Moody's Talks - Inside Economics - The AI Series: Adoption Ramps Up

Episode Date: September 8, 2026

In this episode of our AI series, Ara Kharazian, lead economist at fintech company Ramp, joins the Inside Economics team to discuss the state of AI adoption. Drawing on Ramp’s real-time business spending data, Ara offers a unique view into which firms and industries are moving fastest, which are lagging, which LLMs are gaining traction, how token pricing is evolving, and what data center constraints and government regulation could mean for the AI boom.Guest: Ara Kharazian, Lead Economist, RampView our latest articles and research on AI- https://www.economy.com/ai-insight-hub Questions or Comments, please email us at [e-mail]. We would love to hear from you. To stay informed and follow the insights of Moody's Analytics economists, visit Economic View. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Welcome to Inside Economics. I'm Mark Zandi, the chief economist of Moody's Analytics, and I'm joined by my two trusty co-host, Chris DeRides and Marissa Dina Talley. Hi, guys. Hey, Mark. Good afternoon. Yeah, good afternoon. I understand you're off to France here in an hour or two. Yeah, that's right. That's exciting. My last podcast for a couple weeks, I'll miss you guys. Oh, really?

You're going to muddle through without me. Chris, do you notice how much, she's like traveling the world all the time. And I, what's, what's that all about? What's wrong with us? It's what you're asking. Is that what it is? What's wrong with you? Yeah.

Yes. Yes. Well, you go off to Italy for a month or so, don't you? In the summertime? A little less than that. A little less than a month. Yeah. All right.

But I do bring my microphone with me, so. Oh, you do? Yes, you do. Yeah. Dedicated. It's a little different situation, Chris. How so? Because he's going to his mother-in-law's house and sitting in his mother-in-law's house and

Working. All right. I'm going to traveling all over the French countryside. I can't tow the pod mic with me. I'm sorry. So all of the places you're going in France, where are you most excited to go? I'm most excited about Paris just because I've never been. But I'm really looking forward to the countryside and the chateaus and the wineries

and I'll be in Bordeaux. And I'm excited about all of it. But I mean, Paris obviously is, I think, just very. iconic and I've never been. And you were saying the weather is now more reasonable. It looks more reasonable. Yes. Yeah. Good. Good. Well, we'll miss you.

Thank you. Yeah, we'll miss you. We've gotten to know each other quite well in the last week or so because I don't know how many podcasts we've recorded, but. Quite a few, yeah. Quite a few. But we've got a great guest, Ara Karazian from Ramp. Ara, did I, how badly did I bet your name? Did I really get it? Oh, wow. That's great.

Well, fantastic. So good to have you on board. Thanks for coming on. Thank you. Thank you for having me. I'm excited. Yeah. And we're, you know, I don't know if anyone told you, but we're having a series of podcasts around artificial intelligence AI. We had, our first one was with David Autor. Do you know, David? He's the MIT economist. Yeah, a really good guy. And it was quite fascinating. And then we're releasing a podcast this afternoon with Daryl Spence from Capital Group. And we had Michael Gugas from Construction Connect talking about data centers.

And now you. So it's great to have you on. Would you mind, I don't know you very well, and I don't know Ramp at all. And I know you're doing a lot of great work. I'm starting to follow your AI index very carefully. And you're in the, deeply in the conversation around AI. Could you just give us a brief sense of you? in Ramp?

I have this great privilege of working with a great data set, mostly because of Ramp's product. But I know my name is Rara Krasian. I am the lead economist at Ramp, and I run our economics lab, sort of our equivalent of a think tank. But we produce original research using business spend data from Ramp about new technology frontiers, AI spend especially most recently, and how businesses are investing in it, where they're pulling back, and where they're double-term. down. Now, this work works because if you're trying to understand how AI is developing in the market,

there's not really a lot to go off. The companies that are developing it are all privately held. So you can't really get a great sense of the size of the market. And then if you want to understand how businesses themselves are using it, there's not much in the way of productivity statistics, at least in the international statistics. So you kind of want to start with, well, who's using AI in the first place. For a long time, all we had to work with was surveys, but those surveys are often detached from the reality at most businesses. You come up to the cusp of trying to draw lines about, what does AI adoption really mean? Can we measure the intensity of AI adoption with a survey? And so the privilege of my data set is that it's actual business spend data. So Ramp is a finance

platform for businesses. So you can think about a transaction data set of corporate cards and invoicing and bill pay and token spend management. We pull all those sources together to look at the universe of business spend and understand not only how are firms buying AI and who they're buying it from, but how that's changing over time. And then what's different between firms that are using AI and those that aren't? And then from there, you can start to understand, how are they hiring differently? How are they opting for open source models or not? So it's a really great data. set that allows you to track this fast-changing market in a way that public data sets so far have not been able to do. Are RAMP's clients, I guess, is that the right word, are customers?

Yeah. They're enterprises, their businesses. Are they in a certain set of industries? What are the demographics? What are the characteristics of the Ramp customers? Yeah. 70,000 businesses, $200 billion of annual spend, primarily in the United States is where my research focuses, but that'll change over time. It has a pretty good distribution, I would say, actually, in terms of sector. You know, we have everything from tech to manufacturing to construction, and most of our work is able to therefore break down our estimates of adoption by the sector or size of the company. I do, however, think that we skew toward, regardless of the sector that you're

looking at, relatively tech-forward businesses. Ramp itself is already a sort of AI finance platform, form itself. It's opted for by tech forward businesses versus like a chaser and amax. And that is actually a strength of our data set. If you want to understand, I think, where the market's going to go, you want to look at these early adopters of new technologies. For that reason, our data set was the first two track that Anthropic was starting to overtake Open AI in business adoption. It's also a highly granular data set so we can report these sort of differences between businesses as they come in. So when you talk about adoption,

would it be fair of AI by enterprises, by businesses, would it be fair to say that the, and we're going to get into your data and get a sense of what the data is saying, but would it be fair to say that it kind of overstates the case for the broader economy, or do you make a correction for that? So that's a harder question than you would think,

because there are, The U.S. government estimate lands in like the low 20s, as far as the share of firms that use AI. Our estimate is about 53%, just crossed 50% a couple months ago. But there are other estimates that use survey data of U.S. businesses that land somewhere around 70 to 80%. In many ways, I think our data set is actually likely to understate adoption because we're limited to looking at paid adoption on a business's sort of invoicing and or corporate card platform. So we're not capturing free adoption. You know, Gemini usage happening through Google Workspace. We're not capturing the adoption that is definitely happening on employees' personal accounts.

And that's really the difficulty of doing this research and some of the questions that we've had to deal with, both at Ramp and then other economists externally, is where do you draw the line of adoption? For a long time, that's all that mattered is our business is using it. And the metrics kind of coalesced around that. Turns out answering that question itself is very difficult. because you have to define what it means to use a new nation technology, and we're still trying to figure out how it's applied in enterprise. Increasingly, I think the research is shifting over to a better question, which is how are they using it and to what intensity,

and then you can break down the actual product lines that are adopted. Got it, got it. So just as a fictitious example, if Moody's or a ramp client, and I don't know if they are or we are or not. and we spend, you know, making it up, half a million dollars on our LLM, or we're using Claude,

that would show up in your data. That would be the kind of thing that would show up in your data. Yes. And you would know it's anthropic, it's not open AI, you would know all those things, right? Because you're paying the bill.

You're paying to build anthropic. Well, think about anything that shows up on the receipt because we see that too. To be IRS compliant, businesses have to upload those itemized. receipts. So that's what allows us to see, are they just buying Claude or are they making agent calls through Claude code? Is it a subscription? Is it tokens? Increasingly, we're able to see whether or not, you know, the other software that you're buying outside of those main model

companies is also inclusive of AI features that are charged for as line items. So it's a market that's developing very quickly. And you see that in the story of bills. Literally receipts and invoices are evolving the same way our work is involving. Got it, got it. Hey, Marissa, before we move on to the meat of the matter, the results, the OpenA, the, excuse me, the AI index. And I should say, Arra, Marissa comes from the BLS, so she's, you know, very persnickety when it comes to data sources. Marissa, any thing you want to ask Aura about the data before we move on? What about AI that's built into existing software platforms?

Like if people are using copilot but it's in their, you know, Microsoft, do you have any insight into that or does it have to be a specific AI product they're purchasing separately? That specific example is the type of adoption that I think we underestimate. I see. Okay. Because you buy, and that's a sort of Gemini example, right? That Gemini is free through Google Workspace copilot as well through Microsoft's equivalent. Now, the most advanced businesses that are using AI are typically not primarily relying on Copilot and Gemini.

So for our research, we're able to focus on the segment of businesses that's spending on more advanced tools. But that's a great example of why it's hard to track these trends over time. Is that for many people who are using AI today, they're using versions of AI that are very different than the frontier that, you know, We're increasingly getting used to at companies that are paying for it. Hey, Chris, anything for ARA? What about open source or open weight models? I hear there's an expansion into that area. So in theory, that's free.

That's not charged for, correct? Or maybe do you have any insight into that usage? That's a good point. So open source models or open weight models, theoretically businesses should be able to create their own infrastructure on site that allows them to call this internally hosted model such that no payments are going in and out of the company. It's all hosted on a company's own computers. The market has not adopted that path very popularly, mostly because there are fairly cheap alternatives to using open source

models that are hosted by another company. So this is like using a router product. It's just a cheaper way to access not just one open source model, but a lot of open source models. And you still get most of the price savings, but you're making tool calls and small payments to another vendor. So you can proxy, and this is what we do, you can essentially track open source adoption by tracking payments to those kinds of companies, those routers. We do that too. Just one last question before we get on to the data, because we've worked with third-party

data sources as well over the years. And one of the problems and tried to tease out, you know, information from that data. And one of the issues that would come up is the growth rate of the company itself, right? I mean, if you're growing fast, then some of what you're observing is simply that you're getting a larger share of the marketplace. Or conversely, if your business isn't performing well, just the opposite. Is that an issue here at all or something to be wary of when looking at the data? Yeah. Well, I think my frustration with private data source is exactly what you're describing is that, you know, this is a call to my counterparts of their companies, is that there is oftentimes pressure to use these private sector data sets in ways that they're just not well suited to do.

I mean, I think that's the best life advice is the same as the best economics data set advice is that, you know, don't try to be something that you're not. And so for all of our research, we do a combination of things. One, it's picking a metric that accounts for exactly what you're describing. So if we're comparing two different groups of businesses, we compare like for like businesses to try to track those changes. I see. Right. If we are unable to do that and sometimes we are able to that, it's actually just helpful to show that there are large differences between the fast growing businesses and the slower growing businesses, we want to highlight that change. And we'll be transparent about our methodology,

and we'll show here's what's different about these firms versus others. And then we publish all of our data online so that researchers can download them segment by segment. And if they want to make their own weights, you know, to make tech a little heavier, tech a little bit lighter in this sample, let's make it more construction manufacturing focused. You can do that too.

So we've seen actually a couple of researchers, you know, produce similar results. Got it, got it. Well, let's move on to the AI index. And I know you do a lot of different things, but this feels like kind of your flagship set of data. You want to describe the index and what broadly it's saying at this point about adoption and intensity of use?

AI index does a couple of things, but it starts with our spend data, and it tracks first the share of businesses that are using AI at all, and then the respective share of businesses that are using OpenEye versus Anthropic versus the other model providers. I'll start with that metric because it's the broadest image of the market, which is where

you see, you know, pretty mediocre adoption rates of AI from most of 23 to 2024 and then this rapid rise in adoption in 2025 into the beginning of 26, especially driven by adoption of the more advanced tools like coding agents. That's where you saw the share of firms using AI cross about 50% in our sample. It's actually simultaneously around the same. time that you start to see this intense competition between Open Eye versus Anthropic. So remember for most of the last five years, well, I mean less than five years, right? This is much younger market than I even realized. Open AI was kind of this default. You know, it was it was the AI that most of us were familiar with as far as a vendor that you could buy from. And despite AI being this thing

that businesses could buy, it was also something that consumers were learning about for the first time. And so if you're a business buying AI for the first time, you're probably going to buy Open AI because that's what you're familiar with as a consumer. That started to change at the end of 2025 and the beginning of 26, where Anthropics start to see this really meteoric rise, ultimately crossing Open AI in business adoption. And today, having about 53 percent, not 53, about 43 percent of businesses are using Anthropic today. So you start to see in this metric, like a story of a market that is becoming increasingly complex and expanding and starting to behave differently from how we'd expect software to behave. Because remember, if you're typically buying software, you are buying one vendor. I mean, there's some try before you buy period where you experiment, but software purchasing companies typically follows this cycle where as you become more advanced, you stick to one vendor and you sign one long-term contract at a lower price. AI is nothing like that.

And that we saw as firms became more advanced, they actually start to consider more than one vendor and sign on for more than one vendor. And then the contracts they start to sign with those vendors start to get bigger and bigger and bigger. That's Open AI versus Anthropics story. But then you start to see that in the rest of the metrics that we track. Per employee per month's spend

tells a similar story where if you want to see, well, how far is this market going to go? where should it land? Because no one really knows how much to spend on AI at all. You see every single segment of business,

no matter how you break it up versus tech, manufacturing, or retail, every segment is growing their AI spend. And in virtually every segment, if you divide it by the top 1%, the top 10%, and the median, you observe these power law effects where the vast majority of spend,

90% of the spend is, driven by less than 10% of the customers. Overall today, the top 1% spends about $7.4,000 per employee per month on AI. Top 10% is about $6.50 and the top, and the median is about $12 per employee per month. And so there you start to see these distributional effective AI. And you start to see how the adoption metric itself is lacking to simply track whether or not businesses use AI, fails to track the fact that some businesses are using it much more than others. Some businesses aren't really using very complex versions of it at all.

You know, $12 per month maps to like a simple chat GPT subscription for everyone. And that starts to explain the macro statistics that we're all familiar with. Why doesn't AI show up in productivity metrics in the U.S.? It's because the vast majority of firms that are using AI are not getting ROI from it. They just aren't. They're using extremely simple models. They're not even using multiple models on average. You're not seeing that yet.

The median firm, and by the way, this is our very tech forward data set that we just talked about. Yeah, yeah. Is spending a chat GPT subscription, which, like, if you've used chat GPT, even the paid versions, you know that it's like nice to have, but it's not particularly productivity enhancing. So the gains from AI end up being concentrated amongst a very small set of, by the way,

fast-growing firms. And yet the metrics that the market is tracking are typically reported at the aggregate level. You know, the average spend or the total, amount of open eye and anthropic revenues. This is a really fast-changing market. It's a very dynamic market and it's a very heterogeneous market. And it's so I think that's the place where Ramp Bay Eye Index has been so helpful to analysts is showing how complex its development has been. It's not tracking the same way as software has historically because as firms become more advanced,

their behavior becomes more complex and difficult to track and spread between multiple vendors. I don't think we've figured out yet. I don't think most firms have even figured out where it's going to land in terms of how much they expect to spend on AI. So there's a lot to unpack there. So the lead statistic, the stat you led with, was 53% of the enterprises that you have business with

our customers use AI in some way or spending some money on artificial intelligence. And of course, that's, excuse the pun, ramped up quite a bit since 2020 when chat GPT was put on the planet was close to zero and now it's 53 percent and it's still rising. Do you know, just quickly kind of a sidebar, you know, what the total spend is? Do you have an estimate of the total spend by enterprises on AI? Do you have any sense of that? Like, is it $300 billion? Is it a trillion?

It's not a trillion dollars. Is it $100 billion? From the public metrics, it's in like the low hundreds of billions. I think the late assessments are somewhere between $200 to $300 billion. Okay, $2,300 billion, which in the grand scheme and things is pretty small. And you're saying that's got to be one reason why it's not showing up in the macroeconomic data, at least not yet. But that's the other thing you're saying.

You're saying that the intensity of use is very concentrated among a few firms. So you've got these big, big power users that are using it very aggressively, spending a lot of money on it. But most of the other enterprises captured by that 53%, they're not. They're casual users.

Who knows what they're doing with it? As you say, the ROI is probably not there. Is that, do I have that right? Exactly. Yeah. Got it. Got it. And that's another reason why we're not seeing it in your view in the macroeconomic data in the productivity day because, first of all, it's only a couple 300 billion dollars.

Second of all, it's very concentrated in a few businesses. It's not broad-based in terms of intensive use. I mean, even in our data set, that's one of the adjustments that we have to make for our research is that realization that to do research on all AI adopters doesn't really. get at the quantifiable impact that AI adoption is going to have on a firm. Because all AI adopters is including this group of firms that is high intensity, highly invested in it, doubling down on it every month. And it's also including these businesses that look fairly casual in their adoption by comparison.

But it requires you to come up with a definition of what does it mean to be a good adopter of AI? Right. And this is, of course, what businesses are trying to address and why they're hiring, you know, the model companies, open ionanthropic. No one's really come up with a definition, at least no one's shared publicly the playbook. Yeah. You know, the sort of practices and habits that use AI to drive productivity gains throughout a large complex organization. We have some ideas of it, but it's not known to the public. Now, there's two different ways of kind of thinking about what you're saying, or at least my interpretation of what you're. you're saying. One is a more optimistic perspective, and that is, well, we got a lot of room to run here. I mean, it's early days. Adoption is still pretty low, only 53%. You can envision, it should be at 100% at

some point. And more importantly, intensity should broaden out to more firms. So there's a lot of opportunity for growth here and for the impact to become meaningful from a macroeconomic perspective. other way of looking at it is just the opposite, that it's not working. You know, it's not working for the median company in the United States. They're floundering in trying to figure out how to use it. Maybe it's premature to conclude that this is not going to work, but maybe it's not going to work. So, Chris, you were shaking your head, yes. Is that, you're thinking the same lines? Yeah, I was thinking very, yeah, it's kind of this bimodal distribution. Right. Right. Either we haven't discovered how to really use it. This new hammer and takes time just to figure out how to use it. Or it's not really useful for most people. It's really a surgical instrument and you have to be a highly specialized firm to really take advantage of it.

Are, is that a good frame? And if it's a good frame, where do you land, oh? I think it's, I think it's a good frame insofar that it shows how difficult it is for AI to proliferate through a market point. place through traditional mechanisms. You know, we found, for example, that firms that small businesses, for example, are less likely to use AI in the first place. But when they do, they tend to be more intense adopters. Intensity defined as far as their spend per employee is higher. They're more likely to use multiple models than even in enterprise, at least, you know, when they're high intensity.

they're more likely to sort of reinvest in it year over year. That is an economically suboptimal outcome because what it means is that AI adoption proliferates through, oftentimes through what we found in our research is the labor pool you can hire from and the other businesses that you're around. Tech companies in California

are more likely to adopt AI and use it more intensely than tech companies, in New York. There is no economic reason for that, other than these are the companies that observe AI be used in a certain way, and then they learned from it, and they applied it in their own organizations. Now, I'm not going to go on a show and say, like, oh, we need to help out, like, the New York-based tech companies be more productive.

I'm in New York, and I work at a New York-based tech company, actually, right? But, like, that's not the policy concern I have. So much as the policy concern I have is that there are companies that are upon AI adoption through these uneven mechanisms, then enjoying the growth that is bequeathed by intense AI adoption. And then you can walk through the economic stories that are likely to occur, right? I think if you want to go into historic technological transformation and theory and how this tends to go, it's that even if a technology doesn't make a firm that much more productive, a doctor's office is not going to become that much more productive when it uses AI, the doctor's offices that use computers are likely to put out of business the doctor's offices that don't.

you know, because they, you can imagine, they're more likely to get a Google Maps listing. They're more likely to advertise and then drive more patients over to them. Even though the computer itself and the internet didn't do that much for, you know, my day-to-day experience going to a doctor. Yeah, it might be something similar there, where through these methods of proliferate, through these sort of uneven proliferation mechanics, you will have some firms be winners and losers simply by the fact that they were part of these networks rather than they were the ones that were best positioned to succeed. Yeah, so, okay, it's complex, but in my kind of simplistic way of thinking about this, you know, I'm an economist. So like all economists, I look at history and I say, this is how technology diffuses to the economy and how it diffuses and it takes a long time.

There's a lot of frictions, and you're saying the same thing. And the productivity gains are hard to come by, certainly early on. It takes a long time. In fact, it's not until new companies form and optimize around the technology that you start to see the real lift, because those companies can organize themselves, get the right people, don't have the same frictions, and, you know, we see the technology develop more fully. That's, that, am I reading too much in your data when that makes sense? It makes sense. That's what your data is kind of sort of saying, at least so far. We wrote a paper. A good example

is that we wrote a paper a couple months ago about AI's impact on jobs. Everyone wants to write that paper, but we had a very good data set for it in that, you know, previous work so far has worked, has focused on surveys to try to estimate AI adoption or it's

try to estimate jobs impact through AI exposure scores. Like, is this job exposed to AI? Is it not? Yeah. And, you know, economists disagree on the ability of those scores and even the surveys to get at the answer. You know, what you really want is a data set that shows, hey, these firms adopted AI,

these firms didn't. What happened to jobs over time pre and post adoption? And so we used our data set and we joined it with a workforce data set from Ravello Labs. Yeah, great. We've had their chief, who's our chief economist again? Lisa Simon. Yeah, she's great. Yeah.

Lisa Simon was on the paper with, co-authored the paper with us. And we found, you know, essentially that kind of story where firms that adopt AI tend to grow faster following adoption, but it's essentially exclusively concentrated amongst the firms that adopt AI intensely. So if you divide AI adoption to two different groups, firms that adopt AI intensely versus firms that adopt it sort of very lightly. So using agents, multiple models versus maybe having a simple. chat subscription for everybody. Firms that adopt AI intensely grow their headcount 10% over the two years following adoption. Firms that adopt AI lightly see no change from the control group. And by the way, this is comparing like for like firms.

So firms that were growing at otherwise similar rates pre-adoption, there are obviously firms that adopt AI were fast-growing firms as is. Tends to be the case with new technologies. their headcount growth accelerated upon adoption. Interesting. And so, and today the results were concentrated amongst the sort of high-intensity groups of businesses. But you saw some interesting growth, too, in unexpected places. You saw outsized hiring growth for entry-level hires.

That surprised us when we were first reading it, but such that it was about 12% overall growth over two years. such that the final firm had a 1% higher share of entry-level workers two years following adoption. And that was statistically significant. But, you know, it's surprising to many of us on the outside, but if you talk to many of these firms and these sort of high-intensity group, they'll say that they are hiring differently today. And the main difference is that they want to hire people who know how to use the AI models and know how to use them well. Right. And what better place to look than young people and recent college grads? So it sounds like you're kind of landing where most economists have landed.

That is adoption, but slower adoption. It's going to take a while for it to really come on and change the macroeconomic picture. And that the impact on jobs, of course, a lot of churn, a lot of job disruption over here, but job creation over here, the net of all that is, it feels like the job market's going to be okay, just fine. That's kind of what you're saying. I think that I've landed somewhere.

Sometimes it feels like I'm too econ for the tech people and two tech for the econ people. I know how that feels. And really that's because of the data set that I work with, in that I am not a, I mean, if you talk about what tech people think, is going to happen to the job market. You know, you get very disparate views on each side, right?

You get either massive job boom or you see massive job loss and none of us have anything to do after. And also, we all die, is apparently one of the mainstream opinions. So I'm not in that school of thought. And then you see a lot of the econ people who, I think, underestimate the effects of AI adoption, and the productivity gains that are already happening at firms that use it intensely. I agree that you don't really see them

in national productivity statistics. I mean, you know, you can see like the impact of the data center boom, but you don't actually see the, clearly see the impact of AI as far as it's applied to white-collar work. But if you have a data set, again, it's not otherwise available

to work so granularly to people in the public that we do want to make that available. And you can track specifically what's happening, at the specific segment of companies that are applying it intensely, which is still a very small group, you do see productivity gains. You see hiring growth and you see ROI in the form of the revealed preferences of those businesses. It's not bubble-like behavior when a firm expands their AI vendor list to experiment with new ones

and to double down on the old ones as well and to increase contract sizes. retention rates for AI-enabled software has risen each year since we started tracking it. As in I'm not just talking about like the model companies open-ionanthropic, but I'm talking about the fact that in every area of software, you're seeing AI-native entrance, CRM, right? This like Salesforce and HubSpot dominate the market, right? But there are also new entrants that are AI-native and challenging those large incumbents. that are growing their business adoption rates,

like whole percentage points month over month, growing very quickly. Salesforce and Upspot are still so large and not losing a massive amount of market share. But you can start to see how there is increasingly a better matching happening between software vendor and business buyer and that the companies that are buying AI,

whether it's through open-inanthropic or it's through some vertical-specific software category, are finding such use for it that they can both expand the vendor list and double down on the ones that they've bought previously. That doesn't happen in bubbles where there's just spend without question. You know, you are seeing some actual conscious decision-making about who you want to go with and who you want to buy from. Got it, got it.

So there's really two really big concerns about it. There's lots of concerns about AI. I mean, there's a lot of worries. But from a macro perspective in the near term, there's two. We talked about one, and that's job loss. And the other is around what's going on in the equity market and the credit markets more broadly. Stock prices have risen very significantly corporate credit spreads or paper thin. There's a lot of euphoria around, investor euphoria, around the success of AI.

these companies that are in that ecosystem and that they're going to be able to generate the kind of profits that are necessary to support those higher valuations. How do you have a view on that? How do you think about that in the context of what you're observing? Well, so far, across these segments that we track,

you see every group buy more and more from the model companies. every month. You know, the median top 10%, top 1%, even the threats that we have talked about, open source models and Chinese models, cheaper Chinese models that may eat away at the revenues of open-anthropic. Adoption of those tools is relatively marginal. Only about 6% of businesses that buy AI are buying open-source models, are using open-source models. Typically through a router platform. That's how we can track it. amongst those businesses, they continue to increase their spend on open-Aanthropic.

So I imagine, you know, it's had some marginal effect on how much they would have spent and some counterfactual on open-anthropic, but not enough that it's led to a net decrease amongst the companies that, again, specifically the companies that are already using open-source models. So that's some reprieve for investors that are banking on open-ionanthropics growth, is that companies, regardless of their cost-conscious posture or their exploration of cheaper models, continue to spend more and more on American model companies and at the frontier.

However, there are cracks in that AI thesis. So one of them being what's revenue, it's price times quantity. So quantity is addressed. Quantity will be fine. companies will buy more and more AI that continues to happen at the top 1% within it, the median, everyone's buying more AI.

Maybe if you asked me a month ago, I would say there's no problem really on price either because companies keep going for the frontier as well. But that's starting to appear not the case. There's two main signs of that. So Fable came out over a month ago now. Fable is

Fable 5 is anthropopic. I'm a big fan. Big fan of Fable. I know. I'm in the minority. But yeah, you're exactly. That's what I was. So, and remember Fable was so good, highly performed that the government briefly banned it.

Right. We all know that it's a good model. It's also Anthropics' most expensive models. Essentially the most expensive frontier model to date for sort of normal business usage. And Fables only cracked about 5 to 10% of business adoption slash model spend in today's market across open-9 Anthropic. So lagging pretty significantly behind GPD 5.6 Seoul, that's Open AI's frontier model,

lagging behind even Opus's latest model, which is Anthropics are their best model. And, you know, there's a couple of reasons going around about it that, well, Fable 5 because of its sort of everything that happened with the government, they have to do safety checks on every prompt that's going through the model. It's one of the things the government imposed on Anthropa to get it approved. And therefore,

Anthropic has to keep your data for 30 days to do these safety checks. That is a non-starter for many businesses. So they're not willing to use Fable for that reason. But the second reason is price. It is the most expensive model. And there are other highly performant models

that are pretty much just as good. and they're half the price. Like 5.6 sole was half the price. Actually, if you asked me two weeks ago, it was half the price. But this week, Open I just cut the price of 5.6 sold by 20%. Did they? I didn't know that.

Forget threats from Chinese model companies. These are two companies that are competing with each other. Yeah. And Anthropic has yet to respond, actually, with its own price cuts. But Open up is very... I kind of sort of did. Do you remember after the... the government allowed it to go forward.

Then they gave you a week. You have this for free for a week. And you'll see it's not just price cuts, right? They are also increasingly starting to introduce cheaper models and highly performant cheap models. Right. To compete both of each other and the Chinese model companies are doing price cuts on those two.

So if you look at a chart of these sort of effective price per million tokens to the frontier model, to open an anthropic, it's fallen like in a month, like 30, 40 percent. Where do you get that data, Aura? RAMP, my data set. You have that data? Yeah, because we have itemized receipts. We see the price per token.

Do you publish that data? I just haven't noticed it. I hope by the time that this one goes, this podcast goes up, we'll have prices posted on our website. Oh, okay, okay, great. I thought I might have missed it. Yeah.

Ramp.com slash data has our dashboard of metrics that I track. Because I've been looking for that data. I just haven't seen it. Yeah, and I'm increasingly adding more to it as we, Thank you. Like I want listeners to know that this is a scrappy team. That website is vibe coded by me. Like, I see.

It looks pretty good. There's data that we have that we use to inform sort of our outlook that isn't on the website because I haven't made the charts yet. And so. You could, I could, you need a little lesson on chart making. I'm just saying. I need fable five. You could do better.

I'm just saying. You have some cool charts like that first chart where you show market kind of the adoption rate by LLM is a pretty cool chart. I made that before the vibe coding era. I wrote that by name. Oh, is that why? Okay, that makes sense. I'm not joking.

Like, that was probably the last piece of code that I truly wrote myself. I see. The job of an economist has changed a lot over the last year and a half as well. So. But the picture of your painting, okay, you say, okay,

we're going to get the quantity, we're going to get demand, it's going to happen, but it's only going to happen at the cost, the price is going to be, is falling, the effective price is falling. So what does that mean for overall revenue? Well, the whole other lever that the model companies had to push their profits up is now

totally gone. There's no price lever anymore, clearly. No price lever. So it's all going to be a volume play. And they're going to have to, they're going to have to increase volume faster than the price can fall. And so far, that's actually been happening just fine. but is going to become increasingly challenging.

I am not an investor, so I don't actually know where the valuation should end up. But I can say as these levers start to deteriorate, especially the price lever, I would re-evaluate my price. I mean, part of the challenge here, too, is even tracking the volumes, there's a chart going around that use open-anthropics public disclosures to try to estimate. their total revenue for the year. And then it had like a revenue figure for Open AI for like and Anthropic for like, this is a couple weeks ago, like July and August of this year.

And I was like, how did they already get that? Like, I don't think that was announced. And you look at the chart notes and said like, oh, we can, we just continued the previous trend. I saw that. Yeah, I saw that. That's on like, that's fine. Almost always in every market.

Because most markets just don't change that much. much. They're also not that fast growing. The methods are fine. It's really the best the market can do. But I saw that and I was like, oh, I can make that chart with the ramp data. Because we only get a subset of the market, but we correlate surprisingly closely. And so I published my own chart. And I think the shock that came to many analysts and watchers is that it was the first chart to show and confirm what open AIS CFO sort of said in remarks to the company, which is, that they saw an inflection point in July,

and they're now growing faster. And that's exactly what our chart showed. Is the first chart to show that actually, after several months of anthropic beating Open AI in growth, open eye was now growing faster. And that's entirely something, as entirely a result of the fact that previous methods of financial analysis

that we're all used to, which is just, yeah, sure, impute the next month. It's fine. It's one month. How much can change in the market? That doesn't work for AI model companies, in part because these models aren't very sticky. You know, we saw the reasoning behind OpenEye's recent growth is that they came out with a better model than Anthropic. Yeah. That doesn't go well for these valuations either, right? I mean, it's a contestable

market. I mean, yeah, right? I mean, very contestable. People switch, you know, I switch very, quickly. I mean, I switch rapidly. Yeah. The model companies, but it's not like they can give much forward guidance about it and say, oh, we're going to have the better model in December of 27. They can't do that because they don't know what's going to come out of their models until it does and until it's digested by users. Forget the market, like literally engineers who use it and talk about which one works for them. And in this most recent cycle, it was open AI one. Open Eye beat Fable, beat Anthropics Fable. Interesting. And so you have all of these models sort of all of the sort of financial models fall apart.

I know, we're going to lose you pretty soon. So let me hand the baton back over to Chris and Marissa. Hey, Chris, there's so much to unpack here. We go on forever. Where do you want to take the conversation? Well, I'm feeling, I came into this hoping to have a little bit better. I actually am feeling a little bit worse about the outlook here. So maybe you can either confirm.

In what way, Chris, in what way are you feeling worse? Very consistent with our job. Well, I'm hearing. what I'm hearing is there's a race to the bottom. The prices are heavily subsidized and there's highly competition. Things are going down. The prices are coming down. Token costs are coming down, which, okay, great for consumers for the time being.

But then if I go back to my previous, there were our previous podcast, right, on the data centers, the costs are going up. Yeah. Right. Regulations, cost of electricity, so on and so forth. So it seems like something's going to crack here, right? The two can't continue indefinitely. Or am I missing something? Is there some innovation on the horizon here that makes all the math work out here?

It depends on who you are in the market. I mean, if you are a business or a consumer of these models, it's great. Price comes down. For now, right? Just like ride sharing, right? Uber was highly subsidized until it wasn't, right? Until the market sorted itself out.

Now it's like a taxi, right? Well, but the increased competition, I think, still Uber and Lyft is a story like there's only so many companies that can enter

that market and so many sort of fixed resources for that I think AI is a little bit difference at the end of the

day I mean there are sort fixed resources as far as power and compute but you know the malls do get

more efficient over time as well and so you can imagine there's going to be more volume to support demand anyway

for that at least for the users it sounds like you're saying well the demand is going to

material either perhaps it's still a nascent technology. Going back to your earlier point, just a few firms really are leveraging it, but that's destined to grow. So the demand is going to be sufficient to grow into the supply, even if the prices come in here,

there's still so much demand, so much future opportunity that... Not to worry, Chris. He's not saying that. He's actually... I don't know if the demand is defined by profits. I don't know about. Yeah, yeah.

Demand as far as like, is there enough demand for tokens in general that business would be able to employ into productivity enhancing workflows? Yes. Will that be sufficient given the realities of an ever-declining price war to support high valuations? I'm not an investor and then I'll pause on that point. I was going to ask. I was going to ask, okay, not an investor. Here you. Okay.

Okay, very interesting. Yeah, very interesting. Marissa, anything you want to ask Aro about? I'm a little more interested in the demographics of these firms that you're looking at that are these. You said that it looks like they're small businesses, right, and they're very intense adopters. Can you tell are these like startups with one or two people? It sounds like you have employee accounts there. So what more can you glean about what kinds of? of firms these are.

It varies. I think the high-intensity adopt, I think the stat you're referring to is that small businesses are less likely to adopt AI in the first place, but when they do, they're more likely to be intense. Yeah. Yeah. So if you're an intense adopter, yes, you do skew toward the smaller side,

but it's, on average, it's still in like the sort of tens to low hundreds in terms of employee count. You do see some very intense adopters in enterprise with, you know, a thousand, two thousand employees and certainly very, you know, some examples in pure tech and software development with more than that. But I think it's hard to give you an image of the profile of the segments because there is so much heterogeneity amongst the users.

You'll find examples in any group. But I do think that the, um, today the productivity gains have been concentrated amongst tech, not only because of the sort of proliferation mechanics of how AIs spread amongst firms, but because the most advanced productivity enhancing version of AI is made by tech for tech, it's coding agents. Now, open question whether or not that technology will, you know, proliferate through other types of firms, if they'll be able to make this sort of equivalent coding agent, but for manufacturing. There's a lot of really interesting work happening in robotics with AI. I'm bullish on that work happening, and I'm bullish on the gains that will come from it.

But today, it's not nearly as developed as what's developed for tech. Yeah. Hey, Ara, what do you vibe code in? What LLM do you use? I, at Ramp, Ramp is one of the very tech forward companies, and internally sometimes it feels like we're babies in terms of using AI because we have a pretty good setup for employees. We have our own harness. You can use codex, you can use Claude, but there's an internal version called Ramp Glass that connects to all of the frontier models and it connects to our internal company tools. So it just, it makes it

easy for me to connect between things and make direct edits to our website. I don't have to log into a million accounts. That's one of the challenges to using these coding agents is that you can have to make your own setup. And so we have a pretty good system here at Ramp. Got it. Does it optimize across the models?

Yeah. Right? So, in terms of the cost, right? That's an emerging area, right? Because you can add-hoc select the model and you can set an auto model that follows the firm defaults, the company-level

defaults to sort of keep you from spinning up an agent that blows thousands of dollars on a task that really doesn't need it. I mean, I think that's where the market's going to go. Firms that are trying to figure out how much to spend on AI, you know, you can't leave it in the ends of your employees who have to track in their head what model is good for what and how much it's going to cost. This market is changing way too quickly. So I think increasingly firms are going to figure out how to have their own sort of implementations of this. You're going to see products come to the market, but you're going to have companies set defaults and sort of monitor and track spend that way instead of free-for-all we're in right now at most

firms. You know, this is one other thing that struck me was, again, going back to that first chart, is anthropic and open AI dominate to a significant degree. I didn't realize. I thought Gemini perhaps or some of the others would have a higher adoption. Right, but they don't. It's really those two companies that are kind of dominating things here. I think there's probably a lot of free adoption of Gemini because of its integration.

Free adoption, yeah. But no, it's not a major player in enterprise spend. Yeah. So to end, what didn't we ask you? Because, you know, you look at this data, you live this data, you know, it's your baby. What is it that you see in there that we should have? asked you about that we didn't ask you about. Anything? There's a couple things I'm thinking about today. I mean, I'm thinking about,

uh, government is trying to figure out what to do about AI. I think there's a lot of discourse about data centers. I understand why it's there. But the real question is, how do you get more firms to use AI in a productivity enhancing way? That's ultimately going to be what vaults us into a, right? You know, next boom of productivity. And today, I'm not sure that we have those systems set up. We found that there are uneven mechanisms that distribute AI through the economy. That's going to lead to some economically suboptimal outcomes. Is there a role for government there to get businesses outside of tech using AI more effectively? Probably, I don't think it's going to be in the form of subsidies. I hope not. I hope not. I don't want to do that.

It's hard for the government to get any of this stuff through because there's this not unreasonably that AI is going to disproportionately impact white-collar work. I actually think our evidence is that it's probably the opposite. It's more likely going to impact blue-collar work once we get the sort of manufacturing automation wave going. And that it's going to work like every other technological development and paradigm of the last century. And yet today the focus is on data centers and making sure software engineers keep their

jobs, even though software engineers are doing pretty well, you know, if you really look at it. So I do think that the discourse, which tends to be informed by a lot of punditry and vibes and people who, you know, think they have a viewpoint because they are on Twitter a lot, would be better informed by observed behaviors of businesses. And so my goal is to publish that kind of work so that people can make these more informed decisions and ideally better policy decisions with it. Well, that's great. I mean, thank you. Good luck with all that. That's a heavy burden. And on the government, I'd say, I'd just settle for bright yellow lines around the use of these models. I mean, that would be kind of nice as opposed to kind of an ad hoc quasi licensing regime,

which you have no idea how it's being implemented or will be implemented. Who's under the auspices of the government's oversight who's not. I mean, let's just nail that down. I think that would go a long way to getting adoption and use in a place where we want it. But thank you for, you know, spending time with us. And really, thank you for providing this information to the marketplace. You know, we are going to be, we are using it and we're going to use it a lot more intensely going for it. And I can't wait to see that price data. That'll be really cool when you release that. Thank you, guys. Thank you for having me. This is really fun. Take care now. And with that, dear listener,

We're going to call this a podcast. Hope you enjoyed it. Take care now.