous bubbles in history to expand before they pop, whether it's the railroads, the 19th century, or the.com boom. In the last few months, I've actually gotten less worried about the supply side story than you might expect for three reasons. First, the company's building AI are already rich. The railroad companies, the 1800s, were often small, thinly funded, and deep in debt. Google and Microsoft, by contrast, are like the most profitable businesses ever built. And I think creditors want to lend the most profitable companies ever built their money. Second, past bubbles left behind huge amounts of stuff that nobody used. Railroads went to towns that never grew. By the mid 1890s, about a quarter of American railroad mileage was in bankruptcy, often because it went to nowhere's villa Nebraska. In 2001, after the.com boom, less than 3% of the fiber optic cable laid across the US was actually in use. That means 95% of the rest of it sat dark. The AI build out looks nothing like that, at least so far. Practically every AI chip in the country is running. Companies are paying premium prices to rent them, and almost everybody agrees that we need more chips and more data centers to meet all of the incoming demand. Third, and this part is really, really important. The supply side is mostly visible to us. We know who buys the chips. We know who sells the chips. We know roughly what they cost. We know who's building the data centers. We know who's building the power that feeds those data centers. Most of the companies on the supply side are public, and they report their numbers every few months. The demand side, by contrast, is a mess of uncertainty. As strange as it sounds, there's actually no official measure of how much money the AI models actually make. The two most important AI companies, open AI and Anthropic, are private, at least for now. They publish quarterly results, so the public relies on leaks. A few weeks ago, one outlet reported that open AI was bringing in money at a pace of about $70 billion a year, then a week later, the Financial Times reported that that number was actually closer to $50 billion. The stock market moved on both the $70 billion and $50 billion number, even though one of those was off reality by 30%. The biggest tech bet in American history currently rests on revenue figures that almost nobody can check. There's lots of people doing estimates, I really like Azim Azar and his folks at Exponential View, but these are essentially estimates. That's why I think demand is the most important question in the AI debate, spending. All those chips and data centers are only worth what customers will pay to use them. The math is simple, really. If businesses keep paying more for AI every year, the debt is manageable and the build-out is gonna look visionary. If they don't, then today's AI chips will be like the dark fiber of the 2020s. The stakes, it should be clear, reach well beyond Silicon Valley or some abstract question of bubble or no bubble. The 10 largest companies in the S&P 500 mostly ride the AI boom. They make up about 40% of the index. If you have 401K, you are in on the AI bet whether or not you want to be. These guests has one of the clearest views of the demand side of anyone. Aura Karazian is the lead economist at Ramp, a company that oversees corporate cards and expense software for tens of thousands of American businesses. That means Ramp can see month by month by month which companies pay for AI. How much they pay, what they buy. Over the past three months, Aura has published a series called Cracks in the AI thesis. AI revenue exploded in 2025 and 2026 but his data shows something more complicated happening underneath that explosion. About 1% of companies now account for most AI spending. New business customers are slowing down their spending and prices are falling fast. For the AI bet to pay off, spending on AI models will have to rise from about $200 billion this year to $400 billion the year after that to $800 billion the year after that. It is a very big deal. If economists with a bird's eye view of this industry say they're seeing signs that spending growth is slowing down. Today we talk about who's actually paying for AI. How much and what happens if customers stop showing up? I'm Derek Thompson. This is Plane English. This episode of Plane English is presented by Comcast Business. Did you know Comcast Business detected 79.3 billion cyber threats last year. That's why they're introducing security edge preferred. Its big business cybersecurity designed for small business. Get started with Comcast Business Internet for $50 a month for one year on a qualifying internet package and find out how to get one month of security edge preferred on them. Plus ask about a $500 prepaid card. Edit time offer, restriction supply requires 300 megabits per second internet security edge, security edge preferred and wireless connect call today. Our Akarazian welcome to the show. Thank you Derek. Thanks for having me. What is ramp and what is your job there? Well, I am the lead economist at ramp and specifically a run ramp economics lab. Ramp itself is an AI finance platform so you can think about the data set that I work with in terms of our products with a corporate card. We have bill pay and a stage in invoicing and for 85,000 plus businesses in the U.S. We cover more than $275 billion in business spend every year. And so you can imagine that data set being for the businesses on our platform, we see the full universe of business spend where they're pulling back, where they're investing more. We cover more than $10 billion a year in AI spend and that's been especially helpful in understanding how is AI proliferating through business spend? Where are they getting ROI if they're even able to measure that at the business level? And the kinds of metrics that we'd want to use to measure AI's actual impact on the economy all start with seeing how businesses use the technology in the first place. Right, so you've got this side, this bird's eye view on the state of AI spending and that's why I wanted to talk to you because I think that there's a lot of public data that we know about the so-called supply side, the hyperscalers. But there's not a lot of public data on what is open AI earning, what is anthropic earning, what are all of the AI models earning? But it's companies like yours that can help us understand because you have this perspective into how these thousands and thousands of companies are using AI. So most of this conversation is going to focus on trying to see the murky present and the much murkier future of AI spending. But I think we want to set the table here by talking about the last 12 months of revenue growth paced by the Frontier Labs. How would you briefly characterize what we've seen over the last year? Well, we can even go back a little bit before 12 months ago. Let's think about when ChatGPT came out. It was primarily a consumer facing product. No one had any sense that enterprise spend would become the significant driver of AI's growth in the US economy. There was no proven business case. Then over the following few years, you had open eyes start to make some inroads into enterprise spend. The main player in the market where if you were business buying for the AI for the first time, you're going to buy it from the consumer default. That was open AI. Then around 12 to 24 months ago, you see anthropic slowly and then very quickly start to rise in terms of both market share. And then via the introduction of cloud code, this highly commercially advanced productivity enhancing coding agent where if you're software engineer, it's basically taken over your job or at least augmented it. Start to now rise as this massive revenue driver for the American AI model companies. That then started to precipitate more and more investment into the actual infrastructure to support the spend. Then now or the last couple months, we're seeing some threats to that revenue model where we're seeing increased competition amongst different American AI companies. The SEX AI is now emerging and video is starting to put out its own models. And you see increased competition from overseas, which are all driving down margins for the American AI companies and posing a threat to their growth models. Let's talk about who's actually paying for it. for AI right now. I mean, again, there is no like official register of who is actually paying for AI models. So it's companies like yours that we rely on to understand this. Your data shows something really extraordinary, which is that 1% of companies accounts for 80% of open AI and anthropic spending or revenue. Is that right? Who are these companies? And what are they doing that causes them to consume so much AI? It's one of the most defining aspects of AI's proliferation through our economy is that it has grown so quickly as an industry, but it's also extremely concentrated. Now, there is both a product issue there and maybe even a policy issue there. There are probably many companies out there that could take advantage of AI and could grow more quickly with it and work more productively with it that today do not because they are barely using it. That's all very true. If you look at the median company on our platform, they spend of businesses that spend on AI. They spend about $12 to $15 per employee per month on AI. The top 10% company spends about $800 per employee per month and then the top 1% company spends upwards of $10,000 per employee per month on AI. But what I would caution people about is it's not even really a concentration issue because it's normal that the top 1% of spenders will drive a disproportionate amount of revenue in a specific industry. Yeah, it's 80% for anthropic and open AI comes from 1% of businesses. But if you'll get digital advertising, about 60 to 70% of it is coming from the top percentile firms. So that's not what's unusual. What is unusual about AI, though, is that it's also highly correlated within that top 1%. All of the companies in that top 1% are extremely technically advanced. They are concentrated in the information sector, software and media, or finance or technical services. That's all fine because they're all the early adopters. But you can imagine that is a potential crack in the AI thesis that in the event of a market correction when these companies all start to turn down their spend all at the same time. What does that mean for open AI and anthropic or the other American model companies? That's the big threat there. So you said the median company on your platform is spending about $12 per employee per month. And the top 1% is spending closer to $10,000 per month. I think the next question is, are those numbers going up, right? Is the median company spending more and more per month and is the top 1% of companies on your platform spending more and more per month? Because one of those two things, hopefully both has to be true for people in the AI industry to feel really confident that the revenue is going to continue to grow. Well, over the last couple of months, AI spent a start to slow down amongst that top 1% group. Not because they don't see value in AI. It looks like they actually do. And that if you look at how much they use AI, they're using it more and more every month. They're using AI for multiple models and multiple model companies. They're using AI software solutions produced by companies other than open-iron and anthropic. So they're spreading out their spend in a way that looks more and more advanced and is doubling down on AI's technology. But they're able to spend less and less on it every month because of a few factors. One, it's a highly competitive market where the price of AI continues to come down, open-iron and anthropic themselves, including sticker price cuts, so that you can pay less for AI, trying to battle for market share, offering discounts for long-term contracts? And then alongside that, it's getting easier for businesses to control their AI spend. A couple months ago, we were in kind of a Wild West situation where businesses were encouraging their employees to use as much AI as possible. And we saw Token spend rise as a result. Nowadays, it's a lot easier to control that AI spend. You can impose controls on model types, make sure that your employees are using a highly efficient cost-efficient model for very simple tasks and devote only the most performant models to the most complex tasks, having a system that allows you to route those models effectively and blocking certain models for certain tasks. So businesses are also becoming better at optimizing their AI spend as the market advances. So that is, again, one of the threats to the AI thesis is ironically, as businesses become more advanced, they are spreading their AI spend across multiple different providers, which is actually not how typical software spend proliferates, typically you may lock in with one contract. And that is also therefore allowing businesses to switch their AI spend really quickly, which is making it even more competitive, which is then driving further price cuts from open-iron anthropic. The cost of AI spend continues to fall and fall and fall, despite the fact we see rising usage from businesses. Yeah, there's a couple of pieces here that I want to make sure I understand really clearly, because this seems really, really important, that usage of AI is growing, but people are finding ways to essentially be discount shoppers among AI models. They're saying, I don't need the most expensive models to do every single part of my job. I can use the best anthropic model over here, maybe an open AI model over here, maybe an open-weight model over here. And as a result, you're seeing spending moderate, which might be difficult for the overall AI demand thesis. It seems like there's two ways that revenue could significantly grow for spending on AI models the next year. One is that software companies, the 1%, the early adopters, that they're spending more and more every month. The other possibility is that more companies outside of software are finding reasons to increase their spending from $12 to $120 to $1,200. Is there any evidence of the ladder of high spending on AI broadening throughout the economy, broadening throughout businesses? - Well, there's places where it can grow, and there's places where I'm a little bit bearish about further growth. One is that there's still a lot of room to grow amongst the sort of more old-school, large enterprise firms in the US. These are firms that tend to have a lot of ability to spend, but they're much less likely to adopt a new and nation technology right from the get-go. You can assume that AI has a lot of room to grow in certain enterprise tasks, offering AI customer service chatbots, and yet companies wouldn't want to roll that out unless they're absolutely sure that it's going to work and it's not going to be a bad experience for thousands to millions of customers if they roll it out all at once. So that's one of the main headwinds we hear from businesses as far as why they're less likely to implement AI in the first place. But then on the other side, there is also a point at which, you know, most of the businesses that exist in our economy may not actually end up benefiting from that much AI use at the end of the day. That's a product issue that can change, right? But let's think about coding agents. Coding agents absolutely deliver ROI. They make you a faster software engineer. They make it easier for companies to push software and fix changes and ultimately help the end user get a better result from software. And yet, coding agents really only help about 3% of Americans who are software engineers. The number of Americans who would ever use or be enhanced by this technology is relatively low. And if you look at the top 10 occupations in the US, it's a list that includes healthcare workers and cashiers and food service workers. Very few of them would be augmented by AI and the businesses they work at similarly. So the challenge for the AI companies, therefore, is will there be another coding agent revolution? Will there be a coding agent for finance tasks? Or will there be a coding agent for manufacturing automation? We're already seeing that in some ways. Self-driving cars, for example, will absolutely impact productivity, will absolutely impact the job market. But we're not necessarily going to see that across sectors of the economy that are less technical and are just less likely to take advantage of the AI tools that exist today. One of the takes I've heard a lot about people who believe that AI is going to prove to be a bubble is that the companies that are paying thousands of dollars per employee on AI models are not yet seeing any kind of clear ROI, return on investment. What are you seeing in terms of the ROI picture? - Well, I would say that we are, look, if you put tech people on one side and you put economics people on the other side, you will get two very disreviews on how AI is going to impact the economy and the ROI of the technology. You'll have tech people who are projecting massive economic growth and productivity growth and are saying that's already occurring at many businesses. Either that or we all die, right? Like dramatic impacts from AI adoption. And then you'll have econ people where the general consensus is AI will have a pretty modest impact on the economy and we're also not seeing it just yet. 'Cause you don't really see it in that many productivity statistics. I'll put myself a little bit closer to the econ side in that I believe the effects of AI in our economy will be meaningful, but really. relatively modest, and yet I do think they're happening today. I think they're happening in large ways today, but I think they're relatively concentrated in specific sectors and job categories in our economy, such that we're not likely to observe them in aggregate statistics just yet. But if you look at certain businesses and certain parts of our economy, you do see that growth. We wrote a paper a couple months ago that found that firms that adopt AI grow 10% in head account over the two years following adoption, so they're actually hiring more. But it's not just firms that use AI. The impact was entirely concentrated amongst businesses that use it with high intensity, so they're spending at least $50 per employee per month on AI. They're using the coding agents. They're using multiple models. They're using them for a sustained period. They tend to be highly technical already. Things that use AI with relatively low intensity, otherwise the average AI spender saw no difference from the control. So there are already signs that at certain kinds of businesses, highly concentrated pockets of the economy, we are seeing growth and we are seeing ROI. But we're not likely not to see them in aggregate statistics, because most businesses just don't use AI all that much, and they're not in the sectors that are that enhanced by the technology at least today. I think somebody thinking about investing in the anthropic IPO is going to listen to the arguments that you're making, and they're going to say this is a business where the top one percent of early adopters aren't spending more and more and more on a month-to-month basis because once you hit $10,000 per employee, you are highly incentivized to say there's gotta be some way for me to moderate cost increases. And then you're also saying, it seems to me, that spending at that level isn't broadening as quickly as the frontier labs wanted to. When you put it together, I mean, isn't this a pretty strong bear case for anthropic and open AI? Well, there are many ways to look at it, because at the same time, I hear those concerns, and by the way, I have no dog in the race, so who knows how the IPOs are going to go. But this is an extremely competitive market, and it's a market that's changed a lot, even over the last two years that I've been studying it. We first, coding agents weren't in existence in any meaningful way when we started studying AI spend. And no one was at the time assuming that we would get any businesses spend tens of thousands of dollars per employee per month on AI, because there was no commercial version of AI that was able to drive those levels of spend. Now yes, we are now seeing the slow down in AI spend for that top 1% of businesses, but I wouldn't discount the fact that there's further growth here, one further growth from more old-school businesses that have yet to adopt AI in that comprehensive way. Yeah, maybe they'll hit some top, but they're still room to grow. And then, frankly, new product offerings as well, they're increasingly coming out from the AI companies. It's really hard for anyone to be overly confident about this side or that side or being bearish or being bullish when so much has changed just over the last two years as far as how advanced this market is. Consumers spend now being a possible driver of future growth for the AI model companies. We weren't projecting that a year ago, and now with the introduction of new products like Grockbott and Muse and Meta's entry into this, it's not unreasonable to think that that could be a significant driver for the companies. Obviously, the AI labs aren't the only companies in this ecosystem. There are people who are selling the chips, there's people who are building the data centers, there's people who are supplying power to the data centers, and I know that you've been in conversation with some people who are more focused on the rest of the AI ecosystem. They essentially say, "If everything that you say is true, if AI is getting cheaper, then people might just use it more." This is an idea in economics called Jevin's Paradox. People might just use much more AI, and that might be bearish for open AI and anthropic, but it's bullish for Nvidia, it's bullish for basically everybody else in the AI industry. How do you feel about this idea that the very phenomenon that you're talking about, rising usage with flat lining or falling token prices, are going to be rough for the AI labs, but good for pretty much everybody else in the AI industry? That's probably where I land, in many ways, in that we are seeing increasingly a market that is so competitive, that is driving down profits for, driving down potential profits for the actual model developers, and yet the need to support more AI spend that is distributed across multiple different model providers and a greater number of customers will then drive more investment and compute. All of that therefore has implications on who is the winner and then who are the relative losers in terms of the AI stack. Really though, the winner here is the ultimate end user of AI, consumers or businesses are routinely seeing price cuts and better and better models for them to use. As they advance along their own adoption curve, they will then drive the productivity gains associated with it. It's a really great time to be a user of AI, if you are interested in better and better models every month at a lower and lower cost. The product is getting better, but what does that look like for the actual developers? I think they are still room to grow, that's my sort of most clear understanding is that this is a market that is changing so much that I wouldn't discount the fact that there are new products coming on to the market that are going to drive more value, but that's why my latest series of posts was called cracks in the AI thesis, that if your impression is based on the product we have today that we are going to see continuously rising spend in AI driven by the top 1%, that is the latest crack that is starting to form, which is that the top 1% is starting to slow its pace of spend, and the products are getting pretty good such that they allow businesses to outright cut their spend and optimize it with cheaper and cheaper models every month. The folks who run Anthropic and OpenAI are really, really smart, right? They are surely looking at the same numbers you are, if in fact your numbers are correct. They are also seeing that sometimes they are debuting models, advanced models like Fable that are seeing low pickup. They are also seeing that it's getting harder and harder to charge more and more for the next advanced model because of all of this competition. How might they react? What would be a rational response by Anthropic or OpenAI to strengthen their business in the face of all of this competition? The frontier models are cutting their prices. There's more competition from Chinese open-weight models, from domestic open-weight models. What should they do? What do you think they will do? Look, I think that they have way more competitive threats to deal with today than they did a year or two ago. So far, their responses have been introducing better and better models, investing more and more into computes such that they can train better and better models and then providing those models at relative price cuts in a way that benefits businesses. I also wouldn't overrate the extent to which they have all the information on how to invest and how far spend is going to go and where it's going to land. Anthropic, for example, only sees how much spend is going to anthropic. OpenAI only sees how much spend is going to open AI. They can use data sets like what we provide at ramp to track general market behavior and how much spend is going across the tech stack. But we're all kind of figuring this out for the first time. You saw this even a couple months ago, when Fable came out, this was the model that was described as so dangerous that it had to be restricted by the Department of Defense. And that was not an unreasonable decision by the Department of Defense to block the usage of the model. Given everything anthropic itself was saying, and they imposed some checks on those models. So one of the main responses was, okay, you can release the model, but you need to collect all the data from the model and do checks on everything that's being prompted in the model so that we know that no one's misusing it. That ended up being kind of a non-starter for most businesses, which have obligations to their customers about data retention and making sure that their software vendors don't store any data. And so despite this model being so powerful it had to be restricted, businesses didn't really end up using Fable. I think it topped out at like 5% to 10% of business spend at its peak in that period. And why would they? Because it actually was at the best model on the market at the time, yes. But there was also a really good other model available at the time, which was GPT Seoul. Open AI's competitor that had just come out. And so for relatively low trade-off in performance, and actually a benefit in that Seoul was is actually a little bit cheaper. Businesses were able to still use a highly performant model, fulfill all of their obligations to customers, and continue to drive gains from AI spend, such that we kind of figured out that, oh, there is a limit in terms of whether or not businesses are willing to pay or use the frontier model, that businesses have other drivers to their business, to their actual AI spend allocations, and they're able to be kind of picky, and they're therefore able to be in a sort of the strong, negotiating position in the market where they have the sort of competitive advantage, where open and enthropic have to cater to the business and to the customer and drive further for the price cuts. So, I think it's a very challenging environment for a model maker in that you need to, and one that ends up actually driving outside gains to the users of the models. Eventually, somebody outside of the AI capital loop has to obtain enough value from AI to pay the bill, right? Revenue right now is around $200 billion, $220 billion a year. It probably has to roughly double by next year and then roughly double from that in 2027, but several things are standing in the way. One, coordinate data, a tiny group of buyers accounts for most of AI model spend. Two, their spending growth is leveling out. Three, the number of businesses paying for AI also seems to be growing slowly, or growing more slowly. And four, prices are falling because of competition. Is all of that right? Is there something that I miss there? Is that the nut of your thesis? That's pretty much the case. What's the strongest argument against it? Well, there's more growth that we just haven't seen yet. Either more growth that will come from firms that are using AI in low intensity ways today, but will end up increasing spend in some point, landing somewhere around that top 1% firm. So you'll get more firms, the median end up somewhere around the top 1%. That's going to happen. It's just the question is to what extent is that going to happen? Typical restaurant or doctor's office is not going to spend tens of thousands of dollars a month per employee per month on AI. But there will be some businesses in their Fortune 500 businesses that end up becoming more technical advanced and using the software. So that is valid. And then the other growth trajectory is what else can the models do? The models are becoming so increasingly advanced and highly technical advanced in solving math and physics problems every month that maybe there will be another coding agent revolution. And there will be some other form in which AI augments a new sector. I mean, I think one of the interesting and underrated parts about AI's impact today is that it's not unlikely that in a couple years, AI will start to accelerate the pace of manufacturing automation that we've already seen over the last century. AI's sort of implementation as far as vision models. It's a ability to be used in manufacturing. Is that going to be a paradigm shift? It's likely going to be a continuation of what we've already seen over the last 100 years and one that will accelerate it and therefore drive greater gains. But we're all unsure about where it's going to land. But those are some of the great growth trajectories for the market today. It's funny. One thing that reminds me of this conversation is I'm sure you remember in 2020 and early 2021, a lot of investors were looking at companies like Peloton or Zoom or Netflix. And they were saying, oh my God, look at the revenue increase in the last 12 months. If you continue this rate of growth, like these companies are going to absolutely eat all of fitness and all of entertainment. But rather than sort of start a new normal, what ended up happening is that the COVID and post COVID period sort of allowed a lot of companies to like eat several years worth of revenue growth very, very quickly, which meant that later earnings reports were disappointing because those quarters, those years, had already essentially been concentrated in 2020 and 2021. And in a similar way, what I'm taking to be your thesis is saying is that a lot of people in March and April of this year were looking at the revenue growth of anthropic and open AI with the rise of coding agents and saying, oh my God, if this continues at this rate for the next two, three, four years, you're going to see a trillion dollars of AI revenue coming into this system. And it's not that the revenue was fake. It's just that these models saturated their early demand. They saturated the relevant market such that you saw this explosion followed by a little bit of a peak, a little bit of a flattening of the S curve, I guess you could say. And so it's interesting that we just saw like four or five years ago how strange periods can create acceleration along an S curve followed by a flattening. And you're suggesting that we might be in a similar period for now until the AI labs come out with some new technology that creates a new adoption S curve because that's what it's going to take to continue to grow this revenue by 2x year after year. That's the thing. I think that this take is a little bit unsatisfying to most people because it is suggesting that we should invest more and more into a technology, which is providing some return, but that the further investment is based on a promise of future returns that we're not even sure of yet, because we're not sure what the technology may produce. And I think that that is extremely unsatisfying to most people who process information, especially if you're someone who is used to understanding investments or economic changes in terms of one big number. We probably will never get one true number or one big number that says, hey, this is the value of AI in our economy and this is how much we can spend on it. In the same way that we never really got a number for the productivity impact of computers. I mean, look at the last 50 years of economic growth in our economy, and if you are going to look at productivity growth, you could probably say something like computers boost our productivity by like one to two percent, pretty reasonably per year. But that's a pretty marginal growth number if you express it to most people. It's actually a big deal, but it's marginal at the end of the day. And the truth is that it's really hard to express these rapid technological changes in the form of one big number. I'll say from my part, though, we try to produce our research by looking at several different indicators and thinking about what that says in terms of the assumptions that market participants are making. So when we hear that market participants expect or investors or whoever expect that AI spend is going to rise more and more and more and more, I think it would be a mistake to just look at AI spend at the average. First of all, because it's so concentrated, so you really do want to look at it broken out, the top 1%, the top 10, the median. That's then where you see the really interesting results. And then you could drive some other theses around why is the top 1% slowing its spend and whether or not that's likely to hold. And if the drivers there are because of price that is making it increasingly competitive and driving ultimately more value to the users or because they're not using AI at all and they're not seeing the value of it. And that's what's really difficult to sort of express about the economy today is that we actually are seeing some gains because of AI, but they are highly concentrated in specific pockets and it is not yet clear whether or not those gains will proliferate across the rest of the economy. If that there is the policy and product challenge, a policy challenge on the part of government making sure that more firms or more consumers can actually gain from AI's adoption and then a product responsibility on behalf of the AI model companies developing something that actually works for more than 3% of Americans. - Yeah, no, I think it's a really compelling piece and it's again, it's not particularly anti-AI. You're just trying to see reality as clearly as reality as trying to describe itself. Like for this bet to pay off, it seems to me like the AI companies need the typical American or the typical employee to be spending thousands of dollars per person on this technology. And right now what I'm hearing from you is that this technology is worth thousands of dollars per employee at software companies, at tech companies. It's not worth thousands of dollars per employee for nurses. It's not yet worth thousand dollars per employee for waiters and waitresses cashiers. There's a lot of individuals that aren't seeing and aren't going to pay this amount of money for this technology. And that poses a risk, a challenge to AI because it has to continue to prove its value to employees beyond tech and software. - Or the ultimate winners end up being the median company that could drive gains from AI but doesn't need to spend thousands of dollars on the technology per month. That really is. is one of the ways in which AI could grow our economy more quickly, that more businesses could benefit from it without having to shell out thousands of dollars. And that is, frankly, the case today. AI getting cheaper does allow more and more firms to take advantage of it. I mean, one of the early concerns that we, one of the frequent concerns that we hear from businesses who haven't adopted it yet, is that they don't want to adopt it one because of concerns around its performance. They're not sure it will do well at scale or if it's reliable, but once you get past that, it ends up being a question of cost. It's really hard to control AI spend, especially if you roll it out to your employees and they're using the APIs or they're spinning out tokens. It's actually gotten a lot easier, both because of the price coming down and because of new products to make it easier to control that spend and setting better defaults on it. And the ultimate winner there, it may not be the model companies, but it doesn't need to be for our economy to gain from it. It can be the ultimate end user. - Yeah, we'll keep following the story because obviously it's one of the most important stories in the world, whether or not this bad actually pays off. Aura Karazian, thank you so much. - Thank you. - So today's interview is jam packed with information and data and I want to make sure that the bottom line here is as clear to listeners as it is to me or maybe as clear to me as it should be to listeners. I think if you stop the clock right now, AI is not a bubble, period. It's just not. Data centers cannot get built fast enough to service all the demand for AI use. And the question therefore is really whether AI spending can continue to grow in the next few years at anything like the rate it's grown in the 12 months after cloud code was launched. And today's interview, I think casts some major doubts on that thesis. And I think it's worth breaking down Aura's thesis into two parts. Number one is that AI use is continuing to rise even though spending is stagnating, which suggests that many AI users are shopping for discounts on their AI models rather than just upgrading to the most advanced models available. That matters because it's really important that the AI frontier labs are able to replenish this flywheel of we spend all this money training, the newest best AI and then people spend a lot of money to pay back all of that investment. And then we build the next model that's going to break all sorts of land speed records for capabilities. And then we keep that flywheel going. It matters that a lot of customers now, a lot of companies aren't remotely token maxing anymore. They're bargain hunting. People like shopping for their favorite shirt by opening seven different tabs and seeing which company is selling the green shirt at the price they want to pay for. That kind of discount shopping is not good, I think, for the businesses that the frontier labs have built. The second piece of our thesis is that AI per capita spending is extremely concentrated among software workers. And that means that you either really, really want software workers to be increasing their spending significantly or you want that level of AI spend to spread through all sorts of other industries. If neither of those things is happening, then the AI frontier labs are probably going to have a lot of problem seeing revenue growth that justifies their current valuations. So when you put both of those two things together, when you put together the fact that customers are discount shopping and the fact that the AI labs are seeing spending flatline from their power users, I think that that should give some people some pause about the next year of the AI bubble thesis. Now, as Ethan Mollick explained in our Tuesday show, AI is growing in capabilities and changing so rapidly that I don't think anybody should be certain about the future. And I think Aura is right to hedges analysis by saying it's hard to predict the future of this technology. But look, I think sometimes the best way to predict the future is to just describe this head of the present. And the state of the present, according to ramp, is that the year of hyper growth for the AI labs might be winding down, which just very simply means that 2027 could look very different from 2026. Thanks for listening. [MUSIC PLAYING]